23444 lines
814 KiB
JSON
23444 lines
814 KiB
JSON
{
|
|
"context": {
|
|
"date": "2023-10-09T16:29:18-04:00",
|
|
"host_name": "frontier10439",
|
|
"executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope",
|
|
"num_cpus": 128,
|
|
"mhz_per_cpu": 1837,
|
|
"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": [2.11,2.17,4.33],
|
|
"library_build_type": "release"
|
|
},
|
|
"benchmarks": [
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9033,
|
|
"real_time": 7.7622128175128732e+04,
|
|
"cpu_time": 2.7965941226613527e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3192114466247069e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9033,
|
|
"real_time": 7.7436884263685599e+04,
|
|
"cpu_time": 2.7887871947304334e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3223672539730653e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9033,
|
|
"real_time": 7.7400070028531904e+04,
|
|
"cpu_time": 2.7894172689029103e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3229962190247685e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9033,
|
|
"real_time": 7.7558241222929573e+04,
|
|
"cpu_time": 2.7920338724676182e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3202981190053873e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9033,
|
|
"real_time": 7.7307865440764654e+04,
|
|
"cpu_time": 2.7880830942101195e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3245741480013002e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7465037826208092e+04,
|
|
"cpu_time": 2.7909831105944875e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3218894373258457e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7436884263685599e+04,
|
|
"cpu_time": 2.7894172689029103e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3223672539730653e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2557950414638132e+02,
|
|
"cpu_time": 3.4738868230188717e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1427128879792785e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6211120225374117e-03,
|
|
"cpu_time": 1.2446821372125480e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6209471287658831e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9301,
|
|
"real_time": 7.7195294497975920e+04,
|
|
"cpu_time": 2.7875992635200504e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6530114475483984e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9301,
|
|
"real_time": 7.7382355819046192e+04,
|
|
"cpu_time": 2.7877881238576508e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6465981531876855e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9301,
|
|
"real_time": 7.7345920500265973e+04,
|
|
"cpu_time": 2.7866944285560661e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6478448853588309e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9301,
|
|
"real_time": 7.7405324886844028e+04,
|
|
"cpu_time": 2.7871900688098074e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6458128080902644e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9301,
|
|
"real_time": 7.6994885602807641e+04,
|
|
"cpu_time": 2.7849192312654556e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6599169333986502e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7264756261387942e+04,
|
|
"cpu_time": 2.7868382232018060e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6506368455167662e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7345920500265973e+04,
|
|
"cpu_time": 2.7871900688098074e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6478448853588309e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7163328424413587e+02,
|
|
"cpu_time": 1.1520086694981407e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.8978000185980367e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.2213657629812180e-03,
|
|
"cpu_time": 4.1337479151358654e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2250501907017015e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9061,
|
|
"real_time": 7.7477862720084086e+04,
|
|
"cpu_time": 2.7899607493654115e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2866713874106653e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9061,
|
|
"real_time": 7.7454629817347697e+04,
|
|
"cpu_time": 2.7908812040613568e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2882571508754529e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9061,
|
|
"real_time": 7.7450707062008179e+04,
|
|
"cpu_time": 2.7873792583600042e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2885249927036583e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9061,
|
|
"real_time": 7.7566102187136581e+04,
|
|
"cpu_time": 2.7905493521686381e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2806572516921870e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9061,
|
|
"real_time": 7.7625824017308114e+04,
|
|
"cpu_time": 2.7926438141485455e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2765945506571643e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7515025160776946e+04,
|
|
"cpu_time": 2.7902828756207915e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2841410666678265e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7477862720084086e+04,
|
|
"cpu_time": 2.7905493521686381e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2866713874106653e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7541459171466514e+01,
|
|
"cpu_time": 1.9064230854563536e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.2832364701951396e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0003410178947209e-03,
|
|
"cpu_time": 6.8323649265567964e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.9982880917423014e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8950,
|
|
"real_time": 7.7925954951726511e+04,
|
|
"cpu_time": 2.8070843877094961e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0512543612811381e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8950,
|
|
"real_time": 7.7644784487855446e+04,
|
|
"cpu_time": 2.8051572581005574e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0550612064975625e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8950,
|
|
"real_time": 7.7769835398261886e+04,
|
|
"cpu_time": 2.8044154391061427e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0533647085722244e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8950,
|
|
"real_time": 7.7674976165026586e+04,
|
|
"cpu_time": 2.8042990290502791e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0546511121671222e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8950,
|
|
"real_time": 7.7907380543248568e+04,
|
|
"cpu_time": 2.8056875642458157e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0515049977135852e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7784586309223814e+04,
|
|
"cpu_time": 2.8053287356424582e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0531672772463265e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7769835398261886e+04,
|
|
"cpu_time": 2.8051572581005574e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0533647085722244e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2926720324041651e+02,
|
|
"cpu_time": 1.1330140032552897e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7500095674024185e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6618614223456738e-03,
|
|
"cpu_time": 4.0387922772117265e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6616634462647728e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8826,
|
|
"real_time": 7.9748117373028523e+04,
|
|
"cpu_time": 2.8367841162474547e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0544685617294842e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8826,
|
|
"real_time": 7.9658727413601839e+04,
|
|
"cpu_time": 2.8369197337412217e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0567740073139063e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8826,
|
|
"real_time": 7.9485271903024346e+04,
|
|
"cpu_time": 2.8288674461817415e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0612623707180905e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8826,
|
|
"real_time": 7.9229827018406868e+04,
|
|
"cpu_time": 2.8284198855653702e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0679080867100251e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8826,
|
|
"real_time": 7.9537469873559312e+04,
|
|
"cpu_time": 2.8320961975980096e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0599096282601944e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9531882716324195e+04,
|
|
"cpu_time": 2.8326174758667604e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0600645309463403e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9537469873559341e+04,
|
|
"cpu_time": 2.8320961975980102e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0599096282601944e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9766665935008120e+02,
|
|
"cpu_time": 4.1177825613286751e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.1264656252626918e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4853763371240982e-03,
|
|
"cpu_time": 1.4537023076399203e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4884975923097547e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8566,
|
|
"real_time": 8.2053540718342745e+04,
|
|
"cpu_time": 2.8806988337613840e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9934900691829425e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8566,
|
|
"real_time": 8.1889746469591642e+04,
|
|
"cpu_time": 2.8799961218771833e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.0014777689130878e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8566,
|
|
"real_time": 8.1521729606258043e+04,
|
|
"cpu_time": 2.8705518841933366e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.0195418029360056e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8566,
|
|
"real_time": 8.1553944989880620e+04,
|
|
"cpu_time": 2.8695835127247160e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.0179540062796372e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8566,
|
|
"real_time": 8.2152258628065669e+04,
|
|
"cpu_time": 2.8767395773990185e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9886913089453965e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.1834244082427758e+04,
|
|
"cpu_time": 2.8755139859911276e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.0042309912514138e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.1889746469591657e+04,
|
|
"cpu_time": 2.8767395773990185e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.0014777689130878e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8659102385985091e+02,
|
|
"cpu_time": 5.2025770105251468e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4028288047655709e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.5020916619108913e-03,
|
|
"cpu_time": 1.8092685467262404e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5033663338366870e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8020,
|
|
"real_time": 8.7139039878866126e+04,
|
|
"cpu_time": 2.9624378179551085e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.5208540386838114e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8020,
|
|
"real_time": 8.7058481544874870e+04,
|
|
"cpu_time": 2.9634466720698238e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.5278133545459366e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8020,
|
|
"real_time": 8.7131180006601644e+04,
|
|
"cpu_time": 2.9623326820448897e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.5215324749457717e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8020,
|
|
"real_time": 8.7097663537380256e+04,
|
|
"cpu_time": 2.9607293615960213e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.5244268718957663e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8020,
|
|
"real_time": 8.7138580650722884e+04,
|
|
"cpu_time": 2.9581844364089699e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.5208936742598093e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.7112989123689156e+04,
|
|
"cpu_time": 2.9614261940149625e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.5231040828662193e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.7131180006601644e+04,
|
|
"cpu_time": 2.9623326820448897e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.5215324749457717e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4894218977096223e+01,
|
|
"cpu_time": 2.0568181074801777e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.0143642381105840e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.0056275565921573e-04,
|
|
"cpu_time": 6.9453633915881610e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0068091640201590e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7191,
|
|
"real_time": 9.4524010053624123e+04,
|
|
"cpu_time": 3.1246118217215902e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3866529776470757e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7191,
|
|
"real_time": 9.4389951583834350e+04,
|
|
"cpu_time": 3.1213037240995682e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3886223883014255e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7191,
|
|
"real_time": 9.6790214781456438e+04,
|
|
"cpu_time": 3.1172400917813869e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3541864773825405e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7191,
|
|
"real_time": 9.4456254043549328e+04,
|
|
"cpu_time": 3.1189074718397990e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3876476611021316e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7191,
|
|
"real_time": 9.4252980360317699e+04,
|
|
"cpu_time": 3.1179628605200851e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3906403754971743e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.4882682164556405e+04,
|
|
"cpu_time": 3.1200051939924865e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3815499759860697e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.4456254043549328e+04,
|
|
"cpu_time": 3.1189074718397990e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3876476611021316e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0710363086187599e+03,
|
|
"cpu_time": 2.9970447795394568e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5367480476476291e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1288006242922668e-02,
|
|
"cpu_time": 9.6058967636022312e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1123361980088980e-02,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 6112,
|
|
"real_time": 1.1474508862564449e+05,
|
|
"cpu_time": 3.4596641966623027e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2845770842118044e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 6112,
|
|
"real_time": 1.1461435848338209e+05,
|
|
"cpu_time": 3.4553172005890211e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2871828928659768e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 6112,
|
|
"real_time": 1.1517446763366539e+05,
|
|
"cpu_time": 3.4614308147905738e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2760600103991761e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 6112,
|
|
"real_time": 1.1482692212328651e+05,
|
|
"cpu_time": 3.4580333949607343e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2829489387388024e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 6112,
|
|
"real_time": 1.1473427873232197e+05,
|
|
"cpu_time": 3.4583089692408493e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2847923296889215e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1481902311966009e+05,
|
|
"cpu_time": 3.4585509152486961e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2831122511809363e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1474508862564449e+05,
|
|
"cpu_time": 3.4583089692408487e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2845770842118044e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1268446742406960e+02,
|
|
"cpu_time": 2.2536910008626359e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.2219454735275777e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8523452094033042e-03,
|
|
"cpu_time": 6.5162868961279354e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8492062627861519e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4419,
|
|
"real_time": 1.5870898509992743e+05,
|
|
"cpu_time": 4.1804784679791855e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3034550606564217e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4419,
|
|
"real_time": 1.5855842671720329e+05,
|
|
"cpu_time": 4.1796369155917771e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3065918403384089e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4419,
|
|
"real_time": 1.5850332219376092e+05,
|
|
"cpu_time": 4.1755643267707585e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3077413945878620e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4419,
|
|
"real_time": 1.5872892396030121e+05,
|
|
"cpu_time": 4.1786641977822955e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3030400945143857e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4419,
|
|
"real_time": 1.5896887240375424e+05,
|
|
"cpu_time": 4.1875464426340733e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2980544686031146e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5869370607498946e+05,
|
|
"cpu_time": 4.1803780701516184e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3037765717400389e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5870898509992746e+05,
|
|
"cpu_time": 4.1796369155917771e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3034550606564217e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8147781089069761e+02,
|
|
"cpu_time": 4.4174211280471008e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.7760359466085597e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1435728320878819e-03,
|
|
"cpu_time": 1.0567037368194032e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1429453126183379e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2801,
|
|
"real_time": 2.4744380099413119e+05,
|
|
"cpu_time": 5.6404454551945871e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2376329323556981e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2801,
|
|
"real_time": 2.5035907400804228e+05,
|
|
"cpu_time": 5.6490175758658000e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1882883780210686e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2801,
|
|
"real_time": 2.4752509000835661e+05,
|
|
"cpu_time": 5.6449455515887914e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2362412633183947e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2801,
|
|
"real_time": 2.5014097339381231e+05,
|
|
"cpu_time": 5.6418074973224127e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1919401918579822e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2801,
|
|
"real_time": 2.4764967856585240e+05,
|
|
"cpu_time": 5.6446983577294613e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2341100786899419e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4862372339403900e+05,
|
|
"cpu_time": 5.6441828875402105e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2176425688486171e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4764967856585240e+05,
|
|
"cpu_time": 5.6446983577294624e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2341100786899419e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4884106312292374e+03,
|
|
"cpu_time": 3.3103042298076798e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5194174185810495e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.9865993916850971e-03,
|
|
"cpu_time": 5.8649839946103204e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.9735204618556158e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1636,
|
|
"real_time": 4.2767761120971746e+05,
|
|
"cpu_time": 8.5474152689486207e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9035814478762455e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1636,
|
|
"real_time": 4.2763192383521225e+05,
|
|
"cpu_time": 8.5416843459658534e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9041053371126156e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1636,
|
|
"real_time": 4.2781677805963211e+05,
|
|
"cpu_time": 8.5471158618581621e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9019863351588430e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1636,
|
|
"real_time": 4.2804718465307466e+05,
|
|
"cpu_time": 8.5448043337408616e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8993477242461195e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1636,
|
|
"real_time": 4.2775496344502177e+05,
|
|
"cpu_time": 8.5420986430317280e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9026947182800875e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2778569224053173e+05,
|
|
"cpu_time": 8.5446236907090456e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9023431125347824e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2775496344502189e+05,
|
|
"cpu_time": 8.5448043337408616e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9026947182800875e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6249238474118579e+02,
|
|
"cpu_time": 2.6949986678597696e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8615972703031127e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.7984530031878895e-04,
|
|
"cpu_time": 3.1540285042513494e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.7973622563120923e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0810976120170089e+05,
|
|
"cpu_time": 1.4774104466357152e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1902652354586754e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0760229337612586e+05,
|
|
"cpu_time": 1.4771326264501070e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1935266088287106e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0779497822305828e+05,
|
|
"cpu_time": 1.4774436856148303e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1922877872135239e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0807581215744652e+05,
|
|
"cpu_time": 1.4774725046403713e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1904832899301977e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0792582742416079e+05,
|
|
"cpu_time": 1.4775482691415397e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1914468601311245e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0790173447649856e+05,
|
|
"cpu_time": 1.4774015064965128e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1916019563124466e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0792582742416079e+05,
|
|
"cpu_time": 1.4774436856148303e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1914468601311245e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0929673719675043e+02,
|
|
"cpu_time": 1.5869427561490158e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3450783692855968e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.5906212137589957e-04,
|
|
"cpu_time": 1.0741445363165141e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5908734540985401e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 441,
|
|
"real_time": 1.5720614119139321e+06,
|
|
"cpu_time": 2.7389811995465439e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3360561721231689e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 441,
|
|
"real_time": 1.5800190115951491e+06,
|
|
"cpu_time": 2.7384551065759715e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3091816860678549e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 441,
|
|
"real_time": 1.5723204452443083e+06,
|
|
"cpu_time": 2.7334466258503790e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3351770788025160e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 441,
|
|
"real_time": 1.5804724878147920e+06,
|
|
"cpu_time": 2.7327479251700821e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3076583519643154e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 441,
|
|
"real_time": 1.5725954414647317e+06,
|
|
"cpu_time": 2.7339006507936586e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3342441284115410e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5754937596065828e+06,
|
|
"cpu_time": 2.7355063015873274e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3244634834738798e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5725954414647317e+06,
|
|
"cpu_time": 2.7339006507936586e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3342441284115410e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3450207969251524e+03,
|
|
"cpu_time": 2.9664535116946076e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4669512154440584e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.7578787731981557e-03,
|
|
"cpu_time": 1.0844257642445454e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7551155529513829e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1852891848591920e+06,
|
|
"cpu_time": 5.3633389409091072e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2670935121834631e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1872164659117432e+06,
|
|
"cpu_time": 5.3664091090908526e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2639085482387114e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1888375344516877e+06,
|
|
"cpu_time": 5.3676288136363775e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2612326023955927e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1878797263330356e+06,
|
|
"cpu_time": 5.3663101590909818e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2628133556652555e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1877968439155002e+06,
|
|
"cpu_time": 5.3673866727273576e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2629501883165541e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1874039510942320e+06,
|
|
"cpu_time": 5.3662147390909353e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2635996413599167e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1877968439155007e+06,
|
|
"cpu_time": 5.3664091090908526e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2629501883165541e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3175047444720735e+03,
|
|
"cpu_time": 1.7095419804529120e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1763194071459272e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.1334727718454875e-04,
|
|
"cpu_time": 3.1857502235225815e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.1346598438927773e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 6.6909265285357833e+06,
|
|
"cpu_time": 1.0885648220000234e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.0149156259443970e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 6.7031313618645072e+06,
|
|
"cpu_time": 1.0899702510000112e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.0057846383405323e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 6.6902225138619542e+06,
|
|
"cpu_time": 1.0892897329999870e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.0154433474336843e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 6.7804312193766236e+06,
|
|
"cpu_time": 1.1132999339999968e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.9487165217620058e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 6.7067730193957686e+06,
|
|
"cpu_time": 1.0915768220000073e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.0030665870101280e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.7142969286069274e+06,
|
|
"cpu_time": 1.0945403124000054e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.9975853440981493e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.7031313618645072e+06,
|
|
"cpu_time": 1.0899702510000112e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 5.0057846383405323e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7685463214902433e+04,
|
|
"cpu_time": 1.0545925856459734e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7859833086673006e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.6127191894567215e-03,
|
|
"cpu_time": 9.6350273598745928e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.5746587938860937e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.4025020632235443e+07,
|
|
"cpu_time": 2.2827394745098002e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7849387006073551e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.4010997014302833e+07,
|
|
"cpu_time": 2.2794314647059057e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7897279495166063e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.3996195212444838e+07,
|
|
"cpu_time": 2.2801981784313589e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7947933692957907e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.3964530774483494e+07,
|
|
"cpu_time": 2.2709169705882348e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.8056655167120829e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.4014482425124038e+07,
|
|
"cpu_time": 2.2827237980392039e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7885367410852518e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4002245211718127e+07,
|
|
"cpu_time": 2.2792019772549007e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7927324554434175e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4010997014302835e+07,
|
|
"cpu_time": 2.2801981784313589e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7897279495166063e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3473626090760365e+04,
|
|
"cpu_time": 4.8632171762732876e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.0449538371926034e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6764187268421692e-03,
|
|
"cpu_time": 2.1337368187661046e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6785735302323056e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 27,
|
|
"real_time": 2.6435841081870928e+07,
|
|
"cpu_time": 4.2795387962963067e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0771120761519232e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 27,
|
|
"real_time": 2.6404617354273796e+07,
|
|
"cpu_time": 4.2740501703704059e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0831158126317558e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 27,
|
|
"real_time": 2.6409630767173238e+07,
|
|
"cpu_time": 4.2769515333333172e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0821508707660751e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 27,
