3859 lines
133 KiB
JSON
3859 lines
133 KiB
JSON
{
|
|
"context": {
|
|
"date": "2023-10-09T16:55:39-04:00",
|
|
"host_name": "frontier10465",
|
|
"executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope",
|
|
"num_cpus": 128,
|
|
"mhz_per_cpu": 1796,
|
|
"cpu_scaling_enabled": false,
|
|
"caches": [
|
|
{
|
|
"type": "Data",
|
|
"level": 1,
|
|
"size": 32768,
|
|
"num_sharing": 2
|
|
},
|
|
{
|
|
"type": "Instruction",
|
|
"level": 1,
|
|
"size": 32768,
|
|
"num_sharing": 2
|
|
},
|
|
{
|
|
"type": "Unified",
|
|
"level": 2,
|
|
"size": 524288,
|
|
"num_sharing": 2
|
|
},
|
|
{
|
|
"type": "Unified",
|
|
"level": 3,
|
|
"size": 33554432,
|
|
"num_sharing": 16
|
|
}
|
|
],
|
|
"load_avg": [6.47,11.33,11.88],
|
|
"library_build_type": "release"
|
|
},
|
|
"benchmarks": [
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 52235,
|
|
"real_time": 1.3454085324331463e+04,
|
|
"cpu_time": 2.0377894783191343e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 1.9027677752052665e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 52235,
|
|
"real_time": 1.3337203987187244e+04,
|
|
"cpu_time": 2.0289086359720488e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 1.9194427876032602e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 52235,
|
|
"real_time": 1.3186539730989720e+04,
|
|
"cpu_time": 2.0082922427491154e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 1.9413735917268258e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 52235,
|
|
"real_time": 1.3334940758000001e+04,
|
|
"cpu_time": 2.0219355929932044e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 1.9197685587498281e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 52235,
|
|
"real_time": 1.3183338191850033e+04,
|
|
"cpu_time": 2.0146454235665744e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 1.9418450492171984e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3299221598471693e+04,
|
|
"cpu_time": 2.0223142747200160e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 1.9250395525004759e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3334940758000002e+04,
|
|
"cpu_time": 2.0219355929932040e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 1.9197685587498281e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1492112619808184e+02,
|
|
"cpu_time": 1.1603227839795277e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6615992098446153e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.6411919184268809e-03,
|
|
"cpu_time": 5.7375987426097324e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.6315068575413302e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 52205,
|
|
"real_time": 1.3247311615410497e+04,
|
|
"cpu_time": 2.0172652945120200e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 3.8649351269460157e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 52205,
|
|
"real_time": 1.3286441682410798e+04,
|
|
"cpu_time": 2.0238151498898569e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 3.8535524577495351e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 52205,
|
|
"real_time": 1.3318225952499288e+04,
|
|
"cpu_time": 2.0290479858251147e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 3.8443558611041471e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 52205,
|
|
"real_time": 1.3323327234518472e+04,
|
|
"cpu_time": 2.0287575653673044e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 3.8428839207183562e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 52205,
|
|
"real_time": 1.3326873080040150e+04,
|
|
"cpu_time": 2.0313490144622177e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 3.8418614548586778e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3300435912975841e+04,
|
|
"cpu_time": 2.0260470020113033e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 3.8495177642753467e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3318225952499290e+04,
|
|
"cpu_time": 2.0287575653673044e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 3.8443558611041471e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3754449304474456e+01,
|
|
"cpu_time": 5.6250320904968305e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.7883351717426907e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.5378453402075171e-03,
|
|
"cpu_time": 2.7763581421915351e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5427432138594892e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 52479,
|
|
"real_time": 1.3428744710109755e+04,
|
|
"cpu_time": 2.0394118275881789e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 7.6254335167239219e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 52479,
|
|
"real_time": 1.3314139275841755e+04,
|
|
"cpu_time": 2.0264469902246623e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 7.6910717154508665e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 52479,
|
|
"real_time": 1.3351280570019146e+04,
|
|
"cpu_time": 2.0291210712856559e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 7.6696762878269106e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 52479,
|
|
"real_time": 1.3333394843860455e+04,
|
|
"cpu_time": 2.0274975132910356e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 7.6799645700998262e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 52479,
|
|
"real_time": 1.3319397647958092e+04,
|
|
"cpu_time": 2.0283799386421284e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 7.6880353531376287e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3349391409557842e+04,
|
|
"cpu_time": 2.0301714682063324e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 7.6708362886478320e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3333394843860460e+04,
|
|
"cpu_time": 2.0283799386421284e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 7.6799645700998262e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.6633528753775536e+01,
|
|
"cpu_time": 5.2611530224362411e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6699078353475052e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.4933074717089386e-03,
