15327 lines
535 KiB
JSON
15327 lines
535 KiB
JSON
{
|
|
"context": {
|
|
"date": "2023-10-09T16:05:43-04:00",
|
|
"host_name": "frontier10341",
|
|
"executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope",
|
|
"num_cpus": 128,
|
|
"mhz_per_cpu": 1827,
|
|
"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": [0.38,0.35,3.05],
|
|
"library_build_type": "release"
|
|
},
|
|
"benchmarks": [
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 101391,
|
|
"real_time": 6.8832135380240470e+03,
|
|
"cpu_time": 7.2836224173479617e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7191930569320895e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 101391,
|
|
"real_time": 6.8831901123572707e+03,
|
|
"cpu_time": 7.2776635869976056e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7192057145190232e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 101391,
|
|
"real_time": 6.8815539094073747e+03,
|
|
"cpu_time": 7.2929069942424619e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7200900170241661e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 101391,
|
|
"real_time": 6.8839979062861476e+03,
|
|
"cpu_time": 7.2996353076306963e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7187692890817516e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 101391,
|
|
"real_time": 6.8807073345413510e+03,
|
|
"cpu_time": 7.2980100856532168e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7205477221051469e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8825325601232398e+03,
|
|
"cpu_time": 7.2903676783743882e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7195611599324360e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8831901123572707e+03,
|
|
"cpu_time": 7.2929069942424631e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7192057145190232e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3539341607010642e+00,
|
|
"cpu_time": 9.4556325833544414e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.3176079382199750e+03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9672034223937193e-04,
|
|
"cpu_time": 1.2970035258170771e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9673309897538815e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 101632,
|
|
"real_time": 6.8833271613560100e+03,
|
|
"cpu_time": 7.2831583630834901e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4382633281538859e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 101632,
|
|
"real_time": 6.8800538991193253e+03,
|
|
"cpu_time": 7.2689925844951665e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4418021647408620e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 101632,
|
|
"real_time": 6.8776049510008643e+03,
|
|
"cpu_time": 7.2875970072851241e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4444520097870871e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 101632,
|
|
"real_time": 6.8815903366087359e+03,
|
|
"cpu_time": 7.2941785835706663e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4401406499927580e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 101632,
|
|
"real_time": 6.8833082273653345e+03,
|
|
"cpu_time": 7.3055121075360525e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4382837886655822e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8811769150900536e+03,
|
|
"cpu_time": 7.2878877291941011e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4405883882680342e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8815903366087350e+03,
|
|
"cpu_time": 7.2875970072851253e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4401406499927580e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4164571536953048e+00,
|
|
"cpu_time": 1.3506215948210873e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6133492925745690e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.5116916532056358e-04,
|
|
"cpu_time": 1.8532414946661642e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5122884860761466e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 100747,
|
|
"real_time": 6.9217955489801152e+03,
|
|
"cpu_time": 7.2625362894242978e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4793849265757066e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 100747,
|
|
"real_time": 6.9471354139926707e+03,
|
|
"cpu_time": 7.2736093697190452e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4739888298959827e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 100747,
|
|
"real_time": 6.9219968121989623e+03,
|
|
"cpu_time": 7.2513956090677404e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4793419121421096e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 100747,
|
|
"real_time": 6.9462243846049150e+03,
|
|
"cpu_time": 7.2690350772906268e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4741821503340954e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 100747,
|
|
"real_time": 6.9246305006402572e+03,
|
|
"cpu_time": 7.2694550711015558e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4787792646919137e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9323565320833850e+03,
|
|
"cpu_time": 7.2652062833206534e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4771354167279616e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9246305006402572e+03,
|
|
"cpu_time": 7.2690350772906268e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4787792646919137e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3127094176459869e+01,
|
|
"cpu_time": 8.6773744969125628e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7952613376033376e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8935976699563626e-03,
|
|
"cpu_time": 1.1943741386715948e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8923527971424507e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 99696,
|
|
"real_time": 7.0134417029681144e+03,
|
|
"cpu_time": 7.2563256637428439e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9201069699250209e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 99696,
|
|
"real_time": 7.0167152805706191e+03,
|
|
"cpu_time": 7.2431951106793475e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9187446235291034e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 99696,
|
|
"real_time": 7.0146015609048018e+03,
|
|
"cpu_time": 7.2556992092616868e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9196241329148734e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 99696,
|
|
"real_time": 7.0156757197062589e+03,
|
|
"cpu_time": 7.2345733066593945e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9191771139697832e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 99696,
|
|
"real_time": 7.0102073833546483e+03,
|
|
"cpu_time": 7.2424339343158472e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9214542281058091e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0141283295008889e+03,
|
|
"cpu_time": 7.2464454449318242e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9198214136889184e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0146015609048000e+03,
|
|
"cpu_time": 7.2431951106793475e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9196241329148734e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5077856203292166e+00,
|
|
"cpu_time": 9.3655677273045364e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0441676302203588e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.5753346709977655e-04,
|
|
"cpu_time": 1.2924361051879349e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5761352571942124e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 96318,
|
|
"real_time": 7.2715229882754275e+03,
|
|
"cpu_time": 7.1284072239173138e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6329327523331392e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 96318,
|
|
"real_time": 7.2637745524600678e+03,
|
|
"cpu_time": 7.1014859489183414e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6389415315936291e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 96318,
|
|
"real_time": 7.2708004800537037e+03,
|
|
"cpu_time": 7.1409159461823308e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6334925036613667e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 96318,
|
|
"real_time": 7.2724851450131036e+03,
|
|
"cpu_time": 7.1015756229565668e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6321875099445391e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 96318,
|
|
"real_time": 7.2729789290610734e+03,
|
|
"cpu_time": 7.1440497530179262e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6318051240783477e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.2703124189726750e+03,
|
|
"cpu_time": 7.1232868989984965e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6338718843222046e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.2715229882754284e+03,
|
|
"cpu_time": 7.1284072239173138e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6329327523331392e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7509463699712158e+00,
|
|
"cpu_time": 2.0704818706262945e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9084289986176387e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.1592643531833801e-04,
|
|
"cpu_time": 2.9066383257950702e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.1623981842951398e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 76950,
|
|
"real_time": 9.1016529673389869e+03,
|
|
"cpu_time": 6.4125409891139269e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.0005628970877600e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 76950,
|
|
"real_time": 9.0394521733708898e+03,
|
|
"cpu_time": 6.3790720360982823e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.0624960925537276e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 76950,
|
|
"real_time": 9.0727790614569039e+03,
|
|
"cpu_time": 6.4149996028094625e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.0292069767259729e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 76950,
|
|
"real_time": 9.0905429992215850e+03,
|
|
"cpu_time": 6.3946022714386475e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.0115628964094603e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 76950,
|
|
"real_time": 9.1010498345801061e+03,
|
|
"cpu_time": 6.4221199957488120e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.0011593705089879e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.0810954071936958e+03,
|
|
"cpu_time": 6.4046669790418267e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.0209976466571808e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.0905429992215832e+03,
|
|
"cpu_time": 6.4125409891139269e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.0115628964094603e+08,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6045009004561770e+01,
|
|
"cpu_time": 1.7534041063583076e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5930989254712206e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.8680470622442652e-03,
|
|
"cpu_time": 2.7376975447685600e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.8745145792517957e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 71214,
|
|
"real_time": 9.8288677364313953e+03,
|
|
"cpu_time": 6.1999552916587448e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6669264903495996e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 71214,
|
|
"real_time": 9.8260664630183692e+03,
|
|
"cpu_time": 6.1690916029448724e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6674017076582203e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 71214,
|
|
"real_time": 9.8312995963441426e+03,
|
|
"cpu_time": 6.2225834676590300e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6665141611687369e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 71214,
|
|
"real_time": 9.8282672740485523e+03,
|
|
"cpu_time": 6.1799529594162047e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6670283319687283e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 71214,
|
|
"real_time": 9.8287645411260382e+03,
|
|
"cpu_time": 6.1849234735099447e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6669439919376640e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.8286531221937003e+03,
|
|
"cpu_time": 6.1913013590377593e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6669629366165900e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.8287645411260364e+03,
|
|
"cpu_time": 6.1849234735099435e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6669439919376640e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8642681808305432e+00,
|
|
"cpu_time": 2.0713508473029789e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1618159744045825e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8967687206509627e-04,
|
|
"cpu_time": 3.3455823375150716e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8967524141970868e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 65436,
|
|
"real_time": 1.0702317257733945e+04,
|
|
"cpu_time": 6.0085641069241750e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.0617668315074916e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 65436,
|
|
"real_time": 1.0697540117576669e+04,
|
|
"cpu_time": 5.9920447637640011e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.0631341074533858e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 65436,
|
|
"real_time": 1.0695315493146962e+04,
|
|
"cpu_time": 6.0088095880269742e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.0637712390060997e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 65436,
|
|
"real_time": 1.0696274031581914e+04,
|
|
"cpu_time": 5.9857337200108433e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.0634966814844975e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 65436,
|
|
"real_time": 1.0701570458494207e+04,
|
|
"cpu_time": 6.0015443958464932e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.0619804940863523e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0698603471706740e+04,
|
|
"cpu_time": 5.9993393149144971e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.0628298707075653e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0697540117576671e+04,
|
|
"cpu_time": 6.0015443958464932e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.0631341074533858e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1608160965725620e+00,
|
|
"cpu_time": 1.0222132012184173e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.0482924939460261e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.9544193360671588e-04,
|
|
"cpu_time": 1.7038762896393798e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9542262795862438e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 56001,
|
|
"real_time": 1.2474600264486029e+04,
|
|
"cpu_time": 5.7444350403202879e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.2535551128299170e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 56001,
|
|
"real_time": 1.2479256499438323e+04,
|
|
"cpu_time": 5.7487078433301461e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.2515949169687872e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 56001,
|
|
"real_time": 1.2463638488247168e+04,
|
|
"cpu_time": 5.7350523399519670e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.2581756171601467e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 56001,
|
|
"real_time": 1.2461678599760717e+04,
|
|
"cpu_time": 5.7427741474294806e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.2590025874410210e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 56001,
|
|
"real_time": 1.2480424912036558e+04,
|
|
"cpu_time": 5.7535397741248631e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.2511032646648741e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2471919752793758e+04,
|
|
"cpu_time": 5.7449018290313482e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.2546862998129501e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2474600264486025e+04,
|
|
"cpu_time": 5.7444350403202879e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.2535551128299170e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.7580147004177213e+00,
|
|
"cpu_time": 6.9075537654859177e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.6905300142716630e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.0221865390497661e-04,
|
|
"cpu_time": 1.2023797744600633e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.0233117710624020e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 44954,
|
|
"real_time": 1.5570322001624018e+04,
|
|
"cpu_time": 5.4155702783158743e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4180661123340225e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 44954,
|
|
"real_time": 1.5567311346595274e+04,
|
|
"cpu_time": 5.3988865154220462e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4196941322604656e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 44954,
|
|
"real_time": 1.5574241228749710e+04,
|
|
"cpu_time": 5.4524614054208958e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4159477225795088e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 44954,
|
|
"real_time": 1.5566706794679476e+04,
|
|
"cpu_time": 5.4032071859205699e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4200211212816648e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 44954,
|
|
"real_time": 1.5570461523521220e+04,
|
|
"cpu_time": 5.4167785188102865e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4179906807514076e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5569808579033943e+04,
|
|
"cpu_time": 5.4173807807779348e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4183439538414145e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5570322001624018e+04,
|
|
"cpu_time": 5.4155702783158755e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4180661123340225e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.0081372411736669e+00,
|
|
"cpu_time": 2.1078540664433767e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6263297952383459e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9320322571109687e-04,
|
|
"cpu_time": 3.8909099281381671e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9318880342210628e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 35495,
|
|
"real_time": 1.9723436549717331e+04,
|
|
"cpu_time": 5.3878092180714750e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3290990104042339e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 35495,
|
|
"real_time": 1.9720745215069102e+04,
|
|
"cpu_time": 5.3728590838341129e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3292803955485891e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 35495,
|
|
"real_time": 1.9725746003490709e+04,
|
|
"cpu_time": 5.3804046056245291e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3289434019560549e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 35495,
|
|
"real_time": 1.9719958163029311e+04,
|
|
"cpu_time": 5.3721997128953385e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3293334490509405e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 35495,
|
|
"real_time": 1.9721247229598866e+04,
|
|
"cpu_time": 5.3825436642507362e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3292465580298494e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9722226632181068e+04,
|
|
"cpu_time": 5.3791632569352400e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3291805629979336e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9721247229598870e+04,
|
|
"cpu_time": 5.3804046056245291e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3292465580298494e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3531709151382918e+00,
|
|
"cpu_time": 6.6323881111275242e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5858137817537088e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1931568169379953e-04,
|
|
"cpu_time": 1.2329776573664181e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1930762651064844e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 22745,
|
|
"real_time": 3.0776497849898456e+04,
|
|
"cpu_time": 5.1923099526234138e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7035336592130474e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 22745,
|
|
"real_time": 3.0770789758224681e+04,
|
|