|
|
"real_time": 2.6428807181892570e+07,
|
|
"cpu_time": 4.2785779518518895e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0784633251234245e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 27,
|
|
"real_time": 2.6422140223008614e+07,
|
|
"cpu_time": 4.2772966222223110e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0797447469119902e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6420207321643829e+07,
|
|
"cpu_time": 4.2772830148148462e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0801173663170347e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6422140223008610e+07,
|
|
"cpu_time": 4.2772966222223110e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0797447469119902e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3009791982424274e+04,
|
|
"cpu_time": 2.0812712676840558e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5016138335402608e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.9241823972238311e-04,
|
|
"cpu_time": 4.8658722382300654e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9243229105824218e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.0380702689290047e+07,
|
|
"cpu_time": 8.0095505099998832e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3281403726245394e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.0419471785426140e+07,
|
|
"cpu_time": 8.0073222300001130e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3240434001847649e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.0439391285181046e+07,
|
|
"cpu_time": 8.0002854300002471e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3219408315672045e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.0455615669488907e+07,
|
|
"cpu_time": 8.0122515700000465e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3202295212963982e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.0412702560424805e+07,
|
|
"cpu_time": 8.0015644299999163e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3247582923818159e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.0421576797962189e+07,
|
|
"cpu_time": 8.0061948340000421e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3238224836109447e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.0419471785426140e+07,
|
|
"cpu_time": 8.0073222300001130e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.3240434001847649e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8416001908235583e+04,
|
|
"cpu_time": 5.1375415080759994e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.0007790783594851e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.6356829184651827e-04,
|
|
"cpu_time": 6.4169578864834960e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.6365122759017530e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.9116587804423437e+07,
|
|
"cpu_time": 1.5694365122222269e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4165596686939240e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.9276623792118490e+07,
|
|
"cpu_time": 1.5710115988888952e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4078280615604677e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.9274385306570262e+07,
|
|
"cpu_time": 1.5719449599999774e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4079499998119688e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.9264661471048996e+07,
|
|
"cpu_time": 1.5721402488888872e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4084797554725046e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.9306651287608683e+07,
|
|
"cpu_time": 1.5716372833333340e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4061928887837734e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.9247781932353988e+07,
|
|
"cpu_time": 1.5712341206666642e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4094020748645287e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.9274385306570262e+07,
|
|
"cpu_time": 1.5716372833333343e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4079499998119688e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5003352269169103e+04,
|
|
"cpu_time": 1.0919950902214705e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0916603473504493e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.5571817131682027e-04,
|
|
"cpu_time": 6.9499196577919539e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.5639789587149885e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7572510987520218e+08,
|
|
"cpu_time": 2.8233627925000346e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1103494245219612e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7582659423351288e+08,
|
|
"cpu_time": 2.8211971549999505e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1068226264678612e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7586971446871758e+08,
|
|
"cpu_time": 2.8231693549999905e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1053253383827457e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7594934999942780e+08,
|
|
"cpu_time": 2.8251226924999881e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1025620384701157e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7602631449699402e+08,
|
|
"cpu_time": 2.8252795099999875e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.0998937975170536e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7587941661477092e+08,
|
|
"cpu_time": 2.8236263009999907e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1049906450719471e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7586971446871758e+08,
|
|
"cpu_time": 2.8233627925000346e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.1053253383827457e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1528721907682721e+05,
|
|
"cpu_time": 1.6696761803393063e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0018957299759821e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.5549011530633555e-04,
|
|
"cpu_time": 5.9132335597950362e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.5551218054795562e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9324,
|
|
"real_time": 7.5152164992015300e+04,
|
|
"cpu_time": 2.7379194583869632e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3625688629313568e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9324,
|
|
"real_time": 7.5292586983273286e+04,
|
|
"cpu_time": 2.7420790744315792e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3600276481766896e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9324,
|
|
"real_time": 7.5097374841526151e+04,
|
|
"cpu_time": 2.7395141419991601e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3635629769494481e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9324,
|
|
"real_time": 7.4959766057453089e+04,
|
|
"cpu_time": 2.7378741151866247e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3660661630335838e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9324,
|
|
"real_time": 7.5115221369512554e+04,
|
|
"cpu_time": 2.7387774517374422e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3632390097909195e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5123422848756076e+04,
|
|
"cpu_time": 2.7392328483483533e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3630929321763996e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5115221369512539e+04,
|
|
"cpu_time": 2.7387774517374422e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3632390097909195e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1934396033804019e+02,
|
|
"cpu_time": 1.7290650726512894e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1652365380311225e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5886384806814795e-03,
|
|
"cpu_time": 6.3122237808072620e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5884731605012214e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9066,
|
|
"real_time": 7.5136350816257342e+04,
|
|
"cpu_time": 2.7398657180675032e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7257112938693207e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9066,
|
|
"real_time": 7.5027600659598145e+04,
|
|
"cpu_time": 2.7370817229208030e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7296621269975305e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9066,
|
|
"real_time": 7.5161817887789584e+04,
|
|
"cpu_time": 2.7390526516655734e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7247877413735464e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9066,
|
|
"real_time": 7.5155571221360427e+04,
|
|
"cpu_time": 2.7414776472534629e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7250142161356173e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9066,
|
|
"real_time": 7.4894356776528948e+04,
|
|
"cpu_time": 2.7387086190160981e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7345184445750128e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5075139472306910e+04,
|
|
"cpu_time": 2.7392372717846878e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7279387645902056e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5136350816257342e+04,
|
|
"cpu_time": 2.7390526516655728e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7257112938693207e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1475143179772707e+02,
|
|
"cpu_time": 1.6104484014496822e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.1744976786135601e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5284877604530541e-03,
|
|
"cpu_time": 5.8791854872813952e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5302754346249625e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9320,
|
|
"real_time": 7.5055235975639051e+04,
|
|
"cpu_time": 2.7404586437768210e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4573141323937126e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9320,
|
|
"real_time": 7.5227442818955882e+04,
|
|
"cpu_time": 2.7424614710300998e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4448215259124644e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9320,
|
|
"real_time": 7.4968970093245080e+04,
|
|
"cpu_time": 2.7407171255364758e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4635937974144071e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9320,
|
|
"real_time": 7.4838514929811427e+04,
|
|
"cpu_time": 2.7395978766094596e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4731176905922070e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9320,
|
|
"real_time": 7.5037107614970460e+04,
|
|
"cpu_time": 2.7404948980686505e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4586325755216308e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5025454286524386e+04,
|
|
"cpu_time": 2.7407460030043015e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4594959443668842e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5037107614970460e+04,
|
|
"cpu_time": 2.7404948980686505e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4586325755216308e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4138502703566732e+02,
|
|
"cpu_time": 1.0497008736444836e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0285774595284498e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8844941144336668e-03,
|
|
"cpu_time": 3.8299823204844281e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8840154292810447e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9010,
|
|
"real_time": 7.5905982766047964e+04,
|
|
"cpu_time": 2.7638487014428136e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0792298184517024e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9010,
|
|
"real_time": 7.5814047302671141e+04,
|
|
"cpu_time": 2.7635072674806265e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0805385402121082e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9010,
|
|
"real_time": 7.5837705329999371e+04,
|
|
"cpu_time": 2.7616268867924850e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0802014597294867e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9010,
|
|
"real_time": 7.5798828179706936e+04,
|
|
"cpu_time": 2.7618348690344236e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0807554940794170e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9010,
|
|
"real_time": 7.5805755934700501e+04,
|
|
"cpu_time": 2.7628844350721227e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0806567257315704e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5832463902625197e+04,
|
|
"cpu_time": 2.7627404319644946e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0802764076408571e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5814047302671155e+04,
|
|
"cpu_time": 2.7628844350721233e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0805385402121082e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3635835033303657e+01,
|
|
"cpu_time": 9.8704887218108723e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.2125785105703093e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.7542420208494724e-04,
|
|
"cpu_time": 3.5727166430877074e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.7509156606849732e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9265,
|
|
"real_time": 7.7490174436382091e+04,
|
|
"cpu_time": 2.7937923961144482e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1143325743124944e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9265,
|
|
"real_time": 7.7426883800117837e+04,
|
|
"cpu_time": 2.7894457323260029e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1160608816824245e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9265,
|
|
"real_time": 7.7667860305140319e+04,
|
|
"cpu_time": 2.7924453696708527e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1094954767172402e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9265,
|
|
"real_time": 7.7377732675605497e+04,
|
|
"cpu_time": 2.7889191991365561e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1174050251236302e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9265,
|
|
"real_time": 7.7498276538343242e+04,
|
|
"cpu_time": 2.7926115434430278e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1141115301956183e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7492185551117800e+04,
|
|
"cpu_time": 2.7914428481381777e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1142810976062813e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7490174436382105e+04,
|
|
"cpu_time": 2.7924453696708521e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1143325743124944e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0984823998316847e+02,
|
|
"cpu_time": 2.1359160604890303e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9941661922812503e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4175395777256399e-03,
|
|
"cpu_time": 7.6516560670895786e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4161627778213339e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8787,
|
|
"real_time": 7.9439838367629418e+04,
|
|
"cpu_time": 2.8283462831456109e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1248825115123200e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8787,
|
|
"real_time": 7.9484920661395241e+04,
|
|
"cpu_time": 2.8305167554341676e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1225429587570786e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8787,
|
|
"real_time": 7.9418495194291711e+04,
|
|
"cpu_time": 2.8299043302606524e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1259910452641308e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8787,
|
|
"real_time": 7.9441346464894814e+04,
|
|
"cpu_time": 2.8295980539433140e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1248042056387103e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8787,
|
|
"real_time": 7.9468860589282849e+04,
|
|
"cpu_time": 2.8301037248207675e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1233760943615544e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9450692255498812e+04,
|
|
"cpu_time": 2.8296938295209024e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1243193631067586e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9441346464894799e+04,
|
|
"cpu_time": 2.8299043302606524e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1248042056387103e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6189089133495280e+01,
|
|
"cpu_time": 8.2380003173886934e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3594206486588321e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.2962694710419880e-04,
|
|
"cpu_time": 2.9112691385355550e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.2961090763708716e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8246,
|
|
"real_time": 8.4696899785613845e+04,
|
|
"cpu_time": 2.9201278098472086e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7377094280765617e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8246,
|
|
"real_time": 8.4725037966082062e+04,
|
|
"cpu_time": 2.9185642893524340e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7351396438719797e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8246,
|
|
"real_time": 8.4599807737781142e+04,
|
|
"cpu_time": 2.9177455032743211e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7465897089423871e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8246,
|
|
"real_time": 8.4760043176111052e+04,
|
|
"cpu_time": 2.9190100945913396e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7319450939674366e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8246,
|
|
"real_time": 8.4659220755574381e+04,
|
|
"cpu_time": 2.9196456560756330e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7411532276222599e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.4688201884232505e+04,
|
|
"cpu_time": 2.9190186706281872e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7385074204961264e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.4696899785613845e+04,
|
|
"cpu_time": 2.9190100945913396e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7377094280765617e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.1739680457131286e+01,
|
|
"cpu_time": 9.2881531669333327e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.6428388812724396e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.2902339503592842e-04,
|
|
"cpu_time": 3.1819437334857731e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.2918956778756565e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7354,
|
|
"real_time": 9.1966116547431040e+04,
|
|
"cpu_time": 3.0816895580636361e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4252205586217213e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7354,
|
|
"real_time": 9.4885934788595288e+04,
|
|
"cpu_time": 3.0777875197171140e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3813638479931383e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7354,
|
|
"real_time": 9.1913557432849062e+04,
|
|
"cpu_time": 3.0729356663040316e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4260355453629310e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7354,
|
|
"real_time": 9.1937948933060805e+04,
|
|
"cpu_time": 3.0766336429154535e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4256572125122385e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7354,
|
|
"real_time": 9.4675557681871054e+04,
|
|
"cpu_time": 3.0746958430786361e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3844333554435279e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.3075823076761473e+04,
|
|
"cpu_time": 3.0767484460157744e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4085421039867115e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1966116547431055e+04,
|
|
"cpu_time": 3.0766336429154535e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4252205586217213e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5582620761006217e+03,
|
|
"cpu_time": 3.3263309417701311e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3436124014036108e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6741856527183137e-02,
|
|
"cpu_time": 1.0811189150278283e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6638568309532909e-02,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 6184,
|
|
"real_time": 1.1316918665687587e+05,
|
|
"cpu_time": 3.4268131516817742e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3163902449419327e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 6184,
|
|
"real_time": 1.1327700101908488e+05,
|
|
"cpu_time": 3.4272374142949370e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3141855596603770e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 6184,
|
|
"real_time": 1.1297093010721942e+05,
|
|
"cpu_time": 3.4250612597024674e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3204553574198432e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 6184,
|
|
"real_time": 1.1307153780989825e+05,
|
|
"cpu_time": 3.4261759152651206e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3183906850257058e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 6184,
|
|
"real_time": 1.1326781771812777e+05,
|
|
"cpu_time": 3.4266434411384532e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3143731845558949e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1315129466224124e+05,
|
|
"cpu_time": 3.4263862364165502e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3167590063207507e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1316918665687589e+05,
|
|
"cpu_time": 3.4266434411384526e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3163902449419327e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3095071819006762e+02,
|
|
"cpu_time": 8.3257326611190564e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6821700639892318e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1573064062674490e-03,
|
|
"cpu_time": 2.4298873759855032e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1577251050590674e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4431,
|
|
"real_time": 1.5789855943665479e+05,
|
|
"cpu_time": 4.1628248792597128e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3204102803124819e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4431,
|
|
"real_time": 1.5777160236391795e+05,
|
|
"cpu_time": 4.1603898375085159e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3230821779363737e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4431,
|
|
"real_time": 1.5777719608597693e+05,
|
|
"cpu_time": 4.1619881832543097e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3229643637113547e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4431,
|
|
"real_time": 1.5791993607861301e+05,
|
|
"cpu_time": 4.1626227691265865e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3199608169737859e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4431,
|
|
"real_time": 1.5761199650776942e+05,
|
|
"cpu_time": 4.1621984924395819e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3264472985351434e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5779585809458641e+05,
|
|
"cpu_time": 4.1620048323177412e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3225729874938278e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5777719608597693e+05,
|
|
"cpu_time": 4.1621984924395813e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3229643637113547e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2317122830532601e+02,
|
|
"cpu_time": 9.6180982726987438e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5943917153120884e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.8057326594399199e-04,
|
|
"cpu_time": 2.3109291459766539e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.8083814112658606e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2852,
|
|
"real_time": 2.4834617865204110e+05,
|
|
"cpu_time": 5.6295093969143729e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2222352914443851e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2852,
|
|
"real_time": 2.4542003848437621e+05,
|
|
"cpu_time": 5.6266944214585959e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2725769520517530e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2852,
|
|
"real_time": 2.4837091855598983e+05,
|
|
"cpu_time": 5.6292190322581318e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2218147200821395e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2852,
|
|
"real_time": 2.4549606088360318e+05,
|
|
"cpu_time": 5.6281518022439245e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2712538695158963e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2852,
|
|
"real_time": 2.4839319223175567e+05,
|
|
"cpu_time": 5.6269007187937782e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2214361455673800e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4720527776155324e+05,
|
|
"cpu_time": 5.6280950743337604e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2418633957323108e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4834617865204107e+05,
|
|
"cpu_time": 5.6281518022439233e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2222352914443851e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5953068438970563e+03,
|
|
"cpu_time": 1.2898761065119669e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7439054801400140e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.4533688695588513e-03,
|
|
"cpu_time": 2.2918520200454487e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4686323536505804e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1644,
|
|
"real_time": 4.2588470688224555e+05,
|
|
"cpu_time": 8.4994227737226116e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9242247164790764e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1644,
|
|
"real_time": 4.2585988942286745e+05,
|
|
"cpu_time": 8.4994459975667833e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9245116811590223e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1644,
|
|
"real_time": 4.2588256343116035e+05,
|
|
"cpu_time": 8.5003572810220323e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9242494999187346e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1644,
|
|
"real_time": 4.2575428679366224e+05,
|
|
"cpu_time": 8.5018429927006806e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9257331401019220e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1644,
|
|
"real_time": 4.2561852151398838e+05,
|
|
"cpu_time": 8.4963109489052126e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9273043676297693e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2579999360878486e+05,
|
|
"cpu_time": 8.4994759987834655e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9252046810577049e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2585988942286745e+05,
|
|
"cpu_time": 8.4994459975667833e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9245116811590223e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1467656341599775e+02,
|
|
"cpu_time": 2.0248822331756710e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3267350081157880e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.6932025631113538e-04,
|
|
"cpu_time": 2.3823612578769486e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6937662372051246e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 866,
|
|
"real_time": 8.0451549339488870e+05,
|
|
"cpu_time": 1.4714187424942004e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2134533572509661e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 866,
|
|
"real_time": 8.0416426863779442e+05,
|
|
"cpu_time": 1.4706190219399640e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2157303719859343e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 866,
|
|
"real_time": 8.0493268255503557e+05,
|
|
"cpu_time": 1.4719703083141122e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2107512726236258e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 866,
|
|
"real_time": 8.0450404510407452e+05,
|
|
"cpu_time": 1.4713251374134433e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2135275459769812e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 866,
|
|
"real_time": 8.0443328529043566e+05,
|
|
"cpu_time": 1.4707835369514623e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2139861399266605e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0450995499644580e+05,
|
|
"cpu_time": 1.4712233494226367e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2134897375528336e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0450404510407464e+05,
|
|
"cpu_time": 1.4713251374134435e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2135275459769812e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7574512300962073e+02,
|
|
"cpu_time": 5.3970676029928882e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7866729665162563e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.4274917457154265e-04,
|
|
"cpu_time": 3.6684216608653611e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.4270192451839467e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 446,
|
|
"real_time": 1.5753160680815445e+06,
|
|
"cpu_time": 2.7306748071747981e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3250316999659863e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 446,
|
|
"real_time": 1.5718174465063275e+06,
|
|
"cpu_time": 2.7267315627802936e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3368843936968164e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 446,
|
|
"real_time": 1.5749218445605361e+06,
|
|
"cpu_time": 2.7300084461882305e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3263646249955635e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 446,
|
|
"real_time": 1.5734147059630952e+06,
|
|
"cpu_time": 2.7306240874439217e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3314666300041285e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 446,
|
|
"real_time": 1.5755861135191314e+06,
|
|
"cpu_time": 2.7304507376681161e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3241190234050274e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5742112357261269e+06,
|
|
"cpu_time": 2.7296979282510718e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3287732744135046e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5749218445605361e+06,
|
|
"cpu_time": 2.7304507376681161e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3263646249955635e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5796039763092067e+03,
|
|
"cpu_time": 1.6788784055713259e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.3504075583678670e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0034256778637413e-03,
|
