25273 lines
885 KiB
JSON
25273 lines
885 KiB
JSON
{
|
|
"context": {
|
|
"date": "2023-10-09T15:03:52-04:00",
|
|
"host_name": "frontier10485",
|
|
"executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope",
|
|
"num_cpus": 128,
|
|
"mhz_per_cpu": 1796,
|
|
"cpu_scaling_enabled": false,
|
|
"caches": [
|
|
{
|
|
"type": "Data",
|
|
"level": 1,
|
|
"size": 32768,
|
|
"num_sharing": 2
|
|
},
|
|
{
|
|
"type": "Instruction",
|
|
"level": 1,
|
|
"size": 32768,
|
|
"num_sharing": 2
|
|
},
|
|
{
|
|
"type": "Unified",
|
|
"level": 2,
|
|
"size": 524288,
|
|
"num_sharing": 2
|
|
},
|
|
{
|
|
"type": "Unified",
|
|
"level": 3,
|
|
"size": 33554432,
|
|
"num_sharing": 16
|
|
}
|
|
],
|
|
"load_avg": [1.86,2.42,2.52],
|
|
"library_build_type": "release"
|
|
},
|
|
"benchmarks": [
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8637057530223122e+06,
|
|
"cpu_time": 2.0665387037037041e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1977718281749398e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8671588987725228e+06,
|
|
"cpu_time": 2.0667462645502645e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1937072429629453e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8739381494159470e+06,
|
|
"cpu_time": 2.0730184841269834e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1857711799488184e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8683880865210472e+06,
|
|
"cpu_time": 2.0676043571428584e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1922640320549160e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8635952986538764e+06,
|
|
"cpu_time": 2.0678864179894177e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1979020890204264e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8673572372771413e+06,
|
|
"cpu_time": 2.0683588455026455e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1934832744324091e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8671588987725228e+06,
|
|
"cpu_time": 2.0676043571428582e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1937072429629453e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2395728667930953e+03,
|
|
"cpu_time": 2.6654270785914614e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9731964261298881e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.2703598337589469e-03,
|
|
"cpu_time": 1.2886676237960624e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2672597890753300e-03,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8629308771286455e+06,
|
|
"cpu_time": 2.0684725782493372e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3973719586560363e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8679243958272925e+06,
|
|
"cpu_time": 2.0687235517241363e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3856164726473382e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8629113614470551e+06,
|
|
"cpu_time": 2.0689651167108782e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3974180251049055e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8678853422695454e+06,
|
|
"cpu_time": 2.0684756312997309e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3857081666728193e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8731127194518319e+06,
|
|
"cpu_time": 2.0736159946949594e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3734687800301714e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8669529392248741e+06,
|
|
"cpu_time": 2.0696505745358083e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3879166806222545e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8678853422695459e+06,
|
|
"cpu_time": 2.0687235517241366e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3857081666728193e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2505096827767875e+03,
|
|
"cpu_time": 2.2260922948151015e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.9823362968964302e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.2767096017650684e-03,
|
|
"cpu_time": 1.0755884699591768e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2749603111153027e-03,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8738656192568664e+06,
|
|
"cpu_time": 2.0765487665782480e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7434231311087999e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8682567012211895e+06,
|
|
"cpu_time": 2.0702982281167121e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7696728127834722e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8639770245239132e+06,
|
|
"cpu_time": 2.0705536684350127e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7898079131016713e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8683870033160609e+06,
|
|
"cpu_time": 2.0705153740053021e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7690612121157218e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8637766974549852e+06,
|
|
"cpu_time": 2.0708473740053065e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7907526810334064e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8676526091546032e+06,
|
|
"cpu_time": 2.0717526822281168e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7725435500286128e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8682567012211897e+06,
|
|
"cpu_time": 2.0705536684350125e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7696728127834722e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.1242687358012035e+03,
|
|
"cpu_time": 2.6882155114591037e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9351984989332330e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.2082633117023096e-03,
|
|
"cpu_time": 1.2975561873379584e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2059719486110969e-03,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8716086091184050e+06,
|
|
"cpu_time": 2.0755924350132644e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7507933998783488e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8754222987893478e+06,
|
|
"cpu_time": 2.0789770079575600e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7472331442978423e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8758420435732123e+06,
|
|
"cpu_time": 2.0796438010610035e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7468421774778873e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8709020005382809e+06,
|
|
"cpu_time": 2.0745898620689707e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7514546454369202e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8666112887415544e+06,
|
|
"cpu_time": 2.0747421405835503e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7554806508264378e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8720772481521603e+06,
|
|
"cpu_time": 2.0767090493368700e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7503608035834875e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8716086091184050e+06,
|
|
"cpu_time": 2.0755924350132649e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7507933998783488e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7696349326532609e+03,
|
|
"cpu_time": 2.4167600343725330e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5270188410992007e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0136107825540271e-03,
|
|
"cpu_time": 1.1637451260417282e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0150238932901080e-03,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8714762595266015e+06,
|
|
"cpu_time": 2.0802470026525266e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5018344297125965e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8670684023033357e+06,
|
|
"cpu_time": 2.0805393554376618e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5101017144926541e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8677792800950140e+06,
|
|
"cpu_time": 2.0806227214854066e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5087657679051980e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8724116509633919e+06,
|
|
"cpu_time": 2.0805347294429706e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5000850355892874e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8767881089942197e+06,
|
|
"cpu_time": 2.0857709496021164e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4919232323525898e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8711047403765128e+06,
|
|
"cpu_time": 2.0815429517241367e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5025420360104658e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8714762595266018e+06,
|
|
"cpu_time": 2.0805393554376620e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5018344297125965e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.9210149109476524e+03,
|
|
"cpu_time": 2.3678044375250279e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.3345096340433694e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0955614222636443e-03,
|
|
"cpu_time": 1.1375236987369304e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0940532786289318e-03,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8795514765673594e+06,
|
|
"cpu_time": 2.0964044601063889e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9735786241607964e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8749812780755593e+06,
|
|
"cpu_time": 2.0910561250000035e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9905764677570283e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8698948935049516e+06,
|
|
"cpu_time": 2.0903210771276567e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0095918468613595e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8767365901944591e+06,
|
|
"cpu_time": 2.0936844893617060e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9840381801486015e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8691617025708300e+06,
|
|
"cpu_time": 2.0894307340425544e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0123414052259162e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8740651881826320e+06,
|
|
"cpu_time": 2.0921793771276618e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9940253048307404e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8749812780755595e+06,
|
|
"cpu_time": 2.0910561250000033e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9905764677570283e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.4584329782528384e+03,
|
|
"cpu_time": 2.8454657815994524e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6639566186502579e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3790170194540501e-03,
|
|
"cpu_time": 1.3600486711163229e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3791115218027175e-03,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8738475006927426e+06,
|
|
"cpu_time": 2.1087236879999978e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3989612276510656e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8736764288429790e+06,
|
|
"cpu_time": 2.1088695066666692e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3990889566875616e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8735868667329974e+06,
|
|
"cpu_time": 2.1084762079999940e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3991558366178378e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8771964905317873e+06,
|
|
"cpu_time": 2.1088502693333263e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3964654276853979e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8735651510845248e+06,
|
|
"cpu_time": 2.1087042906666701e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3991720536019593e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8743744875770067e+06,
|
|
"cpu_time": 2.1087247925333316e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3985687004487643e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8736764288429793e+06,
|
|
"cpu_time": 2.1087236879999978e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3990889566875616e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5814608750967111e+03,
|
|
"cpu_time": 1.5724377969826753e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1786925370935374e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.4372727306006843e-04,
|
|
"cpu_time": 7.4568184646495097e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.4278486764027090e-04,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 369,
|
|
"real_time": 1.8973187801363778e+06,
|
|
"cpu_time": 2.1582894607045953e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7633100219580102e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 369,
|
|
"real_time": 1.8981521569639142e+06,
|
|
"cpu_time": 2.1593367371273790e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7620968007043034e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 369,
|
|
"real_time": 1.8947510162953963e+06,
|
|
"cpu_time": 2.1581336693766932e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7670548557091379e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 369,
|
|
"real_time": 1.8925626350814088e+06,
|
|
"cpu_time": 2.1556046937669232e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7702544173786235e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 369,
|
|
"real_time": 1.8920644174921794e+06,
|
|
"cpu_time": 2.1556611951219616e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7709838795811880e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8949698011938557e+06,
|
|
"cpu_time": 2.1574051512195105e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7667399950662529e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8947510162953965e+06,
|
|
"cpu_time": 2.1581336693766932e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7670548557091379e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7353546522690676e+03,
|
|
"cpu_time": 1.6827652833482593e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.9932701694225497e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4434819227967426e-03,
|
|
"cpu_time": 7.7999502429900423e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4433124097470264e-03,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.9078648546704333e+06,
|
|
"cpu_time": 2.2344598086253395e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4960706332688963e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.9041874813577386e+06,
|
|
"cpu_time": 2.2317536522911065e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5066846634887874e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.9091664603377902e+06,
|
|
"cpu_time": 2.2345259973045839e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4923235966259050e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.9045377028114451e+06,
|
|
"cpu_time": 2.2318096603773688e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5056720507664955e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.9090931942294834e+06,
|
|
"cpu_time": 2.2340756091644266e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4925343779417169e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9069699386813785e+06,
|
|
"cpu_time": 2.2333249455525652e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4986570644183612e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9078648546704333e+06,
|
|
"cpu_time": 2.2340756091644270e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4960706332688963e+08,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4388335438792383e+03,
|
|
"cpu_time": 1.4194177536187278e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.0345950541591237e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2789050809923256e-03,
|
|
"cpu_time": 6.3556257518430134e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2793296566319383e-03,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 361,
|
|
"real_time": 1.9505058756740533e+06,
|
|
"cpu_time": 2.4047546288088541e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0751836362837253e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 361,
|
|
"real_time": 1.9524946140958681e+06,
|
|
"cpu_time": 2.4068352326869760e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0740884942062273e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 361,
|
|
"real_time": 1.9507576737919217e+06,
|
|
"cpu_time": 2.4053017202216126e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0750448547120228e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 361,
|
|
"real_time": 1.9520128462166935e+06,
|
|
"cpu_time": 2.4054533324099649e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0743535853591378e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 361,
|
|
"real_time": 1.9510839104786557e+06,
|
|
"cpu_time": 2.4058945484764525e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0748650986955807e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9513709840514385e+06,
|
|
"cpu_time": 2.4056478925207723e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0747071338513389e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9510839104786557e+06,
|
|
"cpu_time": 2.4054533324099653e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0748650986955807e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.4874977650948608e+02,
|
|
"cpu_time": 7.7881717766649945e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6737949942204356e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.3495049554714144e-04,
|
|
"cpu_time": 3.2374529127386609e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.3489010605813548e-04,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 2.0102761516214479e+06,
|
|
"cpu_time": 2.7083918371428493e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0864317554665110e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 2.0105623354070953e+06,
|
|
"cpu_time": 2.7083039885714301e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0861347724146760e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 2.0124988344364932e+06,
|
|
"cpu_time": 2.7115545742857093e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0841274182275090e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 2.0094999827311509e+06,
|
|
"cpu_time": 2.7088186114285812e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0872376392357264e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 2.0083059365528501e+06,
|
|
"cpu_time": 2.7073485657142880e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0884786145677080e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0102286481498077e+06,
|
|
"cpu_time": 2.7088835154285715e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0864820399824262e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0102761516214483e+06,
|
|
"cpu_time": 2.7083918371428489e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0864317554665110e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5407078499262168e+03,
|
|
"cpu_time": 1.5868551833563915e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5987969861367641e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.6643413242780481e-04,
|
|
"cpu_time": 5.8579675881904273e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.6626443721999654e-04,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 332,
|
|
"real_time": 2.1187884028978557e+06,
|
|
"cpu_time": 3.3496945873493967e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9591532540611162e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 332,
|
|
"real_time": 2.1158705524404543e+06,
|
|
"cpu_time": 3.3467632108433880e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9646130479601140e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 332,
|
|
"real_time": 2.1203076978864889e+06,
|
|
"cpu_time": 3.3508787951807156e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9563163442559390e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 332,
|
|
"real_time": 2.1142748914552424e+06,
|
|
"cpu_time": 3.3470243042168580e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9676051746640058e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 332,
|
|
"real_time": 2.1205144704330759e+06,
|
|
"cpu_time": 3.3506921807228960e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9559305616465716e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1179512030226234e+06,
|
|
"cpu_time": 3.3490106156626507e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9607236765175495e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1187884028978557e+06,
|
|
"cpu_time": 3.3496945873493967e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9591532540611162e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7689657398971035e+03,
|
|
"cpu_time": 1.9863178627779587e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.1804194199929414e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3073793843528536e-03,
|
|
"cpu_time": 5.9310587236969278e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3079476992315215e-03,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 216,
|
|
"real_time": 3.2404116333324322e+06,
|
|
"cpu_time": 5.5651034675926017e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1774952994926596e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 216,
|
|
"real_time": 3.2426983848886769e+06,
|
|
"cpu_time": 5.5695314537037127e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1738441287612896e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 216,
|
|
"real_time": 3.2435974334711554e+06,
|
|
"cpu_time": 5.5699233009259049e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1724100613946295e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 216,
|
|
"real_time": 3.2413723475420503e+06,
|
|
"cpu_time": 5.5656653101851996e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1759607354959545e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 216,
|
|
"real_time": 3.2426026968837338e+06,
|
|
"cpu_time": 5.5650260555555578e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1739968069858055e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2421364992236099e+06,
|
|
"cpu_time": 5.5670499175925953e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1747414064260674e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2426026968837334e+06,
|
|
"cpu_time": 5.5656653101851996e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1739968069858055e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2476452584458166e+03,
|
|
"cpu_time": 2.4604967013991381e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9915694570865461e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.8482194032996101e-04,
|
|
"cpu_time": 4.4197496660190733e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.8486357107881503e-04,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 108,
|
|
"real_time": 6.4752646075147726e+06,
|
|
"cpu_time": 1.0791743944444470e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1819398949440460e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 108,
|
|
"real_time": 6.4792202535533793e+06,
|
|
"cpu_time": 1.0790454768518558e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1787762549973269e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 108,
|
|
"real_time": 6.4819447601352027e+06,
|
|
"cpu_time": 1.0797768675926041e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1765994993299065e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 108,
|
|
"real_time": 6.4812114902047645e+06,
|
|
"cpu_time": 1.0805205685185134e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1771851683457251e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 108,
|
|
"real_time": 6.4907196124463724e+06,
|
|
"cpu_time": 1.0828757203703716e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1696012158123760e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.4816721447708989e+06,
|
|
"cpu_time": 1.0802786055555584e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1768204066858768e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.4812114902047645e+06,
|
|
"cpu_time": 1.0797768675926041e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1771851683457251e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.6835617635047947e+03,
|
|
"cpu_time": 1.5648626933941498e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.5368786241820492e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.7686659191641152e-04,
|
|
"cpu_time": 1.4485732526280872e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.7638323676878167e-04,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 53,
|
|
"real_time": 1.3288090023089129e+07,
|
|
"cpu_time": 2.2032856528301824e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0503017275916195e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 53,
|
|
"real_time": 1.3273665681481361e+07,
|
|
"cpu_time": 2.1997827962264154e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0557898330697260e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 53,
|
|
"real_time": 1.3281047027909530e+07,
|
|
"cpu_time": 2.2037747773584943e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0529799238699856e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 53,
|
|
"real_time": 1.3270097731981637e+07,
|
|
"cpu_time": 2.2034948377358567e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0571491902628632e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 53,
|
|
"real_time": 1.3258061424460052e+07,
|
|
"cpu_time": 2.2023639037735753e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0617403141751604e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3274192377784345e+07,
|
|
"cpu_time": 2.2025403935849048e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0555921977938719e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3273665681481361e+07,
|
|
"cpu_time": 2.2032856528301828e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0557898330697260e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1373015887496633e+04,
|
|
"cpu_time": 1.6298443272543740e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.3323108805163093e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.5677648506364014e-04,
|
|
"cpu_time": 7.3998385319126974e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.5693440274055657e-04,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 28,
|
|
"real_time": 2.5632657908967562e+07,
|
|
"cpu_time": 4.2015903178571425e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2362001816847887e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 28,
|
|
"real_time": 2.5611840321549349e+07,
|
|
"cpu_time": 4.2012037000000238e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2404562231739197e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 28,
|
|
"real_time": 2.5645903311669827e+07,
|
|
"cpu_time": 4.2056258249999702e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2334958285102015e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 28,
|
|
"real_time": 2.5600480514445476e+07,
|
|
"cpu_time": 4.1998232499999747e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2427815924886847e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 28,
|
|
"real_time": 2.5637046367462192e+07,
|
|
"cpu_time": 4.2031203678571366e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2353038675447931e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5625585684818886e+07,
|
|
"cpu_time": 4.2022726921428502e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2376475386804771e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5632657908967566e+07,
|
|
"cpu_time": 4.2015903178571433e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2362001816847887e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8797434350808478e+04,
|
|
"cpu_time": 2.2119109238842309e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.8429009588903016e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.3354164786736782e-04,
|
|
"cpu_time": 5.2636063528669274e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.3370743840820666e-04,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 14,
|
|
"real_time": 5.0024046695658140e+07,
|
|
"cpu_time": 8.0855187642856866e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3661283668859797e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 14,
|
|
"real_time": 5.0035932234355383e+07,
|
|
"cpu_time": 8.0903674999999806e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3648536963939772e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 14,
|
|
"real_time": 5.0016686320304871e+07,
|
|
"cpu_time": 8.0859939999999359e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3669180377314482e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 14,
|
|
"real_time": 4.9957604812724248e+07,
|
|
"cpu_time": 8.0851532142857656e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3732651316307545e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 14,
|
|
"real_time": 5.0030336316142760e+07,
|
|
"cpu_time": 8.0852397285714313e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3654537579709759e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.0012921275837086e+07,
|
|
"cpu_time": 8.0864546414285600e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3673237981226273e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.0024046695658140e+07,
|
|
"cpu_time": 8.0855187642856866e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3661283668859797e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1743698016587317e+04,
|
|
"cpu_time": 2.2118004374465865e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.4092372274395926e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.3470993508878994e-04,
|
|
"cpu_time": 2.7351917935890973e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.3518381891401244e-04,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.8500934739907578e+07,
|
|
"cpu_time": 1.5723903466666734e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4504143886310463e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.8667867481708527e+07,
|
|
"cpu_time": 1.5737451211111084e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4411930216240606e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.8673563864496022e+07,
|
|
"cpu_time": 1.5737583888888827e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4408789038699436e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.8686150378651083e+07,
|
|
"cpu_time": 1.5749301011111230e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4401849696240864e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.8810971611075938e+07,
|
|
"cpu_time": 1.5758711333333349e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4333127510692444e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.8667897615167841e+07,
|
|
"cpu_time": 1.5741390182222244e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4411968069636765e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.8673563864496022e+07,
|
|
"cpu_time": 1.5737583888888830e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4408789038699436e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1035402198929476e+05,
|
|
"cpu_time": 1.3213554336765135e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.0878370677803131e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1184389721133623e-03,
|
|
"cpu_time": 8.3941470123064747e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1188415497099945e-03,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7548662051558495e+08,
|
|
"cpu_time": 2.8295565350000018e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1186534953224030e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7585363611578941e+08,
|
|
"cpu_time": 2.8338289299999887e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1058835501871758e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7552291974425316e+08,
|
|
"cpu_time": 2.8288711749999964e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1173881198222017e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7636067420244217e+08,
|
|