"cpu_time": 5.1855789041886455e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7038496708062679e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 22745,
|
|
"real_time": 3.0771133711646755e+04,
|
|
"cpu_time": 5.1980252169316763e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7038306255240736e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 22745,
|
|
"real_time": 3.0774971735143052e+04,
|
|
"cpu_time": 5.1919315263507980e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7036181365563906e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 22745,
|
|
"real_time": 3.0774256068737206e+04,
|
|
"cpu_time": 5.1910043651659685e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7036577548095827e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.0773529824730031e+04,
|
|
"cpu_time": 5.1917699930521011e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7036979693818726e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.0774256068737206e+04,
|
|
"cpu_time": 5.1919315263507986e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7036577548095827e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4831950638948475e+00,
|
|
"cpu_time": 4.4260942850547965e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3747663008962651e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.0692565267547484e-05,
|
|
"cpu_time": 8.5252125787121306e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.0693076214386237e-05,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 13294,
|
|
"real_time": 5.2697902649681113e+04,
|
|
"cpu_time": 5.2249273284225696e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.9897869692663097e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 13294,
|
|
"real_time": 5.2695674484535877e+04,
|
|
"cpu_time": 5.2312670451031107e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.9898711047103420e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 13294,
|
|
"real_time": 5.2714376618794471e+04,
|
|
"cpu_time": 5.2262932515811300e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.9891651334185501e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 13294,
|
|
"real_time": 5.2694081461670190e+04,
|
|
"cpu_time": 5.2271069954378331e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.9899312615644260e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 13294,
|
|
"real_time": 5.2692677383010727e+04,
|
|
"cpu_time": 5.2321397542297572e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.9899842863898273e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.2698942519538476e+04,
|
|
"cpu_time": 5.2283468749548799e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.9897477510698910e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.2695674484535877e+04,
|
|
"cpu_time": 5.2271069954378337e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 1.9898711047103420e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.8438484129024015e+00,
|
|
"cpu_time": 3.1765270171865739e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.3384973397742882e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6781832784639817e-04,
|
|
"cpu_time": 6.0755858269522086e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6778495354394406e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7654,
|
|
"real_time": 9.1562330388719362e+04,
|
|
"cpu_time": 5.0718666077462524e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2904091574523453e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7654,
|
|
"real_time": 9.1422632890395413e+04,
|
|
"cpu_time": 5.0712631572040826e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2939089957234436e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7654,
|
|
"real_time": 9.1577222157901051e+04,
|
|
"cpu_time": 5.0736712377176052e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2900367040878441e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7654,
|
|
"real_time": 9.1496049839896659e+04,
|
|
"cpu_time": 5.0725626994054937e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2920683501306099e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7654,
|
|
"real_time": 9.1446112857155429e+04,
|
|
"cpu_time": 5.0698742236060029e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2933200050568394e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1500869626813583e+04,
|
|
"cpu_time": 5.0718475851358873e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2919486424902161e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1496049839896645e+04,
|
|
"cpu_time": 5.0718666077462518e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2920683501306099e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8464690435449569e+01,
|
|
"cpu_time": 1.4206662846229406e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7148959627844483e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.4824087153141708e-04,
|
|
"cpu_time": 2.8010823684578007e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.4822617356783507e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4017,
|
|
"real_time": 1.7425761721233142e+05,
|
|
"cpu_time": 4.9151879122554475e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4069559007508270e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4017,
|
|
"real_time": 1.7444207502565192e+05,
|
|
"cpu_time": 4.9140932142121673e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4044107474548340e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4017,
|
|
"real_time": 1.7432378120987589e+05,
|
|
"cpu_time": 4.9106986333258677e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4060423488349518e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4017,
|
|
"real_time": 1.7442885363438091e+05,
|
|
"cpu_time": 4.9179079442742801e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4045929974358780e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4017,
|
|
"real_time": 1.7429346882058054e+05,
|
|
"cpu_time": 4.9139473601418501e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4064607976318714e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7434915918056414e+05,
|
|
"cpu_time": 4.9143670128419226e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4056925584216724e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7432378120987589e+05,
|
|
"cpu_time": 4.9140932142121679e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4060423488349518e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2325496931373124e+01,
|
|
"cpu_time": 2.5942317566836681e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1358534651631786e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.7218751910419588e-04,
|
|
"cpu_time": 5.2788726399647832e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.7215237923352505e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2061,
|
|
"real_time": 3.3944288959156419e+05,
|
|
"cpu_time": 4.8628747069918007e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4712869991454590e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2061,
|
|
"real_time": 3.3943976909891475e+05,
|
|
"cpu_time": 4.8681254121008945e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4713097178532166e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2061,
|
|
"real_time": 3.3959667002357455e+05,
|
|
"cpu_time": 4.8615527415477914e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4701679199085396e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2061,
|
|
"real_time": 3.3943201619716006e+05,
|
|
"cpu_time": 4.8622901869529355e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4713661645657650e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2061,
|
|
"real_time": 3.3948861067836283e+05,
|
|
"cpu_time": 4.8714117211790258e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4709541752337334e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3947999111791526e+05,
|
|
"cpu_time": 4.8652509537544888e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4710169953413429e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3944288959156419e+05,
|
|
"cpu_time": 4.8628747069918013e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4712869991454590e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8891565127864197e+01,
|
|
"cpu_time": 4.3100762258374848e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.0134397446862767e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0293262321883162e-04,
|
|
"cpu_time": 8.8588980646752053e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0288973139959027e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1042,
|
|
"real_time": 6.7178899960345705e+05,
|
|
"cpu_time": 4.8210122745105827e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4973936771669735e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1042,
|
|
"real_time": 6.7177456438285788e+05,
|
|
"cpu_time": 4.8289281563147855e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4974473416409863e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1042,
|
|
"real_time": 6.7149712427049130e+05,
|
|
"cpu_time": 4.8237056245585269e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4984792032023403e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1042,
|
|
"real_time": 6.7183692567988927e+05,
|
|
"cpu_time": 4.8208986555662155e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4972155233982857e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1042,
|
|
"real_time": 6.7160889024143736e+05,
|
|
"cpu_time": 4.8249874631189430e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4980634181255016e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.7170130083562655e+05,
|
|
"cpu_time": 4.8239064348138106e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4977198327068176e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.7177456438285788e+05,
|
|
"cpu_time": 4.8237056245585263e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.4974473416409863e+10,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4289900870814608e+02,
|
|
"cpu_time": 3.3109490197259450e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.3142524920669701e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.1274189662931023e-04,
|
|
"cpu_time": 6.8636261181001502e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1276415482948032e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 70,
|
|
"real_time": 9.7514665419501923e+06,
|
|
"cpu_time": 6.9276807379999685e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4409626342510595e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 70,
|
|
"real_time": 9.7755970300308298e+06,
|
|
"cpu_time": 6.9318396921427894e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4324688197477980e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 70,
|
|
"real_time": 9.7856407585952971e+06,
|
|
"cpu_time": 6.9412097684285903e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4289458225336132e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 70,
|
|
"real_time": 9.7389092269752715e+06,
|
|
"cpu_time": 6.9264137014285922e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4453993992529902e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 70,
|
|
"real_time": 9.7605185583233833e+06,
|
|
"cpu_time": 6.9299511062856972e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4377714462092910e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.7624264231749941e+06,
|
|
"cpu_time": 6.9314190012571263e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4371096243989506e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.7605185583233833e+06,
|
|
"cpu_time": 6.9299511062856972e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4377714462092910e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8635904296332563e+04,
|
|
"cpu_time": 5.8570141043280135e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.5610362912210748e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9089418438117849e-03,
|
|
"cpu_time": 8.4499495749221730e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9088818828024454e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.9139317392900184e+07,
|
|
"cpu_time": 7.4161301632431936e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5063352899353838e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.9085670343121968e+07,
|
|
"cpu_time": 7.4130449381080580e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5161910896247077e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.9182415074996047e+07,
|
|
"cpu_time": 7.4208961432432485e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.4984575058786664e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.9188439574193310e+07,
|
|
"cpu_time": 7.4277342489189029e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.4973591125281110e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.9252271164913435e+07,
|
|
"cpu_time": 7.4416517748648024e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.4857634938314948e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9169622710024990e+07,
|
|
"cpu_time": 7.4238914536756420e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5008212983596725e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9182415074996047e+07,
|
|
"cpu_time": 7.4208961432432485e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.4984575058786664e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.1876688317856038e+04,
|
|
"cpu_time": 1.1369394197703972e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1302060757649465e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.2278511295632788e-03,
|
|
"cpu_time": 1.5314601875105369e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.2284026502424163e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7719552454195522e+07,
|
|
"cpu_time": 7.9518064710525715e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5583064821086187e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7875746229761526e+07,
|
|
"cpu_time": 7.9823271278948033e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5436325712451863e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7632276353083156e+07,
|
|
"cpu_time": 7.9496637047368777e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5665588427526450e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7486869253610313e+07,
|
|
"cpu_time": 7.9333624152631378e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5803930995670881e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7652436251703061e+07,
|
|
"cpu_time": 7.9477399468420112e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5646492328615041e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7673376108470716e+07,
|
|
"cpu_time": 7.9529799331578803e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5627080457070084e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7652436251703061e+07,
|
|
"cpu_time": 7.9496637047368777e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5646492328615041e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4141993786207883e+05,
|
|
"cpu_time": 1.7929979699535605e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3366052741479063e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.7538429647212086e-03,
|
|
"cpu_time": 2.2544982950077896e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.7516553615963254e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.3759940266609192e+07,
|
|
"cpu_time": 9.1436034199998915e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6393122747893491e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.4049559980630875e+07,
|
|
"cpu_time": 9.1490360210001719e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6250783403738604e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.3754198104143143e+07,
|
|
"cpu_time": 9.1294294910001397e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6395956148958607e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.4159944802522659e+07,
|
|
"cpu_time": 9.1635879339998782e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6196825215391579e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.4050534516572952e+07,
|
|
"cpu_time": 9.1483502580001068e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6250306328298349e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.3954835534095764e+07,
|
|
"cpu_time": 9.1468014248000371e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6297398768856125e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4049559980630875e+07,
|
|
"cpu_time": 9.1483502580001068e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6250783403738604e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8603788028236612e+05,
|
|
"cpu_time": 1.2259593327879980e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.1354210650894474e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.5155607329637847e-03,
|
|
"cpu_time": 1.3403148006077975e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5168252753494328e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4423488080501556e+08,
|
|
"cpu_time": 1.0929790742000136e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7221988814603786e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4405136108398438e+08,
|
|
"cpu_time": 1.0915879555999992e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7269409185727525e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4433465003967285e+08,
|
|
"cpu_time": 1.0923560807999933e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7196259654381800e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4400748908519745e+08,
|
|
"cpu_time": 1.0920196993999979e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7280763341577144e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4442668557167053e+08,
|
|
"cpu_time": 1.0933128597999997e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7172556434079652e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4421101331710818e+08,
|
|
"cpu_time": 1.0924511339600010e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7228195486073985e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4423488080501556e+08,
|
|
"cpu_time": 1.0923560807999933e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7221988814603786e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7977845118309054e+05,
|
|
"cpu_time": 7.0027496083199920e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6411401330276597e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2466346851593955e-03,
|
|
"cpu_time": 6.4101261746471775e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2466734077303798e-03,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8553931415081024e+08,
|
|
"cpu_time": 1.3159162894999809e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7603992542788463e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8564266860485077e+08,
|
|
"cpu_time": 1.3177685255000000e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7590386241818137e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8561739623546600e+08,
|
|
"cpu_time": 1.3165796574999859e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7593712363192186e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8537715971469879e+08,
|
|
"cpu_time": 1.3160098759999955e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7625359544311676e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8524251282215118e+08,
|
|
"cpu_time": 1.3169921995000067e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7643120353152914e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8548381030559540e+08,
|
|
"cpu_time": 1.3166533095999939e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7611314209052677e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8553931415081024e+08,
|
|
"cpu_time": 1.3165796574999859e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7603992542788463e+09,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7011809196020965e+05,
|
|
"cpu_time": 7.6177144269919046e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2418261704601450e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.9589400806338961e-04,
|
|
"cpu_time": 5.7856645872148428e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.9605100688573103e-04,
|
|
"hip_id": NaN,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 101143,