|
"cpu_time": 6.1504182869310728e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0040598995003673e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1792140153067354e+06,
|
|
"cpu_time": 5.3560008181818835e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2771584168992500e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1801726108162920e+06,
|
|
"cpu_time": 5.3577139409092274e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2755677295433340e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1795086095702243e+06,
|
|
"cpu_time": 5.3553246590909837e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2766694669424992e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1767579853873360e+06,
|
|
"cpu_time": 5.3539669181816401e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2812383181762543e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1803027906624433e+06,
|
|
"cpu_time": 5.3584681272726897e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2753517838801069e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1791912023486067e+06,
|
|
"cpu_time": 5.3562948927272847e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2771971430882902e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1795086095702248e+06,
|
|
"cpu_time": 5.3560008181818826e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2766694669424992e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4335797596203142e+03,
|
|
"cpu_time": 1.8154851125881312e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3807234708130215e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.5092593316226677e-04,
|
|
"cpu_time": 3.3894420470634957e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.5113407861427525e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.1065133670344949e+06,
|
|
"cpu_time": 1.1948709510000413e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7216448160994682e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.1104654856026173e+06,
|
|
"cpu_time": 1.1938254599999709e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7190204449963989e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.1095663262531161e+06,
|
|
"cpu_time": 1.1921918859999893e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7196172678065243e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.1137839974835515e+06,
|
|
"cpu_time": 1.1942890999999918e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7168190673022451e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.0096231997013092e+06,
|
|
"cpu_time": 1.1602646649999999e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7869095162532854e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0899904752150178e+06,
|
|
"cpu_time": 1.1870884123999987e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7328022224915838e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1095663262531161e+06,
|
|
"cpu_time": 1.1938254599999709e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7196172678065243e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5001323935543456e+04,
|
|
"cpu_time": 1.5028011234890897e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.0295744859892122e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.3471628195916150e-03,
|
|
"cpu_time": 1.2659555158581644e-02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4012277369035987e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4392152808758678e+07,
|
|
"cpu_time": 2.3773366755102526e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6628787848305330e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4390190284014964e+07,
|
|
"cpu_time": 2.3765079775510188e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6635147051909695e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4379852090258988e+07,
|
|
"cpu_time": 2.3741692428571623e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6668674739331999e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4290685746438649e+07,
|
|
"cpu_time": 2.3493958857142974e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6959862662100773e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4400401019624302e+07,
|
|
"cpu_time": 2.3767605142856322e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6602079975791416e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4370656389819119e+07,
|
|
"cpu_time": 2.3708340591836728e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6698910455487843e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4390190284014964e+07,
|
|
"cpu_time": 2.3765079775510188e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6635147051909695e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5299770311501765e+04,
|
|
"cpu_time": 1.2044905257427278e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4779153066072494e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.1522408637920229e-03,
|
|
"cpu_time": 5.0804505742483670e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1647747071443110e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6610708365646694e+07,
|
|
"cpu_time": 4.3352412423075750e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0437487854802637e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6803725112516146e+07,
|
|
"cpu_time": 4.3448223692306645e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0074281629356918e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6795497235770408e+07,
|
|
"cpu_time": 4.3418737230767258e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0089657534261856e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6732672292452592e+07,
|
|
"cpu_time": 4.3413806423077315e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0207374156864052e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6816506655170366e+07,
|
|
"cpu_time": 4.3370240538461775e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0050414741146793e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6751821932311244e+07,
|
|
"cpu_time": 4.3400684061537758e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0171843183286457e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6795497235770412e+07,
|
|
"cpu_time": 4.3413806423077315e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0089657534261856e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.5242790050731011e+04,
|
|
"cpu_time": 3.8778502300728083e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6034895217857832e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.1864293305486424e-03,
|
|
"cpu_time": 8.9349979474388276e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1959948450128841e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.1034783199429512e+07,
|
|
"cpu_time": 8.0989923599997848e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2598529702973375e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.1072863116860390e+07,
|
|
"cpu_time": 8.1064759899999216e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2559312248813982e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.1090016588568687e+07,
|
|
"cpu_time": 8.1035881799999744e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2541665461126900e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.1011231169104576e+07,
|
|
"cpu_time": 8.1000039500003129e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2622814593540020e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.1019582524895668e+07,
|
|
"cpu_time": 8.1058970100002617e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2614200805938272e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.1045695319771767e+07,
|
|
"cpu_time": 8.1029914980000526e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2587304562478514e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.1034783199429512e+07,
|
|
"cpu_time": 8.1035881799999744e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2598529702973375e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4246146406524596e+04,
|
|
"cpu_time": 3.3859364173007707e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5273471591891833e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.7089195654975982e-04,
|
|
"cpu_time": 4.1786251634800241e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.7076021266660910e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 8.2808495395713389e+07,
|
|
"cpu_time": 1.5344536566666672e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.4832830186622534e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 8.2801898320515946e+07,
|
|
"cpu_time": 1.5342250722222340e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.4837995612351646e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 8.2858967284361526e+07,
|
|
"cpu_time": 1.5355789011111179e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.4793338560149651e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 8.2826433082421616e+07,
|
|
"cpu_time": 1.5342590700000376e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.4818789367127886e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 8.2753898368941411e+07,
|
|
"cpu_time": 1.5341328888889113e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.4875603757839937e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2809938490390792e+07,
|
|
"cpu_time": 1.5345299177777940e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.4831711496818333e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2808495395713389e+07,
|
|
"cpu_time": 1.5342590700000378e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.4832830186622534e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8350737490895430e+04,
|
|
"cpu_time": 5.9793458239469866e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.0028052948934268e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.6311757006492195e-04,
|
|
"cpu_time": 3.8965325828289388e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6316921542951893e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7992122471332550e+08,
|
|
"cpu_time": 2.7860801674999893e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9678441257324076e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7943850159645081e+08,
|
|
"cpu_time": 2.7833930799999964e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9838987421707163e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7965969815850258e+08,
|
|
"cpu_time": 2.7829260549999899e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9765313813045835e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7994965612888336e+08,
|
|
"cpu_time": 2.7867571625000662e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9669012272575045e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.7952642217278481e+08,
|
|
"cpu_time": 2.7828715775000036e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9809682107215366e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7969910055398941e+08,
|
|
"cpu_time": 2.7844056085000092e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9752287374373493e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7965969815850258e+08,
|
|
"cpu_time": 2.7833930799999964e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.9765313813045835e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2989096710513532e+05,
|
|
"cpu_time": 1.8642356138665709e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.6434176347495234e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2793106164494469e-03,
|
|
"cpu_time": 6.6952731605466622e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2791841066870397e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8934,
|
|
"real_time": 7.8131708223486072e+04,
|
|
"cpu_time": 2.8120728800089529e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3106074643485012e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8934,
|
|
"real_time": 7.8202055855512925e+04,
|
|
"cpu_time": 2.8119730378330580e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3094284911025291e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8934,
|
|
"real_time": 7.8411000949025300e+04,
|
|
"cpu_time": 2.8107535146631126e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3059392019057360e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8934,
|
|
"real_time": 7.8338919414926830e+04,
|
|
"cpu_time": 2.8104769498544786e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3071408281448483e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8934,
|
|
"real_time": 7.8163335267924005e+04,
|
|
"cpu_time": 2.8086876158496132e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3100771563674821e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.8249403942175035e+04,
|
|
"cpu_time": 2.8107927996418427e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3086386283738194e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.8202055855512939e+04,
|
|
"cpu_time": 2.8107535146631126e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3094284911025291e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2002973868928011e+02,
|
|
"cpu_time": 1.3751973615781944e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0062321522765316e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5339380575726815e-03,
|
|
"cpu_time": 4.8925604254907188e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5330681127528516e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9016,
|
|
"real_time": 7.8272066854153585e+04,
|
|
"cpu_time": 2.8113004026175541e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6165145272273090e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9016,
|
|
"real_time": 7.8221748438384107e+04,
|
|
"cpu_time": 2.8103028604702838e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6181976763319552e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9016,
|
|
"real_time": 7.8537176412197325e+04,
|
|
"cpu_time": 2.8137978737799363e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6076822386015046e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9016,
|
|
"real_time": 7.8563104539748034e+04,
|
|
"cpu_time": 2.8106313986247254e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6068216270193845e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9016,
|
|
"real_time": 7.8482502010693439e+04,
|
|
"cpu_time": 2.8097309216947528e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6094988660287037e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.8415319651035301e+04,
|
|
"cpu_time": 2.8111526914374507e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6117429870417714e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.8482502010693439e+04,
|
|
"cpu_time": 2.8106313986247254e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6094988660287037e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5747546729428430e+02,
|
|
"cpu_time": 1.5838399765153989e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.2482413974230454e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0082232399878408e-03,
|
|
"cpu_time": 5.6341300184072193e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0094785066763180e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8830,
|
|
"real_time": 7.8043685151673562e+04,
|
|
"cpu_time": 2.8094945990940329e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2483426327699050e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8830,
|
|
"real_time": 7.8225704569690977e+04,
|
|
"cpu_time": 2.8082965209513338e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2361305309188865e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8830,
|
|
"real_time": 7.8367747809973385e+04,
|
|
"cpu_time": 2.8089848505096021e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2266399309215911e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8830,
|
|
"real_time": 7.8519213858680450e+04,
|
|
"cpu_time": 2.8114022763307311e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2165575770689905e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8830,
|
|
"real_time": 7.8453962728834696e+04,
|
|
"cpu_time": 2.8100085164212674e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2208962524394840e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.8322062823770626e+04,
|
|
"cpu_time": 2.8096373526613938e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2297133848237716e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.8367747809973400e+04,
|
|
"cpu_time": 2.8094945990940335e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2266399309215911e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9048115059727476e+02,
|
|
"cpu_time": 1.1722505496337348e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2732751696903541e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4320241797751020e-03,
|
|
"cpu_time": 4.1722485947281958e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4346939803341831e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8737,
|
|
"real_time": 7.8948821625911311e+04,
|
|
"cpu_time": 2.8347757582694339e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0376342333286141e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8737,
|
|
"real_time": 7.9026173352290192e+04,
|
|
"cpu_time": 2.8365889893556002e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0366185850200468e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8737,
|
|
"real_time": 7.9015534070152644e+04,
|
|
"cpu_time": 2.8309324779672513e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0367581636196835e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8737,
|
|
"real_time": 7.9332020824697393e+04,
|
|
"cpu_time": 2.8355881584067672e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0326221259511510e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8737,
|
|
"real_time": 7.9146161748084516e+04,
|
|
"cpu_time": 2.8329476879936183e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0350470343810780e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9093742324227205e+04,
|
|
"cpu_time": 2.8341666143985343e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0357360284601146e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9026173352290192e+04,
|
|
"cpu_time": 2.8347757582694339e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0366185850200468e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5100376283159923e+02,
|
|
"cpu_time": 2.2475077355521859e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9747119895063178e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9091745869425786e-03,
|
|
"cpu_time": 7.9300480223501283e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9065784478331125e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8569,
|
|
"real_time": 8.0586958651769237e+04,
|
|
"cpu_time": 2.8581227191037498e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0330833020759866e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8569,
|
|
"real_time": 8.0839816575512552e+04,
|
|
"cpu_time": 2.8621226735908823e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0267240444188401e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8569,
|
|
"real_time": 8.0429866935637241e+04,
|
|
"cpu_time": 2.8550313805577427e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0370542217993525e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8569,
|
|
"real_time": 8.0588391188393594e+04,
|
|
"cpu_time": 2.8615640424785874e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0330471620532408e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8569,
|
|
"real_time": 8.0711892475251225e+04,
|
|
"cpu_time": 2.8613404854708869e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0299362953264716e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0631385165312779e+04,
|
|
"cpu_time": 2.8596362602403696e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0319690051347783e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0588391188393580e+04,
|
|
"cpu_time": 2.8613404854708863e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0330471620532408e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5358781995847505e+02,
|
|
"cpu_time": 3.0124202481579511e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.8699170086708578e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9048143553975279e-03,
|
|
"cpu_time": 1.0534277698327754e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9045157671655382e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8484,
|
|
"real_time": 8.3216442013707609e+04,
|
|
"cpu_time": 2.9030878948609869e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9376833720675510e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8484,
|
|
"real_time": 8.3269678131304914e+04,
|
|
"cpu_time": 2.9067873962753057e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9351659253839481e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8484,
|
|
"real_time": 8.3108827260205464e+04,
|
|
"cpu_time": 2.9007361232908262e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9427821424319530e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8484,
|
|
"real_time": 8.3395034515875828e+04,
|
|
"cpu_time": 2.9039563920320204e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9292507270036554e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8484,
|
|
"real_time": 8.3166780107375147e+04,
|
|
"cpu_time": 2.9007765051861666e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9400347058878338e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.3231352405693804e+04,
|
|
"cpu_time": 2.9030688623290614e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9369833745549881e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.3216442013707609e+04,
|
|
"cpu_time": 2.9030878948609869e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9376833720675510e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0916158575195384e+02,
|
|
"cpu_time": 2.5155585173878890e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.1605059821688029e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3115440587805006e-03,
|
|
"cpu_time": 8.6651699862528165e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3107766762545992e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7833,
|
|
"real_time": 8.8461863511395291e+04,
|
|
"cpu_time": 2.9942347491383139e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.4083901693477130e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7833,
|
|
"real_time": 8.8519569190682028e+04,
|
|
"cpu_time": 2.9995666807098064e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.4035606588671255e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7833,
|
|
"real_time": 8.8200877497790876e+04,
|
|
"cpu_time": 2.9846402080940834e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.4303115636963415e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7833,
|
|
"real_time": 8.8322149919352567e+04,
|
|
"cpu_time": 2.9881979394866771e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.4201092319244123e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7833,
|
|
"real_time": 8.8399602102668199e+04,
|
|
"cpu_time": 2.9854856070470269e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.4136080300322866e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.8380812444377807e+04,
|
|
"cpu_time": 2.9904250368951814e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.4151959307735765e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.8399602102668214e+04,
|
|
"cpu_time": 2.9881979394866771e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.4136080300322866e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2449260841844156e+02,
|
|
"cpu_time": 6.3424311852437563e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0450385835939264e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4085931660425799e-03,
|
|
"cpu_time": 2.1209129494945663e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4093202571451200e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7292,
|
|
"real_time": 9.8764271379881247e+04,
|
|
"cpu_time": 3.1439947696105659e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3271195966793716e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7292,
|
|
"real_time": 9.6089997309636135e+04,
|
|
"cpu_time": 3.1468632707077247e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3640545704006987e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7292,
|
|
"real_time": 9.5775014587886326e+04,
|
|
"cpu_time": 3.1413840729565802e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3685406425046690e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7292,
|
|
"real_time": 9.8574751912635606e+04,
|
|
"cpu_time": 3.1404121050465934e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3296711120933473e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7292,
|
|
"real_time": 9.6061094464038193e+04,
|
|
"cpu_time": 3.1431976686780038e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3644649869054804e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.7053025930815507e+04,
|
|
"cpu_time": 3.1431703773998935e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3507701817167137e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.6089997309636135e+04,
|
|
"cpu_time": 3.1431976686780044e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3640545704006987e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4822850690192085e+03,
|
|
"cpu_time": 2.5058321099128452e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0520312254318159e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5272940279841033e-02,
|
|
"cpu_time": 7.9723076035914089e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5191564436400715e-02,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 5958,
|
|
"real_time": 1.1740623083200750e+05,
|
|
"cpu_time": 3.5038178113459953e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2327946152627349e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 5958,
|
|
"real_time": 1.1722764336827341e+05,
|
|
"cpu_time": 3.4984756411546806e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2361961092783251e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 5958,
|
|
"real_time": 1.1717272489693336e+05,
|
|
"cpu_time": 3.5004788939241541e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2372442070506191e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 5958,
|
|
"real_time": 1.1702821545290596e+05,
|
|
"cpu_time": 3.4976838032897661e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2400068136174474e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 5958,
|
|
"real_time": 1.1731607596527779e+05,
|
|
"cpu_time": 3.5023831537428679e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2345104696272583e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1723017810307963e+05,
|
|
"cpu_time": 3.5005678606914927e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2361504429672771e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1722764336827339e+05,
|
|
"cpu_time": 3.5004788939241541e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2361961092783251e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4358726781684402e+02,
|
|
"cpu_time": 2.5768465657474201e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7395753324630442e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2248319514672136e-03,
|
|
"cpu_time": 7.3612244307082110e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2251301521680016e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4345,
|
|
"real_time": 1.6150350031502958e+05,
|
|
"cpu_time": 4.2380869666282862e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2462949656033530e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4345,
|
|
"real_time": 1.6135477488517540e+05,
|
|
"cpu_time": 4.2346659677788877e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2492871709132757e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4345,
|
|
"real_time": 1.6094436895818170e+05,
|
|
"cpu_time": 4.2162465431528958e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2575728085039506e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4345,
|
|
"real_time": 1.6094285608146025e+05,
|
|
"cpu_time": 4.2218170540850691e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2576034299691734e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4345,
|
|
"real_time": 1.6089089612542975e+05,
|
|
"cpu_time": 4.2234799148446269e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2586554778790445e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6112727927305532e+05,
|
|
"cpu_time": 4.2268592892979534e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2538827705737596e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6094436895818167e+05,
|
|
"cpu_time": 4.2234799148446263e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2575728085039506e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8135458736466660e+02,
|
|
"cpu_time": 9.1716908952656081e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.6774277138859285e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7461635834355978e-03,
|
|
"cpu_time": 2.1698595263124911e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7448163053780895e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2802,
|
|
"real_time": 2.5237423824526963e+05,
|
|
"cpu_time": 5.6995209386154229e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1548456264420404e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2802,
|
|
"real_time": 2.4953578669801526e+05,
|
|
"cpu_time": 5.7014796716628282e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2021066952972641e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2802,
|
|
"real_time": 2.5235630564280596e+05,
|
|
"cpu_time": 5.7007964846536098e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1551408724622536e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2802,
|
|
"real_time": 2.4951705652387810e+05,
|
|
"cpu_time": 5.6966288258385938e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2024221294052262e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2802,
|
|
"real_time": 2.5231439268891548e+05,
|
|
"cpu_time": 5.6981067630266538e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1558310995473595e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5121955595977692e+05,
|
|
"cpu_time": 5.6993065367594222e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1740692846308289e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5231439268891551e+05,
|
|