|
|
"cpu_time": 2.5914821013047228e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.4805955112074752e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 52128,
|
|
"real_time": 1.3468469297934842e+04,
|
|
"cpu_time": 2.0451500153468442e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 1.5205885351158839e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 52128,
|
|
"real_time": 1.3456059822556450e+04,
|
|
"cpu_time": 2.0469015903161475e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 1.5219908554262882e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 52128,
|
|
"real_time": 1.3475397379251719e+04,
|
|
"cpu_time": 2.0529924320902384e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 1.5198067577237743e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 52128,
|
|
"real_time": 1.3475203561443317e+04,
|
|
"cpu_time": 2.0522792702578263e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 1.5198286175505021e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 52128,
|
|
"real_time": 1.3487248049730617e+04,
|
|
"cpu_time": 2.0534116693523672e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 1.5184713682498819e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3472475622183390e+04,
|
|
"cpu_time": 2.0501469954726846e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 1.5201372268132663e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3475203561443317e+04,
|
|
"cpu_time": 2.0522792702578263e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 1.5198286175505021e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1398816209536109e+01,
|
|
"cpu_time": 3.8341774167241141e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2863856797243975e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.4608178401652839e-04,
|
|
"cpu_time": 1.8701963445504554e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.4622996992258853e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 51576,
|
|
"real_time": 1.3570238159934961e+04,
|
|
"cpu_time": 2.0635467329765728e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.0183700180687404e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 51576,
|
|
"real_time": 1.3617236896091308e+04,
|
|
"cpu_time": 2.0704426516209089e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.0079523704076236e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 51576,
|
|
"real_time": 1.3633417601955982e+04,
|
|
"cpu_time": 2.0761581917946289e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.0043824076894319e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 51576,
|
|
"real_time": 1.3636821748398545e+04,
|
|
"cpu_time": 2.0771142469365615e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.0036324266547072e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 51576,
|
|
"real_time": 1.3637286882233237e+04,
|
|
"cpu_time": 2.0784069799906949e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.0035299802457780e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3619000257722806e+04,
|
|
"cpu_time": 2.0731337606638735e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.0075734406132567e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3633417601955985e+04,
|
|
"cpu_time": 2.0761581917946292e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.0043824076894319e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8463650528154325e+01,
|
|
"cpu_time": 6.1634297363464761e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.3002683755535365e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0899955936202947e-03,
|
|
"cpu_time": 2.9730014788687726e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0948011744208266e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 50322,
|
|
"real_time": 1.3798449863577252e+04,
|
|
"cpu_time": 2.1236514148881161e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 5.9368987683347070e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 50322,
|
|
"real_time": 1.3811272590445869e+04,
|
|
"cpu_time": 2.1222058562855207e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 5.9313868047662199e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 50322,
|
|
"real_time": 1.3793247698380748e+04,
|
|
"cpu_time": 2.1298020507929061e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 5.9391378877084148e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 50322,
|
|
"real_time": 1.3810728769636104e+04,
|
|
"cpu_time": 2.1340698799729616e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 5.9316203631561506e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 50322,
|
|
"real_time": 1.3777194034978558e+04,
|
|
"cpu_time": 2.1381403879019137e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 5.9460583767649245e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3798178591403708e+04,
|
|
"cpu_time": 2.1295739179682838e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 5.9370204401460838e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3798449863577252e+04,
|
|
"cpu_time": 2.1298020507929061e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 5.9368987683347070e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4086329252794323e+01,
|
|
"cpu_time": 6.7641426240481351e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.0638553247913159e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0208832390073666e-03,
|
|
"cpu_time": 3.1762891942729338e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0213633902601338e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 49547,
|
|
"real_time": 1.4118803353911655e+04,
|
|
"cpu_time": 2.2180979009829094e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.1604382885225728e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 49547,
|
|
"real_time": 1.4261388916744452e+04,
|
|
"cpu_time": 2.2408044503198933e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.1488362105294924e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 49547,