|
|
"real_time": 6.9204708956761615e+03,
|
|
"cpu_time": 7.2349408145925009e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.6991702423016645e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 101143,
|
|
"real_time": 6.9217920008940910e+03,
|
|
"cpu_time": 7.2129039708989239e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.6984642122579291e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 101143,
|
|
"real_time": 6.9186296516765087e+03,
|
|
"cpu_time": 7.2209746831171882e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7001546966452613e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 101143,
|
|
"real_time": 6.9147871474079548e+03,
|
|
"cpu_time": 7.1969517861334836e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7022108496277139e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 101143,
|
|
"real_time": 6.9160527396295383e+03,
|
|
"cpu_time": 7.2123636443281245e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7015333693611011e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9183464870568514e+03,
|
|
"cpu_time": 7.2156269798140442e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7003066740387343e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9186296516765096e+03,
|
|
"cpu_time": 7.2129039708989239e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7001546966452613e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.9323640537742919e+00,
|
|
"cpu_time": 1.3860457600646871e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5684311986823011e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.2385330935076589e-04,
|
|
"cpu_time": 1.9208944197672583e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.2386519195459609e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 100881,
|
|
"real_time": 6.9579157771106211e+03,
|
|
"cpu_time": 7.2127755140597880e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3585254033157572e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 100881,
|
|
"real_time": 6.9595788241537539e+03,
|
|
"cpu_time": 7.2116004817977941e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3567670247955889e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 100881,
|
|
"real_time": 6.9648597435455631e+03,
|
|
"cpu_time": 7.2157170089635587e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3511889521462053e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 100881,
|
|
"real_time": 6.9603731612709043e+03,
|
|
"cpu_time": 7.2036792129909480e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3559274501097754e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 100881,
|
|
"real_time": 6.9657687705762382e+03,
|
|
"cpu_time": 7.2105027768201864e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3502296281024158e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9616992553314176e+03,
|
|
"cpu_time": 7.2108549989264536e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3545276916939482e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9603731612709043e+03,
|
|
"cpu_time": 7.2116004817977941e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3559274501097754e+07,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4321677299940951e+00,
|
|
"cpu_time": 4.4581170076738374e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.6254712469553524e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.9300718174024379e-04,
|
|
"cpu_time": 6.1825081884707964e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9295772603451954e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 100549,
|
|
"real_time": 6.9635280794324835e+03,
|
|
"cpu_time": 7.1909079957484055e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4705189500484565e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 100549,
|
|
"real_time": 6.9648359984010194e+03,
|
|
"cpu_time": 7.1901996859192717e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4702428028959891e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 100549,
|
|
"real_time": 6.9632338986138257e+03,
|
|
"cpu_time": 7.1937162898894727e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4705810761345360e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 100549,
|
|
"real_time": 6.9618127468403418e+03,
|
|
"cpu_time": 7.1920003510952306e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4708812736521077e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 100549,
|
|
"real_time": 6.9627264921763935e+03,
|
|
"cpu_time": 7.1921577666730428e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4706882442540410e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9632274430928137e+03,
|
|
"cpu_time": 7.1917964178650856e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4705824693970260e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9632338986138257e+03,
|
|
"cpu_time": 7.1920003510952306e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4705810761345360e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1102017069198984e+00,
|
|
"cpu_time": 1.3413535166889665e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3445803782340241e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5943780610256600e-04,
|
|
"cpu_time": 1.8651160833153184e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5943209082285321e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 99359,
|
|
"real_time": 7.0477241125739592e+03,
|
|
"cpu_time": 7.1408127620180762e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9059026251412565e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 99359,
|
|
"real_time": 7.0430845847710807e+03,
|
|
"cpu_time": 7.1573318022696364e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9078168455172193e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 99359,
|
|
"real_time": 7.0723090647129857e+03,
|
|
"cpu_time": 7.1684422714935648e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.8958010478054720e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 99359,
|
|
"real_time": 7.0474199799723110e+03,
|
|
"cpu_time": 7.1567093057424068e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9060280298607188e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 99359,
|
|
"real_time": 7.0489822899090241e+03,
|
|
"cpu_time": 7.1539411437604582e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9053839487323099e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0519040063878729e+03,
|
|
"cpu_time": 7.1554474570568275e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9041864994113952e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0477241125739583e+03,
|
|
"cpu_time": 7.1567093057424068e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9059026251412565e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1621779503221726e+01,
|
|
"cpu_time": 9.8799228230862343e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.7766467914217815e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6480342745298707e-03,
|
|
"cpu_time": 1.3807554150009839e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6447451953894444e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 96740,
|
|
"real_time": 7.2352855489432613e+03,
|
|
"cpu_time": 7.0695011105305302e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6611449158329833e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 96740,
|
|
"real_time": 7.2338364171991661e+03,
|
|
"cpu_time": 7.0740564616214275e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6622789952249300e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 96740,
|
|
"real_time": 7.2360271317770666e+03,
|
|
"cpu_time": 7.0633670219350517e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6605647344968975e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 96740,
|
|
"real_time": 7.2339251332365347e+03,
|
|
"cpu_time": 7.0782466543204284e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6622095536775422e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 96740,
|
|
"real_time": 7.2368922468838646e+03,
|
|
"cpu_time": 7.0693795467868602e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6598880572854984e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.2351932956079791e+03,
|
|
"cpu_time": 7.0709101590388608e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6612172513035703e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.2352855489432604e+03,
|
|
"cpu_time": 7.0695011105305302e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6611449158329833e+08,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3266106810598171e+00,
|
|
"cpu_time": 5.5890627986559877e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0379946396778742e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8335525076643318e-04,
|
|
"cpu_time": 7.9043046410530286e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8335184706767121e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 92134,
|
|
"real_time": 7.5976074588125648e+03,
|
|
"cpu_time": 6.9129500254365420e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0782341736408074e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 92134,
|
|
"real_time": 7.5959314546184842e+03,
|
|
"cpu_time": 6.9171474555192578e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0784720806056108e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 92134,
|
|
"real_time": 7.5969608578464395e+03,
|
|
"cpu_time": 6.9061844238138604e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0783259455047712e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 92134,
|
|
"real_time": 7.5947105828994154e+03,
|
|
"cpu_time": 6.9179616029489720e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0786454481156225e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 92134,
|
|
"real_time": 7.5973310366665191e+03,
|
|
"cpu_time": 6.9130378782754207e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0782734042341275e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5965082781686842e+03,
|
|
"cpu_time": 6.9134562771988106e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0783902104201877e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5969608578464376e+03,
|
|
"cpu_time": 6.9130378782754207e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.0783259455047712e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1889482467075330e+00,
|
|
"cpu_time": 4.6699560405403021e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6879732794093632e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5651246640833736e-04,
|
|
"cpu_time": 6.7548789683421161e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5652713304506496e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 86986,
|
|
"real_time": 8.0532485548788754e+03,
|
|
"cpu_time": 6.7471322680106556e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0344585030936530e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 86986,
|
|
"real_time": 8.0504065639193004e+03,
|
|
"cpu_time": 6.7466574551020646e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0351767168419294e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 86986,
|
|
"real_time": 8.0511955328298600e+03,
|
|
"cpu_time": 6.7444104524817586e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0349772817207062e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 86986,
|
|
"real_time": 8.0485394708604363e+03,
|
|
"cpu_time": 6.7525110358579588e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0356488353344007e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 86986,
|
|
"real_time": 8.0520311040854103e+03,
|
|
"cpu_time": 6.7470448781629133e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0347661090985036e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0510842453147761e+03,
|
|
"cpu_time": 6.7475512179230702e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0350054892178383e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0511955328298591e+03,
|
|
"cpu_time": 6.7470448781629133e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.0349772817207062e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7697663271882376e+00,
|
|
"cpu_time": 2.9872609689393186e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.4735485118639318e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.1981714179902294e-04,
|
|
"cpu_time": 4.4271779086381093e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1982980073352807e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 78440,
|
|
"real_time": 8.9262184853402341e+03,
|
|
"cpu_time": 6.4698229714859200e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.6709834129442110e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 78440,
|
|
"real_time": 8.9236577547035304e+03,
|
|
"cpu_time": 6.4745786115680456e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.6720368374424109e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 78440,
|
|
"real_time": 8.9270560275090738e+03,
|
|
"cpu_time": 6.4794260990862906e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.6706389989067078e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 78440,
|
|
"real_time": 8.9218562173252831e+03,
|
|
"cpu_time": 6.4771163043698585e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.6727783100077400e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 78440,
|
|
"real_time": 8.9262380279974896e+03,
|
|
"cpu_time": 6.4796465745453203e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.6709753758774872e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.9250053025751222e+03,
|
|
"cpu_time": 6.4761181122110868e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.6714825870357118e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.9262184853402359e+03,
|
|
"cpu_time": 6.4771163043698585e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.6709834129442110e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1754034277287158e+00,
|
|
"cpu_time": 4.0739437325036136e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.9500982859407749e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4374253616421369e-04,
|
|
"cpu_time": 6.2907187020291714e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4377340961780024e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 65526,
|
|
"real_time": 1.0648693458428526e+04,
|
|
"cpu_time": 6.0661068656453264e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.1543700413432150e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 65526,
|
|
"real_time": 1.0682980736342734e+04,
|
|
"cpu_time": 6.0785613472948444e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.1346174459578714e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 65526,
|
|
"real_time": 1.0662438233960220e+04,
|
|
"cpu_time": 6.0785739423940539e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.1464365431225348e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 65526,
|
|
"real_time": 1.0694320407530939e+04,
|
|
"cpu_time": 6.0839051338497341e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.1281126338658752e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 65526,
|
|
"real_time": 1.0675037563208047e+04,
|
|
"cpu_time": 6.0755753628460991e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.1391821445080910e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0672694079894094e+04,
|
|
"cpu_time": 6.0765445304060102e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.1405437617595177e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0675037563208047e+04,
|
|
"cpu_time": 6.0785613472948432e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.1391821445080910e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7749770903408109e+01,
|
|
"cpu_time": 6.5621771901057579e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0215173035409631e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6631012535856589e-03,
|
|
"cpu_time": 1.0799192135052616e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6635616374929255e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 50885,
|
|
"real_time": 1.3758592221073284e+04,
|
|
"cpu_time": 5.6743157690755975e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 9.5265560526784267e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 50885,
|
|
"real_time": 1.3760460135893978e+04,
|
|
"cpu_time": 5.6732773721072388e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 9.5252628695242844e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 50885,
|
|
"real_time": 1.3760500950284200e+04,
|
|
"cpu_time": 5.6708452705277777e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 9.5252346170793247e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 50885,
|
|
"real_time": 1.3758101779116194e+04,
|
|
"cpu_time": 5.6744688192839468e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 9.5268956506018753e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 50885,
|
|
"real_time": 1.3760365747913616e+04,
|
|
"cpu_time": 5.6705912533134544e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 9.5253282072007065e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3759604166856254e+04,
|
|
"cpu_time": 5.6726996968616033e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 9.5258554794169235e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3760365747913616e+04,
|
|
"cpu_time": 5.6732773721072376e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 9.5253282072007065e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1616918772492240e+00,
|
|
"cpu_time": 1.8681289736388411e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.0427476140930841e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.4427710504018314e-05,
|
|
"cpu_time": 3.2931920839602625e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.4430712091649118e-05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 38644,
|
|
"real_time": 1.8115547367903488e+04,
|
|
"cpu_time": 5.4174193946335387e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4470664047637768e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 38644,
|
|
"real_time": 1.8115290019924265e+04,
|
|
"cpu_time": 5.4198609333447289e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4470869619624004e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 38644,
|
|
"real_time": 1.8115484536232561e+04,
|
|
"cpu_time": 5.4167879070473015e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4470714237628532e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 38644,
|
|
"real_time": 1.8116374558653708e+04,
|
|
"cpu_time": 5.4184240811151934e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4470003319443447e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 38644,
|
|
"real_time": 1.8118635430779275e+04,
|
|
"cpu_time": 5.4186621481825197e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4468197729431620e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8116266382698661e+04,
|
|
"cpu_time": 5.4182308928646576e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4470089790753077e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8115547367903488e+04,
|
|
"cpu_time": 5.4184240811151946e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4470664047637768e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3879518141771583e+00,
|
|
"cpu_time": 1.1854400258260302e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1085137546823674e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.6613568428353181e-05,
|
|
"cpu_time": 2.1878728486583187e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.6607247827221412e-05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 23663,
|
|
"real_time": 2.9583180605874262e+04,
|
|
"cpu_time": 5.2134538219596314e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7722502762123333e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 23663,
|
|
"real_time": 2.9582511895465115e+04,
|
|
"cpu_time": 5.2103991006821716e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7722903377936996e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 23663,
|
|
"real_time": 2.9585877649282374e+04,
|
|
"cpu_time": 5.2130014408988595e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7720887181885475e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 23663,
|
|
"real_time": 2.9583964928999125e+04,
|
|