"cpu_time": 5.6995209386154229e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1558310995473595e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5457798793668092e+03,
|
|
"cpu_time": 1.9747164987276739e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5741238413322698e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.1531033022536909e-03,
|
|
"cpu_time": 3.4648364428042884e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.1669408574753339e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1628,
|
|
"real_time": 4.2988252260047756e+05,
|
|
"cpu_time": 8.5769344656025467e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8784304775029030e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1628,
|
|
"real_time": 4.2967014111614507e+05,
|
|
"cpu_time": 8.5728095761672314e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8808418349766464e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1628,
|
|
"real_time": 4.2980742363413086e+05,
|
|
"cpu_time": 8.5766990417691076e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8792828710775814e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1628,
|
|
"real_time": 4.2956839124666393e+05,
|
|
"cpu_time": 8.5759760135139641e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8819979373104925e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1628,
|
|
"real_time": 4.2978274708980444e+05,
|
|
"cpu_time": 8.5777870085994445e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8795630215509634e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2974224513744441e+05,
|
|
"cpu_time": 8.5760412211304589e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8800232284837179e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2978274708980444e+05,
|
|
"cpu_time": 8.5766990417691076e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8795630215509634e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2348135105707345e+02,
|
|
"cpu_time": 1.9187345323899416e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4023521963330039e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.8733817178616086e-04,
|
|
"cpu_time": 2.2373196244234258e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.8736588550393676e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0816061288634804e+05,
|
|
"cpu_time": 1.4772045023202258e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1899386497196770e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0797611587203329e+05,
|
|
"cpu_time": 1.4774140777262801e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1911237443858938e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0791152177730750e+05,
|
|
"cpu_time": 1.4767057227378213e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1915387848078203e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0813464173547958e+05,
|
|
"cpu_time": 1.4778916287702979e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1901054395993690e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0812646436943673e+05,
|
|
"cpu_time": 1.4773516856149007e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1901579578546820e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0806187132812093e+05,
|
|
"cpu_time": 1.4773135234339049e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1905729152734890e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0812646436943661e+05,
|
|
"cpu_time": 1.4773516856149007e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1901579578546820e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1087907161876487e+02,
|
|
"cpu_time": 4.2634510344099380e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.1227322757492436e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3721606668128683e-04,
|
|
"cpu_time": 2.8859486945600170e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3722439491776120e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5800152941910403e+06,
|
|
"cpu_time": 2.7295567011235743e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3091941773227730e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5717226140766165e+06,
|
|
"cpu_time": 2.7307849146067835e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3372064032611055e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5799893926975599e+06,
|
|
"cpu_time": 2.7301415303370734e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3092812133870697e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5709092903338121e+06,
|
|
"cpu_time": 2.7300141887641135e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3399696924686546e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5796162214213877e+06,
|
|
"cpu_time": 2.7295194629214476e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3105354871904707e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5764505625440832e+06,
|
|
"cpu_time": 2.7300033595505985e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3212373947260151e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5796162214213877e+06,
|
|
"cpu_time": 2.7300141887641135e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3105354871904707e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.6987025101864238e+03,
|
|
"cpu_time": 5.1567829980663328e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5877880800217642e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.9805581106226607e-03,
|
|
"cpu_time": 1.8889291765982367e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9838700329277786e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1861429917626083e+06,
|
|
"cpu_time": 5.3590537272730842e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2656820624107218e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1843654980713669e+06,
|
|
"cpu_time": 5.3572608727274034e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2686213345048599e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1854622563431887e+06,
|
|
"cpu_time": 5.3585612681822367e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2668073421970854e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1861051451414824e+06,
|
|
"cpu_time": 5.3604019318186138e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2657446115937862e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1855262265625326e+06,
|
|
"cpu_time": 5.3577212727273861e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2667015766823921e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1855204235762358e+06,
|
|
"cpu_time": 5.3585998145457450e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2667113854777689e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1855262265625326e+06,
|
|
"cpu_time": 5.3585612681822358e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2667015766823921e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1881160248536662e+02,
|
|
"cpu_time": 1.2267589563877473e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1886301919099987e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.2564966062229484e-04,
|
|
"cpu_time": 2.2893274341139451e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2568736065307901e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.0546700945124030e+06,
|
|
"cpu_time": 1.1733475930000167e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7563431812496653e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.1264962712302804e+06,
|
|
"cpu_time": 1.1948753700000906e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7084051857936831e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 6.8996255705133080e+06,
|
|
"cpu_time": 1.1212203989999807e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.8632250630237703e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.0510380715131760e+06,
|
|
"cpu_time": 1.1765065190000996e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7587931960774546e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 100,
|
|
"real_time": 7.1213378477841616e+06,
|
|
"cpu_time": 1.1942133170000488e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7118157735544901e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0506335711106658e+06,
|
|
"cpu_time": 1.1720326396000475e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7597164799398127e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0546700945124049e+06,
|
|
"cpu_time": 1.1765065190000996e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7563431812496653e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1616206802506320e+04,
|
|
"cpu_time": 3.0072417911528522e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.2557143773407049e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2994038887213687e-02,
|
|
"cpu_time": 2.5658345079699014e-02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3143039934638731e-02,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4358119666576385e+07,
|
|
"cpu_time": 2.3649018159999285e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6739312360113335e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4380571041256189e+07,
|
|
"cpu_time": 2.3745463220000148e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6666341557280626e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4390750322490931e+07,
|
|
"cpu_time": 2.3766411699998569e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6633332172483940e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4347553234547377e+07,
|
|
"cpu_time": 2.3665596419998564e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6773734101511478e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4371873345226049e+07,
|
|
"cpu_time": 2.3684911860000286e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6694583502081699e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4369773522019390e+07,
|
|
"cpu_time": 2.3702280271999370e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6701460738694220e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4371873345226049e+07,
|
|
"cpu_time": 2.3684911860000286e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6694583502081699e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7252357514184718e+04,
|
|
"cpu_time": 5.1141534888472001e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.6076597246837290e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2006005166155353e-03,
|
|
"cpu_time": 2.1576630729866074e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2007461085767571e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6888132238617308e+07,
|
|
"cpu_time": 4.3679104384615093e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.9917088628132238e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6818045701545019e+07,
|
|
"cpu_time": 4.3608271499998897e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0047542424863405e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6906218093175154e+07,
|
|
"cpu_time": 4.3764465230769753e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.9883535298498430e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6839264095402680e+07,
|
|
"cpu_time": 4.3674642923073970e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0007976195960703e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6832328607829716e+07,
|
|
"cpu_time": 4.3639296653846212e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0020902010284357e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6856797747313976e+07,
|
|
"cpu_time": 4.3673156138460793e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.9975408911547823e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6839264095402680e+07,
|
|
"cpu_time": 4.3674642923073970e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0007976195960703e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8184125360944730e+04,
|
|
"cpu_time": 5.8590578391117335e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.1020261239733556e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4217676180237695e-03,
|
|
"cpu_time": 1.3415695949558245e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4211041547540573e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.1054961234331131e+07,
|
|
"cpu_time": 8.0973996699992761e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2577741616126165e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.1040897145867348e+07,
|
|
"cpu_time": 8.0992308999998391e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2592229175135994e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.1056816801428795e+07,
|
|
"cpu_time": 8.1009095699994296e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2575830773783760e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.1083281636238098e+07,
|
|
"cpu_time": 8.1061215299996540e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2548592690563145e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.1099408790469170e+07,
|
|
"cpu_time": 8.1077404799998477e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2532008168765230e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.1067073121666908e+07,
|
|
"cpu_time": 8.1022804299996108e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2565280484874859e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.1056816801428795e+07,
|
|
"cpu_time": 8.1009095699994296e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2575830773783760e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3692736486831425e+04,
|
|
"cpu_time": 4.4600141195923708e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4384163894462325e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.6395328806849112e-04,
|
|
"cpu_time": 5.5046405245104380e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6388345443107887e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 8.8904720731079578e+07,
|
|
"cpu_time": 1.5475204137500498e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.0387222139073553e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 8.8871059939265251e+07,
|
|
"cpu_time": 1.5472776562499747e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.0410094395959625e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 8.8820379227399826e+07,
|
|
"cpu_time": 1.5496675762500444e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.0444564262160110e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 8.9015362784266472e+07,
|
|
"cpu_time": 1.5483108575000414e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.0312163564522629e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 8.8900898583233356e+07,
|
|
"cpu_time": 1.5481085275000340e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.0389818388320932e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.8902484253048912e+07,
|
|
"cpu_time": 1.5481770062500289e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.0388772550007372e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.8900898583233356e+07,
|
|
"cpu_time": 1.5481085275000340e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 6.0389818388320932e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1550334924568422e+04,
|
|
"cpu_time": 9.3332049083835504e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.8579240598099101e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.0481817269481548e-04,
|
|
"cpu_time": 6.0285128061618085e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.0444159645522010e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8375889211893082e+08,
|
|
"cpu_time": 2.8148274099999070e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.8432101522742214e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8359233438968658e+08,
|
|
"cpu_time": 2.8163505149998969e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.8485112004780855e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8404061719775200e+08,
|
|
"cpu_time": 2.8175963775001377e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.8342655026322927e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8408229574561119e+08,
|
|
"cpu_time": 2.8151237675001538e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.8329445515164356e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8375937268137932e+08,
|
|
"cpu_time": 2.8161753624999619e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.8431948712720232e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8384670242667201e+08,
|
|
"cpu_time": 2.8160146865000117e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.8404252556346130e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8375937268137932e+08,
|
|
"cpu_time": 2.8161753624999619e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.8431948712720232e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0805491052082609e+05,
|
|
"cpu_time": 1.1004210698286895e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.6089412005736595e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1316760527908280e-03,
|
|
"cpu_time": 3.9077248961239924e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1315856142834141e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9551,
|
|
"real_time": 7.5213892547279727e+04,
|
|
"cpu_time": 2.7433647513349610e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3614506114762109e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9551,
|
|
"real_time": 7.5291537655092165e+04,
|
|
"cpu_time": 2.7424473144173791e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3600466027017634e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9551,
|
|
"real_time": 7.5125689708821854e+04,
|
|
"cpu_time": 2.7419776285205525e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3630490501570115e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9551,
|
|
"real_time": 7.5198531170149319e+04,
|
|
"cpu_time": 2.7410739158202906e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3617287253696855e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9551,
|
|
"real_time": 7.5454320553748214e+04,
|
|
"cpu_time": 2.7421237336404616e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3571124787620032e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5256794327018259e+04,
|
|
"cpu_time": 2.7421974687467294e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3606774936933352e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5213892547279727e+04,
|
|
"cpu_time": 2.7421237336404610e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3614506114762109e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2514720148320457e+02,
|
|
"cpu_time": 8.2781478397116643e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2602231142459976e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6629355874420357e-03,
|
|
"cpu_time": 3.0188007734888031e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6611012710374107e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9311,
|
|
"real_time": 7.4974807966653301e+04,
|
|
"cpu_time": 2.7413886628719221e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7315841887996472e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9311,
|
|
"real_time": 7.5044883284884345e+04,
|
|
"cpu_time": 2.7426807206530450e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7290334934967000e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9311,
|
|
"real_time": 7.5261954194968639e+04,
|
|
"cpu_time": 2.7447813113521674e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7211624012506858e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9311,
|
|
"real_time": 7.5038352846846392e+04,
|
|
"cpu_time": 2.7412467737084994e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7292709958332602e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9311,
|
|
"real_time": 7.5123794557768357e+04,
|
|
"cpu_time": 2.7435537385887362e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7261668717028636e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5088758570224207e+04,
|
|
"cpu_time": 2.7427302414348745e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7274435902166314e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5044883284884345e+04,
|
|
"cpu_time": 2.7426807206530450e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7290334934967000e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1031178826541118e+02,
|
|
"cpu_time": 1.4906372139386821e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0029709384874383e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4690852581115165e-03,
|
|
"cpu_time": 5.4348662927887763e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4676640619978858e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9282,
|
|
"real_time": 7.5258890835269718e+04,
|
|
"cpu_time": 2.7467744947208988e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4425463284670547e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9282,
|
|
"real_time": 7.5483254655089229e+04,
|
|
"cpu_time": 2.7442775856495975e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4263690916828260e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9282,
|
|
"real_time": 7.4993768838753545e+04,
|
|
"cpu_time": 2.7411857218272012e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4617871103490457e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9282,
|
|
"real_time": 7.5261060416557637e+04,
|
|
"cpu_time": 2.7449507185951486e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4423894339640066e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9282,
|
|
"real_time": 7.5273716461958524e+04,
|
|
"cpu_time": 2.7467811387631722e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4414743851129197e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5254138241525739e+04,
|
|
"cpu_time": 2.7447939319112041e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4429132699151710e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5261060416557637e+04,
|
|
"cpu_time": 2.7449507185951492e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.4423894339640066e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7373857317145377e+02,
|
|
"cpu_time": 2.3011951051263196e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2573892850863650e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3086912857050519e-03,
|
|
"cpu_time": 8.3838538054621604e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3101402185413873e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9183,
|
|
"real_time": 7.6388515552889861e+04,
|
|
"cpu_time": 2.7720102471958788e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0724125139371276e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9183,
|
|
"real_time": 7.6105587909530674e+04,
|
|
"cpu_time": 2.7699398867472162e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0763992796084975e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9183,
|
|
"real_time": 7.6046405002489686e+04,
|
|
"cpu_time": 2.7694247489927005e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0772369844086385e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9183,
|
|
"real_time": 7.5986727591799528e+04,
|
|
"cpu_time": 2.7689852488293691e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0780830099708201e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9183,
|
|
"real_time": 7.6200860848326411e+04,
|
|
"cpu_time": 2.7725170314712101e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0750534716800277e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.6145619381007244e+04,
|
|
"cpu_time": 2.7705754326472746e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0758370519210222e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.6105587909530688e+04,
|
|
"cpu_time": 2.7699398867472156e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0763992796084975e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5711682144071969e+02,
|
|
"cpu_time": 1.5878582133117351e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2170756021728262e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0633730832834863e-03,
|
|
"cpu_time": 5.7311495460513127e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0607912677984091e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9015,
|
|
"real_time": 7.7535490609447952e+04,
|
|
"cpu_time": 2.7946059179145494e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1130968374892253e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9015,
|
|
"real_time": 7.7878356874051096e+04,
|
|
"cpu_time": 2.7974091713810584e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1037937441973838e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9015,
|
|
"real_time": 7.7660210753685911e+04,
|
|
"cpu_time": 2.7945113288962707e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1097032625838426e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9015,
|
|
"real_time": 7.7597967088010715e+04,
|
|
"cpu_time": 2.7942844292844733e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1113955190884647e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9015,
|
|
"real_time": 7.7641573822070510e+04,
|
|
"cpu_time": 2.7940363693844568e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1102096716311869e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7662719829453257e+04,
|
|
"cpu_time": 2.7949694433721626e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1096398069980207e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7641573822070524e+04,
|
|
"cpu_time": 2.7945113288962713e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1102096716311869e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2974299213617721e+02,
|
|
"cpu_time": 1.3814995257213826e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5191905086255277e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6705955240956256e-03,
|
|
"cpu_time": 4.9428072603705745e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6681475657369549e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8736,
|
|
"real_time": 7.9659545995277629e+04,
|
|
"cpu_time": 2.8410601064561366e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1135057437990612e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8736,
|
|
"real_time": 7.9581086440104074e+04,
|
|
"cpu_time": 2.8401165762362565e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1175612781640667e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8736,
|
|
"real_time": 7.9871436351296550e+04,
|
|
"cpu_time": 2.8376949107142410e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1025930541523409e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8736,
|
|
"real_time": 7.9906365094707377e+04,
|
|
"cpu_time": 2.8375233928571327e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1007997249233401e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8736,
|
|
"real_time": 7.9801986875291725e+04,
|
|
"cpu_time": 2.8367612557233631e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1061634281371033e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9764084151335468e+04,
|
|
"cpu_time": 2.8386312483974261e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1081246458351827e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9801986875291725e+04,
|
|
"cpu_time": 2.8376949107142410e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1061634281371033e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3930305681562373e+02,
|
|
"cpu_time": 1.8511141920381255e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.1781890610376094e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7464383663119063e-03,
|
|
"cpu_time": 6.5211506182185090e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7473153031797267e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8222,
|
|
"real_time": 8.5215729444221288e+04,
|
|
"cpu_time": 2.9270914522014052e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.6905989571910167e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8222,
|
|
"real_time": 8.5287774909282089e+04,
|
|
"cpu_time": 2.9306485125274782e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.6841024484116948e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8222,
|
|
"real_time": 8.5027086251323912e+04,
|
|
"cpu_time": 2.9355862819265516e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7076615099202681e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8222,
|
|
"real_time": 8.4942084454380922e+04,
|
|
"cpu_time": 2.9272745013378514e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7153745897531891e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8222,
|
|
"real_time": 8.5305174657506010e+04,
|
|
"cpu_time": 2.9302273729020776e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.6825351173738539e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.5155569943342853e+04,
|
|
"cpu_time": 2.9301656241790729e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.6960545245300055e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.5215729444221288e+04,
|
|
"cpu_time": 2.9302273729020776e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.6905989571910167e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6245230236349656e+02,
|
|
"cpu_time": 3.4434593474849919e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4691636553086929e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9077119966618987e-03,
|
|
"cpu_time": 1.1751756689350028e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9089829088735752e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7562,
|
|
"real_time": 9.5624914734477585e+04,
|
|
"cpu_time": 3.0966503834964224e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3706888038953929e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7562,
|
|
"real_time": 9.3088596688505117e+04,
|
|
"cpu_time": 3.1005374173498777e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4080349759551721e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7562,
|
|
"real_time": 9.2516574155783062e+04,
|
|
"cpu_time": 3.0855738971171045e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4167407428997078e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7562,
|
|
"real_time": 9.5322923234877147e+04,
|
|
"cpu_time": 3.0883745305474050e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3750312679462900e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7562,
|
|
"real_time": 9.2550661895009456e+04,
|
|
"cpu_time": 3.0874162774398772e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4162189369179184e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.3820734141730456e+04,
|
|
"cpu_time": 3.0917105011901370e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3973429455228963e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.3088596688505102e+04,