"cpu_time": 2.8379837500000352e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0883290952237196e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7532899603247643e+08,
|
|
"cpu_time": 2.8317925324999750e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1241542944848061e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7571056932210925e+08,
|
|
"cpu_time": 2.8324065844999993e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1108817110080614e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7552291974425316e+08,
|
|
"cpu_time": 2.8317925324999750e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1173881198222017e+09,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.1050824298441241e+05,
|
|
"cpu_time": 3.6785460467934085e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4252171321563601e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3362751857680035e-03,
|
|
"cpu_time": 1.2987351699165664e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3322610378613499e-03,
|
|
"dst_gpu": NaN,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8653959957610201e+06,
|
|
"cpu_time": 2.0670536904761684e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1957804183711493e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8619631477959622e+06,
|
|
"cpu_time": 2.0679179074074042e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1998287156480541e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8666776288622590e+06,
|
|
"cpu_time": 2.0670429338624051e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1942728281886117e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8622311334222404e+06,
|
|
"cpu_time": 2.0680946666666551e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1995121478141872e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8667443865818088e+06,
|
|
"cpu_time": 2.0673297671958148e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1941943575361036e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8646024584846583e+06,
|
|
"cpu_time": 2.0674877931216892e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1967176935116211e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8653959957610201e+06,
|
|
"cpu_time": 2.0673297671958145e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1957804183711493e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3512248370793013e+03,
|
|
"cpu_time": 4.9107302349539930e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7708531705019068e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2609791574500636e-03,
|
|
"cpu_time": 2.3752160720326706e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2613606102805529e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8699633056861805e+06,
|
|
"cpu_time": 2.0732202142857693e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3808346265885457e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8694025817413128e+06,
|
|
"cpu_time": 2.0728949708994767e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3821486500619408e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8668607600572552e+06,
|
|
"cpu_time": 2.0698006455026935e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3881151584913908e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8630634143529090e+06,
|
|
"cpu_time": 2.0708296931217324e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3970591322278194e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8670075489411366e+06,
|
|
"cpu_time": 2.0701158968253722e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3877701537125809e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8672595221557587e+06,
|
|
"cpu_time": 2.0713722841270089e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3871855442164559e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8670075489411366e+06,
|
|
"cpu_time": 2.0708296931217327e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3877701537125809e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7263974476803055e+03,
|
|
"cpu_time": 1.5871659407034608e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4108920711892742e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4601063298007276e-03,
|
|
"cpu_time": 7.6623886148615740e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4612767129579538e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8637051782655697e+06,
|
|
"cpu_time": 2.0736394854111797e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7910900238242261e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8681291539351076e+06,
|
|
"cpu_time": 2.0729346604774226e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7702715658004880e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8715795572425316e+06,
|
|
"cpu_time": 2.0768844111405446e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7541028841644134e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8719916783281865e+06,
|
|
"cpu_time": 2.0772831697612451e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7521756585114766e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8659180500133911e+06,
|
|
"cpu_time": 2.0707349628646953e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7806642954562865e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8682647235569577e+06,
|
|
"cpu_time": 2.0742953379310179e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7696608855513781e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8681291539351079e+06,
|
|
"cpu_time": 2.0736394854111797e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7702715658004880e+06,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5774650009194761e+03,
|
|
"cpu_time": 2.7653343105076010e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6796134219130046e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9148597925182683e-03,
|
|
"cpu_time": 1.3331439645744235e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9152546989362881e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8646944900293739e+06,
|
|
"cpu_time": 2.0770412042440274e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7572851839919265e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8693527596785296e+06,
|
|
"cpu_time": 2.0765771458885854e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7529061773035858e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8652369172242223e+06,
|
|
"cpu_time": 2.0773277559681837e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7567741501044352e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8694970482482947e+06,
|
|
"cpu_time": 2.0763734217506472e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7527708872663576e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8745385391496564e+06,
|
|
"cpu_time": 2.0815745331564941e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7480568852355786e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8686639508660156e+06,
|
|
"cpu_time": 2.0777788122015879e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7535586567803767e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8693527596785296e+06,
|
|
"cpu_time": 2.0770412042440276e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7529061773035858e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.9743073372910030e+03,
|
|
"cpu_time": 2.1548862521036440e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.7263320700245902e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.1268175775795034e-03,
|
|
"cpu_time": 1.0371105140976745e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1250113622468302e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8769245327941936e+06,
|
|
"cpu_time": 2.0901369681697651e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4916694227676801e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8706382617038805e+06,
|
|
"cpu_time": 2.0833231989389537e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5034031614592440e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8656540789466354e+06,
|
|
"cpu_time": 2.0840197347479502e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5127626680398434e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8702893991639966e+06,
|
|
"cpu_time": 2.0836875994695188e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5040566464897916e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8666747537546849e+06,
|
|
"cpu_time": 2.0847538435013462e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5108419325958602e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8700362052726783e+06,
|
|
"cpu_time": 2.0851842689655074e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5045467662704840e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8702893991639963e+06,
|
|
"cpu_time": 2.0840197347479505e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5040566464897916e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.4265681566584908e+03,
|
|
"cpu_time": 2.8184320397755550e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.2840839264217982e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3671029171400629e-03,
|
|
"cpu_time": 1.3516465099623180e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3638103523548259e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8687503875008454e+06,
|
|
"cpu_time": 2.0987034574467754e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0138848332380995e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8774252788948906e+06,
|
|
"cpu_time": 2.1029305585106602e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9814762522614449e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8732116552316856e+06,
|
|
"cpu_time": 2.0982608962765397e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9971804645742789e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8700831623139861e+06,
|
|
"cpu_time": 2.0988155292553273e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0088861629990488e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8781733905733686e+06,
|
|
"cpu_time": 2.1037253537234152e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9786953993628010e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8735287749029554e+06,
|
|
"cpu_time": 2.1004871590425437e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9960246224871352e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8732116552316856e+06,
|
|
"cpu_time": 2.0988155292553273e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9971804645742789e+07,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2296866092492983e+03,
|
|
"cpu_time": 2.6166910967421650e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5793189779537887e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.2576042951186525e-03,
|
|
"cpu_time": 1.2457543886794911e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2574520004938602e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8811295734501092e+06,
|
|
"cpu_time": 2.1257958138297554e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3935456849961245e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8821163567020237e+06,
|
|
"cpu_time": 2.1264830824468075e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3928150566596591e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8807623416657231e+06,
|
|
"cpu_time": 2.1255781968085393e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3938177843768847e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8808742264267725e+06,
|
|
"cpu_time": 2.1253071781915152e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3937348724162868e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8815393867314807e+06,
|
|
"cpu_time": 2.1259181409574733e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3932421603747764e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8812843769952219e+06,
|
|
"cpu_time": 2.1258164824468186e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3934311117647463e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8811295734501090e+06,
|
|
"cpu_time": 2.1257958138297554e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3935456849961245e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.5270442127543367e+02,
|
|
"cpu_time": 4.3911504057555777e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0931033153342221e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.9379100152748328e-04,
|
|
"cpu_time": 2.0656300494487450e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9374278217101149e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.9019490625973421e+06,
|
|
"cpu_time": 2.1834389973118245e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7565827619169843e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8978019544456694e+06,
|
|
"cpu_time": 2.1812802419355218e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7626064920622325e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8955408784833746e+06,
|
|
"cpu_time": 2.1791155618279665e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7659018381048244e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8952828115327987e+06,
|
|
"cpu_time": 2.1788024999999399e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7662784509504694e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8956974467640663e+06,
|
|
"cpu_time": 2.1791814354838775e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7656733984368318e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8972544307646502e+06,
|
|
"cpu_time": 2.1803637473118259e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7634085882942688e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8956974467640663e+06,
|
|
"cpu_time": 2.1791814354838775e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7656733984368318e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8101198024781929e+03,
|
|
"cpu_time": 1.9805584055061086e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0870843933053303e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4811507391476379e-03,
|
|
"cpu_time": 9.0836146397496392e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4790011186250631e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8924553179174601e+06,
|
|
"cpu_time": 2.2583985591397751e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5408230253694892e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8999830579325077e+06,
|
|
"cpu_time": 2.2634749112902619e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5188702637223625e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8923675725286605e+06,
|
|
"cpu_time": 2.2586013010752639e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5410799425126958e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8987959859171703e+06,
|
|
"cpu_time": 2.2616611559139681e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5223205008699727e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8920402925227447e+06,
|
|
"cpu_time": 2.2587787043010220e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5420384235152054e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8951284453637090e+06,
|
|
"cpu_time": 2.2601829263440585e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5330264311979449e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8924553179174601e+06,
|
|
"cpu_time": 2.2587787043010225e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5408230253694892e+08,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.9154448670976649e+03,
|
|
"cpu_time": 2.2737453630389032e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1422302225383462e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0660577791845776e-03,
|
|
"cpu_time": 1.0060005924904416e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0643859861176265e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 361,
|
|
"real_time": 1.9445494260739568e+06,
|
|
"cpu_time": 2.4517262049861229e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0784770867121377e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 361,
|
|
"real_time": 1.9434303120410694e+06,
|
|
"cpu_time": 2.4512840443212888e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0790981220198660e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 361,
|
|
"real_time": 1.9417345914680129e+06,
|
|
"cpu_time": 2.4500562548476579e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0800405004962530e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 361,
|
|
"real_time": 1.9406903765055961e+06,
|
|
"cpu_time": 2.4489722520775483e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0806216310383980e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 361,
|
|
"real_time": 1.9415573025421549e+06,
|
|
"cpu_time": 2.4496631911356612e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0801391219585018e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9423924017261579e+06,
|
|
"cpu_time": 2.4503403894736553e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0796752924450314e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9417345914680131e+06,
|
|
"cpu_time": 2.4500562548476579e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0800405004962530e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5618103833133946e+03,
|
|
"cpu_time": 1.1425956321379379e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.6788451821656548e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.0406532785314185e-04,
|
|
"cpu_time": 4.6630077888214244e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.0383845429226708e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 354,
|
|
"real_time": 1.9760276914421616e+06,
|
|
"cpu_time": 2.7870536610169755e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1225937359910562e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 354,
|
|
"real_time": 1.9739828132340254e+06,
|
|
"cpu_time": 2.7865640423729317e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1247925624683464e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 354,
|
|
"real_time": 1.9721504456053178e+06,
|
|
"cpu_time": 2.7861275706215044e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1267667531888676e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 354,
|
|
"real_time": 1.9741960635769186e+06,
|
|
"cpu_time": 2.7866133446327327e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1245630448683050e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 354,
|
|
"real_time": 1.9738019442732977e+06,
|
|
"cpu_time": 2.7867744406779720e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1249872674252701e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9740317916263442e+06,
|
|
"cpu_time": 2.7866266118644234e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1247406727883692e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9739828132340256e+06,
|
|
"cpu_time": 2.7866133446327327e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1247925624683464e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3788847750292057e+03,
|
|
"cpu_time": 3.3814104250203224e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4840310418316727e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.9851193931035158e-04,
|
|
"cpu_time": 1.2134422353621148e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.9845278571531672e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 348,
|
|
"real_time": 2.0250472693099242e+06,
|
|
"cpu_time": 3.5182854224137696e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1424257730332322e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 348,
|
|
"real_time": 2.0076711712292687e+06,
|
|
"cpu_time": 3.5100135201149266e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1782778575556149e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 348,
|
|
"real_time": 2.0270210676746252e+06,
|
|
"cpu_time": 3.5161627298850259e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1383921133209105e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 348,
|
|
"real_time": 2.0595760380410342e+06,
|
|
"cpu_time": 3.4209819712643903e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.0729780523077092e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 348,
|
|
"real_time": 2.1280676090351208e+06,
|
|
"cpu_time": 3.3558140201148721e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9418897991701717e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0494766310579949e+06,
|
|
"cpu_time": 3.4642515327585973e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.0947927190775280e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0270210676746250e+06,
|
|
"cpu_time": 3.5100135201149271e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1383921133209105e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.7768405757844077e+04,
|
|
"cpu_time": 7.3040300871213694e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.3534440123291478e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3307611823407202e-02,
|
|
"cpu_time": 2.1084006221987989e-02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2842289351428478e-02,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 217,
|
|
"real_time": 3.2330593952591517e+06,
|
|
"cpu_time": 5.5389429447004683e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1892693417886286e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 217,
|
|
"real_time": 3.2336360159059684e+06,
|
|
"cpu_time": 5.5407397327189604e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1883439934100075e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 217,
|
|
"real_time": 3.2328050487035294e+06,
|
|
"cpu_time": 5.5395317926267991e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1896776165727234e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 217,
|
|
"real_time": 3.2336462980076191e+06,
|
|
"cpu_time": 5.5407249539170228e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1883274959098415e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 217,
|
|
"real_time": 3.2388017768220557e+06,
|
|
"cpu_time": 5.5468687880184976e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1800687896565161e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2343897069396651e+06,
|
|
"cpu_time": 5.5413616423963513e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1871374474675436e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2336360159059693e+06,
|
|
"cpu_time": 5.5407249539170237e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1883439934100075e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4933881847506741e+03,
|
|
"cpu_time": 3.1748718659660735e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.9948505644541951e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.7089912183460525e-04,
|
|
"cpu_time": 5.7294074468547168e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.7014550027097413e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 107,
|
|
"real_time": 6.4497282336875098e+06,
|
|
"cpu_time": 1.0729077588785185e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.2024567213145800e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 107,
|
|
"real_time": 6.4598218163597248e+06,
|
|
"cpu_time": 1.0763218383177523e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1943277932252293e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 107,
|
|
"real_time": 6.4698285901936414e+06,
|
|
"cpu_time": 1.0769785102803659e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1862938147787495e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 107,
|
|
"real_time": 6.4693455732314382e+06,
|
|
"cpu_time": 1.0761629336448690e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1866810359984465e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 107,
|
|
"real_time": 6.4616730941511756e+06,
|
|
"cpu_time": 1.0750865691588724e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1928396115198097e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.4620794615246989e+06,
|
|
"cpu_time": 1.0754915220560756e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1925197953673630e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.4616730941511756e+06,
|
|
"cpu_time": 1.0761629336448690e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.1928396115198097e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2258783674999322e+03,
|
|
"cpu_time": 1.5961807230101176e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.6137655908627426e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2729460255753453e-03,
|
|
"cpu_time": 1.4841406838415722e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2737102315456515e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 54,
|
|
"real_time": 1.3215825392830150e+07,
|
|
"cpu_time": 2.1883666537036631e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0779169673661013e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 54,
|
|
"real_time": 1.3218089231048469e+07,
|
|
"cpu_time": 2.1890109759259194e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0770472817179546e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 54,
|
|
"real_time": 1.3238667079282029e+07,
|
|
"cpu_time": 2.1900477888888970e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0691556482315826e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 54,
|
|
"real_time": 1.3229964714911249e+07,
|
|
"cpu_time": 2.1882920148148291e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0724900214104757e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 54,
|
|
"real_time": 1.3275241841458611e+07,
|
|
"cpu_time": 2.1925055870370246e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0551895627557497e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3235557651906101e+07,
|
|
"cpu_time": 2.1896446040740661e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0703598962963734e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3229964714911249e+07,
|
|
"cpu_time": 2.1890109759259194e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 5.0724900214104757e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4032749316137168e+04,
|
|
"cpu_time": 1.7472779541221302e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.1918049511072636e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8157715714136248e-03,
|
|
"cpu_time": 7.9797331076976322e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8128505942588780e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 28,
|
|
"real_time": 2.5522057632250447e+07,
|
|
"cpu_time": 4.1826264464285754e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2588913454375410e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 28,
|
|
"real_time": 2.5580400734075479e+07,
|
|
"cpu_time": 4.1887044607142746e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2468970050656586e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 28,
|
|
"real_time": 2.5570555190954890e+07,
|
|
"cpu_time": 4.1876360607142866e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2489172408535357e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 28,
|
|
"real_time": 2.5561915843614511e+07,
|
|
"cpu_time": 4.1881215107143328e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2506912557388859e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 28,
|
|
"real_time": 2.5565904365586381e+07,
|
|
"cpu_time": 4.1861169249999978e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2498720984291515e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5560166753296342e+07,
|
|
"cpu_time": 4.1866410807142936e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2510537891049547e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5565904365586385e+07,
|
|
"cpu_time": 4.1876360607142873e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.2498720984291515e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2392064631524805e+04,
|
|
"cpu_time": 2.4408616749659683e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6043133185481634e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.7605315128224011e-04,
|
|
"cpu_time": 5.8301192481241472e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.7683606062107606e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 14,
|
|
"real_time": 4.9962869446192466e+07,
|
|
"cpu_time": 8.0656709428570896e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3726989457459335e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 14,
|
|
"real_time": 4.9967598968318529e+07,
|
|
"cpu_time": 8.0639374785714522e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3721904102336178e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 14,
|
|
"real_time": 4.9963975591318950e+07,
|
|
"cpu_time": 8.0544187499999762e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3725800003520870e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 14,
|
|
"real_time": 4.9962455938969336e+07,
|
|
"cpu_time": 8.0624483928571865e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3727434121313429e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 14,
|
|
"real_time": 4.9923163971730642e+07,
|
|
"cpu_time": 8.0566655428571358e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3769720234880056e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.9956012783305988e+07,
|
|
"cpu_time": 8.0606282214285702e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3734369583901978e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.9962869446192473e+07,
|
|
"cpu_time": 8.0624483928571880e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3726989457459335e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8474431642528631e+04,
|
|
"cpu_time": 4.8464833414919281e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9881281207809520e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.6981397459931610e-04,
|
|
"cpu_time": 6.0125379912795375e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.6999189460604854e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 8.2339207331339523e+07,
|
|
"cpu_time": 1.5326703777777576e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.5202341557599497e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 8.2411971357133657e+07,
|
|
"cpu_time": 1.5323738988888952e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.5144772435240135e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 8.2214141885439560e+07,
|
|
"cpu_time": 1.5308496722222269e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.5301528385237808e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 8.2228523161676198e+07,
|
|
"cpu_time": 1.5297504977777660e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.5290107539012270e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 8.2198532091246709e+07,
|
|
"cpu_time": 1.5303727777777630e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.5313929378207369e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2278475165367126e+07,
|
|
"cpu_time": 1.5312034448888820e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.5250535859059410e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2228523161676198e+07,
|
|
"cpu_time": 1.5308496722222269e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.5290107539012270e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.2917299090918968e+04,
|
|
"cpu_time": 1.2696527991905920e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.3643544127642307e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1293026384381752e-03,
|
|
"cpu_time": 8.2918622174516443e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1286274228722308e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7823755741119385e+08,
|
|
"cpu_time": 2.7692496224999982e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0242175644433842e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7819091677665710e+08,
|
|
"cpu_time": 2.7695264149999589e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0257943750624418e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7814448103308678e+08,
|
|
"cpu_time": 2.7687966725000024e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0273650790257931e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7813486233353615e+08,
