25273 lines
873 KiB
JSON
25273 lines
873 KiB
JSON
{
|
|
"context": {
|
|
"date": "2023-10-09T17:45:33-04:00",
|
|
"host_name": "frontier05673",
|
|
"executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope",
|
|
"num_cpus": 128,
|
|
"mhz_per_cpu": 1796,
|
|
"cpu_scaling_enabled": false,
|
|
"caches": [
|
|
{
|
|
"type": "Data",
|
|
"level": 1,
|
|
"size": 32768,
|
|
"num_sharing": 2
|
|
},
|
|
{
|
|
"type": "Instruction",
|
|
"level": 1,
|
|
"size": 32768,
|
|
"num_sharing": 2
|
|
},
|
|
{
|
|
"type": "Unified",
|
|
"level": 2,
|
|
"size": 524288,
|
|
"num_sharing": 2
|
|
},
|
|
{
|
|
"type": "Unified",
|
|
"level": 3,
|
|
"size": 33554432,
|
|
"num_sharing": 16
|
|
}
|
|
],
|
|
"load_avg": [0.29,3.2,19.1],
|
|
"library_build_type": "release"
|
|
},
|
|
"benchmarks": [
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57046,
|
|
"real_time": 1.2267238004456958e+04,
|
|
"cpu_time": 2.8659642726922131e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3389749171833408e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57046,
|
|
"real_time": 1.2267363340961074e+04,
|
|
"cpu_time": 2.8685867457841036e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3389408026444769e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57046,
|
|
"real_time": 1.2266843972453978e+04,
|
|
"cpu_time": 2.8643220593906663e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3390821707668602e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57046,
|
|
"real_time": 1.2267951096180519e+04,
|
|
"cpu_time": 2.8658853626897599e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3387808346213901e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57046,
|
|
"real_time": 1.2270554779584207e+04,
|
|
"cpu_time": 2.8656289660975348e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3380723802439147e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2267990238727349e+04,
|
|
"cpu_time": 2.8660774813308562e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3387702210919970e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2267363340961074e+04,
|
|
"cpu_time": 2.8658853626897599e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3389408026444769e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4875471038461685e+00,
|
|
"cpu_time": 1.5515158688378810e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0478563462652673e+04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2125434361288528e-04,
|
|
"cpu_time": 5.4133772689126270e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2123794326107750e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57079,
|
|
"real_time": 1.2266389029302303e+04,
|
|
"cpu_time": 2.8631468263284209e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6784120252755022e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57079,
|
|
"real_time": 1.2262311124837492e+04,
|
|
"cpu_time": 2.8661623101315730e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6806329708981073e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57079,
|
|
"real_time": 1.2261666860868856e+04,
|
|
"cpu_time": 2.8634291911210770e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6809839909641123e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57079,
|
|
"real_time": 1.2265609271990548e+04,
|
|
"cpu_time": 2.8638970724784984e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6788365896401525e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57079,
|
|
"real_time": 1.2264726635322684e+04,
|
|
"cpu_time": 2.8624457453704552e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6793172351733136e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2264140584464374e+04,
|
|
"cpu_time": 2.8638162290860044e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6796365623902380e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2264726635322684e+04,
|
|
"cpu_time": 2.8634291911210770e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6793172351733136e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0629004981107664e+00,
|
|
"cpu_time": 1.4130424599038241e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1235892381115085e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6820587499820000e-04,
|
|
"cpu_time": 4.9341240738579116e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6821113358740043e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2259492711462217e+04,
|
|
"cpu_time": 2.8882859822742234e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3364337648882899e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2262444678003701e+04,
|
|
"cpu_time": 2.8930858999509579e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3361120421109438e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2255068018272501e+04,
|
|
"cpu_time": 2.8861759160653026e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3369162843952556e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2258835438477276e+04,
|
|
"cpu_time": 2.8871548938555286e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3365054194768710e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2260314938719997e+04,
|
|
"cpu_time": 2.8873243063826751e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3363441381311307e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2259231156987140e+04,
|
|
"cpu_time": 2.8884053997057381e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3364623298004982e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2259492711462217e+04,
|
|
"cpu_time": 2.8873243063826747e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3364337648882899e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6952749399898996e+00,
|
|
"cpu_time": 2.7214181974058071e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9386144054639083e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.1985676797143435e-04,
|
|
"cpu_time": 9.4218706199727291e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1988007742070612e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57059,
|
|
"real_time": 1.2270824070995655e+04,
|
|
"cpu_time": 2.9219640477400560e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6703992992168460e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57059,
|
|
"real_time": 1.2269717017990344e+04,
|
|
"cpu_time": 2.9266209519970598e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6706402398648858e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57059,
|
|
"real_time": 1.2262965151123200e+04,
|
|
"cpu_time": 2.9195636919679571e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6721106678671989e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57059,
|
|
"real_time": 1.2271878894085194e+04,
|
|
"cpu_time": 2.9219176045847249e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6701697664074516e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57059,
|
|
"real_time": 1.2270538420832327e+04,
|
|
"cpu_time": 2.9293355719518288e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6704614643778038e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2269184711005346e+04,
|
|
"cpu_time": 2.9238803736483253e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6707562875468373e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2270538420832327e+04,
|
|
"cpu_time": 2.9219640477400557e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6704614643778038e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5617737699640939e+00,
|
|
"cpu_time": 3.9820995193408926e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.7559356779179128e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.9030240018876043e-04,
|
|
"cpu_time": 1.3619228595088363e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9040222479610641e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57040,
|
|
"real_time": 1.2269548862944950e+04,
|
|
"cpu_time": 2.9962817636746113e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3413536823610620e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57040,
|
|
"real_time": 1.2268704471768180e+04,
|
|
"cpu_time": 2.9999416619915879e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3417212999796772e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57040,
|
|
"real_time": 1.2270383841569075e+04,
|
|
"cpu_time": 2.9890998352033723e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3409902123827591e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57040,
|
|
"real_time": 1.2271850895125537e+04,
|
|
"cpu_time": 2.9900360729312779e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3403517171180220e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57040,
|
|
"real_time": 1.2271198935568042e+04,
|
|
"cpu_time": 2.9889967671809249e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3406354459827108e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2270337401395158e+04,
|
|
"cpu_time": 2.9928712201963550e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3410104715648470e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2270383841569075e+04,
|
|
"cpu_time": 2.9900360729312779e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3409902123827591e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2573099426227816e+00,
|
|
"cpu_time": 4.9723169886469833e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.4728482149608352e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0246743031530849e-04,
|
|
"cpu_time": 1.6613868833022362e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0246840451067982e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57017,
|
|
"real_time": 1.2283274306853786e+04,
|
|
"cpu_time": 3.1169210796779997e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0670770409064692e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57017,
|
|
"real_time": 1.2265457319427969e+04,
|
|
"cpu_time": 3.1219434589683715e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0686270930345781e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57017,
|
|
"real_time": 1.2275615143631721e+04,
|
|
"cpu_time": 3.1178717540382651e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0677428256456610e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57017,
|
|
"real_time": 1.2275160683021224e+04,
|
|
"cpu_time": 3.1186930336566289e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0677823564566156e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57017,
|
|
"real_time": 1.2277971962727957e+04,
|
|
"cpu_time": 3.1194118683901252e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0675378669856323e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2275495883132533e+04,
|
|
"cpu_time": 3.1189682389462781e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0677534366057913e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2275615143631720e+04,
|
|
"cpu_time": 3.1186930336566285e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0677428256456610e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.4717499414752684e+00,
|
|
"cpu_time": 1.9050996974575177e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.6307324319026209e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.2720883971522046e-04,
|
|
"cpu_time": 6.1081086805203969e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.2734388285387052e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 55980,
|
|
"real_time": 1.2502224271107592e+04,
|
|
"cpu_time": 3.2938395105394789e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0967788956226765e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 55980,
|
|
"real_time": 1.2503779237529803e+04,
|
|
"cpu_time": 3.2993425616291497e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0965181407968311e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 55980,
|
|
"real_time": 1.2494847404097687e+04,
|
|
"cpu_time": 3.2943666881029014e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0980168186290123e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 55980,
|
|
"real_time": 1.2497971250537626e+04,
|
|
"cpu_time": 3.2938716720257202e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0974924229300278e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 55980,
|
|
"real_time": 1.2498137452993402e+04,
|
|
"cpu_time": 3.2958429653447783e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0974645301025589e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2499391923253224e+04,
|
|
"cpu_time": 3.2954526795284051e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0972541616162216e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2498137452993404e+04,
|
|
"cpu_time": 3.2943666881029014e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0974645301025589e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5887108353647781e+00,
|
|
"cpu_time": 2.3218948111325087e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.0213650426327754e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.8711083366291808e-04,
|
|
"cpu_time": 7.0457537611032630e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.8710707327873360e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 55774,
|
|
"real_time": 1.2547736020311517e+04,
|
|
"cpu_time": 3.3048482052569365e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 4.1783473859452759e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 55774,
|
|
"real_time": 1.2548934697670797e+04,
|
|
"cpu_time": 3.3088739358841027e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 4.1779482691651337e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 55774,
|
|
"real_time": 1.2548640023849746e+04,
|
|
"cpu_time": 3.3068405565317335e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 4.1780463779624443e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 55774,
|
|
"real_time": 1.2545332714741106e+04,
|
|
"cpu_time": 3.3037689048660708e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 4.1791478306824608e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 55774,
|
|
"real_time": 1.2541593949525428e+04,
|
|
"cpu_time": 3.3054768099831483e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 4.1803936733244263e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2546447481219719e+04,
|
|
"cpu_time": 3.3059616825043988e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 4.1787767074159485e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2547736020311517e+04,
|
|
"cpu_time": 3.3054768099831483e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 4.1783473859452759e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.0599548991389200e+00,
|
|
"cpu_time": 1.9708403038252701e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0193526365077004e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4389014529565005e-04,
|
|
"cpu_time": 5.9614735229849354e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4393565578622234e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 55465,
|
|
"real_time": 1.2622851021522703e+04,
|
|