|
|
"real_time": 1.4237295107339834e+04,
|
|
"cpu_time": 2.2339299553958965e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.1507803888642766e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 49547,
|
|
"real_time": 1.4261502487993177e+04,
|
|
"cpu_time": 2.2314195248955541e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.1488270617905629e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 49547,
|
|
"real_time": 1.4108478324423633e+04,
|
|
"cpu_time": 2.2155082063495243e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.1612875338679962e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4197493638082553e+04,
|
|
"cpu_time": 2.2279520075887558e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.1540338967149801e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4237295107339834e+04,
|
|
"cpu_time": 2.2314195248955544e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.1507803888642766e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7265424585996286e+01,
|
|
"cpu_time": 1.0780728485271615e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.2917380611719685e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.4421876533717111e-03,
|
|
"cpu_time": 4.8388513076362298e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.4519525631627815e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 47214,
|
|
"real_time": 1.4879928993340669e+04,
|
|
"cpu_time": 2.3958850277460191e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.2021610462432261e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 47214,
|
|
"real_time": 1.4875616734282377e+04,
|
|
"cpu_time": 2.4022779620451500e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.2027994257530713e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 47214,
|
|
"real_time": 1.4814705222423083e+04,
|
|
"cpu_time": 2.3909495467446086e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.2118563621774507e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 47214,
|
|
"real_time": 1.4818709830845966e+04,
|
|
"cpu_time": 2.3869183060109364e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.2112586300726118e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 47214,
|
|
"real_time": 1.4838393142940069e+04,
|
|
"cpu_time": 2.3917259414580418e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.2083253681407294e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4845470784766430e+04,
|
|
"cpu_time": 2.3935513568009512e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.2072801664774184e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4838393142940065e+04,
|
|
"cpu_time": 2.3917259414580414e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.2083253681407294e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.0858175194175153e+01,
|
|
"cpu_time": 5.8244866445699948e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.5865187118772333e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0786255715002345e-03,
|
|
"cpu_time": 2.4334078431283734e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0779050985615583e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 40826,
|
|
"real_time": 1.6655007561218190e+04,
|
|
"cpu_time": 2.8376005168275118e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.9349126536935973e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 40826,
|
|
"real_time": 1.6662920323938775e+04,
|
|
"cpu_time": 2.8361932371528070e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.9330440718634262e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 40826,
|
|
"real_time": 1.7221535875586982e+04,
|
|
"cpu_time": 2.8890282565032041e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.8054677860006065e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 40826,
|
|
"real_time": 1.7020580796392260e+04,
|
|
"cpu_time": 2.8736632366629128e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.8503973973609195e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 40826,
|
|
"real_time": 1.6805612904586073e+04,
|
|
"cpu_time": 2.8493063709400918e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.8996495023466792e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6873131492344452e+04,
|
|
"cpu_time": 2.8571583236173057e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.8846942822530460e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6805612904586069e+04,
|
|
"cpu_time": 2.8493063709400914e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 3.8996495023466792e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4466939728546805e+02,
|
|
"cpu_time": 2.3301290707422135e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.5982942972153224e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4500532838049505e-02,
|
|
"cpu_time": 8.1554076002066048e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4411157971402636e-02,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 35095,
|
|
"real_time": 1.9975639513666567e+04,
|
|
"cpu_time": 3.6695036557914420e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.5615921788298979e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 35095,
|
|
"real_time": 1.9885085648935510e+04,
|
|
"cpu_time": 3.6357721384812736e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.5914727406274242e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 35095,
|
|
"real_time": 1.9964696663269748e+04,
|
|
"cpu_time": 3.6701273628722003e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.5651886532862301e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 35095,
|
|
"real_time": 1.9929903344682229e+04,
|
|
"cpu_time": 3.6530998546801646e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.5766500586152172e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 35095,
|
|
"real_time": 1.9898953936944865e+04,
|
|
"cpu_time": 3.5999631514460729e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.5868789090791674e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9930855821499783e+04,