"cpu_time": 5.2109957191620547e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7722032907295551e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 23663,
|
|
"real_time": 2.9583735322770361e+04,
|
|
"cpu_time": 5.2119371985860699e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7722170452101761e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.9583854080478246e+04,
|
|
"cpu_time": 5.2119574562577569e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7722099336268627e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.9583735322770361e+04,
|
|
"cpu_time": 5.2119371985860693e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7722170452101761e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2625617651858552e+00,
|
|
"cpu_time": 1.2925589887665550e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.5631066801943234e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.2677392937081600e-05,
|
|
"cpu_time": 2.4799876046851437e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.2676132983389142e-05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 13477,
|
|
"real_time": 5.1997156418423590e+04,
|
|
"cpu_time": 5.1867734888223225e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0166025841145214e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 13477,
|
|
"real_time": 5.1991302384680239e+04,
|
|
"cpu_time": 5.1834393817688650e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0168296463159451e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 13477,
|
|
"real_time": 5.1979206158054862e+04,
|
|
"cpu_time": 5.1828034993716180e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0172989883907825e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 13477,
|
|
"real_time": 5.1992846806156667e+04,
|
|
"cpu_time": 5.1842328923609394e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0167697374013268e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 13477,
|
|
"real_time": 5.1995388065139261e+04,
|
|
"cpu_time": 5.1868297702211833e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0166711683858486e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.1991179966490934e+04,
|
|
"cpu_time": 5.1848158065089875e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0168344249216854e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.1992846806156675e+04,
|
|
"cpu_time": 5.1842328923609400e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0167697374013268e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0638427761226783e+00,
|
|
"cpu_time": 1.8823046916141076e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7405718882306297e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3586617539119956e-04,
|
|
"cpu_time": 3.6304176693240938e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3588482298625219e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7644,
|
|
"real_time": 9.1370031961040309e+04,
|
|
"cpu_time": 5.0084129911420739e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2952295790968033e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7644,
|
|
"real_time": 9.1326329539382743e+04,
|
|
"cpu_time": 5.0057074669360995e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2963279161412518e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7644,
|
|
"real_time": 9.1533554227321409e+04,
|
|
"cpu_time": 5.0177175533960849e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2911292123452049e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7644,
|
|
"real_time": 9.1374767002720750e+04,
|
|
"cpu_time": 5.0108401147462362e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2951106402685062e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7644,
|
|
"real_time": 9.1362186967191054e+04,
|
|
"cpu_time": 5.0076453024475360e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2954266634982212e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1393373939531259e+04,
|
|
"cpu_time": 5.0100646857336062e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2946448022699978e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1370031961040309e+04,
|
|
"cpu_time": 5.0084129911420739e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2952295790968033e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.0636775061739783e+01,
|
|
"cpu_time": 4.6565690090472565e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0225306410463106e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.8230439019672941e-04,
|
|
"cpu_time": 9.2944289168701854e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.8141338434842033e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4011,
|
|
"real_time": 1.7452878236293010e+05,
|
|
"cpu_time": 4.8791152066392040e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4032162163820091e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4011,
|
|
"real_time": 1.7436710999337866e+05,
|
|
"cpu_time": 4.8743519580328637e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4054444672273762e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4011,
|
|
"real_time": 1.7431509160195559e+05,
|
|
"cpu_time": 4.8723751127525508e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4061622900543774e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4011,
|
|
"real_time": 1.7438011186297706e+05,
|
|
"cpu_time": 4.8742463933583015e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4052651160676884e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4011,
|
|
"real_time": 1.7434740606163113e+05,
|
|
"cpu_time": 4.8756385381250817e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4057163193568420e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7438770037657447e+05,
|
|
"cpu_time": 4.8751454417816001e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4051608818176590e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7436710999337866e+05,
|
|
"cpu_time": 4.8743519580328637e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4054444672273762e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2585322958717740e+01,
|
|
"cpu_time": 2.5057336502423772e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1384571165207760e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.7357309477894487e-04,
|
|
"cpu_time": 5.1398131197633924e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.7333927851861895e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2066,
|
|
"real_time": 3.3900890662086930e+05,
|
|
"cpu_time": 4.8528583798450804e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4744506224378944e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2066,
|
|
"real_time": 3.3902540712629509e+05,
|
|
"cpu_time": 4.8547754719023544e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4743301899125935e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2066,
|
|
"real_time": 3.3902052220471913e+05,
|
|
"cpu_time": 4.8558859693853092e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4743658423529007e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2066,
|
|
"real_time": 3.3898280899352342e+05,
|
|
"cpu_time": 4.8540971467474782e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4746411255799911e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2066,
|
|
"real_time": 3.3897700091985764e+05,
|
|
"cpu_time": 4.8512286911230332e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4746835263856937e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3900292917305289e+05,
|
|
"cpu_time": 4.8537691318006504e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4744942613338150e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3900890662086924e+05,
|
|
"cpu_time": 4.8540971467474782e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4744506224378944e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1952214117791904e+01,
|
|
"cpu_time": 1.7947578905784132e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6023855240484420e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.4755234331871574e-05,
|
|
"cpu_time": 3.6976581329746809e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4756081640080900e-05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1046,
|
|
"real_time": 6.6937279266912502e+05,
|
|
"cpu_time": 4.8207348763289839e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5064084145250103e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1046,
|
|
"real_time": 6.6925561769836955e+05,
|
|
"cpu_time": 4.8222272186998081e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5068472428663891e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1046,
|
|
"real_time": 6.6940430877863802e+05,
|
|
"cpu_time": 4.8243354903823316e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5062904107400917e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1046,
|
|
"real_time": 6.6917638566405524e+05,
|
|
"cpu_time": 4.8246182900573057e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5071440593874481e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1046,
|
|
"real_time": 6.6929386085501197e+05,
|
|
"cpu_time": 4.8213849450956500e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5067040027182354e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.6930059313304001e+05,
|
|
"cpu_time": 4.8226601641128159e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5066788260474350e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.6929386085501197e+05,
|
|
"cpu_time": 4.8222272186998081e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5067040027182354e+10,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1467639864189266e+01,
|
|
"cpu_time": 1.7436551883484953e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.4257443990116953e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3666152518410784e-04,
|
|
"cpu_time": 3.6155464598639425e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3666467213166894e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 69,
|
|
"real_time": 9.8928793913860247e+06,
|
|
"cpu_time": 6.8842938163765788e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.3917761121414938e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 69,
|
|
"real_time": 9.9174197547245715e+06,
|
|
"cpu_time": 6.8847102818842244e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.3833832619633713e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 69,
|
|
"real_time": 9.8771252196984012e+06,
|
|
"cpu_time": 6.8786232708694482e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.3971860489407248e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 69,
|
|
"real_time": 9.9822014867179636e+06,
|
|
"cpu_time": 6.8875750704349780e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.3614260385994596e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 69,
|
|
"real_time": 9.8909060538247004e+06,
|
|
"cpu_time": 6.8863605234785211e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.3924528063862138e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.9121063812703341e+06,
|
|
"cpu_time": 6.8843125926087499e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.3852448536062527e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.8928793913860247e+06,
|
|
"cpu_time": 6.8847102818842256e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.3917761121414938e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.1786266984040245e+04,
|
|
"cpu_time": 3.4410998798355216e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4212719119886948e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.2156798340056683e-03,
|
|
"cpu_time": 4.9984654728345897e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.1984316451279266e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 36,
|
|
"real_time": 1.9502318050298426e+07,
|
|
"cpu_time": 7.3062961344446993e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.4410711499484081e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 36,
|
|
"real_time": 1.9333397607422538e+07,
|
|
"cpu_time": 7.2704660866666400e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.4711365980615511e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 36,
|
|
"real_time": 1.9419135836263496e+07,
|
|
"cpu_time": 7.2925210374999630e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.4558110394737649e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 36,
|
|
"real_time": 1.9341069170170359e+07,
|
|
"cpu_time": 7.2758843833332145e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.4697597847125063e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 36,
|
|
"real_time": 1.9403540218869846e+07,
|
|
"cpu_time": 7.2847800258335674e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.4585886515047894e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9399892176604934e+07,
|
|
"cpu_time": 7.2859895335556185e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.4592734447402039e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9403540218869846e+07,
|
|
"cpu_time": 7.2847800258335686e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.4585886515047894e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8476930546099858e+04,
|
|
"cpu_time": 1.4138360554084908e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2190964486418620e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.5297583060115551e-03,
|
|
"cpu_time": 1.9404859818931528e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5241401644483697e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 18,
|
|
"real_time": 3.8042982005410723e+07,
|
|
"cpu_time": 7.6380392733335924e+,
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.6880042011800565e+07,
|
|
"cpu_time": 7.8282756673684156e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.6393051818393846e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.6982442791524686e+07,
|
|
"cpu_time": 7.8160819778946686e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.6292283004831376e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.6989609661855198e+07,
|
|
"cpu_time": 7.8104904242103076e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.6285251244056621e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.6980808761559032e+07,
|
|
"cpu_time": 7.8093396678945005e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.6293886611673350e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7016827220979489e+07,
|
|
"cpu_time": 7.8109667784212887e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.6258571594685826e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6969946089543790e+07,
|
|
"cpu_time": 7.8150309031578362e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.6304608854728203e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6982442791524686e+07,
|
|
"cpu_time": 7.8109667784212887e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.6292283004831376e+09,
|
|
"hip_id": 5.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.2300585462489340e+04,
|
|
"cpu_time": 7.8430184695151995e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.1437333955173064e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/5/log2(N):27/manu,
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.4983213841915131e+07,
|
|
"cpu_time": 9.2329889690000761e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.5799406593312263e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.4419911950826645e+07,
|
|
"cpu_time": 9.2041688409999537e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6070380757420158e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.4864832311868668e+07,
|
|
"cpu_time": 9.2116240659999561e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.5856015128941083e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.4355527758598328e+07,
|
|
"cpu_time": 9.1933674079997444e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6101613974350233e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.4835710227489471e+07,
|
|
"cpu_time": 9.2237509470000422e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.5869968385947828e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4691839218139663e+07,
|
|
"cpu_time": 9.2131800461999547e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.5939476967994318e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4835710227489471e+07,
|
|
"cpu_time": 9.2116240659999561e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.5869968385947828e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8397820590058091e+05,
|
|
"cpu_time": 1.5654878570232657e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3678563734380888e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.8019977667334498e-03,
|
|
"cpu_time": 1.6991829630736055e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.8059996662061189e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4589905440807343e+08,
|
|
"cpu_time": 1.1027682554000421e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.6797422312168999e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4585969448089600e+08,
|
|
"cpu_time": 1.1021987082000351e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.6807352018025570e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4600231945514679e+08,
|
|
"cpu_time": 1.1020370651999884e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.6771396098603182e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4586667418479919e+08,
|
|
"cpu_time": 1.1008088457999749e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.6805590790384078e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4548526704311371e+08,
|
|
"cpu_time": 1.1005063108000171e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.6902081077453804e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4582260191440585e+08,
|
|
"cpu_time": 1.1016638370800116e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.6816768459327135e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4586667418479919e+08,
|
|
"cpu_time": 1.1020370651999884e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.6805590790384078e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9701295460105615e+05,
|
|
"cpu_time": 9.6384689479897427e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9811587471189871e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3510453936125605e-03,
|
|
"cpu_time": 8.7490109265424862e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3529592507886346e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8961427509784698e+08,
|
|
"cpu_time": 1.3249275459999411e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7074892929129043e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8775908052921295e+08,
|
|
"cpu_time": 1.3222702269999900e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7313916281123056e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8892794251441956e+08,
|
|
"cpu_time": 1.3254376140000374e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7162962317028670e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8946802020072937e+08,
|
|
"cpu_time": 1.3258339365000041e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7093625169903812e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8858387470245361e+08,
|
|
"cpu_time": 1.3247280384999840e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7207270333697228e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8887063860893255e+08,
|
|
"cpu_time": 1.3246394723999915e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7170533406176362e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8892794251441956e+08,
|
|
"cpu_time": 1.3249275459999411e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7162962317028670e+09,
|
|
"hip_id": 1.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4681634858200734e+05,
|
|
"cpu_time": 1.3935501978667879e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.6212719689986929e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.5852968379837067e-03,
|
|
"cpu_time": 1.0520222497536961e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5884137480254709e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 99931,
|
|
"real_time": 6.9468509649910447e+03,
|
|
"cpu_time": 7.2498239925601900e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.6851229613262624e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 99931,
|
|