"cpu_time": 3.3160449526728502e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 8.3069664548216263e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 55465,
|
|
"real_time": 1.2610858879747389e+04,
|
|
"cpu_time": 3.3211697899576422e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 8.3148658628158737e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 55465,
|
|
"real_time": 1.2623699276428095e+04,
|
|
"cpu_time": 3.3150592950509315e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 8.3064082646358566e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 55465,
|
|
"real_time": 1.2624760484783088e+04,
|
|
"cpu_time": 3.3155043757324507e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 8.3057100470450317e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 55465,
|
|
"real_time": 1.2629373329553900e+04,
|
|
"cpu_time": 3.3147827639051509e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 8.3026764087037888e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2622308598407035e+04,
|
|
"cpu_time": 3.3165122354638050e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 8.3073254076044373e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2623699276428093e+04,
|
|
"cpu_time": 3.3155043757324507e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 8.3064082646358566e+10,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8783435261058274e+00,
|
|
"cpu_time": 2.6471237228846189e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.5291015521495208e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.4493545871425535e-04,
|
|
"cpu_time": 7.9816491993566429e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.4519370915741782e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 54855,
|
|
"real_time": 1.2788774196077204e+04,
|
|
"cpu_time": 3.3554695579254527e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.6398381641950290e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 54855,
|
|
"real_time": 1.2789838389519067e+04,
|
|
"cpu_time": 3.3587504712423783e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.6397017195452295e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 54855,
|
|
"real_time": 1.2759325876815481e+04,
|
|
"cpu_time": 3.3483170868653629e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.6436228843489767e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 54855,
|
|
"real_time": 1.2757952457744912e+04,
|
|
"cpu_time": 3.3470453176556257e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.6437998236361914e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 54855,
|
|
"real_time": 1.2784425712863116e+04,
|
|
"cpu_time": 3.3550930051955242e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.6403959372926230e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2776063326603955e+04,
|
|
"cpu_time": 3.3529350877768687e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.6414717058036102e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2784425712863116e+04,
|
|
"cpu_time": 3.3550930051955234e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 1.6403959372926230e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6042084194528442e+01,
|
|
"cpu_time": 5.0228019038168142e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0619341163463086e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2556359329499846e-03,
|
|
"cpu_time": 1.4980313582948406e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2561496546398610e-03,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 52432,
|
|
"real_time": 1.3379048322106972e+04,
|
|
"cpu_time": 3.4667116322093199e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.1349793341201331e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 52432,
|
|
"real_time": 1.3352963343092690e+04,
|
|
"cpu_time": 3.4651827509917683e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.1411035080611212e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 52432,
|
|
"real_time": 1.3376617632485670e+04,
|
|
"cpu_time": 3.4658480393652797e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.1355489969407208e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 52432,
|
|
"real_time": 1.3347282063899487e+04,
|
|
"cpu_time": 3.4623636290814677e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.1424405207891510e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 52432,
|
|
"real_time": 1.3374914130522302e+04,
|
|
"cpu_time": 3.4635972135337135e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.1359483575512189e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3366165098421425e+04,
|
|
"cpu_time": 3.4647406530363100e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.1380041434924689e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3374914130522302e+04,
|
|
"cpu_time": 3.4651827509917683e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.1359483575512189e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4854576905107729e+01,
|
|
"cpu_time": 1.7510147611284069e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.4889459878190237e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1113566827677512e-03,
|
|
"cpu_time": 5.0538119197865872e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1118360041220249e-03,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 48505,
|
|
"real_time": 1.4472525933675797e+04,
|
|
"cpu_time": 3.7737375157200426e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.7962293786468433e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 48505,
|
|
"real_time": 1.4416237432042903e+04,
|
|
"cpu_time": 3.7733887949695876e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.8188608779116541e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 48505,
|
|
"real_time": 1.4436217354249269e+04,
|
|
"cpu_time": 3.7691224265539531e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.8108074948946594e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 48505,
|
|
"real_time": 1.4405755091836912e+04,
|
|
"cpu_time": 3.7691980063910960e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.8230949690054382e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 48505,
|
|
"real_time": 1.4464446583381135e+04,
|
|
"cpu_time": 3.7744896732295689e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.7994669561973120e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4439036479037202e+04,
|
|
"cpu_time": 3.7719872833728499e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.8096919353311816e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4436217354249267e+04,
|
|
"cpu_time": 3.7733887949695883e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 5.8108074948946594e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.9165954969134852e+01,
|
|
"cpu_time": 2.6113681129505814e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1733988277707369e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0199377577221579e-03,
|
|
"cpu_time": 6.9230565131055751e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0197264172215480e-03,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 33722,
|
|
"real_time": 2.0767915156319945e+04,
|
|
"cpu_time": 4.9478985499080532e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 8.0784305375469897e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 33722,
|
|
"real_time": 2.0762130330183929e+04,
|
|
"cpu_time": 4.9471432951782197e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 8.0806813815292004e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 33722,
|
|
"real_time": 2.0768493532918430e+04,
|
|
"cpu_time": 4.9477792331415818e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 8.0782055633490295e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 33722,
|
|
"real_time": 2.0764284264084552e+04,
|
|
"cpu_time": 4.9427558181602210e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 8.0798431511646753e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 33722,
|
|
"real_time": 2.0766372425028552e+04,
|
|
"cpu_time": 4.9462770713480895e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 8.0790306831728369e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0765839141707082e+04,
|
|
"cpu_time": 4.9463707935472332e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 8.0792382633525464e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0766372425028556e+04,
|
|
"cpu_time": 4.9471432951782197e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 8.0790306831728369e+11,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6368991176249361e+00,
|
|
"cpu_time": 2.1208094710318925e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0259701812246358e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2698254569105586e-04,
|
|
"cpu_time": 4.2876071357177370e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2698847933207270e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 21432,
|
|
"real_time": 3.2660908862904500e+04,
|
|
"cpu_time": 7.3497603956699895e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.0273575711210642e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 21432,
|
|
"real_time": 3.2659191073034126e+04,
|
|
"cpu_time": 7.3513924318775971e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.0274116075001333e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 21432,
|
|
"real_time": 3.2660043307253120e+04,
|
|
"cpu_time": 7.3457112821948671e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.0273847981257346e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 21432,
|
|
"real_time": 3.2659870793524053e+04,
|
|
"cpu_time": 7.3455791713325161e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.0273902249072377e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 21432,
|
|
"real_time": 3.2648911710485132e+04,
|
|
"cpu_time": 7.3474539520344144e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.0277350834093518e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2657785149440188e+04,
|
|
"cpu_time": 7.3479794466218766e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.0274558570127043e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.2659870793524053e+04,
|
|
"cpu_time": 7.3474539520344144e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.0273902249072377e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.9980324374299716e+00,
|
|
"cpu_time": 2.5496217252552537e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5727559856244606e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5304260269210715e-04,
|
|
"cpu_time": 3.4698269691369431e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5307285221938382e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 11732,
|
|
"real_time": 5.9709155668970896e+04,
|
|
"cpu_time": 1.2458988041254651e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.1239292073070549e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 11732,
|
|
"real_time": 5.9742334951388075e+04,
|
|
"cpu_time": 1.2465228187862219e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.1233050073219604e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 11732,
|
|
"real_time": 5.9725715252046284e+04,
|
|
"cpu_time": 1.2458687700306796e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.1236175861066938e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 11732,
|
|
"real_time": 5.9720244506525865e+04,
|
|
"cpu_time": 1.2466310756904184e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.1237205164601555e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 11732,
|
|
"real_time": 5.9707858378504003e+04,
|
|
"cpu_time": 1.2460508446982823e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.1239536272525310e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.9721061751487025e+04,
|
|
"cpu_time": 1.2461944626662135e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.1237051888896794e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.9720244506525851e+04,
|
|
"cpu_time": 1.2460508446982822e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.1237205164601555e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4061968640061712e+01,
|
|
"cpu_time": 3.5796803605136347e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6455639176895711e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3546079436056870e-04,
|
|
"cpu_time": 2.8724893808747672e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3543220622693957e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 6385,
|
|
"real_time": 1.0961413240540093e+05,
|
|
"cpu_time": 2.2183920438528009e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.2244564186632817e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 6385,
|
|
"real_time": 1.0959466069275090e+05,
|
|
"cpu_time": 2.2183887627251368e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.2246739681623721e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 6385,
|
|
"real_time": 1.0958491457430588e+05,
|
|
"cpu_time": 2.2177868833202700e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.2247828865987886e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 6385,
|
|
"real_time": 1.0962656297729722e+05,
|
|
"cpu_time": 2.2180983680500992e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.2243175773721504e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 6385,
|
|
"real_time": 1.0960308533206001e+05,
|
|
"cpu_time": 2.2181986499608343e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.2245798336184241e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0960467119636301e+05,
|
|
"cpu_time": 2.2181729415818281e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.2245621368830034e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0960308533206003e+05,
|
|
"cpu_time": 2.2181986499608340e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.2245798336184241e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6292742282711835e+01,
|
|
"cpu_time": 2.4992383056286158e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8202715708139816e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4865007216273164e-04,
|
|
"cpu_time": 1.1267103023294267e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4864672979741927e-04,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 3274,
|
|
"real_time": 2.1365967180106134e+05,
|
|
"cpu_time": 4.2165989156994253e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.2563693173222715e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 3274,
|
|
"real_time": 2.1364706544667509e+05,
|
|
"cpu_time": 4.2167810843006015e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.2564434500365264e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 3274,
|
|
"real_time": 2.1363050714057655e+05,
|
|
"cpu_time": 4.2180122877214482e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.2565408358243508e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 3274,
|
|
"real_time": 2.1366651773491304e+05,
|
|