|
|
"cpu_time": 2.7695191499999791e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0276905370131721e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7821199074387550e+08,
|
|
"cpu_time": 2.7701930899999636e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0250818113758183e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7818396165966988e+08,
|
|
"cpu_time": 2.7694569899999803e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0260298733841219e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7819091677665710e+08,
|
|
"cpu_time": 2.7695191499999797e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0257943750624418e+09,
|
|
"dst_gpu": 1.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3806310975324610e+04,
|
|
"cpu_time": 5.0733307397219120e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4814960775648379e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4584878777695132e-04,
|
|
"cpu_time": 1.8318864521243018e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4584944128941953e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 379,
|
|
"real_time": 1.8684413580054671e+06,
|
|
"cpu_time": 2.0710293984169727e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1922015280010812e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 379,
|
|
"real_time": 1.8732680107950969e+06,
|
|
"cpu_time": 2.0758636147757180e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1865531127398470e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 379,
|
|
"real_time": 1.8686072989277872e+06,
|
|
"cpu_time": 2.0708710554090417e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1920068504229314e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 379,
|
|
"real_time": 1.8741013784578014e+06,
|
|
"cpu_time": 2.0766786015830461e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1855808053300721e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 379,
|
|
"real_time": 1.8683835485832468e+06,
|
|
"cpu_time": 2.0706497440633727e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1922693566349931e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8705603189538799e+06,
|
|
"cpu_time": 2.0730184828496303e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1897223306257850e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8686072989277872e+06,
|
|
"cpu_time": 2.0710293984169730e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1920068504229314e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8685061218371716e+03,
|
|
"cpu_time": 2.9862212638114916e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.3559263104792972e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5335010011553103e-03,
|
|
"cpu_time": 1.4405183979385200e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5325807585476973e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8618566833278495e+06,
|
|
"cpu_time": 2.0706044550264676e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3999090119856941e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8619819571473205e+06,
|
|
"cpu_time": 2.0711497407407379e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3996129868791457e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8659908634651436e+06,
|
|
"cpu_time": 2.0703124523809687e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3901608311133217e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8717089861980332e+06,
|
|
"cpu_time": 2.0755895264550245e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3767487683222881e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8663298945828394e+06,
|
|
"cpu_time": 2.0704036031747458e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3893633294831133e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8655736769442372e+06,
|
|
"cpu_time": 2.0716119555555887e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3911589855567133e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8659908634651434e+06,
|
|
"cpu_time": 2.0706044550264678e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3901608311133217e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.0343655632064342e+03,
|
|
"cpu_time": 2.2471423009854007e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.4856227340298410e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.1625334946913611e-03,
|
|
"cpu_time": 1.0847312861653844e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1601638121574980e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8723346030614623e+06,
|
|
"cpu_time": 2.0782423597883661e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7505726664509922e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8676420867060577e+06,
|
|
"cpu_time": 2.0730528068781949e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7725587876937930e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8631218444887381e+06,
|
|
"cpu_time": 2.0733030634920653e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7938424684704170e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8629390051597347e+06,
|
|
"cpu_time": 2.0738852936509138e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7947055457111858e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8675459890339028e+06,
|
|
"cpu_time": 2.0731915343915755e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7730101942365449e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8667167056899793e+06,
|
|
"cpu_time": 2.0743350116402232e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7769379325125869e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8675459890339028e+06,
|
|
"cpu_time": 2.0733030634920653e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7730101942365449e+06,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8826170335066417e+03,
|
|
"cpu_time": 2.2071543939289559e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8242688305759482e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0799176552456799e-03,
|
|
"cpu_time": 1.0640298609161062e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0784798122113553e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8778815844094083e+06,
|
|
"cpu_time": 2.0857112891246832e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7449449567026615e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8712472752204773e+06,
|
|
"cpu_time": 2.0790776206896969e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7511314743872721e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8757884051750812e+06,
|
|
"cpu_time": 2.0839732811670778e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7468921286429171e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8756695726527574e+06,
|
|
"cpu_time": 2.0835638673740905e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7470028025062140e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8709098785090144e+06,
|
|
"cpu_time": 2.0786844986736437e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7514472704646699e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8742993431933478e+06,
|
|
"cpu_time": 2.0822021114058387e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7482837265407469e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8756695726527576e+06,
|
|
"cpu_time": 2.0835638673740909e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7470028025062140e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.0712731985992968e+03,
|
|
"cpu_time": 3.1401350076313875e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.8654132807557184e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6386246998126770e-03,
|
|
"cpu_time": 1.5080836727762535e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6389864169389641e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8663469463524264e+06,
|
|
"cpu_time": 2.0859572499999108e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5114585810576662e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8718707806909424e+06,
|
|
"cpu_time": 2.0866228829786610e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5010963724648476e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8794384871766174e+06,
|
|
"cpu_time": 2.0935917872340295e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4869989333064750e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8726010131355540e+06,
|
|
"cpu_time": 2.0873962260637011e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4997310980978295e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8788699879688136e+06,
|
|
"cpu_time": 2.0932366542552323e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4880540122336447e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8738254430648710e+06,
|
|
"cpu_time": 2.0893609601063072e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4974677994320929e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8726010131355538e+06,
|
|
"cpu_time": 2.0873962260637011e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4997310980978295e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.4360266788412155e+03,
|
|
"cpu_time": 3.7370908479683521e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0151940466759299e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.9010315229522814e-03,
|
|
"cpu_time": 1.7886286378100069e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9026544485721176e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8680988533618646e+06,
|
|
"cpu_time": 2.1000179441490700e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0163310557211921e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8686686347249867e+06,
|
|
"cpu_time": 2.0998503909573988e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0141916851560995e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8734967753574736e+06,
|
|
"cpu_time": 2.1006653031915454e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9961155911245555e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8687984443107431e+06,
|
|
"cpu_time": 2.0998204441488669e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0137044687204048e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8689810011447333e+06,
|
|
"cpu_time": 2.1004678297872599e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0130193896952212e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8696087417799602e+06,
|
|
"cpu_time": 2.1001643824468283e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0106724380834952e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8687984443107429e+06,
|
|
"cpu_time": 2.1000179441490700e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0137044687204048e+07,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1983522069536762e+03,
|
|
"cpu_time": 3.8122508429088157e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.2311639289349594e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1758354343490798e-03,
|
|
"cpu_time": 1.8152154539766523e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1740905029625245e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8811526253897075e+06,
|
|
"cpu_time": 2.1280845493333801e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3935286082685244e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8811859068615984e+06,
|
|
"cpu_time": 2.1279340559999584e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3935039543079370e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8815233825395505e+06,
|
|
"cpu_time": 2.1285209946666630e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3932540112585586e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8802361295480903e+06,
|
|
"cpu_time": 2.1274917546666074e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3942078650674883e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8802199160757787e+06,
|
|
"cpu_time": 2.1275222933333376e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3942198875710389e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8808635920829452e+06,
|
|
"cpu_time": 2.1279107295999890e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3937428652947095e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8811526253897075e+06,
|
|
"cpu_time": 2.1279340559999580e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3935286082685244e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.9807577936227335e+02,
|
|
"cpu_time": 4.2708418410152888e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.4320927336868757e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.1797934835877165e-04,
|
|
"cpu_time": 2.0070587462182372e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1799931278928565e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8929786468800961e+06,
|
|
"cpu_time": 2.1791895456990390e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7696456104462820e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8961580156634052e+06,
|
|
"cpu_time": 2.1824858655914459e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7650016278657472e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8989455442429993e+06,
|
|
"cpu_time": 2.1820745806452739e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7609427852708834e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8977644664431941e+06,
|
|
"cpu_time": 2.1811381747311843e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7626610639550275e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8996152229906279e+06,
|
|
"cpu_time": 2.1837831370967850e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7599694593655437e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8970923792440644e+06,
|
|
"cpu_time": 2.1817342607527459e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7636441093806976e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8977644664431941e+06,
|
|
"cpu_time": 2.1820745806452744e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7626610639550275e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6479469193376867e+03,
|
|
"cpu_time": 1.7109584290775390e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.8609407674814179e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3957922915661153e-03,
|
|
"cpu_time": 7.8421944407070974e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3970470200472421e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 370,
|
|
"real_time": 1.8934762592244945e+06,
|
|
"cpu_time": 2.2667429999998873e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5378354753149235e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 370,
|
|
"real_time": 1.9015273652700204e+06,
|
|
"cpu_time": 2.2713653243243154e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5143881658053350e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 370,
|
|
"real_time": 1.8926610629785354e+06,
|
|
"cpu_time": 2.2660313972972143e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5402207004239082e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 370,
|
|
"real_time": 1.8987613451679717e+06,
|
|
"cpu_time": 2.2690271783784665e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5224212493500018e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 370,
|
|
"real_time": 1.8943688419263659e+06,
|
|
"cpu_time": 2.2668461135135596e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5352261755620563e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8961589749134774e+06,
|
|
"cpu_time": 2.2680026027026894e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5300183532912445e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8943688419263656e+06,
|
|
"cpu_time": 2.2668461135135600e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5352261755620563e+08,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8153922945630570e+03,
|
|
"cpu_time": 2.1890276938873253e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1116827849346234e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0121689927065081e-03,
|
|
"cpu_time": 9.6517865159358607e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0102696119859973e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 364,
|
|
"real_time": 1.9380520192258940e+06,
|
|
"cpu_time": 2.4586348901098287e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0820927298110681e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 364,
|
|
"real_time": 1.9394761820676851e+06,
|
|
"cpu_time": 2.4598195054945457e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0812981460613844e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 364,
|
|
"real_time": 1.9417583997102946e+06,
|
|
"cpu_time": 2.4602398543956289e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0800272579291480e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 364,
|
|
"real_time": 1.9399575180064498e+06,
|
|
"cpu_time": 2.4597001318681720e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0810298578883765e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 364,
|
|
"real_time": 1.9395650080776168e+06,
|
|
"cpu_time": 2.4595008324175589e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0812486259888623e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9397618254175882e+06,
|
|
"cpu_time": 2.4595790428571468e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0811393235357678e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9395650080776170e+06,
|
|
"cpu_time": 2.4597001318681720e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0812486259888623e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3293203989917583e+03,
|
|
"cpu_time": 5.9302149968538220e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.4074755969898414e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.8530083517113485e-04,
|
|
"cpu_time": 2.4110690868324534e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.8515458051829684e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9685012625865219e+06,
|
|
"cpu_time": 2.8028409073033216e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1307093267945755e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9675826253578737e+06,
|
|
"cpu_time": 2.8028866657303418e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1317041256334124e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9675610267185722e+06,
|
|
"cpu_time": 2.8043120646067862e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1317275261317358e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9678302665811370e+06,
|
|
"cpu_time": 2.8026777612359794e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1314358617357216e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9684482288672433e+06,
|
|
"cpu_time": 2.8024048342697588e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1307667321348047e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9679846820222693e+06,
|
|
"cpu_time": 2.8030244466292374e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1312687144860501e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9678302665811367e+06,
|
|
"cpu_time": 2.8028409073033216e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1314358617357216e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.6008386923340839e+02,
|
|
"cpu_time": 7.4410021860781490e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9822938074746256e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3378427354456521e-04,
|
|
"cpu_time": 2.6546333532789150e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3377126373649664e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 2.0003439698900494e+06,
|
|
"cpu_time": 3.5436564400000796e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1935827668983784e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 1.9952031284836787e+06,
|
|
"cpu_time": 3.5391827885713871e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.2043879544110394e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 2.0035233940663082e+06,
|
|
"cpu_time": 3.5427174200000470e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1869279015378308e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 1.9930815320861128e+06,
|
|
"cpu_time": 3.5393054171428117e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.2088634433433523e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 2.0013519413104015e+06,
|
|
"cpu_time": 3.5449319485714375e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1914706888122287e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9987007931673103e+06,
|
|
"cpu_time": 3.5419588028571522e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1970465510005665e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0003439698900494e+06,
|
|
"cpu_time": 3.5427174200000474e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1935827668983784e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3816620416344040e+03,
|
|
"cpu_time": 2.6001816449037306e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.2061155633502677e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.1922551172308550e-03,
|
|
"cpu_time": 7.3410838172546528e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1934747331204965e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 188,
|
|
"real_time": 3.6646647036987101e+06,
|
|
"cpu_time": 6.7922210212765569e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.5781039621624651e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 188,
|
|
"real_time": 3.7328053359933039e+06,
|
|
"cpu_time": 6.9686623297871947e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.4945327950072603e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 188,
|
|
"real_time": 3.6944860934299673e+06,
|
|
"cpu_time": 6.8745376702127447e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.5411501290627413e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 188,
|
|
"real_time": 3.6548331932243989e+06,
|
|
"cpu_time": 6.7903147872340363e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.5904190733253841e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 188,
|
|
"real_time": 3.6396076432132022e+06,
|
|
"cpu_time": 6.7505371702127177e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6096221473994789e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6772793939119168e+06,
|
|
"cpu_time": 6.8352545957446517e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.5627656213914671e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6646647036987096e+06,
|
|
"cpu_time": 6.7922210212765560e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.5781039621624651e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6950385273715510e+04,
|
|
"cpu_time": 8.7162636133903579e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.5602796613639385e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0048294218516647e-02,
|
|
"cpu_time": 1.2751922391913164e-02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.9945516376824754e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 97,
|
|
"real_time": 7.3470689701972548e+06,
|
|
"cpu_time": 1.3861465773195552e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.5670500897855501e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 97,
|
|
"real_time": 7.3321164134390578e+06,
|
|
"cpu_time": 1.3876749206185510e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.5763637820176973e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 97,
|
|
"real_time": 7.3491820595083162e+06,
|
|
"cpu_time": 1.3844524814433115e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.5657369389274454e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 97,
|
|
"real_time": 7.3406179051500624e+06,
|
|
"cpu_time": 1.3893467381443039e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.5710636943054533e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 97,
|
|
"real_time": 7.3555111645039208e+06,
|
|
"cpu_time": 1.3916840824742381e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.5618083161815195e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.3448993025597231e+06,
|
|
"cpu_time": 1.3878609599999920e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.5684045642435331e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.3470689701972548e+06,
|
|
"cpu_time": 1.3876749206185510e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.5670500897855501e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.9077732449482210e+03,
|
|
"cpu_time": 2.8024652916552215e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.5425949392969357e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2127835764669818e-03,
|
|
"cpu_time": 2.0192694891102330e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2132452065822493e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4383576886386288e+07,
|
|
"cpu_time": 2.6741641306122512e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6656589338022690e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4381203002163341e+07,
|
|
"cpu_time": 2.6737716775509935e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6664290873235655e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4367573780520838e+07,
|
|
"cpu_time": 2.6739785163265288e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6708557078011570e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4399286492594650e+07,
|
|
"cpu_time": 2.6759380306122683e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6605687048808384e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4362969587804103e+07,
|
|
"cpu_time": 2.6730821795917924e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6723529970420275e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4378921949893847e+07,
|
|
"cpu_time": 2.6741869069387667e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6671730861699715e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4381203002163341e+07,
|
|
"cpu_time": 2.6739785163265284e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6664290873235655e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4360057771318190e+04,
|
|
"cpu_time": 1.0610392888311982e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6598611195785655e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.9868806725278902e-04,
|
|
"cpu_time": 3.9677080389485792e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.9843332002983289e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7747236908628389e+07,
|
|
"cpu_time": 5.0242324346153460e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8371565227189569e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7716018474445894e+07,
|
|
"cpu_time": 5.0120845192308344e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8426049406680994e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7739046069865044e+07,
|
|
"cpu_time": 5.0101666499999128e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8385848475809898e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7760393630999785e+07,
|
|
"cpu_time": 5.0207765576922938e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8348640074800758e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7749735503815688e+07,
|
|
"cpu_time": 5.0214695538460761e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8367209835763874e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7742486117550962e+07,
|
|
"cpu_time": 5.0177459430768929e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8379862604049015e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7747236908628393e+07,
|
|
"cpu_time": 5.0207765576922946e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8371565227189569e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6644614102662716e+04,
|
|
"cpu_time": 6.2173522922803983e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9038064429159188e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.9996836736750467e-04,
|
|
"cpu_time": 1.2390727555384967e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.0020973326883594e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3590418770909309e+07,
|
|
"cpu_time": 9.3624053399997815e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0090195627602720e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3591524064540863e+07,
|
|
"cpu_time": 9.3503911999999911e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0089162546808748e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3575059026479721e+07,
|
|
"cpu_time": 9.3443534800002232e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0104556276330843e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3546115756034851e+07,
|
|
"cpu_time": 9.3356494299996480e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0131639281369591e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3651123493909836e+07,
|
|
"cpu_time": 9.3570227499998286e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0033519993383036e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.3590848222374924e+07,
|
|
"cpu_time": 9.3499644399998948e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0089814745098982e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.3590418770909309e+07,
|
|
"cpu_time": 9.3503911999999911e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0090195627602720e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8353754056867256e+04,
|
|
"cpu_time": 1.0502145499139026e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5833901852631122e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.1567730926218179e-04,
|
|
"cpu_time": 1.1232283894268096e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.1539298028921690e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.4750837422907352e+07,
|
|
"cpu_time": 1.7327072874999771e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6661336891805019e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.4940037466585636e+07,
|
|
"cpu_time": 1.7331625912499505e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6548420068714733e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.4903376884758472e+07,
|
|
"cpu_time": 1.7318609975000498e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6570264370247269e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.5178016461431980e+07,
|
|
"cpu_time": 1.7377399937500116e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6407028845526609e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.5206435769796371e+07,
|
|
"cpu_time": 1.7377864049999657e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6390191236454086e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.4995740801095963e+07,
|
|
"cpu_time": 1.7346514549999911e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6515448282549553e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.4940037466585636e+07,
|
|
"cpu_time": 1.7331625912499505e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6548420068714733e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9315046295135684e+05,
|
|
"cpu_time": 2.8788078567900800e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1491249973492004e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0332539261500074e-03,
|
|
"cpu_time": 1.6595886444462093e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0332936078010714e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.9021894410252571e+08,
|
|
"cpu_time": 3.0970684724999839e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6447680806243258e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8997678905725479e+08,
|
|
"cpu_time": 3.0744388075000018e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6519632178665676e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8961442261934280e+08,
|
|
"cpu_time": 3.0771750399999577e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6627645153110113e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8953919038176537e+08,
|
|
"cpu_time": 3.0742986949999815e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6650121900241032e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8949230760335922e+08,
|
|