"real_time": 6.9526833767164972e+03,
|
|
"cpu_time": 7.1933747396647263e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.6820316146900341e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 99931,
|
|
"real_time": 6.9352506503035056e+03,
|
|
"cpu_time": 7.1743148283230615e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.6912869182139330e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 99931,
|
|
"real_time": 6.9543363953414873e+03,
|
|
"cpu_time": 7.1790349418671381e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.6811564101427011e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 99931,
|
|
"real_time": 6.9527248485636383e+03,
|
|
"cpu_time": 7.1787240251248395e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.6820096519839555e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9483692471832346e+03,
|
|
"cpu_time": 7.1950545055079913e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.6843215112713777e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9526833767164962e+03,
|
|
"cpu_time": 7.1790349418671393e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.6820316146900341e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.8675377200578795e+00,
|
|
"cpu_time": 3.1448225481560836e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.1762842613081310e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1322854960892101e-03,
|
|
"cpu_time": 4.3708112923239771e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1335287239541121e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 100903,
|
|
"real_time": 6.9363132485674296e+03,
|
|
"cpu_time": 7.2682166377887535e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3814428739322633e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 100903,
|
|
"real_time": 6.9567752670892860e+03,
|
|
"cpu_time": 7.2750670779364729e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3597317771948203e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 100903,
|
|
"real_time": 6.9352290564895438e+03,
|
|
"cpu_time": 7.2545554379083204e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3825968231129020e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 100903,
|
|
"real_time": 6.9602632543805348e+03,
|
|
"cpu_time": 7.2804252540530586e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3560436047841415e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 100903,
|
|
"real_time": 6.9350225309360894e+03,
|
|
"cpu_time": 7.2675468161302674e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3828166774663702e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9447206714925778e+03,
|
|
"cpu_time": 7.2691622447633743e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3725263512980983e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9363132485674287e+03,
|
|
"cpu_time": 7.2682166377887547e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.3814428739322633e+07,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2666043222634503e+01,
|
|
"cpu_time": 9.7286307071153575e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3436811846100251e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8238376778244537e-03,
|
|
"cpu_time": 1.3383427662690783e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8225518914197437e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 100378,
|
|
"real_time": 6.9720638741581715e+03,
|
|
"cpu_time": 7.2247429201464593e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4687186154381594e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 100378,
|
|
"real_time": 6.9715572781115561e+03,
|
|
"cpu_time": 7.2298279535716879e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4688253415273944e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 100378,
|
|
"real_time": 6.9712570517114709e+03,
|
|
"cpu_time": 7.2294225664568448e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4688885984323359e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 100378,
|
|
"real_time": 6.9734606749441691e+03,
|
|
"cpu_time": 7.2347207310151958e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4684244276005733e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 100378,
|
|
"real_time": 6.9686468529608219e+03,
|
|
"cpu_time": 7.2210183991475213e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4694387900642797e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9713971463772377e+03,
|
|
"cpu_time": 7.2279465140675426e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4688591546125486e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9715572781115570e+03,
|
|
"cpu_time": 7.2294225664568448e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4688253415273944e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7544271618776162e+00,
|
|
"cpu_time": 5.2409529385646561e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.6970473083259298e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.5166076828506655e-04,
|
|
"cpu_time": 7.2509570019151949e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5169515380125919e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 99409,
|
|
"real_time": 7.0296708235049655e+03,
|
|
"cpu_time": 7.2028132816668463e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9133654354797733e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 99409,
|
|
"real_time": 7.0320981908896156e+03,
|
|
"cpu_time": 7.2114789465468276e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9123597885098815e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 99409,
|
|
"real_time": 7.0199990163369339e+03,
|
|
"cpu_time": 7.1911924232975781e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9173793261706966e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 99409,
|
|
"real_time": 7.0454282376389328e+03,
|
|
"cpu_time": 7.2003031626352668e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9068495638901389e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 99409,
|
|
"real_time": 7.0397093662026546e+03,
|
|
"cpu_time": 7.1978693286080098e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9092110106595612e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0333811269146199e+03,
|
|
"cpu_time": 7.2007314285509062e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9118330249420100e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0320981908896138e+03,
|
|
"cpu_time": 7.2003031626352668e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9123597885098815e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.7425952529120465e+00,
|
|
"cpu_time": 7.4028863366634282e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0341375221972790e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3851937037266590e-03,
|
|
"cpu_time": 1.0280742185871533e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3854288647878838e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 97139,
|
|
"real_time": 7.1961904921359219e+03,
|
|
"cpu_time": 7.1186574698032343e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6919004638303494e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 97139,
|
|
"real_time": 7.1970441174233465e+03,
|
|
"cpu_time": 7.1036118981330514e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6912253602614176e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 97139,
|
|
"real_time": 7.1937777857726460e+03,
|
|
"cpu_time": 7.1211958870937479e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6938094586418617e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 97139,
|
|
"real_time": 7.1983733907347105e+03,
|
|
"cpu_time": 7.1078585640650475e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6901744014447916e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 97139,
|
|
"real_time": 7.1943742018921375e+03,
|
|
"cpu_time": 7.1149803532088161e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6933374398606372e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1959519975917528e+03,
|
|
"cpu_time": 7.1132608344607794e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6920894248078120e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1961904921359210e+03,
|
|
"cpu_time": 7.1149803532088161e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.6919004638303494e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8927088016525606e+00,
|
|
"cpu_time": 7.3710917738651298e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4971377892498739e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.6302410053402077e-04,
|
|
"cpu_time": 1.0362465183555855e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6302077805118521e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 81349,
|
|
"real_time": 8.6277405924061241e+03,
|
|
"cpu_time": 6.5511047706587470e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.4949539943404758e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 81349,
|
|
"real_time": 8.5961960512181777e+03,
|
|
"cpu_time": 6.5435076739246416e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.5297966114198875e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 81349,
|
|
"real_time": 8.6271941203081606e+03,
|
|
"cpu_time": 6.5492019744015837e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.4955554329260695e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 81349,
|
|
"real_time": 8.6079599556731791e+03,
|
|
"cpu_time": 6.5410289167983150e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.5167728964642358e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 81349,
|
|
"real_time": 8.6288911413170281e+03,
|
|
"cpu_time": 6.5496438866788411e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.4936879673622298e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.6175963721845346e+03,
|
|
"cpu_time": 6.5468974444924259e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.5061533805025816e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.6271941203081606e+03,
|
|
"cpu_time": 6.5492019744015849e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 9.4955554329260695e+08,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4776908026437146e+01,
|
|
"cpu_time": 4.3728011081922951e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6316446880862266e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7147366142760342e-03,
|
|
"cpu_time": 6.6791959783498854e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7164089645689716e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 72395,
|
|
"real_time": 9.6864911434207625e+03,
|
|
"cpu_time": 6.2238167465198445e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6914277582474544e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 72395,
|
|
"real_time": 9.6755187503660472e+03,
|
|
"cpu_time": 6.2250405423591614e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6933458993483069e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 72395,
|
|
"real_time": 9.6844801810678109e+03,
|
|
"cpu_time": 6.2182595340294409e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6917789797359574e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 72395,
|
|
"real_time": 9.6777723546929992e+03,
|
|
"cpu_time": 6.2236428097238469e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6929515801283526e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 72395,
|
|
"real_time": 9.6866169030013298e+03,
|
|
"cpu_time": 6.2188198871345830e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6914057987493582e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.6821758665097896e+03,
|
|
"cpu_time": 6.2219159039533758e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6921820032418861e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.6844801810678091e+03,
|
|
"cpu_time": 6.2236428097238469e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.6917789797359574e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.1807913774510510e+00,
|
|
"cpu_time": 3.1350092642319284e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.0563646145680337e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.3508544451988068e-04,
|
|
"cpu_time": 5.0386557977100891e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.3518856702280428e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 66463,
|
|
"real_time": 1.0537148694112646e+04,
|
|
"cpu_time": 6.0318896986986029e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.1097596656587238e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 66463,
|
|
"real_time": 1.0539309920599095e+04,
|
|
"cpu_time": 6.0430031486049366e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.1091219678391757e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 66463,
|
|
"real_time": 1.0536645782898398e+04,
|
|
"cpu_time": 6.0392536905566216e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.1099080936349220e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 66463,
|
|
"real_time": 1.0536977769770885e+04,
|
|
"cpu_time": 6.0492933639775312e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.1098101102582569e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 66463,
|
|
"real_time": 1.0536134138985959e+04,
|
|
"cpu_time": 6.0383528165400147e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.1100591134988837e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0537243261273397e+04,
|
|
"cpu_time": 6.0403585436755407e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.1097317901779928e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0536977769770885e+04,
|
|
"cpu_time": 6.0392536905566227e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.1098101102582569e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2183638740668701e+00,
|
|
"cpu_time": 6.3990219180705084e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5951975433903490e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1562453706887603e-04,
|
|
"cpu_time": 1.0593778286175576e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1561117761813694e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57356,
|
|
"real_time": 1.2188810258673164e+04,
|
|
"cpu_time": 5.7822117046191728e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3767347763385429e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57356,
|
|
"real_time": 1.2198693670569746e+04,
|
|
"cpu_time": 5.7876500490762603e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3723785324743795e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57356,
|
|
"real_time": 1.2179646636829375e+04,
|
|
"cpu_time": 5.7739753925772643e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3807800795984697e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57356,
|
|
"real_time": 1.2188266183960861e+04,
|
|
"cpu_time": 5.7774395020996082e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3769747895924730e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57356,
|
|
"real_time": 1.2185959249634965e+04,
|
|
"cpu_time": 5.7654295967588151e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3779927092701511e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2188275199933623e+04,
|
|
"cpu_time": 5.7773412490262246e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3769721774548035e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2188266183960859e+04,
|
|
"cpu_time": 5.7774395020996082e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3769747895924730e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8674344973026447e+00,
|
|
"cpu_time": 8.4145528569221613e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.0290074384325966e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.6344596627913731e-04,
|
|
"cpu_time": 1.4564749586741894e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.6332957256743346e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 46077,
|
|
"real_time": 1.5191373128868201e+04,
|
|
"cpu_time": 5.4892280992139626e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.6280548103267612e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 46077,
|
|
"real_time": 1.5189789756604110e+04,
|
|
"cpu_time": 5.4895713766990983e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.6289541922733612e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 46077,
|
|
"real_time": 1.5192116185151312e+04,
|
|
"cpu_time": 5.4843091450315320e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.6276328065545616e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 46077,
|
|
"real_time": 1.5189372867480923e+04,
|
|
"cpu_time": 5.4940993190373075e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.6291910234564934e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 46077,
|
|
"real_time": 1.5192991569824446e+04,
|
|
"cpu_time": 5.4855521283637083e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.6271357025122433e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5191128701585802e+04,
|
|
"cpu_time": 5.4885520136691213e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.6281937070246849e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5191373128868203e+04,
|
|
"cpu_time": 5.4892280992139637e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.6280548103267612e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5314105898670933e+00,
|
|
"cpu_time": 3.8493517518479779e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.6980644467605546e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0080953298139258e-04,
|
|
"cpu_time": 7.0134194633871526e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0080979567808015e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 35392,
|
|
"real_time": 1.9778104666587667e+04,
|
|
"cpu_time": 5.3958604618936265e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3254252842682926e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 35392,
|
|
"real_time": 1.9777143567107174e+04,
|
|
"cpu_time": 5.4072814316786623e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3254896952661608e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 35392,
|
|
"real_time": 1.9781900817964033e+04,
|
|
"cpu_time": 5.3944250651946270e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3251709348474028e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 35392,
|
|
"real_time": 1.9780397640265182e+04,
|
|
"cpu_time": 5.4042009810756898e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3252716389602652e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 35392,
|
|
"real_time": 1.9779532523415819e+04,
|
|
"cpu_time": 5.3918636360826254e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3253296036681517e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9779415843067974e+04,
|
|
"cpu_time": 5.3987263151850486e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3253374314020548e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9779532523415819e+04,
|
|
"cpu_time": 5.3958604618936265e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3253296036681517e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8731699088346030e+00,
|
|
"cpu_time": 6.6515452300622850e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2551238075026702e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.4702994451228276e-05,
|
|
"cpu_time": 1.2320582377649747e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.4702207737005768e-05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 22717,
|
|
"real_time": 3.0815428278965413e+04,
|
|
"cpu_time": 5.1999826942006099e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7013815133566668e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 22717,
|
|
"real_time": 3.0834276954320099e+04,
|
|
"cpu_time": 5.2408304596832561e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7003414763923744e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 22717,
|
|
"real_time": 3.0811403137988491e+04,
|
|
"cpu_time": 5.1829136498940986e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7016037784841627e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 22717,
|
|
"real_time": 3.0819117982746393e+04,
|
|
"cpu_time": 5.1951313165844136e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7011778218101974e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 22717,
|
|
"real_time": 3.0825437069540509e+04,
|
|