|
|
"cpu_time": 3.6456932326542308e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.5763565080875883e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9929903344682229e+04,
|
|
"cpu_time": 3.6530998546801653e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 6.5766500586152172e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.9573753304519713e+01,
|
|
"cpu_time": 2.9191135025073629e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3057698686748749e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9855521337839781e-03,
|
|
"cpu_time": 8.0070190118056508e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9855521322012302e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 29537,
|
|
"real_time": 2.3694661690657071e+04,
|
|
"cpu_time": 4.9952309747096944e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.1063420251463846e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 29537,
|
|
"real_time": 2.3691333689518771e+04,
|
|
"cpu_time": 4.9621267190303617e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.1064974367229252e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 29537,
|
|
"real_time": 2.3680747889745449e+04,
|
|
"cpu_time": 4.9774011510986318e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.1069920646953768e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 29537,
|
|
"real_time": 2.3671143419944528e+04,
|
|
"cpu_time": 4.9751365880082965e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.1074412222061316e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 29537,
|
|
"real_time": 2.3688962868969942e+04,
|
|
"cpu_time": 4.9867550834546651e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.1066081763477335e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3685369911767153e+04,
|
|
"cpu_time": 4.9793301032603296e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.1067761850237104e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3688962868969942e+04,
|
|
"cpu_time": 4.9774011510986325e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.1066081763477335e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.4679630673641206e+00,
|
|
"cpu_time": 1.2502308779866603e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.4252141622118130e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.9973887267263377e-04,
|
|
"cpu_time": 2.5108415229752349e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.9982918155372234e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 20203,
|
|
"real_time": 3.4650176731843450e+04,
|
|
"cpu_time": 7.3905771766569480e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.5130889636074503e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 20203,
|
|
"real_time": 3.4657723620469842e+04,
|
|
"cpu_time": 7.4117631638865700e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.5127594810939646e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 20203,
|
|
"real_time": 3.4646981458664792e+04,
|
|
"cpu_time": 7.3870178636836426e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.5132285062855942e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 20203,
|
|
"real_time": 3.4693106967712120e+04,
|
|
"cpu_time": 7.4108546206008526e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.5112166243511711e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 20203,
|
|
"real_time": 3.4672253574392598e+04,
|
|
"cpu_time": 7.3661698807107678e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.5121255354085669e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4664048470616559e+04,
|
|
"cpu_time": 7.3932765411077562e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.5124838221493494e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4657723620469842e+04,
|
|
"cpu_time": 7.3905771766569465e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.5127594810939646e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8943546023390205e+01,
|
|
"cpu_time": 1.8920707900371826e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.2627146697632009e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.4648971655598603e-04,
|
|
"cpu_time": 2.5591776251259880e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.4630102806794208e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 12625,
|
|
"real_time": 5.5517649586084255e+04,
|
|
"cpu_time": 1.2279796744554563e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.8887254914747513e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 12625,
|
|
"real_time": 5.5582090171450509e+04,
|
|
"cpu_time": 1.2245718106930668e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.8865357469744747e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 12625,
|
|
"real_time": 5.5574984751620999e+04,
|
|
"cpu_time": 1.2271627999999929e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.8867769459341423e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 12625,
|
|
"real_time": 5.5525695301401014e+04,
|
|
"cpu_time": 1.2244084308910898e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.8884518137200932e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 12625,
|
|
"real_time": 5.5555832405150453e+04,
|
|
"cpu_time": 1.2261762526732622e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.8874273943968285e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.5551250443141456e+04,
|
|
"cpu_time": 1.2260597937425735e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.8875834785000580e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.5555832405150446e+04,
|
|
"cpu_time": 1.2261762526732622e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.8874273943968285e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8798799788759300e+01,
|
|
"cpu_time": 1.5698105188034907e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.7862123256426435e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.1841856950161404e-04,
|
|