"cpu_time": 3.0709498850001180e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6664137852367640e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8976833075284961e+08,
|
|
"cpu_time": 3.0787861800000083e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6581843578125553e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8961442261934280e+08,
|
|
"cpu_time": 3.0744388075000018e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6627645153110113e+09,
|
|
"dst_gpu": 2.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1578832222947828e+05,
|
|
"cpu_time": 1.0455652535274522e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.4077054440963455e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6640728248843296e-03,
|
|
"cpu_time": 3.3960307484797439e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6626721310532464e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8670792902373655e+06,
|
|
"cpu_time": 2.0685603749999690e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1938007782622166e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8627699108115034e+06,
|
|
"cpu_time": 2.0694891648935315e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1988759729405358e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8678640003890132e+06,
|
|
"cpu_time": 2.0688252606383476e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1928791384955980e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8630592680795980e+06,
|
|
"cpu_time": 2.0688544361701256e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1985344589826553e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8657218585584371e+06,
|
|
"cpu_time": 2.0669270957446946e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1953969082855699e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8652988656151835e+06,
|
|
"cpu_time": 2.0685312664893339e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1958974513933151e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8657218585584369e+06,
|
|
"cpu_time": 2.0688252606383476e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1953969082855699e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3097705943831552e+03,
|
|
"cpu_time": 9.5951180775990247e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7194757768688733e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2382844577677814e-03,
|
|
"cpu_time": 4.6386139929533923e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2384347798861661e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8696124626095062e+06,
|
|
"cpu_time": 2.0728049074073674e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3816567143364241e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8704966945127428e+06,
|
|
"cpu_time": 2.0731026084656040e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3795853924959674e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8653411385630968e+06,
|
|
"cpu_time": 2.0676225634920604e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3916899866961781e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8609660735857985e+06,
|
|
"cpu_time": 2.0684620502645874e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.4020146934840474e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8653491543008140e+06,
|
|
"cpu_time": 2.0678324285714678e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3916711148217153e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8663531047143913e+06,
|
|
"cpu_time": 2.0699649116402171e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3893235803668667e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8653491543008140e+06,
|
|
"cpu_time": 2.0684620502645872e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3916711148217153e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8355197369979942e+03,
|
|
"cpu_time": 2.7478804795926508e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.0248031396249517e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0550879291327587e-03,
|
|
"cpu_time": 1.3275009948913871e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0560806179777350e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8626591914056840e+06,
|
|
"cpu_time": 2.0713108862434549e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7960267104126364e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8685038509718929e+06,
|
|
"cpu_time": 2.0725636640211826e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7685128352708202e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8749948920606312e+06,
|
|
"cpu_time": 2.0791793068783721e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7381571381209902e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8733178530676505e+06,
|
|
"cpu_time": 2.0769698439154120e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7459797455997076e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8689673446782830e+06,
|
|
"cpu_time": 2.0725102539682805e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7663382919193171e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8696886264368284e+06,
|
|
"cpu_time": 2.0745067910053406e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7630029442646950e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8689673446782832e+06,
|
|
"cpu_time": 2.0725636640211828e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7663382919193171e+06,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.8131236723243910e+03,
|
|
"cpu_time": 3.3865448499849563e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2578507197889685e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.5742915714778848e-03,
|
|
"cpu_time": 1.6324578278887098e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5765719059431687e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8677341489084475e+06,
|
|
"cpu_time": 2.0785402739361571e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7544252761641946e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8726145788489101e+06,
|
|
"cpu_time": 2.0783888191490246e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7498528725618690e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8682298418656152e+06,
|
|
"cpu_time": 2.0785900212765888e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7539597787003476e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8729038936223160e+06,
|
|
"cpu_time": 2.0788642313830114e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7495825659598898e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8769345608930923e+06,
|
|
"cpu_time": 2.0828368191489324e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7458253837261204e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8716834048276762e+06,
|
|
"cpu_time": 2.0794440329787428e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7507291754224844e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8726145788489103e+06,
|
|
"cpu_time": 2.0785900212765888e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7498528725618690e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7899550600108487e+03,
|
|
"cpu_time": 1.9043870735942025e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5435769779189926e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0248910954894029e-03,
|
|
"cpu_time": 9.1581549846582000e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0240577627112771e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8797907300625972e+06,
|
|
"cpu_time": 2.0919673563830149e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4863455251647957e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8728583699680227e+06,
|
|
"cpu_time": 2.0853926276596151e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4992501862871222e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8683745366867331e+06,
|
|
"cpu_time": 2.0854675531914933e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5076478892833628e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8736826393077092e+06,
|
|
"cpu_time": 2.0854255079786025e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4977107982499287e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8683728094982414e+06,
|
|
"cpu_time": 2.0857108563830254e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5076511318744756e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8726158171046611e+06,
|
|
"cpu_time": 2.0867927803191501e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4997211061719365e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8728583699680227e+06,
|
|
"cpu_time": 2.0854675531914930e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4992501862871222e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.7081814585897619e+03,
|
|
"cpu_time": 2.8953792890491959e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.7876115771721496e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.5142271124620223e-03,
|
|
"cpu_time": 1.3874781034111025e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5109462470237266e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8812494213597483e+06,
|
|
"cpu_time": 2.1059939707446336e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9672845350438684e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8709361679902165e+06,
|
|
"cpu_time": 2.0990629335107072e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0056906399323717e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8754710984060869e+06,
|
|
"cpu_time": 2.0996693404254797e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9887507256920472e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8794996233717687e+06,
|
|
"cpu_time": 2.1039775531914197e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9737710170359373e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8711566314797234e+06,
|
|
"cpu_time": 2.0988910159575311e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0048652151769549e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8756625885215090e+06,
|
|
"cpu_time": 2.1015189627659544e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9880724265762359e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8754710984060869e+06,
|
|
"cpu_time": 2.0996693404254802e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9887507256920472e+07,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.7067745316853488e+03,
|
|
"cpu_time": 3.2569022944638891e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7532298977530014e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.5093929795739350e-03,
|
|
"cpu_time": 1.5497848709284331e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5088891338408549e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8763221378127735e+06,
|
|
"cpu_time": 2.1236109786667232e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3971161706037375e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8760042761374887e+06,
|
|
"cpu_time": 2.1234445893333033e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3973528916454771e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8763899917636686e+06,
|
|
"cpu_time": 2.1231572213332583e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3970656481364191e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8765960197197271e+06,
|
|
"cpu_time": 2.1234133839999838e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3969122669201422e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8763144139666108e+06,
|
|
"cpu_time": 2.1234805493333321e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3971219218308732e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8763253678800543e+06,
|
|
"cpu_time": 2.1234213445333201e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3971137798273298e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8763221378127735e+06,
|
|
"cpu_time": 2.1234445893333033e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3971161706037375e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1251999317121854e+02,
|
|
"cpu_time": 1.6570385158742909e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5824868087917826e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1326393428839688e-04,
|
|
"cpu_time": 7.8036255976247153e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1326828434741824e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 369,
|
|
"real_time": 1.8985204291050411e+06,
|
|
"cpu_time": 2.1833960867208987e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7615610133157659e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 369,
|
|
"real_time": 1.8951526401642719e+06,
|
|
"cpu_time": 2.1810286612465950e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7664684568866950e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 369,
|
|
"real_time": 1.8962279669934162e+06,
|
|
"cpu_time": 2.1799643631436876e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7648996277134877e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 369,
|
|
"real_time": 1.9024098803873521e+06,
|
|
"cpu_time": 2.1867526531166257e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7559150391568041e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 369,
|
|
"real_time": 1.8970071383705237e+06,
|
|
"cpu_time": 2.1807198672087826e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7637639806160605e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8978636110041211e+06,
|
|
"cpu_time": 2.1823723262873180e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7625216235377628e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8970071383705239e+06,
|
|
"cpu_time": 2.1810286612465950e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7637639806160605e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8221540937959262e+03,
|
|
"cpu_time": 2.7644932541816129e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.1032175677143905e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4870162836953186e-03,
|
|
"cpu_time": 1.2667376784806499e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4853159999738553e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8916318952118424e+06,
|
|
"cpu_time": 2.2629475201072870e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5432349319874990e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8976737290836780e+06,
|
|
"cpu_time": 2.2663356729222396e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5255863214501131e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8922821727073216e+06,
|
|
"cpu_time": 2.2644794798928574e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5413300147503042e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8982677927256960e+06,
|
|
"cpu_time": 2.2674526193029354e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5238570870675969e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8929380253419259e+06,
|
|
"cpu_time": 2.2645184557640627e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5394100914138126e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8945587230140928e+06,
|
|
"cpu_time": 2.2651467495978763e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5346836893338656e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8929380253419261e+06,
|
|
"cpu_time": 2.2645184557640627e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5394100914138126e+08,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1557932509370430e+03,
|
|
"cpu_time": 1.7611917923159622e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.2143052929670177e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6657141383912481e-03,
|
|
"cpu_time": 7.7751774476802441e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6648296108997727e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9407968059921067e+06,
|
|
"cpu_time": 2.4588264104682789e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0805623718697159e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9426396794581255e+06,
|
|
"cpu_time": 2.4597346969697606e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0795373028646123e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9430901344349871e+06,
|
|
"cpu_time": 2.4596828650139044e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0792870401813920e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9432020874901807e+06,
|
|
"cpu_time": 2.4594154490357400e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0792248595763190e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9409730983962186e+06,
|
|
"cpu_time": 2.4591508787878877e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0804642278313022e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9421403611543239e+06,
|
|
"cpu_time": 2.4593620600551148e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0798151604646683e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9426396794581257e+06,
|
|
"cpu_time": 2.4594154490357405e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0795373028646123e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1668647829772362e+03,
|
|
"cpu_time": 3.7952024922896987e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4888064184409135e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.0081382701078435e-04,
|
|
"cpu_time": 1.5431654224204173e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.0091825490287028e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9738303623005245e+06,
|
|
"cpu_time": 2.8019953286516694e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1249566731314666e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9760834085205672e+06,
|
|
"cpu_time": 2.8026008764045541e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1225338879496720e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9750415908010527e+06,
|
|
"cpu_time": 2.8022730140449903e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1236535066073425e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9766600400973321e+06,
|
|
"cpu_time": 2.8064233033706443e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1219147020311441e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9719615906314775e+06,
|
|
"cpu_time": 2.8021695898876823e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1269704338697922e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9747153984701908e+06,
|
|
"cpu_time": 2.8030924224719079e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1240058407178838e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9750415908010527e+06,
|
|
"cpu_time": 2.8022730140449903e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1236535066073425e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8792439054526342e+03,
|
|
"cpu_time": 1.8750366726268039e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0221609800211259e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.5165303663934640e-04,
|
|
"cpu_time": 6.6891717789786679e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.5205057408771733e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 1.9999032883372689e+06,
|
|
"cpu_time": 3.5363140914284936e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1945068288648782e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 2.0029872123684201e+06,
|
|
"cpu_time": 3.5387059228569991e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1880487045551033e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 1.9966365443542600e+06,
|
|
"cpu_time": 3.5362213457142389e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.2013695600833511e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 2.0053163396992853e+06,
|
|
"cpu_time": 3.5380761999999774e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1831843853912568e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 350,
|
|
"real_time": 1.9996321996274805e+06,
|
|
"cpu_time": 3.5370156342857429e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1950754751612563e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0008951168773430e+06,
|
|
"cpu_time": 3.5372666388570904e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1924369908111687e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9999032883372691e+06,
|
|
"cpu_time": 3.5370156342857429e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1945068288648782e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3405912108659563e+03,
|
|
"cpu_time": 1.0942937597584432e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.9982924057572791e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6695483849645172e-03,
|
|
"cpu_time": 3.0936140005323866e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6692659713421771e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 193,
|
|
"real_time": 3.6235311613876587e+06,
|
|
"cpu_time": 6.6986900362694236e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6300736085225325e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 193,
|
|
"real_time": 3.6225562789742812e+06,
|
|
"cpu_time": 6.7116815129535394e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6313196284559679e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 193,
|
|
"real_time": 3.5965673775059879e+06,
|
|
"cpu_time": 6.6555278031087564e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6647856800708771e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 193,
|
|
"real_time": 3.6447673174673258e+06,
|
|
"cpu_time": 6.7583387046631398e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6030965871528234e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 193,
|
|
"real_time": 3.6350113769192151e+06,
|
|
"cpu_time": 6.7423268963730298e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6154507538898573e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6244867024508938e+06,
|
|
"cpu_time": 6.7133129906735765e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6289452516184120e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6235311613876582e+06,
|
|
"cpu_time": 6.7116815129535394e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6300736085225325e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8070697392518072e+04,
|
|
"cpu_time": 4.0077244357433010e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3139263654731970e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.9857259457728427e-03,
|
|
"cpu_time": 5.9698161568081266e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9988199032256478e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.2283375387390452e+06,
|
|
"cpu_time": 1.3609573505050896e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6420676704941807e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.2289904805295393e+06,
|
|
"cpu_time": 1.3589767555555280e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6416483865036249e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.2282680846524965e+06,
|
|
"cpu_time": 1.3653700262626568e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6421122746187067e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.2317170120324148e+06,
|
|
"cpu_time": 1.3650137303030612e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6398983732591887e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.2333864783021538e+06,
|
|
"cpu_time": 1.3597001131312931e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6388274842845144e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.2301399188511316e+06,
|
|
"cpu_time": 1.3620035951515257e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6409108378320436e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.2289904805295393e+06,
|
|
"cpu_time": 1.3609573505050894e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6416483865036249e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2967867189982244e+03,
|
|
"cpu_time": 2.9981739138543089e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4740366936002644e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.1766836392886622e-04,
|
|
"cpu_time": 2.2012966225105706e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1761797308927535e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4320635179779967e+07,
|
|
"cpu_time": 2.6716827979591511e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6861653241997547e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4323028830849394e+07,
|
|
"cpu_time": 2.6653343346938577e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6853821766705379e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4295906949864358e+07,
|
|
"cpu_time": 2.6659620448979191e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6942711809296398e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4315603376955403e+07,
|
|
"cpu_time": 2.6654741693877839e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6878124681792145e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4320383665665071e+07,
|
|
"cpu_time": 2.6699676408163805e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6862476290284023e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4315111600622838e+07,
|
|
"cpu_time": 2.6676841975510191e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6879757558015108e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4320383665665071e+07,
|
|
"cpu_time": 2.6659620448979191e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6862476290284023e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1068509224213660e+04,
|
|
"cpu_time": 2.9400173491959675e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.6280083305747798e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.7320453608842820e-04,
|
|
"cpu_time": 1.1020859785033608e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.7389656422284406e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7537315367506102e+07,
|
|
"cpu_time": 4.9884253153845310e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8740309724736738e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7598084595340949e+07,
|
|
"cpu_time": 4.9912863000000790e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8632986661204147e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7612626337661192e+07,
|
|
"cpu_time": 4.9890672115382656e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8607374886661482e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7461407634501275e+07,
|
|
"cpu_time": 4.9731827884615906e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8875035754312496e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7600921618823823e+07,
|
|
"cpu_time": 4.9920327769231014e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8627987809096756e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7562071110766672e+07,
|
|
"cpu_time": 4.9867988784615144e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8696738967202330e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7598084595340949e+07,
|
|
"cpu_time": 4.9890672115382656e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8632986661204147e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.3458936114597273e+04,
|
|
"cpu_time": 7.7575637936477040e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1232227334223609e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3024008558561507e-03,
|
|
"cpu_time": 1.5556199443200731e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3065666351475013e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 15,
|
|
"real_time": 4.5961602528889976e+07,
|
|
"cpu_time": 9.4062428800001875e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.8404285584096022e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 15,
|
|
"real_time": 4.5911916593710579e+07,
|
|
"cpu_time": 9.3949280466665640e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.8467490777061710e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 15,
|
|
"real_time": 4.5977101226647697e+07,
|
|
"cpu_time": 9.4076618733333811e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.8384597731972389e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 15,
|
|
"real_time": 4.5911958316961922e+07,
|
|
"cpu_time": 9.4025856133331820e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.8467437643762627e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 15,
|
|
"real_time": 4.5990934967994690e+07,
|
|
"cpu_time": 9.4034116999997273e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.8367036066304264e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5950702726840973e+07,
|
|
"cpu_time": 9.4029660226666078e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.8418169560639410e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5961602528889969e+07,
|
|
"cpu_time": 9.4034116999997258e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.8404285584096022e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6877544389991221e+04,
|
|
"cpu_time": 4.9432075910834610e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6889212850377429e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.0254581979330875e-04,
|
|
"cpu_time": 5.2570726929858727e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.0264775844620295e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.9065597686502665e+07,
|
|
"cpu_time": 1.7557119044444865e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4193476296277037e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.9016334447595805e+07,
|
|
"cpu_time": 1.7433187711111486e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4220438980614643e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.8973720437950552e+07,
|
|
"cpu_time": 1.7495808200000182e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4243784069588413e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.8947587112585708e+07,
|
|
"cpu_time": 1.7493660744443989e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4258110547873316e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.8818984296586782e+07,
|
|
"cpu_time": 1.7428602655555400e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4328721937546120e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.8964444796244308e+07,
|
|
"cpu_time": 1.7481675671111184e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4248906366379910e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.8973720437950566e+07,
|
|
"cpu_time": 1.7493660744443992e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4243784069588413e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.2808161157102702e+04,
|
|
"cpu_time": 5.2921876811289147e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.0902762701920215e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.3779297552957878e-04,
|
|
"cpu_time": 3.0272771218804611e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.3831868900985942e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.9238923862576485e+08,
|
|
"cpu_time": 3.1406285850002289e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.5810908742595549e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.9199131801724434e+08,
|
|
"cpu_time": 3.1226287100000149e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.5926582258451824e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.9233843684196472e+08,
|
|
"cpu_time": 3.1326795974999303e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.5825649913243408e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.9228979945182800e+08,
|
|
"cpu_time": 3.1257211150000328e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.5839770339403334e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.9205844029784203e+08,
|
|
"cpu_time": 3.1155585950000387e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.5907036542359371e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9221344664692879e+08,
|
|
"cpu_time": 3.1274433205000496e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.5861989559210701e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9228979945182800e+08,
|
|
"cpu_time": 3.1257211150000328e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.5839770339403334e+09,
|
|