"cpu_time": 5.2234109093560517e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7008290874099686e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.0821132684712182e+04,
|
|
"cpu_time": 5.2084538059436876e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7010667354906738e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.0819117982746393e+04,
|
|
"cpu_time": 5.1999826942006093e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.7011778218101974e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.9800181050386190e+00,
|
|
"cpu_time": 2.3310462660187325e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9555974060853655e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.9135912027960164e-04,
|
|
"cpu_time": 4.4755053090009787e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9132292711937135e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 13947,
|
|
"real_time": 5.0220634011934730e+04,
|
|
"cpu_time": 5.3114511841874307e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0879385946238953e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 13947,
|
|
"real_time": 5.0234640356590098e+04,
|
|
"cpu_time": 5.3125384899404556e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0873564388173054e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 13947,
|
|
"real_time": 5.0261956967551137e+04,
|
|
"cpu_time": 5.3158193716267276e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0862219922653538e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 13947,
|
|
"real_time": 5.0231222910978402e+04,
|
|
"cpu_time": 5.3107464896477294e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0874984506316410e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 13947,
|
|
"real_time": 5.0182076211573461e+04,
|
|
"cpu_time": 5.3085082034607065e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0895428789735237e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.0226106091725567e+04,
|
|
"cpu_time": 5.3118127477726108e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0877116710623440e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.0231222910978395e+04,
|
|
"cpu_time": 5.3114511841874301e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0874984506316410e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8952006914617005e+01,
|
|
"cpu_time": 2.6817617380819388e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2037039223888904e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.7643343606496832e-04,
|
|
"cpu_time": 5.0486752177144722e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.7656617006714280e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7647,
|
|
"real_time": 9.1445739975841730e+04,
|
|
"cpu_time": 5.0893247577940094e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2933293563527714e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7647,
|
|
"real_time": 9.1477877017632389e+04,
|
|
"cpu_time": 5.0884945308251888e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2925236880997723e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7647,
|
|
"real_time": 9.1397781555941969e+04,
|
|
"cpu_time": 5.0881037285929888e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2945327165478226e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7647,
|
|
"real_time": 9.1411468100341765e+04,
|
|
"cpu_time": 5.0870168994951719e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2941891685821850e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7647,
|
|
"real_time": 9.1418623072681934e+04,
|
|
"cpu_time": 5.0904359453495461e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2940096115128201e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1430297944487975e+04,
|
|
"cpu_time": 5.0886751724113828e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2937169082190746e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1418623072681934e+04,
|
|
"cpu_time": 5.0884945308251888e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2940096115128201e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1821398902780594e+01,
|
|
"cpu_time": 1.2875685393432202e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.9815709549035523e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.4803997819301652e-04,
|
|
"cpu_time": 2.5302627810159031e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.4797541607262838e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4022,
|
|
"real_time": 1.7408934545908510e+05,
|
|
"cpu_time": 4.9270534173992813e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4092824227349148e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4022,
|
|
"real_time": 1.7409073495714078e+05,
|
|
"cpu_time": 4.9279720368918884e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4092631931461437e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4022,
|
|
"real_time": 1.7405532706478654e+05,
|
|
"cpu_time": 4.9287974945201325e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4097533070267963e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4022,
|
|
"real_time": 1.7412379052986571e+05,
|
|
"cpu_time": 4.9298195015417194e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4088058198345924e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4022,
|
|
"real_time": 1.7403046825811177e+05,
|
|
"cpu_time": 4.9272101109574610e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4100975202682640e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7407793325379799e+05,
|
|
"cpu_time": 4.9281705122620964e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4094404526021423e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7408934545908507e+05,
|
|
"cpu_time": 4.9279720368918884e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4092824227349148e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5919971455466175e+01,
|
|
"cpu_time": 1.1534661318083483e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9718429967166102e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0634419759049202e-04,
|
|
"cpu_time": 2.3405564578951463e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0634844871750740e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2067,
|
|
"real_time": 3.3877316291419673e+05,
|
|
"cpu_time": 4.8824139507207346e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4761725302675869e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2067,
|
|
"real_time": 3.3884151756153390e+05,
|
|
"cpu_time": 4.8800447544267660e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4756730109015114e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2067,
|
|
"real_time": 3.3882224057536584e+05,
|
|
"cpu_time": 4.8785725500628287e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4758138620873920e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2067,
|
|
"real_time": 3.3879220504216838e+05,
|
|
"cpu_time": 4.8816656907401931e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4760333547095329e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2067,
|
|
"real_time": 3.3885645462730812e+05,
|
|
"cpu_time": 4.8813806293419278e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4755638812388050e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3881711614411464e+05,
|
|
"cpu_time": 4.8808155150584900e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4758513278409653e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3882224057536584e+05,
|
|
"cpu_time": 4.8813806293419278e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4758138620873920e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4357148153701409e+01,
|
|
"cpu_time": 1.5183966315003525e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5106321473286366e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0140322467973466e-04,
|
|
"cpu_time": 3.1109486249085493e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0140480242478862e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1046,
|
|
"real_time": 6.6940435724792362e+05,
|
|
"cpu_time": 4.8463720939194781e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5062902292681545e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1046,
|
|
"real_time": 6.6950762008688983e+05,
|
|
"cpu_time": 4.8457237224092776e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5059036666113857e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1046,
|
|
"real_time": 6.6957844873795903e+05,
|
|
"cpu_time": 4.8427772593787277e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5056385897159901e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1046,
|
|
"real_time": 6.6938326103313616e+05,
|
|
"cpu_time": 4.8443575993308067e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5063692172561638e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1046,
|
|
"real_time": 6.6936109632380086e+05,
|
|
"cpu_time": 4.8443861004396760e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5064522112417610e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.6944695668594190e+05,
|
|
"cpu_time": 4.8447233550955933e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5061307828186909e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.6940435724792362e+05,
|
|
"cpu_time": 4.8443861004396760e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5062902292681545e+10,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.2485212672424751e+01,
|
|
"cpu_time": 1.3922895731211343e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.4620305074565709e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3815166645953162e-04,
|
|
"cpu_time": 2.8738267824038891e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3814245175037363e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 71,
|
|
"real_time": 9.7161723484455701e+06,
|
|
"cpu_time": 6.9548810798592114e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4534620009460993e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 71,
|
|
"real_time": 9.7455674598754290e+06,
|
|
"cpu_time": 6.9662771101407552e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4430454807429862e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 71,
|
|
"real_time": 9.7286073805790544e+06,
|
|
"cpu_time": 6.9636898254930520e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4490478120212526e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 71,
|
|
"real_time": 9.7122040354240108e+06,
|
|
"cpu_time": 6.9552773847888350e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4548730522561655e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 71,
|
|
"real_time": 9.7204361793021075e+06,
|
|
"cpu_time": 6.9609572388734674e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4519471535081964e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.7245974807252362e+06,
|
|
"cpu_time": 6.9602165278310657e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4504750998949404e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.7204361793021075e+06,
|
|
"cpu_time": 6.9609572388734674e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4519471535081964e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3207158024059720e+04,
|
|
"cpu_time": 5.0548305233427428e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6815300946250465e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3581187345015704e-03,
|
|
"cpu_time": 7.2624615960301510e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3567784027096990e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.8948359062542785e+07,
|
|
"cpu_time": 7.3275471194594157e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5416715388648691e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.8915684239284411e+07,
|
|
"cpu_time": 7.3228552645943451e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5477893979974132e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.8956462827485960e+07,
|
|
"cpu_time": 7.3263082975673056e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5401574972465520e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.8896350486053001e+07,
|
|
"cpu_time": 7.3199655083784986e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5514193097514591e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.8940705000548750e+07,
|
|
"cpu_time": 7.3263980129729736e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5431027513524828e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8931512323182981e+07,
|
|
"cpu_time": 7.3246148405945086e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5448280990425553e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8940705000548750e+07,
|
|
"cpu_time": 7.3263082975673056e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5431027513524828e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4888394078850757e+04,
|
|
"cpu_time": 3.1368040001064300e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6628469858488813e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3146542998772024e-03,
|
|
"cpu_time": 4.2825514629405803e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3153943874198860e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7407093730412029e+07,
|
|
"cpu_time": 7.8870921968418896e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5880287564516339e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7324373071130954e+07,
|
|
"cpu_time": 7.8750799521055257e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5959807749272695e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7196305237318337e+07,
|
|
"cpu_time": 7.8668126668422866e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.6083618290491371e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7204052664731681e+07,
|
|
"cpu_time": 7.8668728021051502e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.6076104184003148e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7285753378742620e+07,
|
|
"cpu_time": 7.8784588800001347e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5997054058862147e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7283515616467126e+07,
|
|
"cpu_time": 7.8748632995789969e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5999374369429140e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7285753378742620e+07,
|
|
"cpu_time": 7.8750799521055245e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5997054058862147e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.7843469246882625e+04,
|
|
"cpu_time": 8.5320250015886535e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.4760059960337449e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3560940483864814e-03,
|
|
"cpu_time": 1.0834505536172026e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3544870277611308e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.3058352619409561e+07,
|
|
"cpu_time": 9.0799446999998856e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6742610033706703e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.3467407375574112e+07,
|
|
"cpu_time": 9.0970657960000956e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6538033066516976e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.3472016304731369e+07,
|
|
"cpu_time": 9.1046235060003376e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6535741020995173e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.3133391886949539e+07,
|
|
"cpu_time": 9.0747465910000074e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6704909901478481e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10,
|
|
"real_time": 7.3598980903625488e+07,
|
|
"cpu_time": 9.1075316110000134e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6472713712096639e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.3346029818058029e+07,
|
|
"cpu_time": 9.0927824408000684e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6598801546958795e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.3467407375574112e+07,
|
|
"cpu_time": 9.0970657960000956e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6538033066516976e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3588260763534499e+05,
|
|
"cpu_time": 1.4715599250921644e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1780008914468613e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.2160242104511284e-03,
|
|
"cpu_time": 1.6183824199830770e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.2186870653002245e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4363502860069275e+08,
|
|
"cpu_time": 1.0895625713999722e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7377436216656322e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4328728318214417e+08,
|
|
"cpu_time": 1.0878594270000122e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7468147910763268e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4357064068317413e+08,
|
|
"cpu_time": 1.0894475270000043e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7394199081742969e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4301749169826508e+08,
|
|
"cpu_time": 1.0866443969999864e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7538828686261506e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4365528225898743e+08,
|
|
"cpu_time": 1.0899903119999747e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7372166449967909e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4343314528465271e+08,
|
|
"cpu_time": 1.0887008468799901e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7430155669078393e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4357064068317413e+08,
|
|
"cpu_time": 1.0894475270000043e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7394199081742969e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7525004546019703e+05,
|
|
"cpu_time": 1.4053286242467305e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.1916968550127307e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9190128258983990e-03,
|
|
"cpu_time": 1.2908308359216728e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9213643989607818e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8568084537982941e+08,
|
|
"cpu_time": 1.3095175205000374e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7585362874868193e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8608894348144531e+08,
|
|
"cpu_time": 1.3160597450000181e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7531748376345029e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8573444485664368e+08,
|
|
"cpu_time": 1.3160235745000365e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7578312427075739e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8603863716125488e+08,
|
|
"cpu_time": 1.3156852230000310e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7538349177446117e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8553828597068787e+08,
|
|
"cpu_time": 1.3141140364999728e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7604127949070401e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8581623136997223e+08,
|
|
"cpu_time": 1.3142800199000192e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7567580160961099e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8573444485664368e+08,
|
|
"cpu_time": 1.3156852230000310e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7578312427075739e+09,
|
|
"hip_id": 2.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3775422012796212e+05,
|
|
"cpu_time": 2.7789814163428601e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1247057669355045e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.3184296073165738e-04,
|
|
"cpu_time": 2.1144515432519962e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.3175593252146390e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 102255,
|
|
"real_time": 6.8484361358507349e+03,
|
|
"cpu_time": 7.2839379970801342e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7380796859573677e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 102255,
|
|
"real_time": 6.8462456280277884e+03,
|
|
"cpu_time": 7.2935433273026943e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7392757126908168e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 102255,
|
|
"real_time": 6.8538754311041830e+03,
|
|
"cpu_time": 7.2699816315656304e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7351131133522436e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 102255,
|
|
"real_time": 6.8490449688680901e+03,