"cpu_time": 4.2174246273671323e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.2563290629046357e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 3274,
|
|
"real_time": 2.1367370671883228e+05,
|
|
"cpu_time": 4.2177047495418414e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.2562867941127979e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1365549376841175e+05,
|
|
"cpu_time": 4.2173043329260900e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.2563938920401165e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1365967180106131e+05,
|
|
"cpu_time": 4.2174246273671323e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.2563693173222715e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7073835874126111e+01,
|
|
"cpu_time": 6.0154110441547957e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0040564159152652e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.9912926988121385e-05,
|
|
"cpu_time": 1.4263639920861787e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.9915735206646956e-05,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1591,
|
|
"real_time": 4.3938768061219191e+05,
|
|
"cpu_time": 8.3659965870522987e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.2218615488991094e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1591,
|
|
"real_time": 4.3965748707792535e+05,
|
|
"cpu_time": 8.3683458705216355e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.2211117239653523e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1591,
|
|
"real_time": 4.3992121742849809e+05,
|
|
"cpu_time": 8.3714079321181134e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.2203796742021416e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1591,
|
|
"real_time": 4.3989123848921340e+05,
|
|
"cpu_time": 8.3709446448774368e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.2204628440517681e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1591,
|
|
"real_time": 4.4068690507346683e+05,
|
|
"cpu_time": 8.3793852357007575e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.2182592807256172e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3990890573625907e+05,
|
|
"cpu_time": 8.3712160540540493e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.2204150143687976e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3989123848921340e+05,
|
|
"cpu_time": 8.3709446448774380e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.2204628440517681e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.8493134812537539e+02,
|
|
"cpu_time": 5.0578354627108746e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3443019152510383e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1023449214190464e-03,
|
|
"cpu_time": 6.0419363567392858e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1015121081137430e-03,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 718,
|
|
"real_time": 9.6656988675817638e+05,
|
|
"cpu_time": 1.7434463676880449e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.1108786221359250e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 718,
|
|
"real_time": 9.6408181269601267e+05,
|
|
"cpu_time": 1.7408987618384555e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.1137455450978044e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 718,
|
|
"real_time": 9.6709043430268229e+05,
|
|
"cpu_time": 1.7440616253481682e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.1102806789462439e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 718,
|
|
"real_time": 9.6376096696170012e+05,
|
|
"cpu_time": 1.7406720947075102e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.1141163222090427e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 718,
|
|
"real_time": 9.6939290407720779e+05,
|
|
"cpu_time": 1.7464233440111112e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.1076435772161184e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time_mean",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.6617920095915569e+05,
|
|
"cpu_time": 1.7431004387186584e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.1113329491210271e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time_median",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.6656988675817638e+05,
|
|
"cpu_time": 1.7434463676880449e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.1108786221359250e+12,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time_stddev",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3215168814992817e+03,
|
|
"cpu_time": 2.3890103543675482e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6689387251170511e+09,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time_cv",
|
|
"family_index": 0,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4027808497581403e-03,
|
|
"cpu_time": 1.3705523223456330e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4015653699712243e-03,
|
|
"own_gpu": NaN,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57071,
|
|
"real_time": 1.2277640844908839e+04,
|
|
"cpu_time": 3.1259668483117526e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3361458049968004e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57071,
|
|
"real_time": 1.2272897514279286e+04,
|
|
"cpu_time": 3.1260137039827576e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3374351861362654e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57071,
|
|
"real_time": 1.2287223178871951e+04,
|
|
"cpu_time": 3.1233036393264665e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3335440728732979e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57071,
|
|
"real_time": 1.2282533430852396e+04,
|
|
"cpu_time": 3.1248157540607521e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3348168951132596e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57071,
|
|
"real_time": 1.2272917350924306e+04,
|
|
"cpu_time": 3.1256994883565676e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3374297918591613e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2278642463967357e+04,
|
|
"cpu_time": 3.1251598868076595e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3358743501957572e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2277640844908841e+04,
|
|
"cpu_time": 3.1256994883565680e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3361458049968004e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.2360423866493466e+00,
|
|
"cpu_time": 1.1440544700576760e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6939376434804205e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.0787718633794432e-04,
|
|
"cpu_time": 3.6607870044893089e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.0779419895746855e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 56885,
|
|
"real_time": 1.2294333634735733e+04,
|
|
"cpu_time": 3.1233921051243571e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6632322201300728e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 56885,
|
|
"real_time": 1.2308327133311632e+04,
|
|
"cpu_time": 3.1220567179397211e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6556567040121329e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 56885,
|
|
"real_time": 1.2287549708702905e+04,
|
|
"cpu_time": 3.1235735483870911e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6669109742830586e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 56885,
|
|
"real_time": 1.2279012985039040e+04,
|
|
"cpu_time": 3.1256319398787084e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6715460029085994e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 56885,
|
|
"real_time": 1.2274431025620728e+04,
|
|
"cpu_time": 3.1261132688757873e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6740364444597340e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2288730897482008e+04,
|
|
"cpu_time": 3.1241535160411331e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6662764691587210e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2287549708702905e+04,
|
|
"cpu_time": 3.1235735483870911e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6669109742830586e+08,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3376495130333689e+01,
|
|
"cpu_time": 1.6836745201320547e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.2531831494868512e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0885172148309120e-03,
|
|
"cpu_time": 5.3892182682033327e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0880411550651149e-03,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57050,
|
|
"real_time": 1.2264766954727169e+04,
|
|
"cpu_time": 3.1442060911481138e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3358590554943378e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57050,
|
|
"real_time": 1.2269391099723452e+04,
|
|
"cpu_time": 3.1461222716914748e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3353555907407084e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57050,
|
|
"real_time": 1.2269643821661888e+04,
|
|
"cpu_time": 3.1412541051708897e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3353280859770577e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57050,
|
|
"real_time": 1.2272824415041428e+04,
|
|
"cpu_time": 3.1484896739702130e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3349820258098016e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57050,
|
|
"real_time": 1.2269082896647336e+04,
|
|
"cpu_time": 3.1457877265556850e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3353891352773492e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2269141837560255e+04,
|
|
"cpu_time": 3.1451719737072755e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3353827786598511e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2269391099723452e+04,
|
|
"cpu_time": 3.1457877265556854e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3353555907407084e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8730893750399202e+00,
|
|
"cpu_time": 2.6726380153239852e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1273525007584289e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3417199125079478e-04,
|
|
"cpu_time": 8.4975894407888131e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3419146560336380e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 56968,
|
|
"real_time": 1.2285648307072537e+04,
|
|
"cpu_time": 3.2065590858025411e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6671771143844557e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 56968,
|
|
"real_time": 1.2286419824413022e+04,
|
|
"cpu_time": 3.2038523170902816e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6670096308194051e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 56968,
|
|
"real_time": 1.2285044529815725e+04,
|
|
"cpu_time": 3.2044383232692071e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6673081990441527e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 56968,
|
|
"real_time": 1.2290796804225180e+04,
|
|
"cpu_time": 3.2081473581659942e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6660598594173665e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 56968,
|
|
"real_time": 1.2290961474093709e+04,
|
|
"cpu_time": 3.2047750035107641e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6660241405090075e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2287774187924035e+04,
|
|
"cpu_time": 3.2055544175677576e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6667157888348780e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2286419824413024e+04,
|
|
"cpu_time": 3.2047750035107641e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6670096308194051e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8766171835801675e+00,
|
|
"cpu_time": 1.7665643207570408e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.2424655129203561e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3410400773861870e-04,
|
|
"cpu_time": 5.5109478443901658e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3408814464055685e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 56912,
|
|
"real_time": 1.2302161673096774e+04,
|
|
"cpu_time": 3.2745299638037795e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3271938494613266e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 56912,
|
|
"real_time": 1.2300826157409980e+04,
|
|
"cpu_time": 3.2722652252600263e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3277722293897562e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 56912,
|
|
"real_time": 1.2301928927486528e+04,
|
|
"cpu_time": 3.2755723327241994e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3272946369874697e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 56912,
|
|
"real_time": 1.2301014995418403e+04,
|
|
"cpu_time": 3.2699490318386303e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3276904405375757e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 56912,
|
|
"real_time": 1.2302790802968519e+04,
|
|
"cpu_time": 3.2722134189626180e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3269214318581228e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2301744511276040e+04,
|
|
"cpu_time": 3.2729059945178509e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3273745176468506e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2301928927486528e+04,
|
|
"cpu_time": 3.2722652252600259e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3272946369874697e+09,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.1827970079837475e-01,
|
|
"cpu_time": 2.2011650374966759e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5436229144760873e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.6517370771952068e-05,
|
|
"cpu_time": 6.7254147879091211e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.6517247900215912e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 56875,
|
|
"real_time": 1.2306820342808471e+04,
|
|
"cpu_time": 3.4105317679120497e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0650354547232204e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 56875,
|
|
"real_time": 1.2305855699178410e+04,
|
|
"cpu_time": 3.4084221415384622e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0651189417794889e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 56875,
|
|
"real_time": 1.2306690897493863e+04,
|
|
"cpu_time": 3.4071219604395359e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0650466570724672e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 56875,
|
|
"real_time": 1.2303706587888404e+04,
|
|
"cpu_time": 3.4047068641758509e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0653049880840416e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 56875,
|
|
"real_time": 1.2304086720885627e+04,
|
|
"cpu_time": 3.4094239665934198e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0652720756389927e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2305432049650957e+04,
|
|
"cpu_time": 3.4080413401318641e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0651556234596420e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2305855699178410e+04,