|
|
"cpu_time": 3.0883745305474044e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4080349759551721e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5298350984723377e+03,
|
|
"cpu_time": 6.5104503184835426e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2666888467330489e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6305938260526606e-02,
|
|
"cpu_time": 2.1057761766431170e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6221421190807508e-02,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 6149,
|
|
"real_time": 1.1386292580699017e+05,
|
|
"cpu_time": 3.4474549048626976e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3022770418209882e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 6149,
|
|
"real_time": 1.1386690716863541e+05,
|
|
"cpu_time": 3.4449239404781582e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3021965425983524e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 6149,
|
|
"real_time": 1.1386485947821775e+05,
|
|
"cpu_time": 3.4380769686126994e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3022379441845961e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 6149,
|
|
"real_time": 1.1388572556625560e+05,
|
|
"cpu_time": 3.4429628899007972e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3018161292522283e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 6149,
|
|
"real_time": 1.1393799398330768e+05,
|
|
"cpu_time": 3.4406765896893624e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3007601839857302e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1388368240068131e+05,
|
|
"cpu_time": 3.4428190587087430e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3018575683683791e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1386690716863540e+05,
|
|
"cpu_time": 3.4429628899007972e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3021965425983524e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1703586331152479e+01,
|
|
"cpu_time": 3.6407374327030686e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4060848873551469e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.7838567969385233e-04,
|
|
"cpu_time": 1.0574873005578622e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7830066357650265e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4411,
|
|
"real_time": 1.5842414877239263e+05,
|
|
"cpu_time": 4.1746662366808759e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3093944582479191e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4411,
|
|
"real_time": 1.5862854996508625e+05,
|
|
"cpu_time": 4.1774214713217912e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3051301302028828e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4411,
|
|
"real_time": 1.5842719783575126e+05,
|
|
"cpu_time": 4.1690007776014722e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3093307661955457e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4411,
|
|
"real_time": 1.5839603501320840e+05,
|
|
"cpu_time": 4.1707573135345348e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3099818436508241e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4411,
|
|
"real_time": 1.5811709634408844e+05,
|
|
"cpu_time": 4.1679016164135042e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3158210726249638e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5839860558610540e+05,
|
|
"cpu_time": 4.1719494831104355e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3099316541844273e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5842414877239263e+05,
|
|
"cpu_time": 4.1707573135345348e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3093944582479191e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8275560847031778e+02,
|
|
"cpu_time": 3.9945200653251260e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.8207344496889599e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1537703112605489e-03,
|
|
"cpu_time": 9.5747086140337786e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1543242727863501e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2799,
|
|
"real_time": 2.4691723686407204e+05,
|
|
"cpu_time": 5.6540773383351916e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2466699097934632e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2799,
|
|
"real_time": 2.4994141837713757e+05,
|
|
"cpu_time": 5.6575304394426872e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1952870668990111e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2799,
|
|
"real_time": 2.4701881003820628e+05,
|
|
"cpu_time": 5.6467824330117949e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2449236956400900e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2799,
|
|
"real_time": 2.4958002872662080e+05,
|
|
"cpu_time": 5.6459011861378921e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2013618050688052e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2799,
|
|
"real_time": 2.4706511257919067e+05,
|
|
"cpu_time": 5.6489361271884677e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2441281533179016e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4810452131704552e+05,
|
|
"cpu_time": 5.6506455048232072e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2264741261438546e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4706511257919070e+05,
|
|
"cpu_time": 5.6489361271884688e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2441281533179016e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5182306151141570e+03,
|
|
"cpu_time": 4.9883279133126450e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5803014849958908e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.1193186123926161e-03,
|
|
"cpu_time": 8.8278903871332414e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.1050923488087296e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1636,
|
|
"real_time": 4.2757288426016219e+05,
|
|
"cpu_time": 8.5247286246948154e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9047824995468168e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1636,
|
|
"real_time": 4.2732818558678933e+05,
|
|
"cpu_time": 8.5242145721269352e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9075910991461468e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1636,
|
|
"real_time": 4.2718157957721851e+05,
|
|
"cpu_time": 8.5125981112469709e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9092753532948465e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1636,
|
|
"real_time": 4.2722960128884390e+05,
|
|
"cpu_time": 8.5168324205382355e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9087235380540609e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1636,
|
|
"real_time": 4.2709278338133357e+05,
|
|
"cpu_time": 8.5122874022001727e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9102960330929756e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2728100681886950e+05,
|
|
"cpu_time": 8.5181322261614283e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9081337046269693e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2722960128884402e+05,
|
|
"cpu_time": 8.5168324205382378e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9087235380540609e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8396646727102240e+02,
|
|
"cpu_time": 6.0618746215126828e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1125769672511341e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.3055147393670221e-04,
|
|
"cpu_time": 7.1164363977528665e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.3042367922038819e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 867,
|
|
"real_time": 8.0566712787520944e+05,
|
|
"cpu_time": 1.4705510738177469e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2060011571548939e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 867,
|
|
"real_time": 8.0599819123100454e+05,
|
|
"cpu_time": 1.4707083713955898e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2038627947713156e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 867,
|
|
"real_time": 8.0531002226422064e+05,
|
|
"cpu_time": 1.4705216332178980e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2083096994213953e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 867,
|
|
"real_time": 8.0563906151070888e+05,
|
|
"cpu_time": 1.4711324279122793e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2061825206625080e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 867,
|
|
"real_time": 8.0578891629451991e+05,
|
|
"cpu_time": 1.4711716366782438e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2052143125619278e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0568066383513273e+05,
|
|
"cpu_time": 1.4708170286043517e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2059140969144077e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0566712787520944e+05,
|
|
"cpu_time": 1.4707083713955900e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2060011571548939e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5091402468197498e+02,
|
|
"cpu_time": 3.1425323432473328e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6214253624758681e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.1143111153691505e-04,
|
|
"cpu_time": 2.1365895839737866e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1145833993628547e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5749198374118698e+06,
|
|
"cpu_time": 2.7204448898874545e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3263714131541719e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5682641372314834e+06,
|
|
"cpu_time": 2.7192282516853595e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3489764898971233e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5752578253617112e+06,
|
|
"cpu_time": 2.7211733123596297e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3252285847707529e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5692248822744477e+06,
|
|
"cpu_time": 2.7203975235955040e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3457016229831142e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5748788576489419e+06,
|
|
"cpu_time": 2.7206113573034662e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3265100101241655e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5725091079856909e+06,
|
|
"cpu_time": 2.7203710669662831e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3345576241858654e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5748788576489419e+06,
|
|
"cpu_time": 2.7204448898874545e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3265100101241655e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4564682940122584e+03,
|
|
"cpu_time": 7.0949165828184323e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1735645669797637e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.1980593158152381e-03,
|
|
"cpu_time": 2.6080694170632304e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1999285595098758e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1770056843842296e+06,
|
|
"cpu_time": 5.3455667090908969e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2808265601991758e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1800143960440024e+06,
|
|
"cpu_time": 5.3495230318183284e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2758302040617085e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1781176941215317e+06,
|
|
"cpu_time": 5.3471278545456156e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2789788216567030e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1743656346489759e+06,
|
|
"cpu_time": 5.3414116818177346e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2852185069270506e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 220,
|
|
"real_time": 3.1790100270882249e+06,
|
|
"cpu_time": 5.3495945818181923e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2774970374556780e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1777026872573933e+06,
|
|
"cpu_time": 5.3466447718181545e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2796702260600634e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1781176941215312e+06,
|
|
"cpu_time": 5.3471278545456165e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2789788216567030e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1706622703028706e+03,
|
|
"cpu_time": 3.3825500273430312e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.6078763466172176e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.8309168098300689e-04,
|
|
"cpu_time": 6.3264910456970142e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.8335259441186416e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 102,
|
|
"real_time": 6.9069803473265730e+06,
|
|
"cpu_time": 1.1297495019607818e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.8580465431594334e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 102,
|
|
"real_time": 7.1055654792443793e+06,
|
|
"cpu_time": 1.1931281529412303e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7222746870764532e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 102,
|
|
"real_time": 7.1135969129482321e+06,
|
|
"cpu_time": 1.1925103450981205e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7169431176124029e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 102,
|
|
"real_time": 7.0308465191035299e+06,
|
|
"cpu_time": 1.1667667127450805e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7724597470346098e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 102,
|
|
"real_time": 6.9537759785923883e+06,
|
|
"cpu_time": 1.1412094862745402e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.8253541821449680e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0221530474430202e+06,
|
|
"cpu_time": 1.1646728398039509e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7790156554055738e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0308465191035299e+06,
|
|
"cpu_time": 1.1667667127450805e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.7724597470346098e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1289875073610368e+04,
|
|
"cpu_time": 2.8979235391878884e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.2266880540804997e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3000268501247176e-02,
|
|
"cpu_time": 2.4881867595329996e-02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3029227152745262e-02,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4371413476765156e+07,
|
|
"cpu_time": 2.3711329000000205e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6696077674264688e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4310549031943083e+07,
|
|
"cpu_time": 2.3568100979998689e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6894681573854313e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4315169826149940e+07,
|
|
"cpu_time": 2.3561119880000662e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6879544437824459e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4338939692825079e+07,
|
|
"cpu_time": 2.3625259600000843e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6801831542383814e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4379909615963697e+07,
|
|
"cpu_time": 2.3742721679998342e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6668488044945593e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4343196328729395e+07,
|
|
"cpu_time": 2.3641706227999747e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6788124654654579e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4338939692825079e+07,
|
|
"cpu_time": 2.3625259600000847e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.6801831542383814e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1676184611617275e+04,
|
|
"cpu_time": 8.2513386055530544e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0330222351759907e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.2084467008354294e-03,
|
|
"cpu_time": 3.4901620576693777e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2078727087285030e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6921875631579988e+07,
|
|
"cpu_time": 4.3262403269230053e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.9854523450275316e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6888331183447286e+07,
|
|
"cpu_time": 4.3265659730768494e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 4.9916719295180988e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6643029987238921e+07,
|
|
"cpu_time": 4.3310815615383871e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0376300317300844e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6807702433031339e+07,
|
|
"cpu_time": 4.3295899576920554e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0066852366513319e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6738797242824849e+07,
|
|
"cpu_time": 4.3339647730769791e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0195873352536936e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6799947295624476e+07,
|
|
"cpu_time": 4.3294885184614561e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0082053756361485e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6807702433031339e+07,
|
|
"cpu_time": 4.3295899576920561e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0066852366513319e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1302736783624697e+05,
|
|
"cpu_time": 3.2278844208624614e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1146663187799629e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.2174473923200775e-03,
|
|
"cpu_time": 7.4555791223336815e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.2224033564345499e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 16,
|
|
"real_time": 4.3017422547563910e+07,
|
|
"cpu_time": 8.1474019562499225e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 6.2401566645047998e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 16,
|
|
"real_time": 4.3008811771869659e+07,
|
|
"cpu_time": 8.1446574999993965e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 6.2414060035848951e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 16,
|
|
"real_time": 4.3006013380363584e+07,
|
|
"cpu_time": 8.1527155437498778e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 6.2418121304535246e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 16,
|
|
"real_time": 4.3025172315537930e+07,
|
|
"cpu_time": 8.1481115937499732e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 6.2390326767630014e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 16,
|
|
"real_time": 4.2943371925503016e+07,
|
|
"cpu_time": 8.1486710749999240e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 6.2509170557373667e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3000158388167620e+07,
|
|
"cpu_time": 8.1483115337498188e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 6.2426649062087173e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3008811771869659e+07,
|
|
"cpu_time": 8.1481115937499732e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 6.2414060035848951e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2625189993348329e+04,
|
|
"cpu_time": 2.9044725819412801e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.7406329058639416e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.5872255396914581e-04,
|
|
"cpu_time": 3.5645085118691508e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.5939249937139635e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.2747760936617851e+07,
|
|
"cpu_time": 1.5140305525000030e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.7885053674437265e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.2776619829237461e+07,
|
|
"cpu_time": 1.5129814837500533e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.7867048076137333e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.2745499685406685e+07,
|
|
"cpu_time": 1.5139755999999237e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.7886464984400263e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.2816980555653572e+07,
|
|
"cpu_time": 1.5130108025000766e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.7841885050127144e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.2949240468442440e+07,
|
|
"cpu_time": 1.5167432274999726e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.7759580314405594e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.2807220295071602e+07,
|
|
"cpu_time": 1.5141483332500061e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.7848006419901524e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.2776619829237461e+07,
|
|
"cpu_time": 1.5139755999999237e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.7867048076137333e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.4478147776588768e+04,
|
|
"cpu_time": 1.5356349833642165e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.2608916128276205e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.1025404605373005e-04,
|
|
"cpu_time": 1.0141905846622642e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.0943352042944549e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8595119938254356e+08,
|
|
"cpu_time": 2.8156681125000203e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.7743205075600014e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8656904250383377e+08,
|
|
"cpu_time": 2.8260201250000703e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.7551982343369532e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8656812608242035e+08,
|
|
"cpu_time": 2.8246645350000674e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.7552265038329868e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8610732257366180e+08,
|
|
"cpu_time": 2.8185950249999791e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.7694764996418123e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8621900677680969e+08,
|
|
"cpu_time": 2.8192582174997938e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.7660162761308193e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8628293946385387e+08,
|
|
"cpu_time": 2.8208412029999864e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.7640476043005142e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8621900677680969e+08,
|
|
"cpu_time": 2.8192582174997938e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.7660162761308193e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7756373997894325e+05,
|
|
"cpu_time": 4.3517561020075966e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.5878442729057446e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4900115962192095e-03,
|
|
"cpu_time": 1.5427157322359978e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4898982212600770e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9015,
|
|
"real_time": 7.7362615567575471e+04,
|
|
"cpu_time": 2.7961024004436290e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3236367365391703e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9015,
|
|
"real_time": 7.7254123510969759e+04,
|
|
"cpu_time": 2.7941708973932732e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3254955897009125e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9015,
|
|
"real_time": 7.7393746579096143e+04,
|
|
"cpu_time": 2.7956084093178360e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3231043143175341e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9015,
|
|
"real_time": 7.7339282778920257e+04,
|
|
"cpu_time": 2.7965451735995122e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3240360696480410e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9015,
|
|
"real_time": 7.7319897091944906e+04,
|
|
"cpu_time": 2.8015313843593263e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3243680326970831e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7333933105701319e+04,
|
|
"cpu_time": 2.7967916530227161e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3241281485805482e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7339282778920271e+04,
|
|
"cpu_time": 2.7961024004436296e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3240360696480410e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.2429799053870553e+01,
|
|
"cpu_time": 2.7959479658136235e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.9800341469979940e+03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.7796628140209326e-04,
|
|
"cpu_time": 9.9969833748317274e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.7818467242951514e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9138,
|
|
"real_time": 7.7511673859869872e+04,
|
|
"cpu_time": 2.8023705044867692e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6421826519996125e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9138,
|
|
"real_time": 7.7485256136519529e+04,
|
|
"cpu_time": 2.8003935095206776e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6430834743472680e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9138,
|
|
"real_time": 7.7333483637860816e+04,
|
|
"cpu_time": 2.7984116447801370e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6482707149084684e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9138,
|
|
"real_time": 7.7351479273860488e+04,
|
|
"cpu_time": 2.7998765222149005e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6476546010828316e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9138,
|
|
"real_time": 7.7348519179733426e+04,
|
|
"cpu_time": 2.7991797647187259e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6477559256707910e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7406082417568832e+04,
|
|
"cpu_time": 2.8000463891442423e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6457894736017942e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7351479273860488e+04,
|
|
"cpu_time": 2.7998765222149005e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.6476546010828316e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.5122915436167716e+01,
|
|
"cpu_time": 1.4978556752302083e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9083408811984198e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0996928507112690e-03,
|
|
"cpu_time": 5.3493959280009895e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0992336730553268e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9116,
|
|
"real_time": 7.7158126017507224e+04,
|
|
"cpu_time": 2.8007515072400711e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.3085789033686675e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9116,
|
|
"real_time": 7.7188033556339913e+04,
|
|
"cpu_time": 2.8014589940763765e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.3065220232749030e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9116,
|
|
"real_time": 7.7394447316991398e+04,
|
|
"cpu_time": 2.8004291761737346e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2923693391383030e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9116,
|
|
"real_time": 7.7316690641734007e+04,
|
|
"cpu_time": 2.7971990982887626e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2976918256626226e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9116,
|
|
"real_time": 7.7287098355554524e+04,
|
|
"cpu_time": 2.7977130210617639e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2997202471706279e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7268879177625437e+04,
|
|
"cpu_time": 2.7995103593681415e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.3009764677230246e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7287098355554524e+04,
|
|
"cpu_time": 2.8004291761737352e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.2997202471706279e+07,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.6420315704793637e+01,
|
|
"cpu_time": 1.9205599678757829e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.6143752572521917e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2478544626374475e-03,
|
|
"cpu_time": 6.8603424218414370e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2477654442584845e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9074,
|
|
"real_time": 7.7725708245404239e+04,
|
|
"cpu_time": 2.8202797575490142e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0539627344578588e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9074,
|
|
"real_time": 7.7859137575426401e+04,
|
|
"cpu_time": 2.8221438968481199e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0521565297411576e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9074,
|
|
"real_time": 7.7715238471154604e+04,
|
|
"cpu_time": 2.8176804981265136e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0541047240099001e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9074,
|
|
"real_time": 7.7741264444046901e+04,
|
|
"cpu_time": 2.8176857780470827e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0537518341878872e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9074,
|
|
"real_time": 7.7988178347545982e+04,
|
|
"cpu_time": 2.8210437194181210e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0504156108754364e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7805905416715628e+04,
|
|
"cpu_time": 2.8197667299977702e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0528782866544481e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7741264444046916e+04,
|
|
"cpu_time": 2.8202797575490142e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0537518341878872e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1714053598199548e+02,
|
|
"cpu_time": 2.0141714130981990e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5834271337829222e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5055481374403399e-03,
|
|
"cpu_time": 7.1430426909810081e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5039033037848172e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8716,
|
|
"real_time": 7.9684315114476543e+04,
|
|
"cpu_time": 2.8523924242771446e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0561135496317342e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8716,
|
|
"real_time": 7.9663877610685435e+04,
|
|
"cpu_time": 2.8524882801745000e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0566410387488329e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8716,
|
|
"real_time": 7.9502316095478716e+04,
|
|
"cpu_time": 2.8501688916933810e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0608204646923181e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8716,
|
|
"real_time": 7.9386849745181855e+04,
|
|