"cpu_time": 1.2803702778733251e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.1845189561730651e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7248,
|
|
"real_time": 9.6243652110317955e+04,
|
|
"cpu_time": 2.1791016956401744e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.1790029305996918e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7248,
|
|
"real_time": 9.6242176138949260e+04,
|
|
"cpu_time": 2.1684709616445893e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.1790363478193230e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7248,
|
|
"real_time": 9.6311357810115631e+04,
|
|
"cpu_time": 2.1788178449227428e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.1774711183437756e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7248,
|
|
"real_time": 9.6281731544218870e+04,
|
|
"cpu_time": 2.1705055974061831e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.1781411347352539e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7248,
|
|
"real_time": 9.6276484023804107e+04,
|
|
"cpu_time": 2.1726458692053054e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.1782598536538628e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.6271080325481176e+04,
|
|
"cpu_time": 2.1739083937637991e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.1783822770303818e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.6276484023804122e+04,
|
|
"cpu_time": 2.1726458692053054e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.1782598536538628e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8951179427801815e+01,
|
|
"cpu_time": 4.8428234125861235e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.5505089147760114e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.0072561074334360e-04,
|
|
"cpu_time": 2.2277035345548736e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.0070520605344841e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 3813,
|
|
"real_time": 1.8357034682521780e+05,
|
|
"cpu_time": 4.2909884552845248e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.2848483279238491e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 3813,
|
|
"real_time": 1.8353737418510145e+05,
|
|
"cpu_time": 4.2625442774717812e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.2852588028038109e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 3813,
|
|
"real_time": 1.8358819036993181e+05,
|
|
"cpu_time": 4.2637528901127534e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.2846262559418667e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 3813,
|
|
"real_time": 1.8354852860403893e+05,
|
|
"cpu_time": 4.2643559874114738e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.2851199254493534e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 3813,
|
|
"real_time": 1.8357622793996782e+05,
|
|
"cpu_time": 4.2621766850249324e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.2847751296925003e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8356413358485157e+05,
|
|
"cpu_time": 4.2687636590610928e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.2849256883622761e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8357034682521780e+05,
|
|
"cpu_time": 4.2637528901127528e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.2848483279238491e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0755647402230917e+01,
|
|
"cpu_time": 1.2455379122251538e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5836292529076226e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1307027683943895e-04,
|
|
"cpu_time": 2.9177954361125432e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1307279121009151e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1994,
|
|
"real_time": 3.5091578963804903e+05,
|
|
"cpu_time": 8.7652850601805793e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.3904903249444553e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1994,
|
|
"real_time": 3.5093890028338274e+05,
|
|
"cpu_time": 8.8224933299899776e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.3903329021736290e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1994,
|
|
"real_time": 3.5096906344072352e+05,
|
|
"cpu_time": 8.8196532998996566e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.3901274709976772e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1994,
|
|
"real_time": 3.5094931681286416e+05,
|
|
"cpu_time": 8.7911928686057520e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.3902619546836266e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1994,
|
|
"real_time": 3.5095378146306012e+05,
|
|
"cpu_time": 8.8155828786358540e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.3902315470229374e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5094537032761594e+05,
|
|
"cpu_time": 8.8028414874623646e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.3902888399644653e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5094931681286410e+05,
|
|
"cpu_time": 8.8155828786358540e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.3902619546836266e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9777286282482969e+01,
|
|
"cpu_time": 2.4379040960243474e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3470608042104114e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.6354315955273612e-05,
|
|
"cpu_time": 2.7694513180733567e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.6355565975467513e-05,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1024,
|
|
"real_time": 6.8427680560034793e+05,
|
|
"cpu_time": 1.7584346923828071e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4518171393052799e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1024,
|
|
"real_time": 6.8385319087838067e+05,
|
|
"cpu_time": 1.7629425781249958e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4533359241111927e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1024,
|
|
"real_time": 6.8385708868845541e+05,
|
|