|
|
"cpu_time": 3.4084221415384622e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0651189417794889e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4559071359044458e+00,
|
|
"cpu_time": 2.2484931109197458e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2602643247827021e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1831418271459577e-04,
|
|
"cpu_time": 6.5976110220328113e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1831738921767545e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 55847,
|
|
"real_time": 1.2536411318272429e+04,
|
|
"cpu_time": 3.5717081759091678e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0910609371751579e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 55847,
|
|
"real_time": 1.2537324150388355e+04,
|
|
"cpu_time": 3.5722505649363477e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0909086887721558e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 55847,
|
|
"real_time": 1.2532403422286310e+04,
|
|
"cpu_time": 3.5649655003850195e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0917296640310081e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 55847,
|
|
"real_time": 1.2531852849090365e+04,
|
|
"cpu_time": 3.5715396171683518e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0918215618772446e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 55847,
|
|
"real_time": 1.2535225937984915e+04,
|
|
"cpu_time": 3.5650308539402431e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0912586761251518e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2534643535604475e+04,
|
|
"cpu_time": 3.5690989424678257e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0913559055961441e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2535225937984917e+04,
|
|
"cpu_time": 3.5715396171683518e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 2.0912586761251518e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4215688324781577e+00,
|
|
"cpu_time": 3.7527457965090896e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0404020748831420e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9319008359509597e-04,
|
|
"cpu_time": 1.0514546828209483e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9319533629219457e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 44386,
|
|
"real_time": 1.5772299782816379e+04,
|
|
"cpu_time": 3.8948603681340777e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3241062319345581e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 44386,
|
|
"real_time": 1.5772660998966487e+04,
|
|
"cpu_time": 3.8997039269138957e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3240301052203827e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 44386,
|
|
"real_time": 1.5771001896235761e+04,
|
|
"cpu_time": 3.8961166110935534e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3243797917818882e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 44386,
|
|
"real_time": 1.5768728159214950e+04,
|
|
"cpu_time": 3.8952763709278101e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3248591434028614e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 44386,
|
|
"real_time": 1.5768194048257274e+04,
|
|
"cpu_time": 3.8946133871040380e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3249717652856079e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5770576977098168e+04,
|
|
"cpu_time": 3.8961141328346755e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3244694075250599e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5771001896235761e+04,
|
|
"cpu_time": 3.8952763709278093e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 3.3243797917818882e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0364331762274590e+00,
|
|
"cpu_time": 2.0864225569340825e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.2929342547772611e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2912864121488650e-04,
|
|
"cpu_time": 5.3551371592291507e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2913141101735047e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 36315,
|
|
"real_time": 1.9276615690533417e+04,
|
|
"cpu_time": 4.2949060415806278e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4396270425982857e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 36315,
|
|
"real_time": 1.9284122964143848e+04,
|
|
"cpu_time": 4.2897540410299087e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4375094057929497e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 36315,
|
|
"real_time": 1.9287484466328995e+04,
|
|
"cpu_time": 4.2904251604020399e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4365617342714905e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 36315,
|
|
"real_time": 1.9284951663558837e+04,
|
|
"cpu_time": 4.2932462839047512e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4372757489530380e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 36315,
|
|
"real_time": 1.9287309239470826e+04,
|
|
"cpu_time": 4.2881111937216490e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4366111259009872e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9284096804807185e+04,
|
|
"cpu_time": 4.2912885441277955e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4375170115033508e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9284951663558833e+04,
|
|
"cpu_time": 4.2904251604020399e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 5.4372757489530380e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.4298568109743064e+00,
|
|
"cpu_time": 2.7438440038192532e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2493678369120922e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.2971554518799248e-04,
|
|
"cpu_time": 6.3939862714986449e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2976807875156792e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 26681,
|
|
"real_time": 2.6190905297728208e+04,
|
|
"cpu_time": 5.0755335931936082e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 8.0071764460234467e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 26681,
|
|
"real_time": 2.6296263931496134e+04,
|
|
"cpu_time": 5.0887085229189055e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 7.9750948859626907e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 26681,
|
|
"real_time": 2.6152483487309735e+04,
|
|
"cpu_time": 5.0699319215922304e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 8.0189401554067520e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 26681,
|
|
"real_time": 2.6223216046795089e+04,
|
|
"cpu_time": 5.0808688879727531e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 7.9973104605386749e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 26681,
|
|
"real_time": 2.6292299064456070e+04,
|
|
"cpu_time": 5.0913097072823271e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 7.9762975267350800e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6231033565557049e+04,
|
|
"cpu_time": 5.0812705265919649e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 7.9949638949333298e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6223216046795085e+04,
|
|
"cpu_time": 5.0808688879727531e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 7.9973104605386749e+10,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.2948328545389053e+01,
|
|
"cpu_time": 8.9127243315764417e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9188047449560010e+08,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3997654681836124e-03,
|
|
"cpu_time": 1.7540346031436854e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4000167732739987e-03,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 19019,
|
|
"real_time": 3.6769659266737552e+04,
|
|
"cpu_time": 6.2624868237026545e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 1.1406969995488202e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 19019,
|
|
"real_time": 3.6809508012879167e+04,
|
|
"cpu_time": 6.2625958567748457e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 1.1394621190080745e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 19019,
|
|
"real_time": 3.6775865334360671e+04,
|
|
"cpu_time": 6.2571832693622106e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 1.1405045025768979e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 19019,
|
|
"real_time": 3.6808070451178763e+04,
|
|
"cpu_time": 6.2648605815236187e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 1.1395066213979384e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 19019,
|
|
"real_time": 3.6779079752126709e+04,
|
|
"cpu_time": 6.2637193175245331e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 1.1404048247720143e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6788436563456577e+04,
|
|
"cpu_time": 6.2621691697775721e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 1.1401150134607491e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6779079752126716e+04,
|
|
"cpu_time": 6.2625958567748450e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 1.1404048247720143e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8892247869440219e+01,
|
|
"cpu_time": 2.9491025746717952e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.8540845616512239e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.1353766656685390e-04,
|
|
"cpu_time": 4.7093946118618584e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.1346438670968027e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10852,
|
|
"real_time": 6.4473072750580308e+04,
|
|
"cpu_time": 9.2191917250274739e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 1.3011025598938118e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10852,
|
|
"real_time": 6.4500143772211566e+04,
|
|
"cpu_time": 9.2154519351273309e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 1.3005564808700540e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10852,
|
|
"real_time": 6.4462797381351353e+04,
|
|
"cpu_time": 9.2072973092519911e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 1.3013099556282623e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10852,
|
|
"real_time": 6.4487198115140185e+04,
|
|
"cpu_time": 9.2121718208626771e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 1.3008175645997771e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10852,
|
|
"real_time": 6.4465273952521304e+04,
|
|
"cpu_time": 9.2083592056765076e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 1.3012599630272591e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.4477697194360953e+04,
|
|
"cpu_time": 9.2124943991891967e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 1.3010093048038330e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.4473072750580301e+04,
|
|
"cpu_time": 9.2121718208626771e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 1.3011025598938118e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5746771053756024e+01,
|
|
"cpu_time": 4.9450687096751139e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1769906909718197e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4422043185396504e-04,
|
|
"cpu_time": 5.3677847664258393e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4419430969795011e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 5811,
|
|
"real_time": 1.2049189893751353e+05,
|
|
"cpu_time": 1.5421081758733073e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 1.3923936918531409e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 5811,
|
|
"real_time": 1.2048655703611216e+05,
|
|
"cpu_time": 1.5417860454311068e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 1.3924554251285925e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 5811,
|
|
"real_time": 1.2052290124824372e+05,
|
|
"cpu_time": 1.5413361400792323e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 1.3920355240572571e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 5811,
|
|
"real_time": 1.2047796726793438e+05,
|
|
"cpu_time": 1.5414054672173219e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 1.3925547036072308e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 5811,
|
|
"real_time": 1.2051687056620503e+05,
|
|
"cpu_time": 1.5416694785751528e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 1.3921051817208914e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2049923901120175e+05,
|
|
"cpu_time": 1.5416610614352243e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 1.3923089052734225e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2049189893751351e+05,
|
|
"cpu_time": 1.5416694785751525e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 1.3923936918531409e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9608476618729636e+01,
|
|
"cpu_time": 3.1084781678178881e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2655806769797452e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6272697470651087e-04,
|
|
"cpu_time": 2.0163174938880667e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6272112233131403e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 3052,
|
|
"real_time": 2.2932132817998584e+05,
|
|
"cpu_time": 2.7450098328964820e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.4632058983045993e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 3052,
|
|
"real_time": 2.2936111668267913e+05,
|
|
"cpu_time": 2.7457633453473327e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.4629520681320419e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 3052,
|
|
"real_time": 2.2931189210689504e+05,
|
|
"cpu_time": 2.7450859600263159e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.4632661085173206e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 3052,
|
|
"real_time": 2.2935257139468478e+05,
|
|
"cpu_time": 2.7450978112712293e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.4630065752459933e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 3052,
|
|
"real_time": 2.2931063191198985e+05,
|
|
"cpu_time": 2.7449119102228485e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.4632741500131705e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2933150805524690e+05,
|
|
"cpu_time": 2.7451737719528412e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.4631409600426251e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2932132817998584e+05,
|
|
"cpu_time": 2.7450859600263153e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 1.4632058983045993e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3688413539330490e+01,
|
|
"cpu_time": 3.3783328897811558e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5112752490486966e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0329332301614592e-04,
|
|
"cpu_time": 1.2306444583935763e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0328979164144713e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1560,
|
|
"real_time": 4.4858054420612269e+05,
|
|