|
|
"cpu_time": 7.2824897000005317e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7377473963688679e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 102255,
|
|
"real_time": 6.8456261915812220e+03,
|
|
"cpu_time": 7.2668517947522902e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7396140664944544e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8486456710864049e+03,
|
|
"cpu_time": 7.2793608901402557e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7379659949727498e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8484361358507349e+03,
|
|
"cpu_time": 7.2824897000005317e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.5600000000000000e+02,
|
|
"bytes_per_second": 3.7380796859573677e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2569616486729140e+00,
|
|
"cpu_time": 1.0912479574331972e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7770160935891941e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):8/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.7556287842764975e-04,
|
|
"cpu_time": 1.4990985800845101e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.7539653811167127e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 102123,
|
|
"real_time": 6.8495124400909181e+03,
|
|
"cpu_time": 7.2719138972379589e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4749845989505768e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 102123,
|
|
"real_time": 6.8517936474987573e+03,
|
|
"cpu_time": 7.2654737067598474e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4724959089640021e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 102123,
|
|
"real_time": 6.8603435568860477e+03,
|
|
"cpu_time": 7.2582593411996472e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4631830863059580e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 102123,
|
|
"real_time": 6.8571205472621796e+03,
|
|
"cpu_time": 7.2476836612865746e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4666909597268865e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 102123,
|
|
"real_time": 6.8548179513394971e+03,
|
|
"cpu_time": 7.2285607977197921e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4691990893784478e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8547176286154809e+03,
|
|
"cpu_time": 7.2543782808407640e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4693107286651745e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8548179513394962e+03,
|
|
"cpu_time": 7.2582593411996472e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.1200000000000000e+02,
|
|
"bytes_per_second": 7.4691990893784478e+07,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2743070977894551e+00,
|
|
"cpu_time": 1.7007566800650614e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6573015537433261e+04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):9/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.2355699087385773e-04,
|
|
"cpu_time": 2.3444554643047210e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.2352494399113361e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 101595,
|
|
"real_time": 6.8738231721046468e+03,
|
|
"cpu_time": 7.2027767068654752e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4897095464363956e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 101595,
|
|
"real_time": 6.8887374931707363e+03,
|
|
"cpu_time": 7.2259923794683099e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4864842810676983e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 101595,
|
|
"real_time": 6.8689065384965234e+03,
|
|
"cpu_time": 7.1906261180905735e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4907758524024621e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 101595,
|
|
"real_time": 6.8880739868229139e+03,
|
|
"cpu_time": 7.2109028359041107e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4866274693897626e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 101595,
|
|
"real_time": 6.8707409000927728e+03,
|
|
"cpu_time": 7.1853838279591227e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4903778426372814e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8780564181375184e+03,
|
|
"cpu_time": 7.2031363736575174e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4887949983867201e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8738231721046468e+03,
|
|
"cpu_time": 7.2027767068654752e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0240000000000000e+03,
|
|
"bytes_per_second": 1.4897095464363956e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.6124206093596936e+00,
|
|
"cpu_time": 1.6236561141599903e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0798541516173678e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):10/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3975489622346834e-03,
|
|
"cpu_time": 2.2540960353018423e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3970050637402247e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 100653,
|
|
"real_time": 6.9667087701692708e+03,
|
|
"cpu_time": 7.1769785341178453e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9396951524216497e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 100653,
|
|
"real_time": 6.9565753642958871e+03,
|
|
"cpu_time": 7.1984533576336956e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9439773060049194e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 100653,
|
|
"real_time": 6.9719639458366664e+03,
|
|
"cpu_time": 7.1830603203338289e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9374793328111953e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 100653,
|
|
"real_time": 6.9561093812928893e+03,
|
|
"cpu_time": 7.1940705447561657e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9441745201817840e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 100653,
|
|
"real_time": 6.9655326508378321e+03,
|
|
"cpu_time": 7.1770669228735065e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9401915153662533e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9633780224865095e+03,
|
|
"cpu_time": 7.1859259359430087e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9411035653571606e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9655326508378321e+03,
|
|
"cpu_time": 7.1830603203338289e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0480000000000000e+03,
|
|
"bytes_per_second": 2.9401915153662533e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8659246736578154e+00,
|
|
"cpu_time": 9.8744722865029122e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.8999858868622105e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):11/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.8600487457179667e-04,
|
|
"cpu_time": 1.3741405595501849e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.8601964276971775e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 98360,
|
|
"real_time": 7.1272402137021882e+03,
|
|
"cpu_time": 7.1045910245265937e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.7469649923197484e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 98360,
|
|
"real_time": 7.1407990328581236e+03,
|
|
"cpu_time": 7.1236857970747590e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.7360527598555946e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 98360,
|
|
"real_time": 7.1374168338831387e+03,
|
|
"cpu_time": 7.1160328804107273e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.7387708961528528e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 98360,
|
|
"real_time": 7.1272451260138196e+03,
|
|
"cpu_time": 7.1187030544560134e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.7469610313386858e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 98360,
|
|
"real_time": 7.1313202549518901e+03,
|
|
"cpu_time": 7.1059993780930352e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.7436769820508265e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1328042922818322e+03,
|
|
"cpu_time": 7.1138024269122255e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.7424853323435426e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1313202549518883e+03,
|
|
"cpu_time": 7.1160328804107273e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 5.7436769820508265e+08,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.1085408180325382e+00,
|
|
"cpu_time": 8.2524175817115302e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9166627144842874e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.5640101252215553e-04,
|
|
"cpu_time": 1.1600571798974640e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.5619073100492678e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 94549,
|
|
"real_time": 7.4070241870890504e+03,
|
|
"cpu_time": 6.9712258698073673e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.1059772174470844e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 94549,
|
|
"real_time": 7.4057946887364305e+03,
|
|
"cpu_time": 6.9771897198414624e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.1061608300402007e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 94549,
|
|
"real_time": 7.4064775102638532e+03,
|
|
"cpu_time": 6.9770830473361874e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.1060588503303459e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 94549,
|
|
"real_time": 7.4058446985120900e+03,
|
|
"cpu_time": 6.9842584715178740e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.1061533604189482e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 94549,
|
|
"real_time": 7.4082446421598852e+03,
|
|
"cpu_time": 6.9735957904288101e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.1057950156478109e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4066771453522624e+03,
|
|
"cpu_time": 6.9766705797863412e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.1060290547768781e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4064775102638541e+03,
|
|
"cpu_time": 6.9770830473361874e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 1.1060588503303459e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0110682153656385e+00,
|
|
"cpu_time": 4.9273454593015619e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5096838238518688e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3650766673421028e-04,
|
|
"cpu_time": 7.0626030037560699e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3649585581243352e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 89931,
|
|
"real_time": 7.7889845918109459e+03,
|
|
"cpu_time": 6.8231781505169857e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1034834267390308e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 89931,
|
|
"real_time": 7.7859605615158689e+03,
|
|
"cpu_time": 6.8379164299266994e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1043004097634623e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 89931,
|
|
"real_time": 7.7883657211345599e+03,
|
|
"cpu_time": 6.8250424296609080e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1036505714594617e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 89931,
|
|
"real_time": 7.7866978973535897e+03,
|
|
"cpu_time": 6.8305448498866165e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1041011499326708e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 89931,
|
|
"real_time": 7.7905507523821607e+03,
|
|
"cpu_time": 6.8380009738194025e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1030605564042017e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7881119048394248e+03,
|
|
"cpu_time": 6.8309365667621219e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1037192228597653e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.7883657211345590e+03,
|
|
"cpu_time": 6.8305448498866153e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 2.1036505714594617e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8304151093580383e+00,
|
|
"cpu_time": 6.9590167008380615e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9441906241567992e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3502681159738394e-04,
|
|
"cpu_time": 1.0187500107524275e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3502141209869910e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 81362,
|
|
"real_time": 8.6076701777321214e+03,
|
|
"cpu_time": 6.5660209113656676e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8068373117699366e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 81362,
|
|
"real_time": 8.6037865211125900e+03,
|
|
"cpu_time": 6.5728230986929035e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8085556771534867e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 81362,
|
|
"real_time": 8.6073227580219864e+03,
|
|
"cpu_time": 6.5567211592792988e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8069909681800151e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 81362,
|
|
"real_time": 8.6024279442364968e+03,
|
|
"cpu_time": 6.5731332374391627e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8091571603286824e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 81362,
|
|
"real_time": 8.6057074537810240e+03,
|
|
"cpu_time": 6.5637967558046961e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8077055461143961e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.6053829709768434e+03,
|
|
"cpu_time": 6.5664990325163448e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8078493327093034e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.6057074537810222e+03,
|
|
"cpu_time": 6.5660209113656676e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 3.8077055461143961e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2572224518600130e+00,
|
|
"cpu_time": 6.8400327543009014e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.9887263812760438e+05,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.6230354412730840e-04,
|
|
"cpu_time": 1.0416559448847942e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6231937003056306e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 67958,
|
|
"real_time": 1.0302569364955414e+04,
|
|
"cpu_time": 6.1650037270971119e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.3611316438133593e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 67958,
|
|
"real_time": 1.0295583253538312e+04,
|
|
"cpu_time": 6.1759700243177187e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.3654480165052385e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 67958,
|
|
"real_time": 1.0308985166918201e+04,
|
|
"cpu_time": 6.1706597862686074e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.3571727904223509e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 67958,
|
|
"real_time": 1.0297161501054032e+04,
|
|
"cpu_time": 6.1724547014867508e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.3644723833156977e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 67958,
|
|
"real_time": 1.0323667581841943e+04,
|
|
"cpu_time": 6.1592305842869616e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.3481315608485622e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0305593373661581e+04,
|
|
"cpu_time": 6.1686637646914315e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.3592712789810410e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0302569364955416e+04,
|
|
"cpu_time": 6.1706597862686074e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 6.3611316438133593e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1386668433271826e+01,
|
|
"cpu_time": 6.5976795406276174e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.0207105131945154e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1049017771624088e-03,
|
|
"cpu_time": 1.0695476025767221e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1040117971377781e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 53497,
|
|
"real_time": 1.3090852139985167e+04,
|
|
"cpu_time": 5.7782888470064843e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0012488002950474e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 53497,
|
|
"real_time": 1.3085423678109963e+04,
|
|
"cpu_time": 5.7836625980563581e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0016641663598915e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 53497,
|
|
"real_time": 1.3088462134745285e+04,
|
|
"cpu_time": 5.7869685774940419e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0014316323080442e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 53497,
|
|
"real_time": 1.3085665950246046e+04,
|
|
"cpu_time": 5.7829706937530529e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0016456212343971e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 53497,
|
|
"real_time": 1.3089028020318825e+04,
|
|
"cpu_time": 5.7821752833771384e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0013883368308910e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3087886384681060e+04,
|
|
"cpu_time": 5.7828131999374175e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0014757114056543e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3088462134745285e+04,
|
|
"cpu_time": 5.7829706937530541e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0014316323080442e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3143993415054251e+00,
|
|
"cpu_time": 3.1180984202165238e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7709499950026821e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7683522560329936e-04,
|
|
"cpu_time": 5.3920095849720152e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7683404348538885e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 39138,
|
|
"real_time": 1.7888081201012868e+04,
|
|
"cpu_time": 5.4520939953593600e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4654674084616560e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 39138,
|
|
"real_time": 1.7884748147024398e+04,
|
|
"cpu_time": 5.4417889453689992e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4657405172552828e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 39138,
|
|
"real_time": 1.7890784688816591e+04,
|
|
"cpu_time": 5.4498379685855150e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4652459607535517e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 39138,
|
|
"real_time": 1.7890662272893282e+04,
|
|
"cpu_time": 5.4460724041128600e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4652559866226015e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 39138,
|
|
"real_time": 1.7890714834102113e+04,
|
|
"cpu_time": 5.4512951031260931e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4652516818406733e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7888998228769855e+04,
|
|
"cpu_time": 5.4482176833105636e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4653923109867531e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7890662272893282e+04,
|
|
"cpu_time": 5.4498379685855138e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.4652559866226015e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6368234094501313e+00,
|
|
"cpu_time": 4.2735385390184785e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1602460209522434e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4739916543842456e-04,
|
|
"cpu_time": 7.8439203193175246e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4741758945750137e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 24145,
|
|
"real_time": 2.9013589618722359e+04,
|
|
"cpu_time": 5.1883634234059668e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.8070428612586391e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 24145,
|
|
"real_time": 2.9001993029025420e+04,
|
|
"cpu_time": 5.2171856186370760e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.8077654162432510e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 24145,
|
|
"real_time": 2.9012961658830627e+04,
|
|
"cpu_time": 5.2240580276229346e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.8070819731029537e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 24145,
|
|
"real_time": 2.9002345490259024e+04,
|
|
"cpu_time": 5.1905080833747369e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.8077434467363750e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 24145,
|
|
"real_time": 2.9016439746548080e+04,
|
|