"cpu_time": 2.8514000114731072e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0638178807434511e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8716,
|
|
"real_time": 7.9443075165940638e+04,
|
|
"cpu_time": 2.8499279256539245e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0623572244373864e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9536086746352637e+04,
|
|
"cpu_time": 2.8512755066544120e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0599500316507447e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9502316095478716e+04,
|
|
"cpu_time": 2.8514000114731066e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0608204646923181e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3263243154899035e+02,
|
|
"cpu_time": 1.2014987774571925e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.4344037220455025e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6675755242015673e-03,
|
|
"cpu_time": 4.2138992694781352e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6672267138894321e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8441,
|
|
"real_time": 8.2085334821090029e+04,
|
|
"cpu_time": 2.8957831548394921e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9919432711604136e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8441,
|
|
"real_time": 8.2143984897463917e+04,
|
|
"cpu_time": 2.8977664328871056e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9890930590843141e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8441,
|
|
"real_time": 8.2079557057364931e+04,
|
|
"cpu_time": 2.8961201753347123e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9922242729817158e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8441,
|
|
"real_time": 8.2225881197256778e+04,
|
|
"cpu_time": 2.8956002701102104e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9851199552840072e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8441,
|
|
"real_time": 8.2201885785523540e+04,
|
|
"cpu_time": 2.8932827840303775e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9862832448268145e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2147328751739842e+04,
|
|
"cpu_time": 2.8957105634403799e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9889327606674528e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2143984897463903e+04,
|
|
"cpu_time": 2.8957831548394926e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.9890930590843141e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.6321919427625119e+01,
|
|
"cpu_time": 1.6051423678556924e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.2202616906083893e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.0735333011325023e-04,
|
|
"cpu_time": 5.5431726779648532e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.0729906564520665e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7836,
|
|
"real_time": 8.8688671075524864e+04,
|
|
"cpu_time": 2.9939864229197567e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.3894443569000292e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7836,
|
|
"real_time": 8.8674069093054204e+04,
|
|
"cpu_time": 2.9921063910158206e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.3906611786616886e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7836,
|
|
"real_time": 8.8733631265959091e+04,
|
|
"cpu_time": 2.9993084137314494e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.3857002204238188e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7836,
|
|
"real_time": 8.8700981270626464e+04,
|
|
"cpu_time": 2.9907412633997132e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.3884188270758617e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7836,
|
|
"real_time": 8.8706349813110268e+04,
|
|
"cpu_time": 2.9918909660540766e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.3879716771204770e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.8700740503654975e+04,
|
|
"cpu_time": 2.9936066914241633e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.3884392520363760e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.8700981270626449e+04,
|
|
"cpu_time": 2.9921063910158206e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.3884188270758617e+08,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2194367543825887e+01,
|
|
"cpu_time": 3.3933437009812303e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8485488427412460e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.5021625995232084e-04,
|
|
"cpu_time": 1.1335302365211169e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5019476775582275e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7146,
|
|
"real_time": 9.8003331283784777e+04,
|
|
"cpu_time": 3.1602160635320115e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3374239251159682e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7146,
|
|
"real_time": 1.0001559941070367e+05,
|
|
"cpu_time": 3.1620324601175857e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3105155672943223e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7146,
|
|
"real_time": 9.7968540130702269e+04,
|
|
"cpu_time": 3.1658422879932815e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3378988788149092e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7146,
|
|
"real_time": 9.7755563400532061e+04,
|
|
"cpu_time": 3.1566256717044982e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3408137137214494e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7146,
|
|
"real_time": 9.9944553700327000e+04,
|
|
"cpu_time": 3.1586792625245819e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3114471489162412e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.8737517585209949e+04,
|
|
"cpu_time": 3.1606791491743922e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3276198467725782e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.8003331283784792e+04,
|
|
"cpu_time": 3.1602160635320109e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3374239251159682e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1385318624662823e+03,
|
|
"cpu_time": 3.5047869935558913e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5247697444007734e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1530894135390178e-02,
|
|
"cpu_time": 1.1088714887341174e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1484987574624341e-02,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 6059,
|
|
"real_time": 1.1537137717456037e+05,
|
|
"cpu_time": 3.4822754563459498e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2721753559669156e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 6059,
|
|
"real_time": 1.1544505176931965e+05,
|
|
"cpu_time": 3.4837973593002831e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2707253016249819e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 6059,
|
|
"real_time": 1.1537269357770208e+05,
|
|
"cpu_time": 3.4785612196731160e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2721494304321609e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 6059,
|
|
"real_time": 1.1516251035783527e+05,
|
|
"cpu_time": 3.4725935005774646e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2762963327688923e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 6059,
|
|
"real_time": 1.1517970947875416e+05,
|
|
"cpu_time": 3.4766839527973585e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2759564265818415e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1530626847163432e+05,
|
|
"cpu_time": 3.4787822977388342e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2734605694749589e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1537137717456036e+05,
|
|
"cpu_time": 3.4785612196731160e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2721753559669156e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2707853435878664e+02,
|
|
"cpu_time": 4.4751015857940837e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5061626898188395e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1020956279584083e-03,
|
|
"cpu_time": 1.2863988610908033e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1023559077594303e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4382,
|
|
"real_time": 1.5944017329201609e+05,
|
|
"cpu_time": 4.2048281857602898e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2883055077954655e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4382,
|
|
"real_time": 1.5933275485948144e+05,
|
|
"cpu_time": 4.2032319169326581e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2905224067855940e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4382,
|
|
"real_time": 1.5937876244306623e+05,
|
|
"cpu_time": 4.2041344796897628e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2895725375411158e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4382,
|
|
"real_time": 1.5948501482950454e+05,
|
|
"cpu_time": 4.2097253879507130e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2873809527527313e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4382,
|
|
"real_time": 1.5937113303198389e+05,
|
|
"cpu_time": 4.2094609356457496e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2897300158792348e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5940156769121045e+05,
|
|
"cpu_time": 4.2062761811958347e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2891022841508284e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5937876244306620e+05,
|
|
"cpu_time": 4.2048281857602904e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2895725375411158e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.0488368574764749e+01,
|
|
"cpu_time": 3.0818342416845314e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2479751618041121e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.7947160401798319e-04,
|
|
"cpu_time": 7.3267519985061285e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.7942728866101864e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2795,
|
|
"real_time": 2.4803616338716034e+05,
|
|
"cpu_time": 5.6801190983903664e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2275125759112582e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2795,
|
|
"real_time": 2.5000300675743859e+05,
|
|
"cpu_time": 5.6804632450801879e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1942535555876894e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2795,
|
|
"real_time": 2.4790781402663860e+05,
|
|
"cpu_time": 5.6740105509836506e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2297012868151326e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2795,
|
|
"real_time": 2.4997883955098368e+05,
|
|
"cpu_time": 5.6803260250450938e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1946590434753213e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2795,
|
|
"real_time": 2.4790672274233110e+05,
|
|
"cpu_time": 5.6795278926655231e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2297199059416685e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4876650929291049e+05,
|
|
"cpu_time": 5.6788893624329648e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2151692735462136e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4803616338716037e+05,
|
|
"cpu_time": 5.6801190983903664e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2275125759112582e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1190024853976104e+03,
|
|
"cpu_time": 2.7506307721743053e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8930096548503920e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.4982039124890376e-03,
|
|
"cpu_time": 4.8436069037905555e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.4909457533073320e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1634,
|
|
"real_time": 4.2859671814888692e+05,
|
|
"cpu_time": 8.5595565055083891e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8930659316702614e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1634,
|
|
"real_time": 4.2855372299246991e+05,
|
|
"cpu_time": 8.5619150795598514e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8935568342661419e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1634,
|
|
"real_time": 4.2837991409765778e+05,
|
|
"cpu_time": 8.5590475459001353e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8955423234944487e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1634,
|
|
"real_time": 4.2842956163451378e+05,
|
|
"cpu_time": 8.5586376866574702e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8949750152606087e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1634,
|
|
"real_time": 4.2864860928115284e+05,
|
|
"cpu_time": 8.5628269155453367e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8924735893041649e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2852170523093629e+05,
|
|
"cpu_time": 8.5603967466342391e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8939227387991247e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2855372299246991e+05,
|
|
"cpu_time": 8.5595565055083891e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8935568342661419e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1330476641351727e+02,
|
|
"cpu_time": 1.8595493738310603e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2940613569379158e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.6440846526655108e-04,
|
|
"cpu_time": 2.1722700814798040e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6442210594756013e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 864,
|
|
"real_time": 8.0888867631545977e+05,
|
|
"cpu_time": 1.4787291909720404e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1852672967376995e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 864,
|
|
"real_time": 8.0864498300393776e+05,
|
|
"cpu_time": 1.4779259212963430e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1868299292714157e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 864,
|
|
"real_time": 8.0833682356266456e+05,
|
|
"cpu_time": 1.4780234525462505e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1888072864403477e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 864,
|
|
"real_time": 8.0829071339394432e+05,
|
|
"cpu_time": 1.4776796296297756e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1891032898157053e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 864,
|
|
"real_time": 8.0837867062655289e+05,
|
|
"cpu_time": 1.4776208807868822e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1885386792171383e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0850797338051186e+05,
|
|
"cpu_time": 1.4779958150462583e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1877092962964621e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0837867062655278e+05,
|
|
"cpu_time": 1.4779259212963430e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1885386792171383e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5343624672296943e+02,
|
|
"cpu_time": 4.4272963897675919e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6258324400749297e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.1346165414214601e-04,
|
|
"cpu_time": 2.9954728861184400e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1340083786800114e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 443,
|
|
"real_time": 1.5709607795713381e+06,
|
|
"cpu_time": 2.7251961331827510e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3397946715696917e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 443,
|
|
"real_time": 1.5747354591357822e+06,
|
|
"cpu_time": 2.7259582460496230e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3269950526189861e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 443,
|
|
"real_time": 1.5709362256616084e+06,
|
|
"cpu_time": 2.7251506636572401e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3398781331604290e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 443,
|
|
"real_time": 1.5746252892463224e+06,
|
|
"cpu_time": 2.7260212279905695e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3273677599926767e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 443,
|
|
"real_time": 1.5714530592351209e+06,
|
|
"cpu_time": 2.7263531422118405e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3381219061567240e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5725421625700346e+06,
|
|
"cpu_time": 2.7257358826184049e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3344315046997023e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5714530592351211e+06,
|
|
"cpu_time": 2.7259582460496230e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3381219061567240e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9631555078462566e+03,
|
|
"cpu_time": 5.3518398136299697e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.6566825853844043e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2483961031848172e-03,
|
|
"cpu_time": 1.9634476868275546e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2478710392137911e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 217,
|
|
"real_time": 3.2241002426907625e+06,
|
|
"cpu_time": 5.3961073179721627e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2036893201552896e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 217,
|
|
"real_time": 3.2250085869194586e+06,
|
|
"cpu_time": 5.3971967327183234e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2022236678835220e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 217,
|
|
"real_time": 3.2232146879683854e+06,
|
|
"cpu_time": 5.3946867281102901e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2051189958354263e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 217,
|
|
"real_time": 3.2211390650114431e+06,
|
|
"cpu_time": 5.3926539723503627e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2084730467668896e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 217,
|
|
"real_time": 3.2235110674293772e+06,
|
|
"cpu_time": 5.3953291290322701e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2046404212842264e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2233947300038859e+06,
|
|
"cpu_time": 5.3951947760366816e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2048290903850708e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2235110674293772e+06,
|
|
"cpu_time": 5.3953291290322701e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2046404212842264e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4349439338259269e+03,
|
|
"cpu_time": 1.7009122650328743e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3176152565600700e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.4516544017064800e-04,
|
|
"cpu_time": 3.1526429269757836e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.4528172132326539e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 105,
|
|
"real_time": 6.7514337599277496e+06,
|
|
"cpu_time": 1.0892355571429236e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.9699712969351683e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 105,
|
|
"real_time": 6.7571068670423254e+06,
|
|
"cpu_time": 1.0894403314285316e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.9657986265780659e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 105,
|
|
"real_time": 6.7529575367059028e+06,
|
|
"cpu_time": 1.0886024171430195e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.9688498435854044e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 105,
|
|
"real_time": 6.7508973536037263e+06,
|
|
"cpu_time": 1.0889326457143918e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.9703661961454887e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 105,
|
|
"real_time": 6.7501735474382127e+06,
|
|
"cpu_time": 1.0884577923811127e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.9708991575089779e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.7525138129435834e+06,
|
|
"cpu_time": 1.0889337487619959e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.9691770241506214e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.7514337599277496e+06,
|
|
"cpu_time": 1.0889326457143916e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.9699712969351683e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7635726262265498e+03,
|
|
"cpu_time": 4.1353072167589817e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0329387513449588e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.0926574943529688e-04,
|
|
"cpu_time": 3.7975746655481978e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0910974623457853e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.3995977009043975e+07,
|
|
"cpu_time": 2.2732559588237066e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7948681222207880e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.4020977943551308e+07,
|
|
"cpu_time": 2.2811074921567827e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7863183488470926e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.3970916342063278e+07,
|
|
"cpu_time": 2.2718445294119600e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.8034690321600704e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.4011377925235851e+07,
|
|
"cpu_time": 2.2773468882352222e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7895977367886438e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.4022938779317865e+07,
|
|
"cpu_time": 2.2796713431373563e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7856490751408997e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4004437599842453e+07,
|
|
"cpu_time": 2.2766452423530061e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7919804630314989e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4011377925235853e+07,
|
|
"cpu_time": 2.2773468882352225e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7895977367886438e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1559806635222329e+04,
|
|
"cpu_time": 4.0029786969434666e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.3849372073484827e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5394982113001716e-03,
|
|
"cpu_time": 1.7582795169291394e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5411033630710234e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 31,
|
|
"real_time": 2.2710633013517626e+07,
|
|
"cpu_time": 4.2304367645160265e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.9099069550422525e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 31,
|
|
"real_time": 2.2522440300353114e+07,
|
|
"cpu_time": 4.2087731354836099e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.9592888785632925e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 31,
|
|
"real_time": 2.2675262043072332e+07,
|
|
"cpu_time": 4.2273408967745245e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.9191257743813257e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 31,
|
|
"real_time": 2.2639406424376272e+07,
|
|
"cpu_time": 4.2275796645159766e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.9285003097733717e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 31,
|
|
"real_time": 2.2703696102384597e+07,
|
|
"cpu_time": 4.2306490645160012e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.9117126742153206e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2650287576740790e+07,
|
|
"cpu_time": 4.2249559051612273e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.9257069183951130e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2675262043072332e+07,
|
|
"cpu_time": 4.2275796645159766e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.9191257743813257e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.6780680962959683e+04,
|
|
"cpu_time": 9.1775025915832972e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0151552376640368e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.3898325000431572e-03,
|
|
"cpu_time": 2.1722126331240551e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.4007001450044897e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 15,
|
|
"real_time": 4.8211543510357536e+07,
|
|
"cpu_time": 7.8690364466668442e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.5678668728440647e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 15,
|
|
"real_time": 4.8127627372741699e+07,
|
|
"cpu_time": 7.8660490400003865e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.5775750988305569e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 15,
|
|
"real_time": 4.8118165632088982e+07,
|
|
"cpu_time": 7.8600270133332148e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.5786718482257795e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 15,
|
|
"real_time": 4.8187468945980072e+07,
|
|
"cpu_time": 7.8680159333331779e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.5706485912535896e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 15,
|
|
"real_time": 4.8188418149948120e+07,
|
|
"cpu_time": 7.8607838399996892e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.5705388619462080e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.8166644722223282e+07,
|
|
"cpu_time": 7.8647824546666607e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.5730602546200409e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.8187468945980072e+07,
|
|
"cpu_time": 7.8660490400003865e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.5706485912535896e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.1219556538068005e+04,
|
|
"cpu_time": 4.1460587354484363e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.7701822946801521e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.5576972977422254e-04,
|
|
"cpu_time": 5.2716763106248703e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.5593589100812128e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.5696569420397282e+07,
|
|
"cpu_time": 1.5219337737499928e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6101374924059553e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.5662770792841911e+07,
|
|
"cpu_time": 1.5229740637499845e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6121196109048100e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.5471829175949097e+07,
|
|
"cpu_time": 1.5184150625000826e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6233437301235504e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.5622089691460133e+07,
|
|
"cpu_time": 1.5228369062501201e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6145072099166555e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8,
|
|
"real_time": 9.5692129805684090e+07,
|
|
"cpu_time": 1.5230681137498435e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6103977734656916e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.5629077777266502e+07,
|
|
"cpu_time": 1.5218455840000048e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6141011633633327e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.5662770792841911e+07,
|
|
"cpu_time": 1.5228369062501201e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.6121196109048100e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.2792585138439084e+04,
|
|
"cpu_time": 1.9702170268881929e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.4530323772081165e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.7033859674529115e-04,
|
|
"cpu_time": 1.2946234805962984e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.7130996013995545e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8888075649738312e+08,
|
|
"cpu_time": 2.8117600350003612e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6847602895686007e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8893163651227951e+08,
|
|
"cpu_time": 2.8123227124996221e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6832293618025837e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8859283998608589e+08,
|
|
"cpu_time": 2.8124326925001240e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6934389666077414e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8940332159399986e+08,
|
|
"cpu_time": 2.8218334275004506e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6690759959407978e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.8916628137230873e+08,
|
|
"cpu_time": 2.8198113575001574e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6761797938328590e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8899496719241142e+08,
|
|
"cpu_time": 2.8156320450001431e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6813368815505171e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8893163651227951e+08,
|
|
"cpu_time": 2.8124326925001240e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 5.6832293618025837e+09,
|
|
"hip_id": 4.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.0613028822608985e+05,
|
|
"cpu_time": 4.7985361660058907e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.2017190833539367e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6197801072365364e-03,
|
|
"cpu_time": 1.7042483141669338e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6196397564867968e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9430,
|
|
"real_time": 7.4131403252224889e+04,
|
|
"cpu_time": 2.7463859268294362e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3813309273479413e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9430,
|
|
"real_time": 7.4016262266150618e+04,
|
|
"cpu_time": 2.7429431187696807e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3834797497850677e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9430,
|
|
"real_time": 7.4198814377630391e+04,
|
|
"cpu_time": 2.7451275143159245e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3800759602281699e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9430,
|
|
"real_time": 7.4321437334310467e+04,
|
|
"cpu_time": 2.7501147433721949e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3777989725815902e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9430,
|
|
"real_time": 7.4104255385124576e+04,
|
|
"cpu_time": 2.7479711834569043e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3818369737044200e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4154434523088188e+04,
|
|
"cpu_time": 2.7465084973488282e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3809045167294377e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4131403252224889e+04,
|
|
"cpu_time": 2.7463859268294368e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.3813309273479413e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1401695965027140e+02,
|
|
"cpu_time": 2.7284850701886791e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1221473587169614e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5375609076321913e-03,
|
|