"cpu_time": 5.1759649038460176e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.4960270762247656e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1560,
|
|
"real_time": 4.4854290330803423e+05,
|
|
"cpu_time": 5.1759570576924336e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.4961526200741910e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1560,
|
|
"real_time": 4.4858454594549799e+05,
|
|
"cpu_time": 5.1767579166666936e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.4960137304452209e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1560,
|
|
"real_time": 4.4857900718806119e+05,
|
|
"cpu_time": 5.1760489615384716e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.4960322022351224e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1560,
|
|
"real_time": 4.4857203105586965e+05,
|
|
"cpu_time": 5.1759048461537284e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.4960554683277075e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.4857180634071713e+05,
|
|
"cpu_time": 5.1761267371794692e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.4960562194614014e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.4857900718806114e+05,
|
|
"cpu_time": 5.1759649038460181e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 1.4960322022351224e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6777485026738752e+01,
|
|
"cpu_time": 3.5659969132278064e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.5957828967178492e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.7402005185308602e-05,
|
|
"cpu_time": 6.8893153013694535e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.7403560266822059e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 788,
|
|
"real_time": 8.8849936202428490e+05,
|
|
"cpu_time": 1.0049560126904177e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.5106114166948779e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 788,
|
|
"real_time": 8.8865387728875072e+05,
|
|
"cpu_time": 1.0051605723350114e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.5103487581632254e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 788,
|
|
"real_time": 8.8853794616897625e+05,
|
|
"cpu_time": 1.0050154403553004e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.5105458194407306e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 788,
|
|
"real_time": 8.8858789070580283e+05,
|
|
"cpu_time": 1.0051104530456451e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.5104609167405066e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 788,
|
|
"real_time": 8.8837205019878782e+05,
|
|
"cpu_time": 1.0048128654822101e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.5108279011025458e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.8853022527732048e+05,
|
|
"cpu_time": 1.0050110687817171e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.5105589624283774e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.8853794616897614e+05,
|
|
"cpu_time": 1.0050154403553003e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 1.5105458194407306e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0564874510727475e+02,
|
|
"cpu_time": 1.3653147594278420e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7961833648079474e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1890281512291893e-04,
|
|
"cpu_time": 1.3585071864759539e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1890852389637274e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 396,
|
|
"real_time": 1.7686724883852282e+06,
|
|
"cpu_time": 1.9812524823232947e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.5177227992338910e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 396,
|
|
"real_time": 1.7686021720609544e+06,
|
|
"cpu_time": 1.9811426767677062e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.5177831410621408e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 396,
|
|
"real_time": 1.7685823879237615e+06,
|
|
"cpu_time": 1.9812230303030692e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.5178001196491135e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 396,
|
|
"real_time": 1.7688535089190635e+06,
|
|
"cpu_time": 1.9813682272727403e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.5175674788583221e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 396,
|
|
"real_time": 1.7685120930598879e+06,
|
|
"cpu_time": 1.9810724722222832e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.5178604492070602e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7686445300697791e+06,
|
|
"cpu_time": 1.9812117777778185e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.5177467976021054e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7686021720609546e+06,
|
|
"cpu_time": 1.9812230303030692e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 1.5177831410621408e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3004733266850400e+02,
|
|
"cpu_time": 1.1229527212991857e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1159291063934125e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.3529378265384503e-05,
|
|
"cpu_time": 5.6680095176837690e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.3525380396550574e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 198,
|
|
"real_time": 3.5298669350926173e+06,
|
|
"cpu_time": 3.9451339949492868e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.5209381029710501e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 198,
|
|
"real_time": 3.5298903839606228e+06,
|
|
"cpu_time": 3.9459596919192625e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.5209279994627420e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 198,
|
|
"real_time": 3.5299461492516026e+06,
|
|
"cpu_time": 3.9454572828283929e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.5209039721861594e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 198,
|
|
"real_time": 3.5300528139082920e+06,
|
|
"cpu_time": 3.9458166414143611e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.5208580163014737e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 198,
|
|
"real_time": 3.5296156285142507e+06,
|
|
"cpu_time": 3.9450992878787224e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.5210463928787320e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5298743821454770e+06,
|
|
"cpu_time": 3.9454933797980053e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.5209348967600314e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5298903839606224e+06,
|
|
"cpu_time": 3.9454572828283934e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 1.5209279994627420e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6144145994309068e+02,
|
|
"cpu_time": 3.8978005452123716e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.9563230040704692e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.5735752172848056e-05,
|
|
"cpu_time": 9.8791207334681281e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.5737151661712559e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.0598146389003349e+06,
|
|
"cpu_time": 7.8621132828281969e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.5209207024835574e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.0598340978977652e+06,
|
|
"cpu_time": 7.8623120303031690e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.5209165103748434e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.0593153512237044e+06,
|
|
"cpu_time": 7.8616253939393843e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.5210282733917972e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.0600135562320547e+06,
|
|
"cpu_time": 7.8624813434347827e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.5208778502304440e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 99,
|
|
"real_time": 7.0593719643474827e+06,
|
|
"cpu_time": 7.8617077676766468e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.5210160753999155e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time_mean",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0596699217202682e+06,
|
|
"cpu_time": 7.8620479636364356e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.5209518823761118e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time_median",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.0598146389003349e+06,
|
|
"cpu_time": 7.8621132828281969e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 1.5209207024835574e+11,
|
|
"own_gpu": 1.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time_stddev",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.0841546763575195e+02,
|
|
"cpu_time": 3.7286535524646695e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.6446313741185069e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time_cv",
|
|
"family_index": 1,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.3686952939096998e-05,
|
|
"cpu_time": 4.7425983277009342e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.3687321414388936e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 56920,
|
|
"real_time": 1.2292831400357622e+04,
|
|
"cpu_time": 3.3593948208011454e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3320232472079945e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 56920,
|
|
"real_time": 1.2298638666061090e+04,
|
|
"cpu_time": 3.3578522189036572e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3304499068691105e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 56920,
|
|
"real_time": 1.2292122141763828e+04,
|
|
"cpu_time": 3.3640695555165403e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3322155058022022e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 56920,
|
|
"real_time": 1.2297850980710149e+04,
|
|
"cpu_time": 3.3649012473647243e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3306632243509859e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 56920,
|
|
"real_time": 1.2292546158547844e+04,
|
|
"cpu_time": 3.3645921626844691e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3321005649848807e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2294797869488106e+04,
|
|
"cpu_time": 3.3621620010541068e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3314904898430347e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2292831400357622e+04,
|
|
"cpu_time": 3.3640695555165395e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3320232472079945e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.1689875490224262e+00,
|
|
"cpu_time": 3.2893551289839152e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.5861099457204720e+04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.5775027639021822e-04,
|
|
"cpu_time": 9.7834522189966882e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5772578285597963e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 56880,
|
|
"real_time": 1.2292663033964811e+04,
|
|
"cpu_time": 3.3616116121659070e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6641377684927845e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 56880,
|
|
"real_time": 1.2296356114191813e+04,
|
|
"cpu_time": 3.3598726969057701e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6621362653487408e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 56880,
|
|
"real_time": 1.2292778448970383e+04,
|
|
"cpu_time": 3.3571382260899525e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6640751999285758e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 56880,
|
|
"real_time": 1.2291695361838667e+04,
|
|
"cpu_time": 3.3579836409986769e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6646624072975647e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 56880,
|
|
"real_time": 1.2292232794298196e+04,
|
|
"cpu_time": 3.3592787201125066e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6643710195595181e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2293145150652776e+04,
|
|
"cpu_time": 3.3591769792545630e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6638765321254361e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2292663033964811e+04,
|
|
"cpu_time": 3.3592787201125066e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6641377684927845e+08,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8446866424351249e+00,
|
|
"cpu_time": 1.7322227285104393e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.9979490696842418e+04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5005815190729857e-04,
|
|
"cpu_time": 5.1566879006620186e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5003202747658663e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 56923,
|
|
"real_time": 1.2290213490692378e+04,
|
|
"cpu_time": 3.3944966252656908e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3330931974784594e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 56923,
|
|
"real_time": 1.2300892077724258e+04,
|
|
"cpu_time": 3.3917151836690886e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3319359194826090e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 56923,
|
|
"real_time": 1.2288445227056443e+04,
|
|
"cpu_time": 3.3947453226288388e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3332850248562000e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 56923,
|
|
"real_time": 1.2290297949809090e+04,
|
|
"cpu_time": 3.3925360065351815e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3330840364414842e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 56923,
|
|
"real_time": 1.2288406840318115e+04,
|
|
"cpu_time": 3.3942855770075621e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3332891897950754e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2291651117120058e+04,
|
|
"cpu_time": 3.3935557430212721e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3329374736107657e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2290213490692378e+04,
|
|
"cpu_time": 3.3942855770075621e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3330931974784594e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.2463379454568839e+00,
|
|
"cpu_time": 1.3472692673624485e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.6862496434820723e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.2682125415597580e-04,
|
|
"cpu_time": 3.9700814407809870e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.2659537720690774e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 56904,
|
|
"real_time": 1.2294262092359062e+04,
|
|
"cpu_time": 3.4505366986503104e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6653083978391399e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 56904,
|
|
"real_time": 1.2299164623591500e+04,
|
|
"cpu_time": 3.4491556973148348e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6642459876621575e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 56904,
|
|
"real_time": 1.2296246221572083e+04,
|
|
"cpu_time": 3.4357238383945252e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6648783221755123e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 56904,