"dst_gpu": 3.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7728734857977860e+05,
|
|
"cpu_time": 9.6002989928890741e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.1538725506321946e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.2234623369213346e-04,
|
|
"cpu_time": 3.0696955976660433e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.2260812607996483e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8639886515899058e+06,
|
|
"cpu_time": 2.0652947857145052e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1974382711537867e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8664766885098452e+06,
|
|
"cpu_time": 2.0679924391533714e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1945090582782249e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8638659039521208e+06,
|
|
"cpu_time": 2.0649462486772868e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1975829866917394e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8625787057442327e+06,
|
|
"cpu_time": 2.0677845343912949e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1991017009739499e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8646553246418035e+06,
|
|
"cpu_time": 2.0659817592592880e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1966526177092986e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8643130548875816e+06,
|
|
"cpu_time": 2.0663999534391493e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1970569269613996e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8639886515899054e+06,
|
|
"cpu_time": 2.0659817592592877e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1974382711537867e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4240441921017978e+03,
|
|
"cpu_time": 1.4109052741708197e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6776742680589196e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.6384392008007913e-04,
|
|
"cpu_time": 6.8278421697727152e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.6360072762392966e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 380,
|
|
"real_time": 1.8637917160192861e+06,
|
|
"cpu_time": 2.0663444157892889e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3953409222660325e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 380,
|
|
"real_time": 1.8673588665957430e+06,
|
|
"cpu_time": 2.0699726157895264e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3869446556538371e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 380,
|
|
"real_time": 1.8637276698044795e+06,
|
|
"cpu_time": 2.0662974868421792e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3954919663018193e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 380,
|
|
"real_time": 1.8663908518838549e+06,
|
|
"cpu_time": 2.0687060052631225e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3892199705819106e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 380,
|
|
"real_time": 1.8638868414104525e+06,
|
|
"cpu_time": 2.0664696500001643e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3951166014997438e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8650311891427631e+06,
|
|
"cpu_time": 2.0675580347368564e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3924228232606687e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8638868414104525e+06,
|
|
"cpu_time": 2.0664696500001638e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3951166014997438e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7184115571057398e+03,
|
|
"cpu_time": 1.6877848845477645e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0453667206196369e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.2138489002727348e-04,
|
|
"cpu_time": 8.1631802164265415e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.2098754682651469e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8685889693362231e+06,
|
|
"cpu_time": 2.0731623085106565e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7681134100989960e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8630114517052672e+06,
|
|
"cpu_time": 2.0711413882980153e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7943635477942228e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8647127153405775e+06,
|
|
"cpu_time": 2.0690714414895605e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7863400432744771e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8674540294941447e+06,
|
|
"cpu_time": 2.0720021276595599e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7734422059310842e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8648582793870680e+06,
|
|
"cpu_time": 2.0692311808509300e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7856542135657668e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8657250890526562e+06,
|
|
"cpu_time": 2.0709216893617448e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7815826841329094e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8648582793870675e+06,
|
|
"cpu_time": 2.0711413882980153e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7856542135657668e+06,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2545028559901111e+03,
|
|
"cpu_time": 1.7689845195933565e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0609536067444526e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2083789134951608e-03,
|
|
"cpu_time": 8.5420155126124310e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2081576236382164e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8717140183674139e+06,
|
|
"cpu_time": 2.0780755013263612e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7506948005113304e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8720959810140501e+06,
|
|
"cpu_time": 2.0784891458883169e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7503376072764549e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8686485232728834e+06,
|
|
"cpu_time": 2.0749204827585774e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7535667939633615e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8662506171294663e+06,
|
|
"cpu_time": 2.0760886445623811e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7558199150337804e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8697693642594453e+06,
|
|
"cpu_time": 2.0756915305038870e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7525156110886615e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8696957008086522e+06,
|
|
"cpu_time": 2.0766530610079046e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7525869455747180e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8697693642594453e+06,
|
|
"cpu_time": 2.0760886445623811e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7525156110886615e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3877583358566435e+03,
|
|
"cpu_time": 1.5523754675648288e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2392531520366898e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2770839312642838e-03,
|
|
"cpu_time": 7.4753722550621054e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2776844867473218e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8744713320941553e+06,
|
|
"cpu_time": 2.0870646525197262e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4962391196873263e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8724966032882296e+06,
|
|
"cpu_time": 2.0848250106098626e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4999262420510881e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8762211725914008e+06,
|
|
"cpu_time": 2.0885628726789861e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4929783842852026e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8782413347733244e+06,
|
|
"cpu_time": 2.0909736312997383e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4892214747211494e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8752102160309714e+06,
|
|
"cpu_time": 2.0874977427056436e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4948615061788678e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8753281317556165e+06,
|
|
"cpu_time": 2.0877847819627915e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4946453453847267e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8752102160309714e+06,
|
|
"cpu_time": 2.0874977427056439e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4948615061788678e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1242549868000492e+03,
|
|
"cpu_time": 2.2436248386008642e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.9582775266680073e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1327377597708172e-03,
|
|
"cpu_time": 1.0746437362626825e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1326693084594652e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8741738105668349e+06,
|
|
"cpu_time": 2.1035854069149285e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9935882819938615e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8774260984850735e+06,
|
|
"cpu_time": 2.1036379547872893e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9814732044986576e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8744180574201969e+06,
|
|
"cpu_time": 2.1035378989361012e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9926769794566154e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8767654669608553e+06,
|
|
"cpu_time": 2.1023755851061614e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9839307205631688e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8788067509957089e+06,
|
|
"cpu_time": 2.1048947819148805e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9763428266656980e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8763180368857342e+06,
|
|
"cpu_time": 2.1036063255318720e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9856024026356012e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8767654669608555e+06,
|
|
"cpu_time": 2.1035854069149285e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9839307205631688e+07,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9893039345038155e+03,
|
|
"cpu_time": 8.9176258630354312e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.4062212303238688e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0602168158046451e-03,
|
|
"cpu_time": 4.2392085224314551e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0602122484854809e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 374,
|
|
"real_time": 1.8849163934841184e+06,
|
|
"cpu_time": 2.1306292727272948e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3907460347111079e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 374,
|
|
"real_time": 1.8847020789853267e+06,
|
|
"cpu_time": 2.1301222032086095e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3909041801510152e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 374,
|
|
"real_time": 1.8878926385759332e+06,
|
|
"cpu_time": 2.1334311336896792e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3885535365916747e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 374,
|
|
"real_time": 1.8858160739696093e+06,
|
|
"cpu_time": 2.1312160989305452e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3900825410199818e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 374,
|
|
"real_time": 1.8896236125878457e+06,
|
|
"cpu_time": 2.1346341417113687e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3872815636601457e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8865901595205669e+06,
|
|
"cpu_time": 2.1320065700534997e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3895135712267852e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8858160739696093e+06,
|
|
"cpu_time": 2.1312160989305452e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3900825410199818e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1132817190626965e+03,
|
|
"cpu_time": 1.9368883095076144e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5556357480142967e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1201594095029831e-03,
|
|
"cpu_time": 9.0848139809391454e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1195541952431917e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 370,
|
|
"real_time": 1.9110426122662486e+06,
|
|
"cpu_time": 2.1948642837837478e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7434657743097758e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 370,
|
|
"real_time": 1.9093984511454364e+06,
|
|
"cpu_time": 2.1946339594595321e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7458281412425095e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 370,
|
|
"real_time": 1.9078443045539127e+06,
|
|
"cpu_time": 2.1926585243243831e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7480649167678684e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 370,
|
|
"real_time": 1.9078737242002594e+06,
|
|
"cpu_time": 2.1930930405405150e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7480225412704951e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 370,
|
|
"real_time": 1.9066027824206881e+06,
|
|
"cpu_time": 2.1919046783781401e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7498543736223125e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9085523749173090e+06,
|
|
"cpu_time": 2.1934308972972636e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7470471494425923e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9078737242002594e+06,
|
|
"cpu_time": 2.1930930405405154e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7480225412704951e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7087984355574804e+03,
|
|
"cpu_time": 1.2788673471747002e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4586569492306162e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.9533745995915698e-04,
|
|
"cpu_time": 5.8304428407136746e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.9501811052988519e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 368,
|
|
"real_time": 1.9222639050934482e+06,
|
|
"cpu_time": 2.2832516929348093e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4549013651121163e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 368,
|
|
"real_time": 1.9147634084220042e+06,
|
|
"cpu_time": 2.2780098749999697e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4762692632827830e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 368,
|
|
"real_time": 1.9153294891546702e+06,
|
|
"cpu_time": 2.2765256494565732e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4746507373140717e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 368,
|
|
"real_time": 1.9157951501256321e+06,
|
|
"cpu_time": 2.2788846141303363e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4733200464112127e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 368,
|
|
"real_time": 1.9156942877958973e+06,
|
|
"cpu_time": 2.2766998070653328e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4736082196415555e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9167692481183310e+06,
|
|
"cpu_time": 2.2786743277174039e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4705499263523483e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9156942877958971e+06,
|
|
"cpu_time": 2.2780098749999697e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4736082196415555e+08,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.0980623228453123e+03,
|
|
"cpu_time": 2.7364913755485304e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.8237131825552892e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6162938370837504e-03,
|
|
"cpu_time": 1.2009137691430135e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6129481133240954e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 359,
|
|
"real_time": 1.9601109346198465e+06,
|
|
"cpu_time": 2.4685241030640788e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0699149537710896e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 359,
|
|
"real_time": 1.9619637005184169e+06,
|
|
"cpu_time": 2.4714361420611469e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0689045875037657e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 359,
|
|
"real_time": 1.9619449216568931e+06,
|
|
"cpu_time": 2.4699144038997153e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0689148185816157e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 359,
|
|
"real_time": 1.9594253878132508e+06,
|
|
"cpu_time": 2.4675389136491558e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0702892863608625e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 359,
|
|
"real_time": 1.9606510185361593e+06,
|
|
"cpu_time": 2.4688701337048477e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0696202333680747e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9608191926289133e+06,
|
|
"cpu_time": 2.4692567392757894e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0695287759170817e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9606510185361591e+06,
|
|
"cpu_time": 2.4688701337048477e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0696202333680747e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1235844373091304e+03,
|
|
"cpu_time": 1.4848791049676195e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.1286799557490356e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.7301787004783249e-04,
|
|
"cpu_time": 6.0134658391298806e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.7302618627478414e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 346,
|
|
"real_time": 2.0320174924227158e+06,
|
|
"cpu_time": 2.8737728959540119e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0641082154264591e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 346,
|
|
"real_time": 2.0318283568906216e+06,
|
|
"cpu_time": 2.8757037427746155e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0643003557735019e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 346,
|
|
"real_time": 2.0314583728477531e+06,
|
|
"cpu_time": 2.8768404710981906e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0646763212382796e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 346,
|
|
"real_time": 2.0318870549426668e+06,
|
|
"cpu_time": 2.8766681820809282e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0642407213516843e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 346,
|
|
"real_time": 2.0327097033096661e+06,
|
|
"cpu_time": 2.8744909450866631e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0634053122149303e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0319801960826847e+06,
|
|
"cpu_time": 2.8754952473988822e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0641461852009709e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0318870549426668e+06,
|
|
"cpu_time": 2.8757037427746155e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.0642407213516843e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5760291471792641e+02,
|
|
"cpu_time": 1.3420125300773536e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6478068944395700e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.2520047961102559e-04,
|
|
"cpu_time": 4.6670657212572776e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2516849473948702e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 327,
|
|
"real_time": 2.1417730470794602e+06,
|
|
"cpu_time": 3.7020312629969595e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9166652187722578e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 327,
|
|
"real_time": 2.1495386934625586e+06,
|
|
"cpu_time": 3.7097361498473673e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9025154678594375e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 327,
|
|
"real_time": 2.1446041509349593e+06,
|
|
"cpu_time": 3.7029459449541336e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9114948072551808e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 327,
|
|
"real_time": 2.1455915654021162e+06,
|
|
"cpu_time": 3.7045710152902422e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9096947132283525e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 327,
|
|
"real_time": 2.1377529732495639e+06,
|
|
"cpu_time": 3.7017068012234550e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9240305615146041e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1438520860257321e+06,
|
|
"cpu_time": 3.7041982348624319e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9128801537259660e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1446041509349593e+06,
|
|
"cpu_time": 3.7029459449541336e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.9114948072551808e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.4002975428615928e+03,
|
|
"cpu_time": 3.2893276780297406e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.0335428113479298e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0525191880279648e-03,
|
|
"cpu_time": 8.8799990429019285e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0531021896231963e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 177,
|
|
"real_time": 3.9675992671505543e+06,
|
|
"cpu_time": 7.0536578248587018e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.2285560789633479e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 177,
|
|
"real_time": 3.9659306370142824e+06,
|
|
"cpu_time": 7.0523577570624296e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.2303352064247360e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 177,
|
|
"real_time": 3.9842628931763484e+06,
|
|
"cpu_time": 7.0695491694915406e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.2108707306271167e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 177,
|
|
"real_time": 3.9656394645361239e+06,
|
|
"cpu_time": 7.0496105706212604e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.2306458138807368e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 177,
|
|
"real_time": 3.9674536953811762e+06,
|
|
"cpu_time": 7.0574188361582719e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.2287112309670238e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.9701771914516971e+06,
|
|
"cpu_time": 7.0565188316384424e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.2258238121725922e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.9674536953811767e+06,
|
|
"cpu_time": 7.0536578248587009e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.2287112309670238e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9229777825477886e+03,
|
|
"cpu_time": 7.8074273029595224e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.4113241438051835e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9956232179276479e-03,
|
|
"cpu_time": 1.1064134439710306e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9904578415162865e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 92,
|
|
"real_time": 7.7330696588868033e+06,
|
|
"cpu_time": 1.3808800032608282e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.3390831170697422e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 92,
|
|
"real_time": 7.8457016953388629e+06,
|
|
"cpu_time": 1.3913975760869868e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.2767917138545694e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 92,
|
|
"real_time": 7.7325444587546848e+06,
|
|
"cpu_time": 1.3845255500000250e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.3393778308005867e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 92,
|
|
"real_time": 7.7584001954401964e+06,
|
|
"cpu_time": 1.3856530902173696e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.3249163686762085e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 92,
|
|
"real_time": 7.6930886134505272e+06,
|
|
"cpu_time": 1.3769375510868849e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.3616333680771246e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7525609243742153e+06,
|
|
"cpu_time": 1.3838787541304190e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.3283604796956472e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7330696588868024e+06,
|
|
"cpu_time": 1.3845255500000250e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.3390831170697422e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.7067557323099114e+04,
|
|
"cpu_time": 5.4146122560091244e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1680599702040996e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.3611233603695413e-03,
|
|
"cpu_time": 3.9126348604227815e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.3193071257938863e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.5111502933692425e+07,
|
|
"cpu_time": 2.6988733276597332e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.4409126143485622e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.5111026354134083e+07,
|
|
"cpu_time": 2.6984515382979866e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.4410526742043781e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.5106583846376298e+07,
|
|
"cpu_time": 2.6987549170211848e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.4423586882680817e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.5117157587504134e+07,
|
|
"cpu_time": 2.7048439361702219e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.4392514671853590e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.5135530422025539e+07,
|
|
"cpu_time": 2.7081101276595060e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.4338627143414669e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5116360228746494e+07,
|
|
"cpu_time": 2.7018067693617266e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.4394876316695700e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5111502933692425e+07,
|
|
"cpu_time": 2.6988733276597332e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.4409126143485622e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1355101024056668e+04,
|
|
"cpu_time": 4.4196483589654214e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.3323553370191483e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.5117957313976204e-04,
|
|
"cpu_time": 1.6358121569180601e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.5061710122749923e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 27,
|
|
"real_time": 2.6227649922172230e+07,
|
|
"cpu_time": 5.0114047481483966e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.1174134319421253e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 27,
|
|
"real_time": 2.6252597846366741e+07,
|
|
"cpu_time": 5.0199289259260833e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.1125503382734833e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 27,
|
|
"real_time": 2.6204787746623710e+07,
|
|
"cpu_time": 5.0047507074071757e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.1218780818895559e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 27,
|
|
"real_time": 2.6195371316538915e+07,
|
|
"cpu_time": 5.0038670481479101e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.1237192394848490e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 27,
|
|
"real_time": 2.6193670652530812e+07,
|
|
"cpu_time": 5.0060795333331920e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.1240519047692919e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6214815496846482e+07,
|
|
"cpu_time": 5.0092061925925508e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.1199225992718611e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6204787746623710e+07,
|
|
"cpu_time": 5.0060795333331920e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.1218780818895559e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5089663067190530e+04,
|
|
"cpu_time": 6.6697913466625032e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.8973387158006541e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.5707952131910648e-04,
|
|
"cpu_time": 1.3315066480045423e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.5652592804765018e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.4834676906466484e+07,
|
|
"cpu_time": 9.4406130099991962e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 4.8953594904530973e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.4790101200342178e+07,
|
|
"cpu_time": 9.4507421200000852e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 4.8993422191073370e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.4820037260651588e+07,
|
|
"cpu_time": 9.4570215999999613e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 4.8966667921744747e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.4824643582105637e+07,
|
|
"cpu_time": 9.4585917400002018e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 4.8962553782586813e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.4862899705767632e+07,
|
|
"cpu_time": 9.4354506199999839e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 4.8928411994195032e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.4826471731066704e+07,
|
|
"cpu_time": 9.4484838179998860e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 4.8960930158826189e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.4824643582105637e+07,
|
|
"cpu_time": 9.4507421200000852e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 4.8962553782586813e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6276806725520517e+04,
|
|
"cpu_time": 1.0148615733838618e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3465545081075784e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.7927225473148779e-04,
|
|
"cpu_time": 1.0740999222017971e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.7927081869063813e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.1086676269769669e+08,
|
|
"cpu_time": 1.8505674387500903e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 4.8424874952279434e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.1094672884792089e+08,
|
|
"cpu_time": 1.8512826562499639e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 4.8389972158251762e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.1101074330508709e+08,
|
|
"cpu_time": 1.8540257287499174e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 4.8362068031968374e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.1111070495098829e+08,
|
|
"cpu_time": 1.8563250799999765e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 4.8318558705645647e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.1100906413048506e+08,
|
|
"cpu_time": 1.8502380837500709e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 4.8362799579044981e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1098880078643560e+08,
|
|
"cpu_time": 1.8524877975000042e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 4.8371654685438042e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1100906413048506e+08,
|
|
"cpu_time": 1.8512826562499639e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 4.8362799579044981e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.0022282909635222e+04,
|
|
"cpu_time": 2.6120549683511429e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.9234635585207110e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.1109339205183320e-04,
|
|
"cpu_time": 1.4100254651481113e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.1110798959330274e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 2.1601998060941696e+08,
|
|
"cpu_time": 3.3632168775000083e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 4.9705671714757681e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 2.1577393636107445e+08,
|
|
"cpu_time": 3.3416782349999607e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 4.9762350453819809e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 2.1595014259219170e+08,
|
|
"cpu_time": 3.3811726425000191e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 4.9721746469401236e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 2.1559766307473183e+08,
|
|
"cpu_time": 3.3454117850001806e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 4.9803036298580513e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 2.1577141061425209e+08,
|
|