"cpu_time": 1.7627120683593734e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4533219407254242e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1024,
|
|
"real_time": 6.8426568236645835e+05,
|
|
"cpu_time": 1.7520782431640625e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4518569953673294e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1024,
|
|
"real_time": 6.8402767544739624e+05,
|
|
"cpu_time": 1.7477040839843778e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4527101171786167e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8405608859620779e+05,
|
|
"cpu_time": 1.7567743332031232e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4526084233375687e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8402767544739624e+05,
|
|
"cpu_time": 1.7584346923828069e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4527101171786167e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0869912424074286e+02,
|
|
"cpu_time": 6.7153447296735594e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.4824469310192904e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.0509066101440125e-04,
|
|
"cpu_time": 3.8225426013764013e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.0508118865697265e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 518,
|
|
"real_time": 1.3490489445371430e+06,
|
|
"cpu_time": 3.1198733146718070e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 2.4872657241885677e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 518,
|
|
"real_time": 1.3496332791090931e+06,
|
|
"cpu_time": 3.4872047451737332e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 2.4861888425091019e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 518,
|
|
"real_time": 1.3491453654150763e+06,
|
|
"cpu_time": 3.1481712200772227e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 2.4870879639924263e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 518,
|
|
"real_time": 1.3491554117427140e+06,
|
|
"cpu_time": 3.4759415617760862e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 2.4870694441834164e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 518,
|
|
"real_time": 1.3492518674553477e+06,
|
|
"cpu_time": 3.4907580096525331e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 2.4868916478346439e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3492469736518748e+06,
|
|
"cpu_time": 3.3443897702702763e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 2.4869007245416317e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3491554117427138e+06,
|
|
"cpu_time": 3.4759415617760858e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 2.4870694441834164e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2758471068356818e+02,
|
|
"cpu_time": 1.9237659075319106e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.1940674816125692e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6867535382909690e-04,
|
|
"cpu_time": 5.7522180118869386e-02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6864635730023323e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 261,
|
|
"real_time": 2.6826534341749558e+06,
|
|
"cpu_time": 6.9875153448275356e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 2.5015853015184265e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 261,
|
|
"real_time": 2.6829316739306939e+06,
|
|
"cpu_time": 7.0195267088122601e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 2.5013258687158642e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 261,
|
|
"real_time": 2.6821377874491201e+06,
|
|
"cpu_time": 6.6247493754790155e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 2.5020662366427010e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 261,
|
|
"real_time": 2.6829359014572082e+06,
|
|
"cpu_time": 6.9878346436781352e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 2.5013219273539307e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 261,
|
|
"real_time": 2.6824712276929752e+06,
|
|
"cpu_time": 7.0353696704980666e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 2.5017552213492374e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6826260049409908e+06,
|
|
"cpu_time": 6.9309991486590030e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 2.5016109111160324e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6826534341749558e+06,
|
|
"cpu_time": 6.9878346436781362e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 2.5015853015184265e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3635325544601767e+02,
|
|
"cpu_time": 1.7244101327110577e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1367316044570981e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2538209009623627e-04,
|
|
"cpu_time": 2.4879676013878801e-02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2538846830731650e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 131,
|
|
"real_time": 5.3476227433171887e+06,
|
|
"cpu_time": 1.3999707854961684e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 2.5098578273444786e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 131,
|
|
"real_time": 5.3501096340820538e+06,
|
|
"cpu_time": 1.3503937290076287e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 2.5086911704572659e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 131,
|
|
"real_time": 5.3470122222681995e+06,
|
|
"cpu_time": 1.3384541061068531e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 2.5101444025326153e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 131,
|
|
"real_time": 5.3500597336763190e+06,
|
|
"cpu_time": 1.4046235312977055e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 2.5087145692067188e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 131,
|
|
"real_time": 5.3472480885968860e+06,
|
|
"cpu_time": 1.3952985511450442e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 2.5100336804312855e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.3484104843881298e+06,
|
|
"cpu_time": 1.3777481406106800e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 2.5094883299944729e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.3476227433171896e+06,