"cpu_time": 9.9343769473950400e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5367806629694415e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9563,
|
|
"real_time": 7.4206926370552144e+04,
|
|
"cpu_time": 2.7463953529228334e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7598501921145692e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9563,
|
|
"real_time": 7.4412321954262981e+04,
|
|
"cpu_time": 2.7473657858412626e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7522323537475269e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9563,
|
|
"real_time": 7.4181377112826827e+04,
|
|
"cpu_time": 2.7477813792742859e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7608007288474519e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9563,
|
|
"real_time": 7.4051294075431273e+04,
|
|
"cpu_time": 2.7461979838963045e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7656505204538818e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9563,
|
|
"real_time": 7.4423664940886709e+04,
|
|
"cpu_time": 2.7518980968316080e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7518128832095101e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4255116890791993e+04,
|
|
"cpu_time": 2.7479277197532589e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7580693356745884e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4206926370552130e+04,
|
|
"cpu_time": 2.7473657858412631e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.7598501921145692e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6001991932352402e+02,
|
|
"cpu_time": 2.3152329059347227e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.9439018424348156e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.1550019180343826e-03,
|
|
"cpu_time": 8.4253777466265062e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1550951477371087e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9532,
|
|
"real_time": 7.4466443511040066e+04,
|
|
"cpu_time": 2.7550964393620490e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.5004641109156035e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9532,
|
|
"real_time": 7.4380819315156579e+04,
|
|
"cpu_time": 2.7510800451111922e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.5067960231050558e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9532,
|
|
"real_time": 7.4127892222387076e+04,
|
|
"cpu_time": 2.7481128902643564e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.5255854135334268e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9532,
|
|
"real_time": 7.4118660248141998e+04,
|
|
"cpu_time": 2.7492183508183016e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.5262736621074833e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9532,
|
|
"real_time": 7.4314580998459729e+04,
|
|
"cpu_time": 2.7497692719262087e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.5117043586438246e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4281679259037090e+04,
|
|
"cpu_time": 2.7506553994964215e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.5141647136610799e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4314580998459729e+04,
|
|
"cpu_time": 2.7497692719262082e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.5117043586438246e+07,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5433304454305281e+02,
|
|
"cpu_time": 2.7026765971710120e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1457710077716227e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0776730693561518e-03,
|
|
"cpu_time": 9.8255731985395431e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0778686660067112e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9396,
|
|
"real_time": 7.5050474547883350e+04,
|
|
"cpu_time": 2.7712085057470755e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0915320721620993e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9396,
|
|
"real_time": 7.5382501606558581e+04,
|
|
"cpu_time": 2.7745953724988160e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0867243492072256e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9396,
|
|
"real_time": 7.5208277193125527e+04,
|
|
"cpu_time": 2.7712975287357101e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0892418103081873e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9396,
|
|
"real_time": 7.5023742365215599e+04,
|
|
"cpu_time": 2.7715121062153857e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0919210028368542e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9396,
|
|
"real_time": 7.5000819605558849e+04,
|
|
"cpu_time": 2.7694269082588016e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0922547304260169e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5133163063668384e+04,
|
|
"cpu_time": 2.7716080842911580e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0903347929880768e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5050474547883350e+04,
|
|
"cpu_time": 2.7712975287357101e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0915320721620993e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6135125674404975e+02,
|
|
"cpu_time": 1.8672666707624040e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3380341262479467e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.1475371216212410e-03,
|
|
"cpu_time": 6.7371237706573489e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1443268079527514e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9030,
|
|
"real_time": 7.6710085437179543e+04,
|
|
"cpu_time": 2.8000450808417075e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1358338876336417e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9030,
|
|
"real_time": 7.6666321751092022e+04,
|
|
"cpu_time": 2.8001684330009919e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1370530926464629e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9030,
|
|
"real_time": 7.6618353717760343e+04,
|
|
"cpu_time": 2.7983959999999712e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1383910257787418e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9030,
|
|
"real_time": 7.6438806939797665e+04,
|
|
"cpu_time": 2.7972008781839418e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1434138830690873e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9030,
|
|
"real_time": 7.6476338989770971e+04,
|
|
"cpu_time": 2.7979721395348397e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1423619666458446e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.6581981367120112e+04,
|
|
"cpu_time": 2.7987565063122904e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1394107711547557e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.6618353717760328e+04,
|
|
"cpu_time": 2.7983959999999718e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1383910257787418e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1885508987269618e+02,
|
|
"cpu_time": 1.3056737918903093e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.3214110003129696e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5519981038741536e-03,
|
|
"cpu_time": 4.6651925201263633e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5524886782355613e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8952,
|
|
"real_time": 7.8926509139704634e+04,
|
|
"cpu_time": 2.8436951988383784e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1517102881110173e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8952,
|
|
"real_time": 7.8889710743489646e+04,
|
|
"cpu_time": 2.8431595844503382e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1536468686702812e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8952,
|
|
"real_time": 7.8910225784707043e+04,
|
|
"cpu_time": 2.8441343621537549e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1525670056250817e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8952,
|
|
"real_time": 7.8954211893166910e+04,
|
|
"cpu_time": 2.8423951407505921e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1502535728351569e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8952,
|
|
"real_time": 7.8908012524880003e+04,
|
|
"cpu_time": 2.8434721235479729e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1526834793447775e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.8917734017189665e+04,
|
|
"cpu_time": 2.8433712819482078e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1521722429172635e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.8910225784707029e+04,
|
|
"cpu_time": 2.8434721235479729e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 4.1525670056250817e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4206523118148258e+01,
|
|
"cpu_time": 6.5097043549226115e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2734224059596250e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.0673109687710565e-04,
|
|
"cpu_time": 2.2894317025184003e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.0668824207180157e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8219,
|
|
"real_time": 8.4251423352324040e+04,
|
|
"cpu_time": 2.9357600365009438e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7786222941232026e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8219,
|
|
"real_time": 8.4223953132370734e+04,
|
|
"cpu_time": 2.9359273695096275e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7811593451331151e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8219,
|
|
"real_time": 8.4332350544549437e+04,
|
|
"cpu_time": 2.9307992079327034e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7711577558104384e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8219,
|
|
"real_time": 8.4207001008062027e+04,
|
|
"cpu_time": 2.9361029018127837e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7827258084782696e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8219,
|
|
"real_time": 8.4353974868516190e+04,
|
|
"cpu_time": 2.9366995279231516e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7691656027059722e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.4273740581164486e+04,
|
|
"cpu_time": 2.9350578087358421e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7765661612502003e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.4251423352324040e+04,
|
|
"cpu_time": 2.9359273695096275e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.7786222941232026e+08,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.5771741762524357e+01,
|
|
"cpu_time": 2.4068860591617130e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.0680596305573662e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.8045357081520832e-04,
|
|
"cpu_time": 8.2004724131767013e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.8030065002132436e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7336,
|
|
"real_time": 9.5096494603131883e+04,
|
|
"cpu_time": 3.0920485305344150e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3783052734699152e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7336,
|
|
"real_time": 9.2613309000664725e+04,
|
|
"cpu_time": 3.0908862009270990e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4152609534668419e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7336,
|
|
"real_time": 9.4887698607651560e+04,
|
|
"cpu_time": 3.0895154048527300e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3813381705247788e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7336,
|
|
"real_time": 9.2536850662310346e+04,
|
|
"cpu_time": 3.0899279021262529e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4164303092431130e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7336,
|
|
"real_time": 9.2436062684101751e+04,
|
|
"cpu_time": 3.0885320651584107e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4179747188922977e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.3514083111572065e+04,
|
|
"cpu_time": 3.0901820207197819e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4018618851193895e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.2613309000664711e+04,
|
|
"cpu_time": 3.0899279021262529e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.4152609534668419e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3527149515417300e+03,
|
|
"cpu_time": 1.3426685159138907e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0171354933322649e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4465360794136217e-02,
|
|
"cpu_time": 4.3449496078588575e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4388974511283440e-02,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 6095,
|
|
"real_time": 1.1387341791879266e+05,
|
|
"cpu_time": 3.4364054126331903e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3020649137531338e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 6095,
|
|
"real_time": 1.1354598417956270e+05,
|
|
"cpu_time": 3.4339715668579348e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3087034023628960e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 6095,
|
|
"real_time": 1.1358365629827550e+05,
|
|
"cpu_time": 3.4351716833470378e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3079376782131290e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 6095,
|
|
"real_time": 1.1355246969122802e+05,
|
|
"cpu_time": 3.4335596242824115e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3085715415333738e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 6095,
|
|
"real_time": 1.1354404303166554e+05,
|
|
"cpu_time": 3.4339803494666400e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3087428719346590e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1361991422390488e+05,
|
|
"cpu_time": 3.4346177273174434e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3072040815594382e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1355246969122805e+05,
|
|
"cpu_time": 3.4339803494666406e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.3085715415333738e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4260926929560898e+02,
|
|
"cpu_time": 1.1668378169410953e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.8911500854573427e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2551432578498192e-03,
|
|
"cpu_time": 3.3972858395872678e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2530968146966938e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4453,
|
|
"real_time": 1.5757481509510826e+05,
|
|
"cpu_time": 4.1631037031216943e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3272322082913618e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4453,
|
|
"real_time": 1.5778575909944941e+05,
|
|
"cpu_time": 4.1655520480573660e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3227840268496671e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4453,
|
|
"real_time": 1.5771968512021087e+05,
|
|
"cpu_time": 4.1638064024250832e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3241760506965117e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4453,
|
|
"real_time": 1.5790224385740812e+05,
|
|
"cpu_time": 4.1655514731639886e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3203328033352871e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4453,
|
|
"real_time": 1.5782212409378123e+05,
|
|
"cpu_time": 4.1630090545698005e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3220183989442253e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5776092545319159e+05,
|
|
"cpu_time": 4.1642045362675871e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3233086976234112e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5778575909944944e+05,
|
|
"cpu_time": 4.1638064024250838e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3227840268496671e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2315712058746423e+02,
|
|
"cpu_time": 1.2678276403541035e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5953002826493895e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.8065668183472682e-04,
|
|
"cpu_time": 3.0445854167635787e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.8093867250591544e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2814,
|
|
"real_time": 2.4633521223270320e+05,
|
|
"cpu_time": 5.6399166524518305e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2567036620385857e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2814,
|
|
"real_time": 2.4889705826110698e+05,
|
|
"cpu_time": 5.6397573240938841e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2128902901696205e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2814,
|
|
"real_time": 2.4638320194623835e+05,
|
|
"cpu_time": 5.6402503269363823e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2558745552336922e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2814,
|
|
"real_time": 2.4898495845346563e+05,
|
|
"cpu_time": 5.6398932444919285e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2114029960407224e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2814,
|
|
"real_time": 2.4638024256036794e+05,
|
|
"cpu_time": 5.6399666204688931e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2559256744910398e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4739613469077647e+05,
|
|
"cpu_time": 5.6399568336885830e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2385594355947323e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4638320194623829e+05,
|
|
"cpu_time": 5.6399166524518305e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2558745552336922e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4107408107458155e+03,
|
|
"cpu_time": 1.8144990595823618e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4119441384761795e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.7023559099220293e-03,
|
|
"cpu_time": 3.2172215374841140e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.6904808700358554e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1641,
|
|
"real_time": 4.2646883163682884e+05,
|
|
"cpu_time": 8.5071650517970289e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9174801167788200e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1641,
|
|
"real_time": 4.2654829202852643e+05,
|
|
"cpu_time": 8.5110001645326242e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9165640542753553e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1641,
|
|
"real_time": 4.2668645110408089e+05,
|
|
"cpu_time": 8.5118111395495990e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9149720938489456e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1641,
|
|
"real_time": 4.2647769413451455e+05,
|
|
"cpu_time": 8.5044867885425966e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9173779281843081e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1641,
|
|
"real_time": 4.2653620131126518e+05,
|
|
"cpu_time": 8.5058697014019347e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9167034206074381e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2654349404304323e+05,
|
|
"cpu_time": 8.5080665691647562e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9166195227389736e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2653620131126512e+05,
|
|
"cpu_time": 8.5071650517970289e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9167034206074381e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.7202391589604389e+01,
|
|
"cpu_time": 3.2047456999868132e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0049799100081554e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0443962411205983e-04,
|
|
"cpu_time": 3.7667144161771947e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0440465351451406e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 861,
|
|
"real_time": 8.0671936622955999e+05,
|
|
"cpu_time": 1.4729545447152157e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1992107485944118e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 861,
|
|
"real_time": 8.0610110608884657e+05,
|
|
"cpu_time": 1.4716147247388437e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2031984180626020e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 861,
|
|
"real_time": 8.0602696088643512e+05,
|
|
"cpu_time": 1.4716997607431312e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2036770524242487e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 861,
|
|
"real_time": 8.0598533578361978e+05,
|
|
"cpu_time": 1.4713986202092627e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2039457962620192e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 861,
|
|
"real_time": 8.0648857008224470e+05,
|
|
"cpu_time": 1.4723444761904667e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2006986280937243e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0626426781414135e+05,
|
|
"cpu_time": 1.4720024253193843e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2021461286874008e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0610110608884646e+05,
|
|
"cpu_time": 1.4716997607431314e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.2031984180626020e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2332857867871633e+02,
|
|
"cpu_time": 6.3836281566756747e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0857587394902604e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.0102059781874083e-04,
|
|
"cpu_time": 4.3366967654897729e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0094197431100160e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5769819344467157e+06,
|
|
"cpu_time": 2.7269252696631034e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3194065301345024e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5708137863430749e+06,
|
|
"cpu_time": 2.7250570494380668e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3402943575693054e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5731077366560865e+06,
|
|
"cpu_time": 2.7258024044942758e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3325069888928537e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5669014736528643e+06,
|
|
"cpu_time": 2.7239281887641787e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3536282536284313e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 445,
|
|
"real_time": 1.5732418583559522e+06,
|
|
"cpu_time": 2.7198846044944385e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3320523830748758e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5722093578909389e+06,
|
|
"cpu_time": 2.7243195033708126e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3355777026599932e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5731077366560865e+06,
|
|
"cpu_time": 2.7250570494380668e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3325069888928537e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7004014521036343e+03,
|
|
"cpu_time": 2.7089870375182604e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2564167361024765e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3536314890451908e-03,
|
|
"cpu_time": 9.9437200158293418e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3547904390486224e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 218,
|
|
"real_time": 3.2158928854954899e+06,
|
|
"cpu_time": 5.3858120871555340e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2169697802030630e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 218,
|
|
"real_time": 3.2145218216651762e+06,
|
|
"cpu_time": 5.3835729908254035e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2191949318636513e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 218,
|
|
"real_time": 3.2168961377865677e+06,
|
|
"cpu_time": 5.3858436788983168e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2153427656336479e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 218,
|
|
"real_time": 3.2150304391382076e+06,
|
|
"cpu_time": 5.3835729908254035e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2183692557813396e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 218,
|
|
"real_time": 3.2169629484329610e+06,
|
|
"cpu_time": 5.3856977110094829e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2152344521631727e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2158608465036806e+06,
|
|
"cpu_time": 5.3848998917428283e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2170222371289759e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2158928854954904e+06,
|
|
"cpu_time": 5.3856977110094829e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.2169697802030630e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0920218749542132e+03,
|
|
"cpu_time": 1.2125059091588769e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7716244607929753e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.3957373377690498e-04,
|
|
"cpu_time": 2.2516777164569495e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.3958537653616233e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.8330636321760640e+06,
|
|
"cpu_time": 1.1052080960395627e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.9105984966971874e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.8273620892057912e+06,
|
|
"cpu_time": 1.1073970524752315e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.9146993467726412e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.8526232504461072e+06,
|
|
"cpu_time": 1.1127624811882079e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.8965820494823790e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.9800810750094382e+06,
|
|
"cpu_time": 1.1460912861386785e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.8071693780368633e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.8491604547984526e+06,
|
|
"cpu_time": 1.1059517277227305e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.8990576613652115e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8684581003271714e+06,
|
|
"cpu_time": 1.1154821287128823e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.8856213864708576e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8491604547984526e+06,
|
|
"cpu_time": 1.1073970524752315e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 4.8990576613652115e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.3293918697629975e+04,
|
|
"cpu_time": 1.7364257464322346e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.4510591589393191e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.2151568478833767e-03,
|
|
"cpu_time": 1.5566594046969076e-02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.1105282354974997e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4167133942246437e+07,
|
|
"cpu_time": 2.3148798780002836e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7369400383715696e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4230987522751093e+07,
|
|
"cpu_time": 2.3347207500000879e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7156856748495493e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4181377310305834e+07,
|
|
"cpu_time": 2.3206355120000806e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7321823918492680e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4205118399113417e+07,
|
|
"cpu_time": 2.3280875899999958e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7242734706236916e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.4167889207601547e+07,
|
|
"cpu_time": 2.3201921620002393e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7366875203960409e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4190501276403666e+07,
|
|
"cpu_time": 2.3237031784001373e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7291538192180243e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4181377310305834e+07,
|
|
"cpu_time": 2.3206355120000809e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 4.7321823918492680e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7355900359201274e+04,
|
|
"cpu_time": 7.7491675728313392e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.1071788913318273e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9277613825164422e-03,
|
|
"cpu_time": 3.3348353803804740e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9257523099212111e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6665634618928798e+07,
|
|
"cpu_time": 4.3360272192304499e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0333596000270901e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6682865662643544e+07,
|
|
"cpu_time": 4.3301535884622283e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0301091980501575e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6750324795452449e+07,
|
|
"cpu_time": 4.3197325461540885e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0174242378850288e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6626219041645527e+07,
|
|
"cpu_time": 4.3230002115382589e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0408106306822157e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 26,
|
|
"real_time": 2.6693887292192534e+07,
|
|
"cpu_time": 4.3298398846157119e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0280323180676727e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6683786282172572e+07,
|
|
"cpu_time": 4.3277506900001481e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0299471969424334e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6682865662643544e+07,
|
|
"cpu_time": 4.3298398846157119e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 5.0301091980501575e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5204318801594767e+04,
|
|
"cpu_time": 6.4322870660360939e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.5172794427516591e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6940743837315080e-03,
|
|
"cpu_time": 1.4862887275135235e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6933138876545422e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.0896696001291275e+07,
|
|
"cpu_time": 8.0828112199992567e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2741234125136461e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.0996154174208641e+07,
|
|
"cpu_time": 8.0783675900011078e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2638372510012045e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.0961642712354660e+07,
|
|
"cpu_time": 8.0795241500004515e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2674019461096182e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.0911192595958710e+07,
|
|
"cpu_time": 8.0835311700002417e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2726216439350939e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 5.0969769805669785e+07,
|
|
"cpu_time": 8.0766976499990046e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2665620626393280e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.0947091057896622e+07,