|
|
"real_time": 1.2303141124065543e+04,
|
|
"cpu_time": 3.4349135526501428e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6633848762332897e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 56904,
|
|
"real_time": 1.2295186545966311e+04,
|
|
"cpu_time": 3.4404491705328248e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6651079979547133e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2297600121510899e+04,
|
|
"cpu_time": 3.4421557915085279e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6645851163729630e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2296246221572084e+04,
|
|
"cpu_time": 3.4404491705328255e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6648783221755123e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6038592759963599e+00,
|
|
"cpu_time": 7.3479739641195323e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.8072010656828864e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.9305386745276477e-04,
|
|
"cpu_time": 2.1347011608964035e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9299874932537566e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 56889,
|
|
"real_time": 1.2282118737244182e+04,
|
|
"cpu_time": 3.5185216087468230e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3358871870591230e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 56889,
|
|
"real_time": 1.2292353370939210e+04,
|
|
"cpu_time": 3.5178188208616659e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3314445185846987e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 56889,
|
|
"real_time": 1.2311689291352342e+04,
|
|
"cpu_time": 3.5083824289405035e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3230713063910818e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 56889,
|
|
"real_time": 1.2313950955627355e+04,
|
|
"cpu_time": 3.5101654502628422e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3220936347850800e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 56889,
|
|
"real_time": 1.2303825006589155e+04,
|
|
"cpu_time": 3.5097219480039326e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3264736750484533e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2300787472350450e+04,
|
|
"cpu_time": 3.5129220513631539e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3277940643736877e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2303825006589157e+04,
|
|
"cpu_time": 3.5101654502628415e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3264736750484533e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3425808177595377e+01,
|
|
"cpu_time": 4.8420315281056155e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.8173709517066209e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0914592425707488e-03,
|
|
"cpu_time": 1.3783486958461586e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0918911056654150e-03,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 45359,
|
|
"real_time": 1.5425051566254304e+04,
|
|
"cpu_time": 3.9742926497498171e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4973459853287525e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 45359,
|
|
"real_time": 1.5432360785901743e+04,
|
|
"cpu_time": 3.9724662999625391e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4933213925209045e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 45359,
|
|
"real_time": 1.5433483098373576e+04,
|
|
"cpu_time": 3.9753569038118367e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4927037639230471e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 45359,
|
|
"real_time": 1.5431807598176016e+04,
|
|
"cpu_time": 3.9759457197028481e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4936258546595821e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 45359,
|
|
"real_time": 1.5425750854371594e+04,
|
|
"cpu_time": 3.9780602702882468e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4969607792449694e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5429690780615445e+04,
|
|
"cpu_time": 3.9752243687030583e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4947915551354523e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5431807598176016e+04,
|
|
"cpu_time": 3.9753569038118367e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.4936258546595821e+09,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.9697873844404810e+00,
|
|
"cpu_time": 2.0649027819176432e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1857352649211660e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.5728236818768814e-04,
|
|
"cpu_time": 5.1944307802463251e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5730298980671266e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 36955,
|
|
"real_time": 1.8927205759773100e+04,
|
|
"cpu_time": 4.5109470382897671e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3850116246802114e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 36955,
|
|
"real_time": 1.8934639961343808e+04,
|
|
"cpu_time": 4.5079583168719357e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3844678353281738e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 36955,
|
|
"real_time": 1.8928166315631352e+04,
|
|
"cpu_time": 4.5131611744014284e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3849413388951202e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 36955,
|
|
"real_time": 1.8936076384124946e+04,
|
|
"cpu_time": 4.5133798511703768e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3843628145679026e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 36955,
|
|
"real_time": 1.8929998381036250e+04,
|
|
"cpu_time": 4.5142496468677215e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3848073027972965e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8931217360381892e+04,
|
|
"cpu_time": 4.5119392055202465e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3847181832537409e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8929998381036250e+04,
|
|
"cpu_time": 4.5131611744014284e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3848073027972965e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.9437302576040105e+00,
|
|
"cpu_time": 2.5366650005274629e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.8844825915231314e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.0831889373671293e-04,
|
|
"cpu_time": 5.6221169767179400e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0830827719365392e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 27431,
|
|
"real_time": 2.5512716152978366e+04,
|
|
"cpu_time": 5.1737283000984942e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.0550065969310539e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 27431,
|
|
"real_time": 2.5519459175741729e+04,
|
|
"cpu_time": 5.1759783748314672e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.0544636012442513e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 27431,
|
|
"real_time": 2.5519085289900027e+04,
|
|
"cpu_time": 5.1780757792280347e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.0544937016512238e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 27431,
|
|
"real_time": 2.5519693624119533e+04,
|
|
"cpu_time": 5.1791431446174000e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.0544447269714775e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 27431,
|
|
"real_time": 2.5511040138091495e+04,
|
|
"cpu_time": 5.1772469468849617e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.0551416059949898e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5516398876166233e+04,
|
|
"cpu_time": 5.1768345091320720e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.0547100465585995e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5519085289900024e+04,
|
|
"cpu_time": 5.1772469468849617e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.0544937016512238e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.1748095957868472e+00,
|
|
"cpu_time": 2.0869927323734405e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.3619955491225743e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6361280508459040e-04,
|
|
"cpu_time": 4.0314070861101131e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6362384341058370e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 18102,
|
|
"real_time": 3.8658933983339870e+04,
|
|
"cpu_time": 6.4984051927961678e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.7123769125446796e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 18102,
|
|
"real_time": 3.8671114038680309e+04,
|
|
"cpu_time": 6.4967595293339626e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.7115226081958092e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 18102,
|
|
"real_time": 3.8672971138233072e+04,
|
|
"cpu_time": 6.5017900121531027e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.7113923992339741e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 18102,
|
|
"real_time": 3.8675233108531691e+04,
|
|
"cpu_time": 6.5000666887636333e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.7112338199939278e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 18102,
|
|
"real_time": 3.8667897323027602e+04,
|
|
"cpu_time": 6.5025142138989257e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.7117481750825623e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8669229918362507e+04,
|
|
"cpu_time": 6.4999071273891590e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.7116547830101906e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8671114038680309e+04,
|
|
"cpu_time": 6.5000666887636318e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.7115226081958092e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.3515158886966843e+00,
|
|
"cpu_time": 2.3733307803977596e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.4545438321605949e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6425245349094986e-04,
|
|
"cpu_time": 3.6513302942392409e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6427400198839597e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10845,
|
|
"real_time": 6.4631122122671142e+04,
|
|
"cpu_time": 9.1186547349011365e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.2448020877922623e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10845,
|
|
"real_time": 6.4574275153049100e+04,
|
|
"cpu_time": 9.1147263992626613e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.2476585993872757e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10845,
|
|
"real_time": 6.4572829296489312e+04,
|
|
"cpu_time": 9.1213701152602691e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.2477313180297924e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10845,
|
|
"real_time": 6.4614537204920634e+04,
|
|
"cpu_time": 9.1211377685572152e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.2456349464347694e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10845,
|
|
"real_time": 6.4617680157628165e+04,
|
|
"cpu_time": 9.1185547810051852e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.2454770813254421e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.4602088786951688e+04,
|
|
"cpu_time": 9.1188887597972949e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.2462608065939087e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.4614537204920627e+04,
|
|
"cpu_time": 9.1186547349011365e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.2456349464347694e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6789478071597465e+01,
|
|
"cpu_time": 2.6789326236948892e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3462926274476882e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.1468439449295958e-04,
|
|
"cpu_time": 2.9377840812200453e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.1472103064333450e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 6161,
|
|
"real_time": 1.1372123013518544e+05,
|
|
"cpu_time": 1.4117586933940114e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.6882330546495552e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 6161,
|
|
"real_time": 1.1362994336573756e+05,
|
|
"cpu_time": 1.4118877130336355e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.6911960666035973e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 6161,
|
|
"real_time": 1.1374693950644032e+05,
|
|
"cpu_time": 1.4119160412270151e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.6873994308765724e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 6161,
|
|
"real_time": 1.1364973368229141e+05,
|
|
"cpu_time": 1.4116518065249579e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.6905533027690193e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 6161,
|
|
"real_time": 1.1376961179661989e+05,
|
|
"cpu_time": 1.4122804642104005e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.6866645967799759e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1370349169725491e+05,
|
|
"cpu_time": 1.4118989436780041e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.6888092903357437e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1372123013518544e+05,
|
|
"cpu_time": 1.4118877130336355e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.6882330546495552e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.0978506613839670e+01,
|
|
"cpu_time": 2.3807734260845073e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9784932415844135e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.3629405485805190e-04,
|
|
"cpu_time": 1.6862208423235874e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.3635010266533392e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 3176,
|
|
"real_time": 2.2040217107966670e+05,
|
|
"cpu_time": 2.5010816152393899e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.8060459926085976e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 3176,
|
|
"real_time": 2.2034609892605850e+05,
|
|
"cpu_time": 2.5008138287154411e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.8070145289093422e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 3176,
|
|
"real_time": 2.2037501889654851e+05,
|
|
"cpu_time": 2.5009590018890746e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.8065149316846550e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 3176,
|
|
"real_time": 2.2031305018867840e+05,
|
|
"cpu_time": 2.5014052707807615e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.8075856118445587e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 3176,
|
|
"real_time": 2.2035371136463949e+05,
|
|
"cpu_time": 2.5019492065490381e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.8068830100703873e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2035801009111837e+05,
|
|
"cpu_time": 2.5012417846347415e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.8068088150235085e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2035371136463952e+05,
|
|
"cpu_time": 2.5010816152393896e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.8068830100703873e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.3246202969606685e+01,