|
|
"cpu_time": 1.3952985511450443e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 2.5098578273444786e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5438563569808107e+03,
|
|
"cpu_time": 3.0888648651025502e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.2431616861367933e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.8865704333788265e-04,
|
|
"cpu_time": 2.2419662738455422e-02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.8863101691143332e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 66,
|
|
"real_time": 1.0676541506792560e+07,
|
|
"cpu_time": 2.7974617409090761e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 2.5142547877439312e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 66,
|
|
"real_time": 1.0677353634188572e+07,
|
|
"cpu_time": 2.6962332712120984e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 2.5140635516695595e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 66,
|
|
"real_time": 1.0677305445300810e+07,
|
|
"cpu_time": 2.7131394393939335e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 2.5140748981583282e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 66,
|
|
"real_time": 1.0676828114259424e+07,
|
|
"cpu_time": 2.7009475348484877e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 2.5141872953961990e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 66,
|
|
"real_time": 1.0677754173450397e+07,
|
|
"cpu_time": 2.7965025484848835e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 2.5139692452130882e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0677156574798353e+07,
|
|
"cpu_time": 2.7408569069696955e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 2.5141099556362213e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0677305445300812e+07,
|
|
"cpu_time": 2.7131394393939335e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 2.5140748981583282e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.7547712676981666e+02,
|
|
"cpu_time": 5.1606272644630505e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1195931786859010e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.4532186396151680e-05,
|
|
"cpu_time": 1.8828517648404580e-02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.4532387144641669e-05,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 33,
|
|
"real_time": 2.1335824533845440e+07,
|
|
"cpu_time": 5.6062930121211961e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 2.5162885603429626e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 33,
|
|
"real_time": 2.1336251248915989e+07,
|
|
"cpu_time": 5.4425615606060192e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 2.5162382357457302e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 33,
|
|
"real_time": 2.1335674449801445e+07,
|
|
"cpu_time": 5.6017643969697289e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 2.5163062609675140e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 33,
|
|
"real_time": 2.1335824816064402e+07,
|
|
"cpu_time": 5.6141968303029791e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 2.5162885270588337e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 33,
|
|
"real_time": 2.1335567996808976e+07,
|
|
"cpu_time": 5.4153427060606077e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 2.5163188159804150e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1335828609087251e+07,
|
|
"cpu_time": 5.5360317012121066e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 2.5162880800190914e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1335824533845440e+07,
|
|
"cpu_time": 5.6017643969697297e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 2.5162885603429626e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5997057369630966e+02,
|
|
"cpu_time": 9.8323099070837663e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.0659914572614193e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2184695446305951e-05,
|
|
"cpu_time": 1.7760573706489063e-02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2184580460430258e-05,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 16,
|
|
"real_time": 4.2654158081859350e+07,
|
|
"cpu_time": 1.1257115531249973e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 2.5173204027127621e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 16,
|
|
"real_time": 4.2658088263124228e+07,
|
|
"cpu_time": 1.1246058049999875e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 2.5170884765790028e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 16,
|
|
"real_time": 4.2654907796531916e+07,
|
|
"cpu_time": 1.1248270693750051e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 2.5172761575803974e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 16,
|
|
"real_time": 4.2654314544051886e+07,
|
|
"cpu_time": 1.1186942206250094e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 2.5173111688175812e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 16,
|
|
"real_time": 4.2655004421249032e+07,
|
|
"cpu_time": 1.1271927299999973e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 2.5172704552929420e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2655294621363282e+07,
|
|
"cpu_time": 1.1242062756249993e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 2.5172533321965370e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2654907796531916e+07,
|
|
"cpu_time": 1.1248270693750051e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 2.5172761575803974e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6039449389435651e+03,
|
|
"cpu_time": 3.2447083462921774e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.4650820142247051e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.7602481782888741e-05,
|
|
"cpu_time": 2.8862215205908638e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.7600832197391686e-05,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
}
|
|
]
|
|
}
|