|
|
"cpu_time": 8.0801863560000136e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2689092632397785e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.0961642712354660e+07,
|
|
"cpu_time": 8.0795241500004515e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 5.2674019461096182e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.1718182261840018e+04,
|
|
"cpu_time": 2.9152696200963117e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.3149677517033657e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.1885307670326432e-04,
|
|
"cpu_time": 3.6079237429116379e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.1894895814739325e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.9134230779276952e+07,
|
|
"cpu_time": 1.5704988599999422e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4155956805207558e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.9137485855155528e+07,
|
|
"cpu_time": 1.5687742355554393e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4154178650888262e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.9097324742211238e+07,
|
|
"cpu_time": 1.5689588044445676e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4176125682161417e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.9218212068080902e+07,
|
|
"cpu_time": 1.5729261344444644e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4110117569102478e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 9.9172240330113307e+07,
|
|
"cpu_time": 1.5699797088889986e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4135200557426653e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.9151898754967600e+07,
|
|
"cpu_time": 1.5702275486666825e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4146315852957268e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.9137485855155528e+07,
|
|
"cpu_time": 1.5699797088889986e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 5.4154178650888262e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5578084338534863e+04,
|
|
"cpu_time": 1.6687742877715878e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4886055453446214e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.5967939001522514e-04,
|
|
"cpu_time": 1.0627595275529229e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.5960754783443001e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.6420438885688782e+08,
|
|
"cpu_time": 2.8615045775001138e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.5390567905941830e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.6459254920482635e+08,
|
|
"cpu_time": 2.8654213499999058e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.5236356638707104e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.6448822245001793e+08,
|
|
"cpu_time": 2.8634273350002104e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.5277732837454157e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.6477622836828232e+08,
|
|
"cpu_time": 2.8654359000000793e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.5163636443974094e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4,
|
|
"real_time": 1.6497995704412460e+08,
|
|
"cpu_time": 2.8666079749996245e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.5083167873102493e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6460826918482783e+08,
|
|
"cpu_time": 2.8644794274999869e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.5230292339835939e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6459254920482635e+08,
|
|
"cpu_time": 2.8654213499999058e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 6.5236356638707104e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.9331715242342890e+05,
|
|
"cpu_time": 2.0180066428868429e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1625860267716967e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7819101912436872e-03,
|
|
"cpu_time": 7.0449332730871989e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7822793445641343e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8815,
|
|
"real_time": 7.9930999779130027e+04,
|
|
"cpu_time": 2.8269975757231913e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.2811049565620050e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8815,
|
|
"real_time": 7.9806262154253098e+04,
|
|
"cpu_time": 2.8244717742486118e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.2831073306262197e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8815,
|
|
"real_time": 7.9752029241268727e+04,
|
|
"cpu_time": 2.8215825263756234e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.2839798682766531e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8815,
|
|
"real_time": 7.9511089906445384e+04,
|
|
"cpu_time": 2.8187596324449137e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.2878706620735077e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8815,
|
|
"real_time": 7.9709553930768976e+04,
|
|
"cpu_time": 2.8244491117413028e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.2846640703690125e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9741987002373236e+04,
|
|
"cpu_time": 2.8232521241067292e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.2841453775814796e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9752029241268727e+04,
|
|
"cpu_time": 2.8244491117413028e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.2839798682766531e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5356052921528433e+02,
|
|
"cpu_time": 3.1590651935910068e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4746657118784500e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9257173665701874e-03,
|
|
"cpu_time": 1.1189454766072401e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9270915544929658e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8764,
|
|
"real_time": 7.9846212602500644e+04,
|
|
"cpu_time": 2.8253962311730004e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.5649306751662508e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8764,
|
|
"real_time": 7.9988776974615714e+04,
|
|
"cpu_time": 2.8258438064810238e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.5603591872018855e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8764,
|
|
"real_time": 7.9533820542063637e+04,
|
|
"cpu_time": 2.8188770880877739e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.5750051814961653e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8764,
|
|
"real_time": 7.9662512288961676e+04,
|
|
"cpu_time": 2.8194851962573238e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.5708453589452997e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8764,
|
|
"real_time": 7.9632535710543918e+04,
|
|
"cpu_time": 2.8184110429026902e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.5718131185025550e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9732771623737121e+04,
|
|
"cpu_time": 2.8216026729803625e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.5685907042624313e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9662512288961661e+04,
|
|
"cpu_time": 2.8194851962573238e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.5708453589452997e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8232713776204986e+02,
|
|
"cpu_time": 3.6904376557498006e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.8687014227222789e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.2867277036656974e-03,
|
|
"cpu_time": 1.3079225119430856e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2847943087948969e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8742,
|
|
"real_time": 7.9963619629171721e+04,
|
|
"cpu_time": 2.8247176721572311e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.1223294030398399e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8742,
|
|
"real_time": 7.9819428825215247e+04,
|
|
"cpu_time": 2.8255865191031864e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.1315826989556946e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8742,
|
|
"real_time": 7.9708807280520879e+04,
|
|
"cpu_time": 2.8208547060170805e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.1387044164202094e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8742,
|
|
"real_time": 7.9857956347466912e+04,
|
|
"cpu_time": 2.8209020521621022e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.1291069635917693e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8742,
|
|
"real_time": 7.9581202082703763e+04,
|
|
"cpu_time": 2.8184642747653776e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.1469441184656695e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9786202833015719e+04,
|
|
"cpu_time": 2.8221050448409963e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.1337335200946368e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.9819428825215247e+04,
|
|
"cpu_time": 2.8209020521621022e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.1315826989556946e+07,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4641092182750921e+02,
|
|
"cpu_time": 2.9669946844171665e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.4248960475036540e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8350406038739825e-03,
|
|
"cpu_time": 1.0513409803228405e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8358755885190380e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8702,
|
|
"real_time": 8.0655506665299486e+04,
|
|
"cpu_time": 2.8466959032406844e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0156777061725971e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8702,
|
|
"real_time": 8.0639028112949934e+04,
|
|
"cpu_time": 2.8462781889222813e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0158852594956358e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8702,
|
|
"real_time": 8.0390973732844155e+04,
|
|
"cpu_time": 2.8464431831764837e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0190198749456267e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8702,
|
|
"real_time": 8.0433557406057938e+04,
|
|
"cpu_time": 2.8461363847391721e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0184803786116029e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8702,
|
|
"real_time": 8.0590637114350204e+04,
|
|
"cpu_time": 2.8503479050793458e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0164952522184874e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0541940606300341e+04,
|
|
"cpu_time": 2.8471803130315937e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0171116942887901e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0590637114350189e+04,
|
|
"cpu_time": 2.8464431831764831e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0164952522184874e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2168863809823263e+02,
|
|
"cpu_time": 1.7829236409310357e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5374422948520703e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5108729338055408e-03,
|
|
"cpu_time": 6.2620678879049692e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5115766571999929e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8522,
|
|
"real_time": 8.2181486635480047e+04,
|
|
"cpu_time": 2.8776045470546960e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.9936363615168005e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8522,
|
|
"real_time": 8.1975165204015459e+04,
|
|
"cpu_time": 2.8762507944145706e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.9986540995952088e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8522,
|
|
"real_time": 8.1913577035179900e+04,
|
|
"cpu_time": 2.8807644297112111e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0001568229603097e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8522,
|
|
"real_time": 8.1941108311944161e+04,
|
|
"cpu_time": 2.8792168199952459e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.9994847931061950e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8522,
|
|
"real_time": 8.2162971297303360e+04,
|
|
"cpu_time": 2.8811910971602425e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.9940856253500330e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2034861696784588e+04,
|
|
"cpu_time": 2.8790055376671941e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.9972035405057096e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.1975165204015459e+04,
|
|
"cpu_time": 2.8792168199952459e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.9986540995952088e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2745026476088799e+02,
|
|
"cpu_time": 2.0896409272722877e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1014622844716330e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5536110151799462e-03,
|
|
"cpu_time": 7.2582039177510082e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5529024566451125e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 8292,
|
|
"real_time": 8.4780401214261714e+04,
|
|
"cpu_time": 2.9190081922332977e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8650442237454027e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 8292,
|
|
"real_time": 8.4494668256846169e+04,
|
|
"cpu_time": 2.9152442486735247e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8781145220183730e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 8292,
|
|
"real_time": 8.4333878750819902e+04,
|
|
"cpu_time": 2.9136380559576669e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8855084676964921e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 8292,
|
|
"real_time": 8.4498803303244829e+04,
|
|
"cpu_time": 2.9167102243126149e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8779247420113081e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 8292,
|
|
"real_time": 8.4504815994216202e+04,
|
|
"cpu_time": 2.9161467836469732e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8776488197125661e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.4522513503877781e+04,
|
|
"cpu_time": 2.9161495009648160e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8768481550368291e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.4498803303244844e+04,
|
|
"cpu_time": 2.9161467836469726e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8779247420113081e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6104337495492871e+02,
|
|
"cpu_time": 1.9756879531930340e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.3779671660966345e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9053311156860032e-03,
|
|
"cpu_time": 6.7749885681079528e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9030838637595661e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7610,
|
|
"real_time": 8.9833033150855903e+04,
|
|
"cpu_time": 3.0071322299604717e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.2953119472149968e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7610,
|
|
"real_time": 8.9849870538069168e+04,
|
|
"cpu_time": 3.0062114835743693e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.2939448446097159e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7610,
|
|
"real_time": 8.9881791415663785e+04,
|
|
"cpu_time": 3.0074579724047327e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.2913544520852721e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7610,
|
|
"real_time": 8.9894737744082493e+04,
|
|
"cpu_time": 3.0102816281211487e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.2903043764999509e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7610,
|
|
"real_time": 9.0046208583801534e+04,
|
|
"cpu_time": 3.0118513823917328e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.2780410225721943e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.9901128286494582e+04,
|
|
"cpu_time": 3.0085869392904907e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.2897913285964262e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.9881791415663785e+04,
|
|
"cpu_time": 3.0074579724047321e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 7.2913544520852721e+08,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.4745723400643868e+01,
|
|
"cpu_time": 2.3749786483307057e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.8648596134225500e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.4265472542878889e-04,
|
|
"cpu_time": 7.8940003937223496e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.4170865858574690e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7015,
|
|
"real_time": 9.6781413510713290e+04,
|
|
"cpu_time": 3.1634006585888023e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3543096266670139e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7015,
|
|
"real_time": 9.9664518699837892e+04,
|
|
"cpu_time": 3.1660107940127968e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3151320220062749e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7015,
|
|
"real_time": 9.6821396517909539e+04,
|
|
"cpu_time": 3.1700416949392605e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3537503559531386e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7015,
|
|
"real_time": 9.6993303981543926e+04,
|
|
"cpu_time": 3.1696348724162957e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3513510172304330e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7015,
|
|
"real_time": 9.9368854012627067e+04,
|
|
"cpu_time": 3.1692280741268600e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3190451002216885e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.7925897344526340e+04,
|
|
"cpu_time": 3.1676632188168028e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3387176244157100e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.6993303981543941e+04,
|
|
"cpu_time": 3.1692280741268600e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.3513510172304330e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4581172677206546e+03,
|
|
"cpu_time": 2.8676325670859353e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9824133113903366e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4890006701604735e-02,
|
|
"cpu_time": 9.0528328581504437e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4808300684437250e-02,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 6002,
|
|
"real_time": 1.1694178033477906e+05,
|
|
"cpu_time": 3.5034565294903488e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2416624687048402e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 6002,
|
|
"real_time": 1.1664264220556423e+05,
|
|
"cpu_time": 3.5032668593799399e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2474113672597761e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 6002,
|
|
"real_time": 1.1764109574363998e+05,
|
|
"cpu_time": 3.5175805398197187e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2283369458854456e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 6002,
|
|
"real_time": 1.1686097558610521e+05,
|
|
"cpu_time": 3.5095460263242910e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2432124897575212e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 6002,
|
|
"real_time": 1.1678643826270479e+05,
|
|
"cpu_time": 3.5080245801399276e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2446441889966817e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1697458642655870e+05,
|
|
"cpu_time": 3.5083749070308456e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2410534921208529e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1686097558610523e+05,
|
|
"cpu_time": 3.5080245801399276e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.2432124897575212e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8852446186801859e+02,
|
|
"cpu_time": 5.8420996400073261e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.4171819406780088e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.3214433471149712e-03,
|
|
"cpu_time": 1.6651868157817613e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.3096853630471147e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4365,
|
|
"real_time": 1.6096995623125124e+05,
|
|
"cpu_time": 4.2273608911796176e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2570549950749941e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4365,
|
|
"real_time": 1.6071479529196254e+05,
|
|
"cpu_time": 4.2251560137458134e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2622261008860583e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4365,
|
|
"real_time": 1.6096850866739897e+05,
|
|
"cpu_time": 4.2296578235968074e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2570842852455668e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4365,
|
|
"real_time": 1.6093775222276925e+05,
|
|
"cpu_time": 4.2283772348227981e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2577067391514397e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4365,
|
|
"real_time": 1.6075271944996002e+05,
|
|
"cpu_time": 4.2258272760591353e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2614564891588235e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6086874637266845e+05,
|
|
"cpu_time": 4.2272758478808345e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2591057219033766e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6093775222276925e+05,
|
|
"cpu_time": 4.2273608911796176e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.2577067391514397e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2462036378641309e+02,
|
|
"cpu_time": 1.8361747552012633e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5254559538269518e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.7467106940535006e-04,
|
|
"cpu_time": 4.3436359993440017e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.7489230768250134e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2774,
|
|
"real_time": 2.4909939978964339e+05,
|
|
"cpu_time": 5.6966258327322640e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2094681917559395e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2774,
|
|
"real_time": 2.5176030215674642e+05,
|
|
"cpu_time": 5.6982819718816655e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1649775243245244e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2774,
|
|
"real_time": 2.4946549084026128e+05,
|
|
"cpu_time": 5.7029779235759575e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2032907897125874e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2774,
|
|
"real_time": 2.5202579451642634e+05,
|
|
"cpu_time": 5.7021997152122424e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1605899983846965e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2774,
|
|
"real_time": 2.4922767546794709e+05,
|
|
"cpu_time": 5.7004793763514922e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2073016089854598e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5031573255420494e+05,
|
|
"cpu_time": 5.7001129639507248e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.1891256226326418e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4946549084026125e+05,
|
|
"cpu_time": 5.7004793763514911e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.2032907897125874e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4489080233379000e+03,
|
|
"cpu_time": 2.6556260383540206e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4198404182596836e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.7883218467866156e-03,
|
|
"cpu_time": 4.6589007185454396e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.7764809085360950e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1628,
|
|
"real_time": 4.2972306977483811e+05,
|
|
"cpu_time": 8.5732849508600484e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8802406654565783e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1628,
|
|
"real_time": 4.2981112731108099e+05,
|
|
"cpu_time": 8.5759601289920707e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8792408263599024e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1628,
|
|
"real_time": 4.2992425059281284e+05,
|
|
"cpu_time": 8.5757788943487697e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8779569822085733e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1628,
|
|
"real_time": 4.3007275832234026e+05,
|
|
"cpu_time": 8.5797480159700860e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8762725827618713e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1628,
|
|
"real_time": 4.2984229061296047e+05,
|
|
"cpu_time": 8.5743679606886755e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8788870843988733e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2987469932280655e+05,
|
|
"cpu_time": 8.5758279901719303e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8785196282371597e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2984229061296053e+05,
|
|
"cpu_time": 8.5757788943487708e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.8788870843988733e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3206621497618215e+02,
|
|
"cpu_time": 2.4487347937683535e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4985746401164008e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.0722025554011368e-04,
|
|
"cpu_time": 2.8553916852980878e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.0717815122492534e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0926739671505254e+05,
|
|
"cpu_time": 1.4794024222736207e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1828406989152889e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0906637250651058e+05,
|
|
"cpu_time": 1.4785273979119295e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1841284504334631e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0898730583994568e+05,
|
|
"cpu_time": 1.4788674675174279e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1846351231002169e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0931807256124390e+05,
|
|
"cpu_time": 1.4787452412991626e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1825161728148642e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 862,
|
|
"real_time": 8.0880037922976131e+05,
|
|
"cpu_time": 1.4779192354988821e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1858333746014423e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0908790537050297e+05,
|
|
"cpu_time": 1.4786923529002047e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1839907639730549e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0906637250651070e+05,
|
|
"cpu_time": 1.4787452412991624e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 5.1841284504334631e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1121629288721783e+02,
|
|
"cpu_time": 5.3911536081128065e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3533697676540583e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.6105481429795479e-04,
|
|
"cpu_time": 3.6458926683018080e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6106716413530493e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 443,
|
|
"real_time": 1.5805256337507651e+06,
|
|
"cpu_time": 2.7354371376979630e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3074798793948631e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 443,
|
|
"real_time": 1.5759599811878919e+06,
|
|
"cpu_time": 2.7326642099319901e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3228559735869827e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 443,
|
|
"real_time": 1.5801131624400481e+06,
|
|
"cpu_time": 2.7349041444694675e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3088653391419849e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 443,
|
|
"real_time": 1.5766975045746558e+06,
|
|
"cpu_time": 2.7338569074490028e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3203661296229343e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 443,
|
|
"real_time": 1.5805826796235570e+06,
|
|
"cpu_time": 2.7363224266365715e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3072883235680466e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5787757923153839e+06,
|
|
"cpu_time": 2.7346369652369996e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3133711290629625e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5801131624400481e+06,
|
|
"cpu_time": 2.7349041444694675e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.3088653391419849e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2562921951934313e+03,
|
|
"cpu_time": 1.4188634957706295e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.5977265783706643e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4291403543022553e-03,
|
|
"cpu_time": 5.1884894185494292e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4299258218220038e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 216,
|
|
"real_time": 3.2298926757734823e+06,
|
|
"cpu_time": 5.4086677129629590e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1943571146624117e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 216,
|
|
"real_time": 3.2288911601494034e+06,
|
|
"cpu_time": 5.4094870185184414e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1959682652244329e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 216,
|
|
"real_time": 3.2304549423753526e+06,
|
|
"cpu_time": 5.4109335324070612e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1934530272890043e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 216,
|
|
"real_time": 3.2310082600883176e+06,
|
|
"cpu_time": 5.4127214259256953e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1925636363249054e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 216,
|
|
"real_time": 3.2306638205665406e+06,
|
|
"cpu_time": 5.4091271805555578e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1931172451913891e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2301821717906194e+06,
|
|
"cpu_time": 5.4101873740739431e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1938918577384300e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2304549423753517e+06,
|
|
"cpu_time": 5.4094870185184404e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 5.1934530272890043e+09,
|
|
"hip_id": 6.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2754610229075183e+02,
|
|
"cpu_time": 1.6509125605178856e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3308550706970312e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_prefetch_managed_HostToGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.5619177441995786e-04,
|
|
"cpu_time": 3.0514886941424485e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5623465161566219e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
} |