|
|
"cpu_time": 4.5164700496712477e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.7434785849135015e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5087358501676129e-04,
|
|
"cpu_time": 1.8056911080792582e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5087383853496812e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1611,
|
|
"real_time": 4.3429186430942989e+05,
|
|
"cpu_time": 4.6921537678460189e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.8631200302767708e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1611,
|
|
"real_time": 4.3441144389147777e+05,
|
|
"cpu_time": 4.6929840657974296e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.8620566368392426e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1611,
|
|
"real_time": 4.3432502654348721e+05,
|
|
"cpu_time": 4.6921515270018799e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.8628250675580551e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1611,
|
|
"real_time": 4.3447978275655792e+05,
|
|
"cpu_time": 4.6958553693359555e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.8614491780392906e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1611,
|
|
"real_time": 4.3435681981357158e+05,
|
|
"cpu_time": 4.6930586902544211e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.8625423234291283e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3437298746290489e+05,
|
|
"cpu_time": 4.6932406840471411e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.8623986472284973e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3435681981357158e+05,
|
|
"cpu_time": 4.6929840657974296e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.8625423234291283e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4207041886842305e+01,
|
|
"cpu_time": 1.5250576812203775e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.5979798337612404e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7083714694201494e-04,
|
|
"cpu_time": 3.2494768197254865e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7082596687671500e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 810,
|
|
"real_time": 8.6374341238600516e+05,
|
|
"cpu_time": 9.1021940000001888e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.8847684993983597e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 810,
|
|
"real_time": 8.6376455922921502e+05,
|
|
"cpu_time": 9.1020751358021214e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.8846733917796394e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 810,
|
|
"real_time": 8.6370253695439501e+05,
|
|
"cpu_time": 9.1022087283952849e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.8849523492567596e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 810,
|
|
"real_time": 8.6375388530956826e+05,
|
|
"cpu_time": 9.1023567901235190e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.8847213969954109e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 810,
|
|
"real_time": 8.6374106906055496e+05,
|
|
"cpu_time": 9.1045422345681244e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.8847790387569931e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.6374109258794761e+05,
|
|
"cpu_time": 9.1026753777778486e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.8847789352374329e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.6374341238600505e+05,
|
|
"cpu_time": 9.1022087283952837e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.8847684993983597e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3485397781227213e+01,
|
|
"cpu_time": 1.0483831532342508e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0563073104357463e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.7190321246451392e-05,
|
|
"cpu_time": 1.1517307931178612e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7190924581431455e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 406,
|
|
"real_time": 1.7255519994470321e+06,
|
|
"cpu_time": 1.7959539285714119e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.8891244089720627e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 406,
|
|
"real_time": 1.7254077694654060e+06,
|
|
"cpu_time": 1.7956824088669752e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.8894495079729912e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 406,
|
|
"real_time": 1.7254050156408965e+06,
|
|
"cpu_time": 1.7957112561576222e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.8894557157104706e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 406,
|
|
"real_time": 1.7253817563757226e+06,
|
|
"cpu_time": 1.7956332290641242e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.8895081480962547e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 406,
|
|
"real_time": 1.7254751616070424e+06,
|
|
"cpu_time": 1.7958247438423259e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.8892975971614304e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7254443405072200e+06,
|
|
"cpu_time": 1.7957611133004918e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.8893670755826416e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7254077694654062e+06,
|
|
"cpu_time": 1.7957112561576224e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.8894495079729912e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9536047182977867e+01,
|
|
"cpu_time": 1.2870613212338361e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5673862112957356e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.0300371069945208e-05,
|
|
"cpu_time": 7.1672190231823282e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0299261572294134e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 203,
|
|
"real_time": 3.4507808592078723e+06,
|
|
"cpu_time": 3.5766368029556745e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.8894885962364395e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 203,
|
|
"real_time": 3.4508391930331737e+06,
|
|
"cpu_time": 3.5766266847291114e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.8894228473749031e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 203,
|
|
"real_time": 3.4507918389504915e+06,
|
|
"cpu_time": 3.5764865172413448e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.8894762206468063e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 203,
|
|
"real_time": 3.4507123037732295e+06,
|
|
"cpu_time": 3.5764318817734816e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.8895658688566345e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 203,
|
|
"real_time": 3.4504293124627303e+06,
|
|
"cpu_time": 3.5759988522166172e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.8898848765054871e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4507107014854997e+06,
|
|
"cpu_time": 3.5764361477832459e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.8895676819240540e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4507808592078723e+06,
|
|
"cpu_time": 3.5764865172413453e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.8894885962364395e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6370539088065300e+02,
|
|
"cpu_time": 2.5997572830923252e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8453486143490612e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.7441065056592351e-05,
|
|
"cpu_time": 7.2691281926107133e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.7443540394602974e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.9035427451207498e+06,
|
|
"cpu_time": 7.1380244950498110e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.8883724764320961e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.9035934007698949e+06,
|
|
"cpu_time": 7.1382870693072947e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.8883439451990875e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.9037185345619619e+06,
|
|
"cpu_time": 7.1383124554458987e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.8882734667720940e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.9037137580758864e+06,
|
|
"cpu_time": 7.1380272772278730e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.8882761569595383e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.9036852697482202e+06,
|
|
"cpu_time": 7.1380707227722229e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.8882922020833946e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9036507416553423e+06,
|
|
"cpu_time": 7.1381444039606210e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.8883116494892426e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9036852697482212e+06,
|
|
"cpu_time": 7.1380707227722229e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.8882922020833946e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.8597678121107862e+01,
|
|
"cpu_time": 1.4328330774299025e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.4268388721524528e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1384944149457638e-05,
|
|
"cpu_time": 2.0072906855665330e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1384989865032936e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.3815276288226539e+07,
|
|
"cpu_time": 1.4239926000000507e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.8860671390084656e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.3815957103289809e+07,
|
|
"cpu_time": 1.4241023941176359e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.8858756435496033e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.3815558679840144e+07,
|
|
"cpu_time": 1.4240406313725615e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.8859877073477280e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.3815762602961531e+07,
|
|
"cpu_time": 1.4240802549020195e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.8859303494757286e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 51,
|
|
"real_time": 1.3815696771238364e+07,
|
|
"cpu_time": 1.4240181980391761e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.8859488659136070e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3815650289111277e+07,
|
|
"cpu_time": 1.4240468156862888e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.8859619410590263e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3815696771238366e+07,
|
|
"cpu_time": 1.4240406313725615e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.8859488659136070e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5356401123822207e+02,
|
|
"cpu_time": 4.4731690068800481e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.1321044075363898e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8353389520728317e-05,
|
|
"cpu_time": 3.1411671004118644e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8353510702661962e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 25,
|
|
"real_time": 2.7648043558001518e+07,
|
|
"cpu_time": 2.8452818040000238e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.8836086963891243e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 25,
|
|
"real_time": 2.7645643875002861e+07,
|
|
"cpu_time": 2.8449220920001607e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.8839457993990700e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 25,
|
|
"real_time": 2.7648011744022369e+07,
|
|
"cpu_time": 2.8451291960000161e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.8836131651750618e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 25,
|
|
"real_time": 2.7646424323320389e+07,
|
|
"cpu_time": 2.8449870919998832e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.8838361570478912e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 25,
|
|
"real_time": 2.7647467628121376e+07,
|
|
"cpu_time": 2.8452427760000773e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.8836895966119263e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time_mean",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7647118225693710e+07,
|
|
"cpu_time": 2.8451125920000326e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.8837386829246147e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time_median",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7647467628121376e+07,
|
|
"cpu_time": 2.8451291960000165e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.8836895966119263e+10,
|
|
"own_gpu": 2.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time_stddev",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0526047265614156e+03,
|
|
"cpu_time": 1.5644229467206942e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4786744818789563e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time_cv",
|
|
"family_index": 2,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.8072855115264158e-05,
|
|
"cpu_time": 5.4986328172726195e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.8073480287954230e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57223,
|
|
"real_time": 1.2218297002052994e+04,
|
|
"cpu_time": 3.1227803558009626e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3523493489409900e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57223,
|
|
"real_time": 1.2241674056273891e+04,
|
|
"cpu_time": 3.1218432920329960e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3459476058348316e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57223,
|
|
"real_time": 1.2253973484322350e+04,
|
|
"cpu_time": 3.1225519598762545e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3425892468597180e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57223,
|
|
"real_time": 1.2297356670181845e+04,
|
|
"cpu_time": 3.1206267602187872e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3307971053094870e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57223,
|
|
"real_time": 1.2245617605830794e+04,
|
|
"cpu_time": 3.1205957691836458e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3448700848291028e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2251383763732378e+04,
|
|
"cpu_time": 3.1216796274225293e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3433106783548260e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2245617605830794e+04,
|
|
"cpu_time": 3.1218432920329960e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3448700848291028e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8907073417635235e+01,
|
|
"cpu_time": 1.0347259310600970e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.8770659959149768e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.3594945660920770e-03,
|
|
"cpu_time": 3.3146448532722660e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3560676089459678e-03,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57301,
|
|
"real_time": 1.2230425359162347e+04,
|
|
"cpu_time": 3.1214217203887565e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6980499528276944e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57301,
|
|
"real_time": 1.2258070462540942e+04,
|
|