"cpu_time": 3.3467933274999952e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 4.9762932954987011e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1582262665033340e+08,
|
|
"cpu_time": 3.3556545735000336e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 4.9751147578309250e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1577393636107445e+08,
|
|
"cpu_time": 3.3467933274999952e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 4.9762350453819809e+09,
|
|
"dst_gpu": 4.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6644168463491328e+05,
|
|
"cpu_time": 1.6485863050275822e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.8370946655822815e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/4/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.7119664058474728e-04,
|
|
"cpu_time": 4.9128605728570706e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.7125751914418090e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 379,
|
|
"real_time": 1.8651022260452958e+06,
|
|
"cpu_time": 2.0669402189973653e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1961262727593379e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 379,
|
|
"real_time": 1.8696540108758104e+06,
|
|
"cpu_time": 2.0714253693931894e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1907796716256035e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 379,
|
|
"real_time": 1.8654636094407460e+06,
|
|
"cpu_time": 2.0669510052769661e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1957008323673247e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 379,
|
|
"real_time": 1.8700166547205301e+06,
|
|
"cpu_time": 2.0713169313984744e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1903548236644655e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 379,
|
|
"real_time": 1.8660478902618843e+06,
|
|
"cpu_time": 2.0670045118734532e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1950133334601396e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8672568782688535e+06,
|
|
"cpu_time": 2.0687276073878896e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1935949867753745e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8660478902618843e+06,
|
|
"cpu_time": 2.0670045118734532e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1950133334601396e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3813127729268099e+03,
|
|
"cpu_time": 2.4136408773696648e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7963528960134786e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2753000407391944e-03,
|
|
"cpu_time": 1.1667272524183527e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2747808564807898e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8648549035812416e+06,
|
|
"cpu_time": 2.0679735119999046e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3928350587856434e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8606232092715800e+06,
|
|
"cpu_time": 2.0687425786666912e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.4028258699444616e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8655226393602788e+06,
|
|
"cpu_time": 2.0678672079999766e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3912627095263684e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8696528168705604e+06,
|
|
"cpu_time": 2.0726634346665379e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3815621414203709e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8648084123657707e+06,
|
|
"cpu_time": 2.0677338879998690e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3929445757954838e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8650923962898864e+06,
|
|
"cpu_time": 2.0689961242665965e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3922860710944654e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8648549035812418e+06,
|
|
"cpu_time": 2.0679735119999046e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3928350587856434e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2051999637651502e+03,
|
|
"cpu_time": 2.0872756832840646e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.5477200578246848e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7185207393162168e-03,
|
|
"cpu_time": 1.0088349894922825e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7184035683595516e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8711508736097519e+06,
|
|
"cpu_time": 2.0750059179894375e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7561084630191363e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8714133419043189e+06,
|
|
"cpu_time": 2.0755244285714258e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7548804067667462e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8660287680667974e+06,
|
|
"cpu_time": 2.0705873624338801e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7801433077442832e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8615732848172742e+06,
|
|
"cpu_time": 2.0710302275131859e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.8011576732571106e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8665286814635335e+06,
|
|
"cpu_time": 2.0704858756615492e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7777917171641886e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8673389899723355e+06,
|
|
"cpu_time": 2.0725267624338958e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7740163135902938e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8665286814635335e+06,
|
|
"cpu_time": 2.0710302275131859e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7777917171641886e+06,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.0849552563674129e+03,
|
|
"cpu_time": 2.5148718896573614e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9204954934168356e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.1875809792992820e-03,
|
|
"cpu_time": 1.2134327697192158e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1888442245566971e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8746223941558679e+06,
|
|
"cpu_time": 2.0818033448276476e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7479786917170193e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8692842521372780e+06,
|
|
"cpu_time": 2.0766075729444970e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7529704196958888e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8746113732995389e+06,
|
|
"cpu_time": 2.0812850530502081e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7479889680987280e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8744564772906543e+06,
|
|
"cpu_time": 2.0810801061007148e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7481334134448923e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8694892260534700e+06,
|
|
"cpu_time": 2.0764142413795153e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7527782210959256e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8724927445873623e+06,
|
|
"cpu_time": 2.0794380636605166e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7499699428104907e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8744564772906541e+06,
|
|
"cpu_time": 2.0810801061007145e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7481334134448923e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8370666372754395e+03,
|
|
"cpu_time": 2.6859545354842626e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6528979581888369e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5151282404037506e-03,
|
|
"cpu_time": 1.2916732565508931e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5159677279532149e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8652612754705471e+06,
|
|
"cpu_time": 2.0839256843501593e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5135024171596184e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8704377289395777e+06,
|
|
"cpu_time": 2.0838542811671549e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5037787671848796e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8759414295644921e+06,
|
|
"cpu_time": 2.0900922413792510e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4934992621392481e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8696097635237291e+06,
|
|
"cpu_time": 2.0836235517240444e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5053304319764383e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8762865145700003e+06,
|
|
"cpu_time": 2.0899620265253065e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4928567407531187e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8715073424136695e+06,
|
|
"cpu_time": 2.0862915570291833e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5017935238426603e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8704377289395779e+06,
|
|
"cpu_time": 2.0839256843501590e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5037787671848796e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.6438089496794209e+03,
|
|
"cpu_time": 3.4122378039111354e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.6920935652393324e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4813201874433117e-03,
|
|
"cpu_time": 1.6355517484670549e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4821833457790922e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8671045489757857e+06,
|
|
"cpu_time": 2.0979116781916218e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0200675196201801e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8674611559269601e+06,
|
|
"cpu_time": 2.0980959601063617e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0187269804248869e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8714126523031774e+06,
|
|
"cpu_time": 2.0976851329787727e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0039069062981695e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8679820252068471e+06,
|
|
"cpu_time": 2.0979832207447235e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0167698741900921e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8685143988593754e+06,
|
|
"cpu_time": 2.0984809627659395e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0147706691482916e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8684949562544290e+06,
|
|
"cpu_time": 2.0980313909574840e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0148483899363235e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8679820252068471e+06,
|
|
"cpu_time": 2.0979832207447235e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0167698741900921e+07,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7159838303355587e+03,
|
|
"cpu_time": 2.9275846981742723e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4361670532623684e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.1837755547138705e-04,
|
|
"cpu_time": 1.3953960416379676e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.1750622329854687e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8795345825678173e+06,
|
|
"cpu_time": 2.1262997872340409e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3947282610882279e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8800073067587677e+06,
|
|
"cpu_time": 2.1262349335104390e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3943775593720973e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8794826666364432e+06,
|
|
"cpu_time": 2.1259200585107952e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3947667869113034e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8794409095673782e+06,
|
|
"cpu_time": 2.1262278962767404e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3947977755807284e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8793460398267191e+06,
|
|
"cpu_time": 2.1259659175530518e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3948681852341062e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8795623010714252e+06,
|
|
"cpu_time": 2.1261297186170137e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3947077136372930e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8794826666364432e+06,
|
|
"cpu_time": 2.1262278962767404e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3947667869113034e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5818832285281144e+02,
|
|
"cpu_time": 1.7350828981279466e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9155676182270363e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3736619568589656e-04,
|
|
"cpu_time": 8.1607574690060222e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3734545234796039e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.9012148767544054e+06,
|
|
"cpu_time": 2.1847029892182150e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7576472623390186e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8991115841119627e+06,
|
|
"cpu_time": 2.1846381239892822e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7607013952535105e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8963535885656669e+06,
|
|
"cpu_time": 2.1813069973044344e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7647164703949142e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8964036459773476e+06,
|
|
"cpu_time": 2.1810521239890684e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7646434930249155e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8957886596824068e+06,
|
|
"cpu_time": 2.1806563126682951e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7655403323692822e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8977744710183584e+06,
|
|
"cpu_time": 2.1824713094338593e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7626497906763285e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8964036459773481e+06,
|
|
"cpu_time": 2.1813069973044344e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7646434930249155e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3165587598575698e+03,
|
|
"cpu_time": 2.0211013887203985e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.3698212071265740e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2206712626998766e-03,
|
|
"cpu_time": 9.2606092001487953e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2197786409625243e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8895772387505819e+06,
|
|
"cpu_time": 2.2630672064344287e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5492624408057272e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8970865201349834e+06,
|
|
"cpu_time": 2.2666657587132524e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5272966671303463e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8899439918794907e+06,
|
|
"cpu_time": 2.2628797774800891e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5481855785431170e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8950356948457847e+06,
|
|
"cpu_time": 2.2655061823054873e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5332783590935552e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8903365102471740e+06,
|
|
"cpu_time": 2.2634388954422888e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5470335271834314e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8923959911716029e+06,
|
|
"cpu_time": 2.2643115640751095e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5410113145512354e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8903365102471740e+06,
|
|
"cpu_time": 2.2634388954422893e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5470335271834314e+08,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4339563503345839e+03,
|
|
"cpu_time": 1.6829538355140235e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0046247399701044e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8146077070310131e-03,
|
|
"cpu_time": 7.4325188380224081e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8130710856551799e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9385193712873380e+06,
|
|
"cpu_time": 2.4587495426997147e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0818318511861539e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9387349209904710e+06,
|
|
"cpu_time": 2.4596245179062192e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0817115724766519e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9390527826038601e+06,
|
|
"cpu_time": 2.4596257079891190e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0815342515760896e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9381112259037832e+06,
|
|
"cpu_time": 2.4595376060607266e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0820596733410139e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9360554690624154e+06,
|
|
"cpu_time": 2.4581431460055732e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0832086340044792e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9380947539695739e+06,
|
|
"cpu_time": 2.4591361041322704e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0820691965168779e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9385193712873380e+06,
|
|
"cpu_time": 2.4595376060607270e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0818318511861539e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1902611897534348e+03,
|
|
"cpu_time": 6.6611923428011551e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.6498164040821453e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.1413983362555496e-04,
|
|
"cpu_time": 2.7087530176177948e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.1454631787759447e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9695095981476475e+06,
|
|
"cpu_time": 2.8057177078651651e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1296184613392107e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9673136386005313e+06,
|
|
"cpu_time": 2.8060245196631253e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1319955891647563e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9673387741727568e+06,
|
|
"cpu_time": 2.8042513876406127e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1319683498657503e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9713589775093487e+06,
|
|
"cpu_time": 2.8055323314606412e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1276206149420645e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9696241625966611e+06,
|
|
"cpu_time": 2.8055670028089006e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1294945907194927e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9690290302053890e+06,
|
|
"cpu_time": 2.8054185898876889e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1301395212062550e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9695095981476470e+06,
|
|
"cpu_time": 2.8055670028089006e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1296184613392107e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7185252661459799e+03,
|
|
"cpu_time": 6.8079862672729882e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8588928166196134e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.7277802398206443e-04,
|
|
"cpu_time": 2.4267274380417988e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.7266247028127059e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 352,
|
|
"real_time": 1.9947554296364121e+06,
|
|
"cpu_time": 3.5499602755683796e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.2053315786833119e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 352,
|
|
"real_time": 1.9932594490951903e+06,
|
|
"cpu_time": 3.5463814886362855e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.2084877629993834e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 352,
|
|
"real_time": 1.9955840361944863e+06,
|
|
"cpu_time": 3.5525892159087909e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.2035854405794911e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 352,
|
|
"real_time": 1.9884335197275504e+06,
|
|
"cpu_time": 3.5476306221592068e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.2187017653722630e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 352,
|
|
"real_time": 1.9994326916771985e+06,
|
|
"cpu_time": 3.5502105227274555e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1954940693519044e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9942930252661675e+06,
|
|
"cpu_time": 3.5493544250000240e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.2063201233972712e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9947554296364118e+06,
|
|
"cpu_time": 3.5499602755683800e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.2053315786833119e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.9905758572850382e+03,
|
|
"cpu_time": 2.4173165543627188e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.4211471364511847e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0009977504446409e-03,
|
|
"cpu_time": 6.8105809251852963e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0020224066183939e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 194,
|
|
"real_time": 3.6265582311898470e+06,
|
|
"cpu_time": 6.6761092628863929e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6262089095135031e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 194,
|
|
"real_time": 3.6076105166149815e+06,
|
|
"cpu_time": 6.6595920773195140e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6505064564846792e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 194,
|
|
"real_time": 3.6097919657231146e+06,
|
|
"cpu_time": 6.6643851443297258e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6476960886689720e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 194,
|
|
"real_time": 3.6105515876996149e+06,
|
|
"cpu_time": 6.6679361907213163e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6467182624273872e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 194,
|
|
"real_time": 3.6093301185377939e+06,
|
|
"cpu_time": 6.6568106237112479e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6482908043880339e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6127684839530699e+06,
|
|
"cpu_time": 6.6649666597936405e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6438841042965155e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6097919657231146e+06,
|
|
"cpu_time": 6.6643851443297267e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.6476960886689720e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7839322070223525e+03,
|
|
"cpu_time": 7.5617613551065424e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.9781103169046994e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.1545615894282880e-03,
|
|
"cpu_time": 1.1345535155821273e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1486561879683788e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.2382099249146199e+06,
|
|
"cpu_time": 1.3781538121212739e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6357362314820948e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.2820323632296287e+06,
|
|
"cpu_time": 1.3688128595960064e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6078389007761002e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.2504750434121098e+06,
|
|
"cpu_time": 1.3610259646464678e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6278942826633215e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.2533518282903563e+06,
|
|
"cpu_time": 1.3660465363635913e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6260587924505672e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.2462423375309119e+06,
|
|
"cpu_time": 1.3733596636363858e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6305975479469481e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.2540622994755264e+06,
|
|
"cpu_time": 1.3694797672727451e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6256251510638075e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.2504750434121089e+06,
|
|
"cpu_time": 1.3688128595960062e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6278942826633215e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6644422683135090e+04,
|
|
"cpu_time": 6.5952875030991097e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0590637142049575e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.2944967931056363e-03,
|
|
"cpu_time": 4.8159072231007223e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2895580156584729e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4361083982702421e+07,
|
|
"cpu_time": 2.6715328857142836e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6729664752905149e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4313158864269450e+07,
|
|
"cpu_time": 2.6653814653061591e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6886130892829475e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4375549474997180e+07,
|
|
"cpu_time": 2.6716058795918055e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6682642716871290e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4347493800581718e+07,
|
|
"cpu_time": 2.6715928571427271e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6773927859985609e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4337142758375527e+07,
|
|
"cpu_time": 2.6696142020406503e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6807697412928448e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4346885776185263e+07,
|
|
"cpu_time": 2.6699454579591252e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6776012727104006e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4347493800581718e+07,
|
|
"cpu_time": 2.6715328857142832e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6773927859985609e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3748670356326933e+04,
|
|
"cpu_time": 2.6893591416728799e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.7460263890900863e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6553188424869125e-03,
|
|
"cpu_time": 1.0072711911233549e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6559826153377348e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7541461114126902e+07,
|
|
"cpu_time": 4.9950500192306451e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8732972968945141e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7556796629841510e+07,
|
|
"cpu_time": 4.9957835884618774e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8705852789382048e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7541135294506185e+07,
|
|
"cpu_time": 4.9992874269229598e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8733549494153681e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7561646098127730e+07,
|
|
"cpu_time": 4.9988129230769120e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8697283000494461e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7547787373455670e+07,
|
|
"cpu_time": 4.9956965384616032e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8721781600989380e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7549765302011602e+07,
|
|
"cpu_time": 4.9969260992308006e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8718287970792942e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7547787373455673e+07,
|
|
"cpu_time": 4.9957835884618782e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8721781600989380e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1915947150385273e+03,
|
|
"cpu_time": 1.9667739674838256e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6252778468188138e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.3363604423763938e-04,
|
|
"cpu_time": 3.9359676897894895e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.3360734018263996e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3412785008549690e+07,
|
|
"cpu_time": 9.3482218299993739e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0256779525170240e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3487505018711090e+07,
|
|
"cpu_time": 9.3461545500008464e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0186572715645542e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3372624516487122e+07,
|
|
"cpu_time": 9.3388184000002638e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0294595484447031e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3348928317427635e+07,
|
|
"cpu_time": 9.3317192599999994e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0316935028722868e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3415488079190254e+07,
|
|
"cpu_time": 9.3334223600004405e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0254236299785452e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.3407466188073158e+07,
|
|
"cpu_time": 9.3396672800001860e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0261823810754232e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.3412785008549690e+07,
|
|
"cpu_time": 9.3388184000002638e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0256779525170240e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.2771866018403824e+04,
|
|
"cpu_time": 7.3850464226993747e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9641140692776190e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.8809903904762894e-04,
|
|
"cpu_time": 7.9071836300995374e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.8765100287018953e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 1.0523140927155812e+08,
|
|
"cpu_time": 1.8242087822222352e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1018124314439373e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 1.0526143511136372e+08,
|
|
"cpu_time": 1.8228022211111590e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1003571386995173e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 1.0527157038450241e+08,
|
|
"cpu_time": 1.8203676633332735e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.0998660895727987e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 1.0538518594370948e+08,
|
|
"cpu_time": 1.8248016511111271e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.0943679340924120e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 1.0541649411122005e+08,
|
|
"cpu_time": 1.8261701744444913e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.0928549324887667e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0531321896447077e+08,
|
|
"cpu_time": 1.8236700984444574e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.0978517052594862e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0527157038450241e+08,
|
|
"cpu_time": 1.8242087822222352e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.0998660895727987e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2088366298419700e+04,
|
|
"cpu_time": 2.2068223979406329e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.9726154009468374e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.7946877994597831e-04,
|
|
"cpu_time": 1.2100995678017612e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.7927245252117962e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7309843003749847e+08,
|
|
"cpu_time": 3.1631299374998891e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.2030708410665207e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7302027344703674e+08,
|
|
"cpu_time": 3.1587986849999082e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.2058728876571980e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7280907183885574e+08,
|
|
"cpu_time": 3.1511488775001341e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.2134575029791431e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7322742566466331e+08,
|
|
"cpu_time": 3.1572473699998230e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1984516590263729e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7332270741462708e+08,
|
|
"cpu_time": 3.1802114125000000e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1950441463585434e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7309558168053630e+08,
|
|
"cpu_time": 3.1621072564999509e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.2031794074175558e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7309843003749847e+08,
|
|
"cpu_time": 3.1587986849999082e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.2030708410665207e+09,