"cpu_time": 5.2022745702263093e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.8068653652189415e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.9009465908677103e+04,
|
|
"cpu_time": 5.2044779446534050e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.8072998125120327e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.9012961658830627e+04,
|
|
"cpu_time": 5.2022745702263099e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.8070819731029537e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.7897709292952513e+00,
|
|
"cpu_time": 1.5847352343602278e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.2303089660401875e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3405363444710452e-04,
|
|
"cpu_time": 3.0449456241585897e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3406791373260449e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 13500,
|
|
"real_time": 5.1772802291250424e+04,
|
|
"cpu_time": 5.1396943129285902e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0253414024243549e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 13500,
|
|
"real_time": 5.1756764818424861e+04,
|
|
"cpu_time": 5.1421504139484298e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0259689794728401e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 13500,
|
|
"real_time": 5.1743254803080126e+04,
|
|
"cpu_time": 5.1425103681124133e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0264979541595848e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 13500,
|
|
"real_time": 5.1800261916435338e+04,
|
|
"cpu_time": 5.1351280668730170e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0242677569692070e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 13500,
|
|
"real_time": 5.1762609208472655e+04,
|
|
"cpu_time": 5.1449032633060032e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0257402322531414e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.1767138607532695e+04,
|
|
"cpu_time": 5.1408772850336903e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0255632650558258e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.1762609208472655e+04,
|
|
"cpu_time": 5.1421504139484292e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.0257402322531414e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1376377005141439e+01,
|
|
"cpu_time": 3.7063881029560952e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.3623105730988011e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.1293333145577687e-04,
|
|
"cpu_time": 7.2096412683225633e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.1283877513785435e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 7664,
|
|
"real_time": 9.1309063604502197e+04,
|
|
"cpu_time": 5.0218371630450851e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2967621364332943e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 7664,
|
|
"real_time": 9.1330230529127221e+04,
|
|
"cpu_time": 5.0230105824518043e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2962298330465424e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 7664,
|
|
"real_time": 9.1287914332428816e+04,
|
|
"cpu_time": 5.0218733449096096e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2972942424373199e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 7664,
|
|
"real_time": 9.1320860050784526e+04,
|
|
"cpu_time": 5.0266113440414649e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2964654503185265e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 7664,
|
|
"real_time": 9.1317477290621144e+04,
|
|
"cpu_time": 5.0243313943044686e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2965505204723721e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1313109161492786e+04,
|
|
"cpu_time": 5.0235327657504863e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2966604365416115e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.1317477290621144e+04,
|
|
"cpu_time": 5.0230105824518043e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.2965505204723721e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5995655629792612e+01,
|
|
"cpu_time": 2.0010543060781184e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0235969935767674e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7517370481278126e-04,
|
|
"cpu_time": 3.9833607132433479e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7519337772176871e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 4014,
|
|
"real_time": 1.7426044831686298e+05,
|
|
"cpu_time": 4.8792193099879640e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4069167963882267e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 4014,
|
|
"real_time": 1.7442562973133035e+05,
|
|
"cpu_time": 4.8911955892895830e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4046374414474129e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 4014,
|
|
"real_time": 1.7429557301855925e+05,
|
|
"cpu_time": 4.8785436617013967e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4064317454312992e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 4014,
|
|
"real_time": 1.7448422920999318e+05,
|
|
"cpu_time": 4.8803888606323403e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4038298584292805e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 4014,
|
|
"real_time": 1.7424793403954618e+05,
|
|
"cpu_time": 4.8864577830687499e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4070896582613644e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7434276286325842e+05,
|
|
"cpu_time": 4.8831610409360063e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4057810999915169e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7429557301855925e+05,
|
|
"cpu_time": 4.8803888606323397e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 2.4064317454312992e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0591826665039991e+02,
|
|
"cpu_time": 5.4764243218697188e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4612297080845298e+07,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.0752889830864045e-04,
|
|
"cpu_time": 1.1214916477176010e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.0738265342997425e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2063,
|
|
"real_time": 3.3931270205283514e+05,
|
|
"cpu_time": 4.8426109110029310e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4722351828413990e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2063,
|
|
"real_time": 3.3933899305380299e+05,
|
|
"cpu_time": 4.8456488892093170e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4720436412298683e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2063,
|
|
"real_time": 3.3920815516867273e+05,
|
|
"cpu_time": 4.8442316224525923e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4729971470847237e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2063,
|
|
"real_time": 3.3930277997439151e+05,
|
|
"cpu_time": 4.8436370201840097e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4723074773018719e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2063,
|
|
"real_time": 3.3932076676234504e+05,
|
|
"cpu_time": 4.8577185519578326e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4721764246970627e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3929667940240947e+05,
|
|
"cpu_time": 4.8467693989613360e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4723519746309853e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3931270205283514e+05,
|
|
"cpu_time": 4.8442316224525917e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 2.4722351828413990e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.1237506455225706e+01,
|
|
"cpu_time": 6.2186136309157859e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.7341501448549172e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5101092809239515e-04,
|
|
"cpu_time": 1.2830430167047840e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5103634851232150e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1045,
|
|
"real_time": 6.6971041994062404e+05,
|
|
"cpu_time": 4.8201730871488458e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5051448358064152e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1045,
|
|
"real_time": 6.7007911445029324e+05,
|
|
"cpu_time": 4.8191839092635822e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5037664416332947e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1045,
|
|
"real_time": 6.6970367425758590e+05,
|
|
"cpu_time": 4.8189723732923561e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5051700692248310e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1045,
|
|
"real_time": 6.6996671679389605e+05,
|
|
"cpu_time": 4.8238196888989627e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5041864885895859e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1045,
|
|
"real_time": 6.7001006072476215e+05,
|
|
"cpu_time": 4.8276667727267593e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5040244891027122e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.6989399723343237e+05,
|
|
"cpu_time": 4.8219631662661016e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5044584648713680e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.6996671679389593e+05,
|
|
"cpu_time": 4.8201730871488458e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 2.5041864885895859e+10,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7532163984461531e+02,
|
|
"cpu_time": 3.7366863092101464e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.5549371510884836e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.6171549613620800e-04,
|
|
"cpu_time": 7.7493049622435377e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6173071915669215e-04,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 71,
|
|
"real_time": 9.6978473306541704e+06,
|
|
"cpu_time": 6.9404968377459455e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4599876504486666e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 71,
|
|
"real_time": 9.7209955368873104e+06,
|
|
"cpu_time": 6.9562010388731992e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4517485243845944e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 71,
|
|
"real_time": 9.7227759633055869e+06,
|
|
"cpu_time": 6.9534157977467251e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4511164431471720e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 71,
|
|
"real_time": 9.7168694630684983e+06,
|
|
"cpu_time": 6.9477946567603242e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4532142401966381e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 71,
|
|
"real_time": 9.7225235224189889e+06,
|
|
"cpu_time": 6.9503368321124971e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4512060498107767e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.7162023632669114e+06,
|
|
"cpu_time": 6.9496490326477396e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4534545815975704e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.7209955368873104e+06,
|
|
"cpu_time": 6.9503368321124983e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.4517485243845944e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0529672694526324e+04,
|
|
"cpu_time": 6.0164722719658841e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.7474047821072116e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0837230741853232e-03,
|
|
"cpu_time": 8.6572318166025053e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0851177259072740e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.9172685573229920e+07,
|
|
"cpu_time": 7.3792731213516283e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5002328569817843e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.9132516422384493e+07,
|
|
"cpu_time": 7.3624571481077754e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5075816750108500e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.9094444761002384e+07,
|
|
"cpu_time": 7.3607924394597733e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5145753039679923e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.9114661055642206e+07,
|
|
"cpu_time": 7.3628402559461391e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5108581734537749e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 37,
|
|
"real_time": 1.9156672463223740e+07,
|
|
"cpu_time": 7.3689100143245602e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5031587103048859e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9134196055096548e+07,
|
|
"cpu_time": 7.3668545958379757e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5072813439438577e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9132516422384497e+07,
|
|
"cpu_time": 7.3628402559461391e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5075816750108500e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1429423635843610e+04,
|
|
"cpu_time": 7.5935290276164142e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.7611421936418125e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6425787394120550e-03,
|
|
"cpu_time": 1.0307694999038670e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6426233394676972e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7568856030702591e+07,
|
|
"cpu_time": 7.9090183836836720e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5725795826817975e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7438060892255686e+07,
|
|
"cpu_time": 7.8904867878950906e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5850608926105957e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7396940157601707e+07,
|
|
"cpu_time": 7.8837894468419063e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5890029353836708e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7417479447628319e+07,
|
|
"cpu_time": 7.8890834536842275e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5870328515275574e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 19,
|
|
"real_time": 3.7475568291388057e+07,
|
|
"cpu_time": 7.8906135178947592e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5814727866540031e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7459380963915274e+07,
|
|
"cpu_time": 7.8925983179999304e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5830298097715249e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7438060892255679e+07,
|
|
"cpu_time": 7.8904867878950906e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5850608926105957e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.7742078198577772e+04,
|
|
"cpu_time": 9.5912333681367652e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4703059741931837e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 19,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8084142464562857e-03,
|
|
"cpu_time": 1.2152187380755095e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8058197441024831e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 7.3962683478991196e+07,
|
|
"cpu_time": 8.4144696300004148e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6293363541392875e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 7.3741830057568014e+07,
|
|
"cpu_time": 8.4004143733339894e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6402060511712356e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 7.3939198421107396e+07,
|
|
"cpu_time": 8.4093002577781641e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6304891279882989e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 7.3666060964266464e+07,
|
|
"cpu_time": 8.3976704033338785e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6439501784982266e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9,
|
|
"real_time": 7.3897884951697454e+07,
|
|
"cpu_time": 8.4129973944445479e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6325187950299239e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.3841531574726120e+07,
|
|
"cpu_time": 8.4069704117781997e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6353001013653946e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.3897884951697454e+07,
|
|
"cpu_time": 8.4093002577781653e+08,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.6325187950299239e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3050155893606119e+05,
|
|
"cpu_time": 7.5409345162084536e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4292221273059156e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 20,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7673192328628276e-03,
|
|
"cpu_time": 8.9698597078961687e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7685533375610841e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4463456571102142e+08,
|
|
"cpu_time": 1.0934778209999876e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7119129121088753e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4455555677413940e+08,
|
|
"cpu_time": 1.0940667435999331e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7139417119663763e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4454233348369598e+08,
|
|
"cpu_time": 1.0936528876000695e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7142814776859660e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4478163421154022e+08,
|
|
"cpu_time": 1.0935118559999409e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7081423684966750e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 5,
|
|
"real_time": 1.4435479044914246e+08,
|
|
"cpu_time": 1.0929844851999631e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7191070024734969e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4457377612590793e+08,
|
|
"cpu_time": 1.0935387586799788e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7134770945462780e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4455555677413940e+08,
|
|
"cpu_time": 1.0935118559999409e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7139417119663763e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5506101102667564e+05,
|
|
"cpu_time": 3.8841013473956374e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.9832378027529311e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 21,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0725389844672417e-03,
|
|
"cpu_time": 3.5518643638056077e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0726436978978088e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8841078281402588e+08,
|
|
"cpu_time": 1.3213358980000293e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7229600555273767e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8932066261768341e+08,
|
|
"cpu_time": 1.3251043865000155e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7112517795484009e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8876855969429016e+08,
|
|
"cpu_time": 1.3234571104999874e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7183474029746709e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8886136412620544e+08,
|
|
"cpu_time": 1.3236674989999528e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7171527845131793e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2,
|
|
"real_time": 2.8888833522796631e+08,
|
|
"cpu_time": 1.3238982144999909e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7168057448657231e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8884994089603430e+08,
|
|
"cpu_time": 1.3234926216999955e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7173035534858704e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8886136412620544e+08,
|
|
"cpu_time": 1.3236674989999528e+09,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7171527845131793e+09,
|
|
"hip_id": 3.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2506394564572454e+05,
|
|
"cpu_time": 1.3643305612190075e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.1827069463494569e+06,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 22,
|
|
"run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1253730730820014e-03,
|
|
"cpu_time": 1.0308561897886192e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1251991897264237e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
}
|
|
]
|
|
}
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5428487739714671e-03,
|
|
"cpu_time": 2.6074273681687640e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5437248099391895e-03,
|
|
"hip_id": 0.0000000000000000e+00,
|
|
"numa_id": NaN
|
|
}
|
|
]
|
|
}
|