|
|
"dst_gpu": 5.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9794965758325611e+05,
|
|
"cpu_time": 1.0993834105757196e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.0964770870059179e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/5/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1435858481274831e-03,
|
|
"cpu_time": 3.4767429482850517e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1440064233061173e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8665508063499560e+06,
|
|
"cpu_time": 2.0695740848805385e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1944219177241344e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8718214993150807e+06,
|
|
"cpu_time": 2.0746848435015006e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1882428434008104e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8669043349981275e+06,
|
|
"cpu_time": 2.0694213342173693e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1940063683038736e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8626848939467785e+06,
|
|
"cpu_time": 2.0701459257293148e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1989763342747292e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8770251101897585e+06,
|
|
"cpu_time": 2.0791693076923778e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1821764545205864e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8689973289599405e+06,
|
|
"cpu_time": 2.0725990992042203e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1915647836448266e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8669043349981278e+06,
|
|
"cpu_time": 2.0701459257293150e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1940063683038736e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.5370150367055267e+03,
|
|
"cpu_time": 4.2657042477899104e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4852070952372242e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.9625590956765917e-03,
|
|
"cpu_time": 2.0581424788941277e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9591674148238376e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8736120451533694e+06,
|
|
"cpu_time": 2.0768588121692725e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3723032317127436e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8735587160027875e+06,
|
|
"cpu_time": 2.0771617354498440e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3724276853610026e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8669791954774938e+06,
|
|
"cpu_time": 2.0703525820104319e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3878367899567485e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8722558559689093e+06,
|
|
"cpu_time": 2.0757880185186800e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3754703577949638e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8668717141269348e+06,
|
|
"cpu_time": 2.0704886825396465e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3880894107558373e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8706555053458989e+06,
|
|
"cpu_time": 2.0741299661375750e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3792254951162590e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8722558559689093e+06,
|
|
"cpu_time": 2.0757880185186800e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3754703577949638e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4483043696020782e+03,
|
|
"cpu_time": 3.4247240988604490e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.0770106430370870e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8433668624434725e-03,
|
|
"cpu_time": 1.6511617665106769e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8443924963545775e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8636640545862194e+06,
|
|
"cpu_time": 2.0734630211640247e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7912840083389711e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8635213974453374e+06,
|
|
"cpu_time": 2.0738160396825501e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7919570027263891e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8677571795744668e+06,
|
|
"cpu_time": 2.0730797883599070e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7720182147728559e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8743108790735872e+06,
|
|
"cpu_time": 2.0797676005293725e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7413460503937807e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8702630252283553e+06,
|
|
"cpu_time": 2.0751193571426382e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7602651493361723e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8679033071815935e+06,
|
|
"cpu_time": 2.0750491613756984e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7713740851136353e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8677571795744668e+06,
|
|
"cpu_time": 2.0738160396825499e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7720182147728559e+06,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5776357322931817e+03,
|
|
"cpu_time": 2.7469735824505719e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1480035885004796e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4506813145484486e-03,
|
|
"cpu_time": 1.3238113262961954e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4488792379132140e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8775307615485755e+06,
|
|
"cpu_time": 2.0849050372341233e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7452710054653466e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8708860888124190e+06,
|
|
"cpu_time": 2.0786579680850899e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7514695414085910e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8674367102574522e+06,
|
|
"cpu_time": 2.0800652898936064e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7547047147575069e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8685200956245721e+06,
|
|
"cpu_time": 2.0812309893617181e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7536873206090383e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8710494742527830e+06,
|
|
"cpu_time": 2.0779417313830669e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7513165980330978e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8710846260991606e+06,
|
|
"cpu_time": 2.0805602031915213e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7512898360547163e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8708860888124190e+06,
|
|
"cpu_time": 2.0800652898936062e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7514695414085910e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.9203769423985700e+03,
|
|
"cpu_time": 2.7407740294323494e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.6629448747515642e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0952429877914057e-03,
|
|
"cpu_time": 1.3173250287245131e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0915697672313342e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8770654348978298e+06,
|
|
"cpu_time": 2.0903324827585267e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4914073202550434e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8718015555099682e+06,
|
|
"cpu_time": 2.0849420477453163e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5012258541555099e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8786204650652884e+06,
|
|
"cpu_time": 2.0916641034484529e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4885173039846770e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8759182900893169e+06,
|
|
"cpu_time": 2.0894089973474026e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4935423544956043e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8672502311281627e+06,
|
|
"cpu_time": 2.0862929920424684e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5097599083120316e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8741311953381132e+06,
|
|
"cpu_time": 2.0885281246684338e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4968905482405730e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8759182900893167e+06,
|
|
"cpu_time": 2.0894089973474026e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4935423544956043e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.6024966679139743e+03,
|
|
"cpu_time": 2.8161042797084565e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.5994580628665193e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4558028164531113e-03,
|
|
"cpu_time": 1.3483678991182029e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4591727834299116e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8678145662558195e+06,
|
|
"cpu_time": 2.0979667526596151e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0173989628287390e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8782775920779852e+06,
|
|
"cpu_time": 2.1032496196807688e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9783082411685377e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8728967328996153e+06,
|
|
"cpu_time": 2.0982842632978032e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9983570208419636e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8689358625356653e+06,
|
|
"cpu_time": 2.0983276170214023e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0131887684026256e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8781541888568022e+06,
|
|
"cpu_time": 2.1035033670211467e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9787667475683197e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8732157885251779e+06,
|
|
"cpu_time": 2.1002663239361471e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9972039481620371e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8728967328996155e+06,
|
|
"cpu_time": 2.0983276170214023e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9983570208419636e+07,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.9396968027350131e+03,
|
|
"cpu_time": 2.8440146961677774e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8450344039136506e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.6370142900750053e-03,
|
|
"cpu_time": 1.3541209815894956e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6368166736061593e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8735877588079211e+06,
|
|
"cpu_time": 2.1211186037233691e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3991551704350924e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8734851832476200e+06,
|
|
"cpu_time": 2.1210729840424000e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3992317758584175e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8738000679227395e+06,
|
|
"cpu_time": 2.1212696063827858e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3989966405039576e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8741460839822702e+06,
|
|
"cpu_time": 2.1211571569147226e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3987383493766111e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8868818952274097e+06,
|
|
"cpu_time": 2.1342680904254811e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3892973410951406e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8763801978375923e+06,
|
|
"cpu_time": 2.1237772882977515e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3970838554538438e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8738000679227393e+06,
|
|
"cpu_time": 2.1211571569147226e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3989966405039576e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.8760702934378342e+03,
|
|
"cpu_time": 5.8649883099735343e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.3568856555457134e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.1315989692332225e-03,
|
|
"cpu_time": 2.7615834966737188e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1185570132655892e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8997033882725944e+06,
|
|
"cpu_time": 2.1814506129034744e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7598413691136098e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8988397120753482e+06,
|
|
"cpu_time": 2.1820103736558240e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7610966669060040e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8947867844666864e+06,
|
|
"cpu_time": 2.1783396962364940e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7670026216040343e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8944027183498559e+06,
|
|
"cpu_time": 2.1780581451614895e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7675635962805623e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8962994808555201e+06,
|
|
"cpu_time": 2.1794922634408753e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7647953569204491e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8968064168040012e+06,
|
|
"cpu_time": 2.1798702182796309e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7640599221649325e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8962994808555201e+06,
|
|
"cpu_time": 2.1794922634408753e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7647953569204491e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3790963853702146e+03,
|
|
"cpu_time": 1.7921264444108936e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.4660189027182182e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2542642012877841e-03,
|
|
"cpu_time": 8.2212529414950764e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2539593931825764e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8986855991902659e+06,
|
|
"cpu_time": 2.2651512291105674e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5226415602835310e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8920125546450964e+06,
|
|
"cpu_time": 2.2610162210243912e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5421196726503313e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8979925368102947e+06,
|
|
"cpu_time": 2.2651656765500437e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5246581831254363e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8919978990229573e+06,
|
|
"cpu_time": 2.2620871509434925e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5421626025139499e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.9001647856838328e+06,
|
|
"cpu_time": 2.2663612614556681e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5183424506135011e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8961706750704895e+06,
|
|
"cpu_time": 2.2639563078168328e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5299848938373494e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8979925368102945e+06,
|
|
"cpu_time": 2.2651512291105678e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5246581831254363e+08,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8826205239264664e+03,
|
|
"cpu_time": 2.2810123519687595e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1329103278900762e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0476113120893673e-03,
|
|
"cpu_time": 1.0075337337973516e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0486680337094567e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 362,
|
|
"real_time": 1.9381477002913740e+06,
|
|
"cpu_time": 2.4480741685083662e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0820393098445086e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 362,
|
|
"real_time": 1.9404009619382988e+06,
|
|
"cpu_time": 2.4537307762432257e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0807828078507650e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 362,
|
|
"real_time": 1.9390511257993316e+06,
|
|
"cpu_time": 2.4527986961326976e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0815351756831553e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 362,
|
|
"real_time": 1.9397583170420907e+06,
|
|
"cpu_time": 2.4530312872928297e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0811408728474569e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 362,
|
|
"real_time": 1.9400606373420181e+06,
|
|
"cpu_time": 2.4536408867401308e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0809723983025630e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9394837484826227e+06,
|
|
"cpu_time": 2.4522551629834501e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0812941129056897e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9397583170420907e+06,
|
|
"cpu_time": 2.4530312872928297e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0811408728474569e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.9744865129603602e+02,
|
|
"cpu_time": 2.3704523220041583e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.0046113775197370e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.6272553301782772e-04,
|
|
"cpu_time": 9.6664179070184153e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6283534866116839e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9720594256101281e+06,
|
|
"cpu_time": 2.7933543960671797e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1268649136687856e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9693609274429365e+06,
|
|
"cpu_time": 2.7927845646066605e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1297792301820374e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9679703976006738e+06,
|
|
"cpu_time": 2.7918885589886950e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1312840910176523e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9671749039108371e+06,
|
|
"cpu_time": 2.7928226966292039e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1321459478064325e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 356,
|
|
"real_time": 1.9659798695663973e+06,
|
|
"cpu_time": 2.7916791460676338e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1334419873409314e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9685091048261947e+06,
|
|
"cpu_time": 2.7925058724718746e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1307032340031679e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9679703976006738e+06,
|
|
"cpu_time": 2.7927845646066600e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1312840910176523e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3344007791938338e+03,
|
|
"cpu_time": 7.0046263689235286e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5251769188870708e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1858724826167085e-03,
|
|
"cpu_time": 2.5083658508918950e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1851377885895285e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 349,
|
|
"real_time": 2.0109855127587572e+06,
|
|
"cpu_time": 3.5154923810890336e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1713915623848243e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 349,
|
|
"real_time": 2.0026650348475424e+06,
|
|
"cpu_time": 3.5121190830945293e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1887224543462424e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 349,
|
|
"real_time": 2.0099017391499016e+06,
|
|
"cpu_time": 3.5158403667619727e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1736408485063581e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 349,
|
|
"real_time": 2.0058100621474849e+06,
|
|
"cpu_time": 3.5120531747849747e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1821547106104784e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 349,
|
|
"real_time": 2.0267462616727652e+06,
|
|
"cpu_time": 3.5337924212032435e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1389532368380947e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0112217221152901e+06,
|
|
"cpu_time": 3.5178594853867507e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1709725625371995e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0099017391499013e+06,
|
|
"cpu_time": 3.5154923810890340e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1736408485063581e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.2912921459366971e+03,
|
|
"cpu_time": 9.0857570928348632e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9187441360723425e+07,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.6197254354257058e-03,
|
|
"cpu_time": 2.5827515654270037e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6002319778032097e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 197,
|
|
"real_time": 3.5496114701183918e+06,
|
|
"cpu_time": 6.4839768730967157e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7264936292986517e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 197,
|
|
"real_time": 3.5474088402266428e+06,
|
|
"cpu_time": 6.4843571827412024e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7294283674751482e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 197,
|
|
"real_time": 3.5486847556995107e+06,
|
|
"cpu_time": 6.4880827055834942e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7277279203384476e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 197,
|
|
"real_time": 3.5473998839290915e+06,
|
|
"cpu_time": 6.4808901218276070e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7294403080990114e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 197,
|
|
"real_time": 3.5477784157347707e+06,
|
|
"cpu_time": 6.4784649086296037e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7289356983489389e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5481766731416816e+06,
|
|
"cpu_time": 6.4831543583757244e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7284051847120390e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5477784157347707e+06,
|
|
"cpu_time": 6.4839768730967157e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7289356983489389e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.5764921617392099e+02,
|
|
"cpu_time": 3.6581972741932159e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2759858216140177e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.6989896625581063e-04,
|
|
"cpu_time": 5.6426194287154563e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6985543153948757e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.1646357566645052e+06,
|
|
"cpu_time": 1.3405315373736840e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6833409456702461e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.1491836382057332e+06,
|
|
"cpu_time": 1.3369780707069946e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6934634355568638e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.1607197898280136e+06,
|
|
"cpu_time": 1.3395559787878400e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6859021138719788e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.1520005644421382e+06,
|
|
"cpu_time": 1.3374278545454374e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6916148422615891e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.1636240769441081e+06,
|
|
"cpu_time": 1.3402284828283321e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6840023484752436e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1580327652169000e+06,
|
|
"cpu_time": 1.3389443848484578e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6876647371671848e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1607197898280146e+06,
|
|
"cpu_time": 1.3395559787878400e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6859021138719788e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0138587671925925e+03,
|
|
"cpu_time": 1.6361741479303208e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.5946528147750180e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.7985843279107436e-04,
|
|
"cpu_time": 1.2219881321773522e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.8015815387676904e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4220062084496021e+07,
|
|
"cpu_time": 2.6268008720001034e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7193087907237806e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4253691025078297e+07,
|
|
"cpu_time": 2.6289347119998183e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7081744568425827e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4217114988714457e+07,
|
|
"cpu_time": 2.6305320740000296e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7202870662065411e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4253707099705935e+07,
|
|
"cpu_time": 2.6343873260000240e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7081691471957150e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4230644721537828e+07,
|
|
"cpu_time": 2.6305629900000438e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7157992707408342e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4235043983906511e+07,
|
|
"cpu_time": 2.6302435948000040e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7143477463418913e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4230644721537832e+07,
|
|
"cpu_time": 2.6305320740000300e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7157992707408342e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7757278252070584e+04,
|
|
"cpu_time": 2.7803442346793105e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.8796044841128560e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2474340277519446e-03,
|
|
"cpu_time": 1.0570672009908343e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2471724192758475e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7305466409486074e+07,
|
|
"cpu_time": 4.8998865576921754e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.9154160557891808e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7396002282889988e+07,
|
|
"cpu_time": 4.9097953076922685e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8991720256872988e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7419442454209693e+07,
|
|
"cpu_time": 4.9204395115387239e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8949838503880157e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7384463697671890e+07,
|
|
"cpu_time": 4.9153214076922655e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.9012363171242466e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7393306892078657e+07,
|
|
"cpu_time": 4.9102060423076816e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8996540844366560e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7379736347267263e+07,
|
|
"cpu_time": 4.9111297653846227e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.9020924666850796e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7393306892078657e+07,
|
|
"cpu_time": 4.9102060423076816e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8996540844366560e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3484729434165267e+04,
|
|
"cpu_time": 7.6368350701157528e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.7983672639444210e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5882084795351005e-03,
|
|
"cpu_time": 1.5550057593555892e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5908241872103805e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3086633235216141e+07,
|
|
"cpu_time": 9.1993131100002751e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0565545343705778e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3266233205795288e+07,
|
|
"cpu_time": 9.2122951600003943e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0395051394547386e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3215737640857697e+07,
|
|
"cpu_time": 9.2416827200008810e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0442870455280895e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3153721615672112e+07,
|
|
"cpu_time": 9.2045481299999207e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0501723649930305e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3138089179992676e+07,
|
|
"cpu_time": 9.2040618199996516e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0516580506073256e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.3172082975506783e+07,
|
|
"cpu_time": 9.2123801880002245e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0484354269907522e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.3153721615672112e+07,
|
|
"cpu_time": 9.2045481299999207e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0501723649930305e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9936348656208502e+04,
|
|
"cpu_time": 1.7028876339358985e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.6385477609971296e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3152832227472451e-03,
|
|
"cpu_time": 1.8484773741253423e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3149713127962506e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.0422693565487862e+08,
|
|
"cpu_time": 1.7837412900000516e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1509804891291590e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.0451261885464191e+08,
|
|
"cpu_time": 1.7852095075001273e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1369003846960335e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.0438843816518784e+08,
|
|
"cpu_time": 1.7836315824999359e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1430112514035044e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.0440433677285910e+08,
|
|
"cpu_time": 1.7841464624999049e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1422280778241072e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.0454489849507809e+08,
|
|
"cpu_time": 1.7831009262499720e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1353142977634201e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0441544558852911e+08,
|
|
"cpu_time": 1.7839659537499985e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1416869001632452e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0440433677285910e+08,
|
|
"cpu_time": 1.7837412900000516e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1422280778241072e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2508946843038130e+05,
|
|
"cpu_time": 7.8891695427967905e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.1631641020761402e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1979977456910206e-03,
|
|
"cpu_time": 4.4222646324686317e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1986657728770811e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7617597430944443e+08,
|
|
"cpu_time": 3.0971456175001323e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0947119958254089e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7633265629410744e+08,
|
|
"cpu_time": 3.0959061174999648e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0892964840788918e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7621233314275742e+08,
|
|
"cpu_time": 3.0941087449997437e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0934544412967634e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7673009634017944e+08,
|
|
"cpu_time": 3.1005951925001794e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0756025500784254e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7635421082377434e+08,
|
|
"cpu_time": 3.1076009124998903e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0885522323760061e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7636105418205261e+08,
|
|
"cpu_time": 3.0990713169999826e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0883235407311001e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7633265629410744e+08,
|
|
"cpu_time": 3.0971456175001323e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0892964840788918e+09,
|
|
"dst_gpu": 6.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1989650568382433e+05,
|
|
"cpu_time": 5.3251789459306828e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.5822328648386952e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/6/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2468541124551868e-03,
|
|
"cpu_time": 1.7183144242984560e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2453728541384650e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8662308527464073e+06,
|
|
"cpu_time": 2.0674684682539222e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1947981376324319e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8618054294998131e+06,
|
|
"cpu_time": 2.0675782222221121e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.2000150687606591e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8671303258559890e+06,
|
|
"cpu_time": 2.0680800105818766e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1937408135247235e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8640767711552880e+06,
|
|
"cpu_time": 2.0691741269840670e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1973343927575718e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8638012103716999e+06,
|
|
"cpu_time": 2.0696458862432707e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1976592660239395e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8646089179258395e+06,
|
|
"cpu_time": 2.0683893428570498e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1967095357398652e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8640767711552877e+06,
|
|
"cpu_time": 2.0680800105818764e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 2.1973343927575718e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1085118889875221e+03,
|
|
"cpu_time": 9.7421770941350883e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4842371453498431e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1308065024879300e-03,
|
|
"cpu_time": 4.7100305983390416e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1308901358745807e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8662278972062275e+06,
|
|
"cpu_time": 2.0686537566138376e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3896032270568628e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8714782807499568e+06,
|
|
"cpu_time": 2.0732478015874962e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3772883096015537e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8663637692471547e+06,
|
|
"cpu_time": 2.0687028730159390e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3892836621579146e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8616721011567665e+06,
|
|
"cpu_time": 2.0686691851851142e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.4003452567774039e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 378,
|
|
"real_time": 1.8667989085659706e+06,
|
|
"cpu_time": 2.0687109285714184e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3882605471913917e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8665081913852151e+06,
|
|
"cpu_time": 2.0695969089947611e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3889562005570261e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8663637692471545e+06,
|
|
"cpu_time": 2.0687028730159388e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 4.3892836621579146e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4739565357941874e+03,
|
|
"cpu_time": 2.0410466829446518e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.1679264337139139e+03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8612061558733457e-03,
|
|
"cpu_time": 9.8620493395306795e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8610179870733909e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 381,
|
|
"real_time": 1.8676533394123751e+06,
|
|
"cpu_time": 2.0712771443567628e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7725059325811192e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 381,
|
|
"real_time": 1.8629079086573434e+06,
|
|
"cpu_time": 2.0710637007874332e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7948523509186599e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 381,
|
|
"real_time": 1.8679426849778164e+06,
|
|
"cpu_time": 2.0711348845145104e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7711470655720755e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 381,
|
|
"real_time": 1.8736901009182464e+06,
|
|
"cpu_time": 2.0773768057744096e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7442421732231118e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 381,
|
|
"real_time": 1.8676650769642817e+06,
|
|
"cpu_time": 2.0710995905512171e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7724508007777762e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8679718221860125e+06,
|
|
"cpu_time": 2.0723904251968667e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7710396646145489e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8676650769642820e+06,
|
|
"cpu_time": 2.0711348845145106e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 8.7724508007777762e+06,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8255164609241656e+03,
|
|
"cpu_time": 2.7886475057199764e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7953428334578435e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0479519099208449e-03,
|
|
"cpu_time": 1.3456187945160329e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0468985457914256e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8709001957498812e+06,
|
|
"cpu_time": 2.0760950134046953e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7514563350006044e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8662455849355427e+06,
|
|
"cpu_time": 2.0769454557639326e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7558246494730089e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8708928955260704e+06,
|
|
"cpu_time": 2.0765002493297546e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7514631691829730e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8663609819549159e+06,
|
|
"cpu_time": 2.0765049517425743e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7557160869103268e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 373,
|
|
"real_time": 1.8709726778447977e+06,
|
|
"cpu_time": 2.0764127426274165e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7513884830079917e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8690744672022418e+06,
|
|
"cpu_time": 2.0764916825736747e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7531697447149809e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8708928955260706e+06,
|
|
"cpu_time": 2.0765002493297546e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 1.7514631691829730e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5302542859802516e+03,
|
|
"cpu_time": 3.0403022509686804e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3745233214071093e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3537471782853621e-03,
|
|
"cpu_time": 1.4641533488833560e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3544172368735144e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8780822733462546e+06,
|
|
"cpu_time": 2.0901844827586075e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4895169892228350e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8782299668518878e+06,
|
|
"cpu_time": 2.0899375490718489e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4892425931125611e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8717068371902448e+06,
|
|
"cpu_time": 2.0831794297081535e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5014030348032951e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8676304325105657e+06,
|
|
"cpu_time": 2.0845208355439203e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.5090454117254399e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 377,
|
|
"real_time": 1.8725743517588419e+06,
|
|
"cpu_time": 2.0847399204245056e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4997809266395420e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8736447723315593e+06,
|
|
"cpu_time": 2.0865124435014073e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4977977911007345e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8725743517588419e+06,
|
|
"cpu_time": 2.0847399204245056e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.4997809266395420e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5218663275038261e+03,
|
|
"cpu_time": 3.2951684985784727e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.4439722702055340e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4134064227536718e-03,
|
|
"cpu_time": 1.5792709546696023e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4140824525903399e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8806865100486903e+06,
|
|
"cpu_time": 2.1038978563829102e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9693699242095694e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8702115833141934e+06,
|
|
"cpu_time": 2.0973138776593683e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0084048868806541e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8705575296328962e+06,
|
|
"cpu_time": 2.0976361329787248e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0071087322143659e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8745201083375118e+06,
|
|
"cpu_time": 2.0976538537231614e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9922962905021116e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 376,
|
|
"real_time": 1.8709324330631443e+06,
|
|
"cpu_time": 2.0976617686171308e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0057046253351405e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8733816328792870e+06,
|
|
"cpu_time": 2.0988326978722597e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.9965768918283686e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8709324330631446e+06,
|
|
"cpu_time": 2.0976538537231619e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 7.0057046253351405e+07,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.4351073819486492e+03,
|
|
"cpu_time": 2.8352760873188727e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6527849848815787e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3674340049614591e-03,
|
|
"cpu_time": 1.3508823691346147e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3622765967339599e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8756399760022762e+06,
|
|
"cpu_time": 2.1213301679999861e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3976242954617101e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8765142098612462e+06,
|
|
"cpu_time": 2.1216119759998643e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3969731677085650e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8768338005368907e+06,
|
|
"cpu_time": 2.1216219146666238e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3967352885748890e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8762966092132654e+06,
|
|
"cpu_time": 2.1214767626667405e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3971351795488107e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 375,
|
|
"real_time": 1.8760785869865990e+06,
|
|
"cpu_time": 2.1213533039999674e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3972975429620025e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8762726365200556e+06,
|
|
"cpu_time": 2.1214788250666363e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3971530948511955e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8762966092132654e+06,
|
|
"cpu_time": 2.1214767626667405e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3971351795488107e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5048639692775453e+02,
|
|
"cpu_time": 1.3789717383492322e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.3546687317826181e+04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4009644875665656e-04,
|
|
"cpu_time": 6.5000495034680275e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4010745451914192e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8994800491174408e+06,
|
|
"cpu_time": 2.1811311940699904e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7601658687786746e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8973413484096981e+06,
|
|
"cpu_time": 2.1788452884097467e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7632771532620871e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8974603317719395e+06,
|
|
"cpu_time": 2.1791337277627243e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7631038774359763e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8975530885821434e+06,
|
|
"cpu_time": 2.1792610269540213e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7629688104892462e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.9001808120100948e+06,
|
|
"cpu_time": 2.1819505471697543e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7591479541643465e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8984031259782636e+06,
|
|
"cpu_time": 2.1800643568732473e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7617327328260660e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8975530885821437e+06,
|
|
"cpu_time": 2.1792610269540213e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7629688104892462e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3284132610361219e+03,
|
|
"cpu_time": 1.3868483020892795e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9319228561202955e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.9975298863437097e-04,
|
|
"cpu_time": 6.3615016580444595e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.9953288135284888e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8923901844837314e+06,
|
|
"cpu_time": 2.2600984408601588e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5410137327787149e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8986783983838065e+06,
|
|
"cpu_time": 2.2637749623654406e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5226625051012814e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8931493139325050e+06,
|
|
"cpu_time": 2.2607742768817237e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5387918548372042e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8989196677300427e+06,
|
|
"cpu_time": 2.2640923897849964e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5219608170863879e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 372,
|
|
"real_time": 1.8932727931921058e+06,
|
|
"cpu_time": 2.2605182043011403e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5384306148089433e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8952820715444386e+06,
|
|
"cpu_time": 2.2618516548386919e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5325719049225068e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8932727931921061e+06,
|
|
"cpu_time": 2.2607742768817237e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.5384306148089433e+08,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2293965112130836e+03,
|
|
"cpu_time": 1.9191533154972756e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.4216254686757736e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7039133961635031e-03,
|
|
"cpu_time": 8.4848770315758986e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7029377350329692e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9410442592660813e+06,
|
|
"cpu_time": 2.4531829834708604e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0804246167951596e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9418270459922017e+06,
|
|
"cpu_time": 2.4534328732783222e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0799890774661825e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9424225478856398e+06,
|
|
"cpu_time": 2.4535437410467044e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0796579777570980e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9406616553399947e+06,
|
|
"cpu_time": 2.4530542286501341e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0806376238894610e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 363,
|
|
"real_time": 1.9443835529175487e+06,
|
|
"cpu_time": 2.4575433168043052e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0785690903696558e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9420678122802936e+06,
|
|
"cpu_time": 2.4541514286500653e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0798556772555115e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9418270459922017e+06,
|
|
"cpu_time": 2.4534328732783222e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.0799890774661825e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4638475473410804e+03,
|
|
"cpu_time": 1.9060591017211730e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.1352923409057653e+05,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.5375717474164452e-04,
|
|
"cpu_time": 7.7666727467164609e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.5336848360902048e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 355,
|
|
"real_time": 1.9729406753240128e+06,
|
|
"cpu_time": 2.7929141690143473e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1259149109038348e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 355,
|
|
"real_time": 1.9712613331495037e+06,
|
|
"cpu_time": 2.7921048478873046e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1277260043947189e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 355,
|
|
"real_time": 1.9710742661886860e+06,
|
|
"cpu_time": 2.7928573887323528e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1279279385602255e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 355,
|
|
"real_time": 1.9708106254922673e+06,
|
|
"cpu_time": 2.7919186676056813e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1282125972668483e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 355,
|
|
"real_time": 1.9728946771708804e+06,
|
|
"cpu_time": 2.7910574478873587e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1259644767325380e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9717963154650703e+06,
|
|
"cpu_time": 2.7921705042254091e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1271491855716333e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9712613331495035e+06,
|
|
"cpu_time": 2.7921048478873046e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.1277260043947189e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0362316197282794e+03,
|
|
"cpu_time": 7.6345833088098777e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1176973644059468e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.2552670455917382e-04,
|
|
"cpu_time": 2.7342826296805357e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.2544380619245828e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 348,
|
|
"real_time": 2.0123664681109246e+06,
|
|
"cpu_time": 3.5156520028736517e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1685290094674783e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 348,
|
|
"real_time": 2.0093071122584199e+06,
|
|
"cpu_time": 3.5129549597701863e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1748759802932148e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 348,
|
|
"real_time": 2.0174142940039476e+06,
|
|
"cpu_time": 3.5173984683905314e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1580988223054519e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 348,
|
|
"real_time": 2.0106216389471772e+06,
|
|
"cpu_time": 3.5142273505749502e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1721464832103024e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 348,
|
|
"real_time": 2.0153531329652786e+06,
|
|
"cpu_time": 3.5181437155173146e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1623514325042720e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0130125292571497e+06,
|
|
"cpu_time": 3.5156752994253272e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1672003455561442e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0123664681109246e+06,
|
|
"cpu_time": 3.5156520028736517e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 4.1685290094674783e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3435693957993049e+03,
|
|
"cpu_time": 2.1540505863707522e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.9190800933072018e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6609779359064211e-03,
|
|
"cpu_time": 6.1269895622125686e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6603665577743655e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 197,
|
|
"real_time": 3.5519911233940856e+06,
|
|
"cpu_time": 6.4838025837559877e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7233271191197748e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 197,
|
|
"real_time": 3.5451979631662066e+06,
|
|
"cpu_time": 6.4720434670048486e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7323777612171240e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 197,
|
|
"real_time": 3.5476824528712607e+06,
|
|
"cpu_time": 6.4710993553304151e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7290636134645100e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 197,
|
|
"real_time": 3.5471423543271073e+06,
|
|
"cpu_time": 6.4738581878167782e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7297836748879614e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 197,
|
|
"real_time": 3.5487333794838279e+06,
|
|
"cpu_time": 6.4732931218275344e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7276631422900209e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5481494546484975e+06,
|
|
"cpu_time": 6.4748193431471130e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7284430621958780e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5476824528712607e+06,
|
|
"cpu_time": 6.4732931218275335e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 4.7290636134645100e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5021506070516216e+03,
|
|
"cpu_time": 5.1355520792510488e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.3333665899567660e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.0519876319570106e-04,
|
|
"cpu_time": 7.9315758588484309e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.0496071246097620e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.1427542902529240e+06,
|
|
"cpu_time": 1.3261001459999306e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6976881237240267e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.1469335211440921e+06,
|
|
"cpu_time": 1.3247606319999931e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6949411101600056e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.1517271408811212e+06,
|
|
"cpu_time": 1.3282232709999561e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6917942112464266e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.1523991553112864e+06,
|
|
"cpu_time": 1.3298100669999259e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6913533866580811e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.1424279222264886e+06,
|
|
"cpu_time": 1.3263534919999529e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6979027811512270e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1472484059631824e+06,
|
|
"cpu_time": 1.3270495215999519e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6947359225879536e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1469335211440939e+06,
|
|
"cpu_time": 1.3263534919999529e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.6949411101600056e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.7466455848916657e+03,
|
|
"cpu_time": 1.9763024435266201e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1177636842839774e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.6412209500531488e-04,
|
|
"cpu_time": 1.4892454361039211e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.6409777582661627e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4288020543754101e+07,
|
|
"cpu_time": 2.6407726299999014e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6968622276607885e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4280122853815556e+07,
|
|
"cpu_time": 2.6429211600000143e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6994598496797209e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4259607717394829e+07,
|
|
"cpu_time": 2.6383650479999684e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7062209094389105e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4290106948465109e+07,
|
|
"cpu_time": 2.6391372999999024e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6961764696385365e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4307553451508284e+07,
|
|
"cpu_time": 2.6420818880001206e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6904499939453640e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4285082302987579e+07,
|
|
"cpu_time": 2.6406556051999815e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6978338900726643e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4288020543754101e+07,
|
|
"cpu_time": 2.6407726299999014e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6968622276607885e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7409648789022576e+04,
|
|
"cpu_time": 1.9191546938628886e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.7271050894217053e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2187293303435520e-03,
|
|
"cpu_time": 7.2677205239626375e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2190948474198012e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7400063350796700e+07,
|
|
"cpu_time": 4.9284043038465619e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8984459007134895e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7396457413068183e+07,
|
|
"cpu_time": 4.9244612538455635e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8990906370243979e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7384765374545868e+07,
|
|
"cpu_time": 4.9223097884613223e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.9011823239776716e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7367841667280748e+07,
|
|
"cpu_time": 4.9279446307694420e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.9042131137605267e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.7417331145933039e+07,
|
|
"cpu_time": 4.9396710576930813e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8953607951702194e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7393291790324915e+07,
|
|
"cpu_time": 4.9285582069231942e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8996585541292620e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7396457413068186e+07,
|
|
"cpu_time": 4.9279446307694413e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.8990906370243979e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8399553968638567e+04,
|
|
"cpu_time": 6.7030472896388586e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.2912641246062280e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.7168101261700633e-04,
|
|
"cpu_time": 1.3600422290281613e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.7173336432443129e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3139927610754967e+07,
|
|
"cpu_time": 9.2058440599998906e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0514832832717571e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3062470629811287e+07,
|
|
"cpu_time": 9.1879854400008291e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0588570945505304e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3090519085526466e+07,
|
|
"cpu_time": 9.1986109799995571e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0561844303605776e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3248919546604156e+07,
|
|
"cpu_time": 9.2517094300001189e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0411437130674877e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.3219959884881973e+07,
|
|
"cpu_time": 9.2577525600017905e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0438868533655853e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.3152359351515777e+07,
|
|
"cpu_time": 9.2203804940004379e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0503110749231873e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.3139927610754967e+07,
|
|
"cpu_time": 9.2058440599998906e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.0514832832717571e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0548824248155957e+04,
|
|
"cpu_time": 3.2065685416815226e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.6520581220165864e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5154327151398388e-03,
|
|
"cpu_time": 3.4776965481717248e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5151657013786957e-03,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.0440095234662294e+08,
|
|
"cpu_time": 1.7829078012499622e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1423947764147587e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.0444433242082596e+08,
|
|
"cpu_time": 1.7837555662498516e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1402589260357914e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.0448521282523870e+08,
|
|
"cpu_time": 1.7849422300000128e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1382477719403877e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.0460723377764225e+08,
|
|
"cpu_time": 1.7839600000002065e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1322541722228928e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 1.0450225602835417e+08,
|
|
"cpu_time": 1.7875780637498906e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1374097785442352e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0448799747973680e+08,
|
|
"cpu_time": 1.7846287322499847e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1381130850316133e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0448521282523870e+08,
|
|
"cpu_time": 1.7839600000002065e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.1382477719403877e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7317281093847225e+04,
|
|
"cpu_time": 1.8005597968991913e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.8005642541917376e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.3996327768499203e-04,
|
|
"cpu_time": 1.0089268229079339e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.3968092786115742e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7942670360207558e+08,
|
|
"cpu_time": 3.0796398924996990e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9842922064783401e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7946901917457581e+08,
|
|
"cpu_time": 3.0724668325001401e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9828812178191805e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7971289530396461e+08,
|
|
"cpu_time": 3.0775671275000602e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9747622572319241e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7959230020642281e+08,
|
|
"cpu_time": 3.0652552250001007e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9787742724261818e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7940849438309669e+08,
|
|
"cpu_time": 3.0800912149999249e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9848995873473234e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7952188253402713e+08,
|
|
"cpu_time": 3.0750040584999859e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9811219082605896e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7946901917457581e+08,
|
|
"cpu_time": 3.0775671275000608e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9828812178191805e+09,
|
|
"dst_gpu": 7.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2860303601241519e+05,
|
|
"cpu_time": 6.2333119448830909e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.2829130106318993e+06,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipManaged_HostToGPUWriteDst/0/7/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.1636412339893651e-04,
|
|
"cpu_time": 2.0270906399791084e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.1607184677455302e-04,
|
|
"dst_gpu": 0.0000000000000000e+00,
|
|
"src_numa": NaN
|
|
}
|
|
]
|
|
}
|