"cpu_time": 3.1207473098200800e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6829441265113282e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57301,
|
|
"real_time": 1.2278142670879843e+04,
|
|
"cpu_time": 3.1197549955498202e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6720189035016060e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57301,
|
|
"real_time": 1.2270865367205246e+04,
|
|
"cpu_time": 3.1190566639325567e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6759757807250488e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57301,
|
|
"real_time": 1.2234233505966160e+04,
|
|
"cpu_time": 3.1194391895430032e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6959650524939549e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2254347473150907e+04,
|
|
"cpu_time": 3.1200839758468432e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6849907632119274e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2258070462540943e+04,
|
|
"cpu_time": 3.1197549955498202e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6829441265113282e+08,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1387800977129206e+01,
|
|
"cpu_time": 9.7582267878611919e+00,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1669460505781746e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7453235289752931e-03,
|
|
"cpu_time": 3.1275526118532257e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7456210366063301e-03,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57243,
|
|
"real_time": 1.2228846770467244e+04,
|
|
"cpu_time": 3.1487548521216675e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3397829171895001e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57243,
|
|
"real_time": 1.2229510060490220e+04,
|
|
"cpu_time": 3.1459743444613272e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3397102515931246e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57243,
|
|
"real_time": 1.2223969864640747e+04,
|
|
"cpu_time": 3.1492761717590834e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3403174403589315e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57243,
|
|
"real_time": 1.2228649486889286e+04,
|
|
"cpu_time": 3.1488322816763684e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3398045317731769e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57243,
|
|
"real_time": 1.2227821930140206e+04,
|
|
"cpu_time": 3.1469686302255403e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3398952073071396e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2227759622525544e+04,
|
|
"cpu_time": 3.1479612560487974e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3399020696443746e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2228649486889288e+04,
|
|
"cpu_time": 3.1487548521216675e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3398045317731769e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2024945622214540e+00,
|
|
"cpu_time": 1.4186796129560184e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4139440258630688e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8012249424367952e-04,
|
|
"cpu_time": 4.5066616059204353e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8015824294560252e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57181,
|
|
"real_time": 1.2241250807470449e+04,
|
|
"cpu_time": 3.1840552561165041e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6768506352310600e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57181,
|
|
"real_time": 1.2239459883710932e+04,
|
|
"cpu_time": 3.1812664433989568e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6772423220741777e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57181,
|
|
"real_time": 1.2242787732554811e+04,
|
|
"cpu_time": 3.1807464857208233e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6765145909429250e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57181,
|
|
"real_time": 1.2241469186674172e+04,
|
|
"cpu_time": 3.1782470173659091e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6768028820977316e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57181,
|
|
"real_time": 1.2240865417118894e+04,
|
|
"cpu_time": 3.1805134695091194e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6769349129656987e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2241166605505852e+04,
|
|
"cpu_time": 3.1809657344222622e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6768690686623187e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2241250807470449e+04,
|
|
"cpu_time": 3.1807464857208233e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6768506352310600e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1969034770715097e+00,
|
|
"cpu_time": 2.0784372839232610e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6173865194120642e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.7776912580632990e-05,
|
|
"cpu_time": 6.5339819961963638e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.7777905914539967e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57070,
|
|
"real_time": 1.2262352202427635e+04,
|
|
"cpu_time": 3.2592112142981889e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3444884731842499e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57070,
|
|
"real_time": 1.2264403995749364e+04,
|
|
"cpu_time": 3.2577696285264454e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3435943583327551e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57070,
|
|
"real_time": 1.2262685809307412e+04,
|
|
"cpu_time": 3.2578722393551761e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3443430761520443e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57070,
|
|
"real_time": 1.2263534835950350e+04,
|
|
"cpu_time": 3.2568962186787670e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3439730776384563e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57070,
|
|
"real_time": 1.2263602209874012e+04,
|
|
"cpu_time": 3.2574574610128213e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3439437188556099e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2263315810661756e+04,
|
|
"cpu_time": 3.2578413523742805e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3440685408326235e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2263534835950350e+04,
|
|
"cpu_time": 3.2577696285264457e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3439730776384563e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.1234012053865379e-01,
|
|
"cpu_time": 8.5487868785279311e+00,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5399703320790699e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.6241474416927541e-05,
|
|
"cpu_time": 2.6240648189629199e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.6241110214644274e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 44426,
|
|
"real_time": 1.5755399022197198e+04,
|
|
"cpu_time": 3.7601110183226338e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3191799722328529e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 44426,
|
|
"real_time": 1.5757520821214399e+04,
|
|
"cpu_time": 3.7634203799577554e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3180597688652496e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 44426,
|
|
"real_time": 1.5755654435766673e+04,
|
|
"cpu_time": 3.7637409084770457e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3190451107162800e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 44426,
|
|
"real_time": 1.5756000223375786e+04,
|
|
"cpu_time": 3.7607599648854841e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3188625375582342e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 44426,
|
|
"real_time": 1.5757668009459176e+04,
|
|
"cpu_time": 3.7604131882230184e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3179820720501757e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5756448502402645e+04,
|
|
"cpu_time": 3.7616890919731872e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3186258922845602e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5756000223375788e+04,
|
|
"cpu_time": 3.7607599648854841e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3188625375582342e+09,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0688743610645428e+00,
|
|
"cpu_time": 1.7456254631007344e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.6430261158353684e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.7837264273199239e-05,
|
|
"cpu_time": 4.6405362602284318e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.7836036731369498e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 35952,
|
|
"real_time": 1.9468445330877657e+04,
|
|
"cpu_time": 4.3319459390297598e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3465071069862478e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 35952,
|
|
"real_time": 1.9469018652335417e+04,
|
|
"cpu_time": 4.3303269692922571e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3464674551973598e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 35952,
|
|
"real_time": 1.9467025964774097e+04,
|
|
"cpu_time": 4.3312613540274753e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3466052825652662e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 35952,
|
|
"real_time": 1.9470745450162925e+04,
|
|
"cpu_time": 4.3304513684914025e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3463480413268229e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 35952,
|
|
"real_time": 1.9466708768529650e+04,
|
|
"cpu_time": 4.3319535185802764e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3466272245454676e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9468388833335950e+04,
|
|
"cpu_time": 4.3311878298842341e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3465110221242329e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9468445330877661e+04,
|
|
"cpu_time": 4.3312613540274753e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3465071069862478e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6304550827885209e+00,
|
|
"cpu_time": 7.8260315360357833e+00,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1276544287679626e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.3748845204728673e-05,
|
|
"cpu_time": 1.8069019039160353e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.3746394217330228e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 26402,
|
|
"real_time": 2.6509280831779801e+04,
|
|
"cpu_time": 5.0500101204454193e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9777526343584324e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 26402,
|
|
"real_time": 2.6512112584354589e+04,
|
|
"cpu_time": 5.0495233732292458e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9775413910598526e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 26402,
|
|
"real_time": 2.6511067602018335e+04,
|
|
"cpu_time": 5.0499602189228361e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9776193394795048e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 26402,
|
|
"real_time": 2.6512376588185783e+04,
|
|
"cpu_time": 5.0545922278613994e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9775216991811619e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 26402,
|
|
"real_time": 2.6510264319638707e+04,
|
|
"cpu_time": 5.0549358419817276e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9776792629397114e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6511020385195443e+04,
|
|
"cpu_time": 5.0518043564881264e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9776228654037323e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6511067602018338e+04,
|
|
"cpu_time": 5.0500101204454193e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9776193394795048e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2876428271862663e+00,
|
|
"cpu_time": 2.7111546952251626e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.6054380331143667e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.8570096830574088e-05,
|
|
"cpu_time": 5.3667056439808008e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.8570625881964675e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 17276,
|
|
"real_time": 4.0515799110269116e+04,
|
|
"cpu_time": 6.5419956066216742e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5880669344473785e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 17276,
|
|
"real_time": 4.0515763606914690e+04,
|
|
"cpu_time": 6.5392902002777962e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5880692023315166e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 17276,
|
|
"real_time": 4.0511681549578105e+04,
|
|
"cpu_time": 6.5430101296598034e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5883299825921940e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 17276,
|
|
"real_time": 4.0515556582419849e+04,
|
|
"cpu_time": 6.5433753067840829e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5880824267264015e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 17276,
|
|
"real_time": 4.0512246085300532e+04,
|
|
"cpu_time": 6.5439841166935272e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5882939143689331e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.0514209386896466e+04,
|
|
"cpu_time": 6.5423310720073772e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5881684920932850e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.0515556582419849e+04,
|
|
"cpu_time": 6.5430101296598034e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5880824267264015e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0616904000191765e+00,
|
|
"cpu_time": 1.8468581580990282e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3170951212421979e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.0888081767331497e-05,
|
|
"cpu_time": 2.8229359501556967e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.0889079488675191e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10289,
|
|
"real_time": 6.8334814465553543e+04,
|
|
"cpu_time": 9.3945417047330135e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0689364073083714e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10289,
|
|
"real_time": 6.8142144219185968e+04,
|
|
"cpu_time": 9.3675261249877629e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0776137499493740e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10289,
|
|
"real_time": 6.7981925743700311e+04,
|
|
"cpu_time": 9.3614845660414183e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0848670099557117e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10289,
|
|
"real_time": 6.8327692831926732e+04,
|
|
"cpu_time": 9.3993252016719198e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0692562752829948e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10289,
|
|
"real_time": 6.8131711486696819e+04,
|
|
"cpu_time": 9.3678368548934741e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0780850124533909e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8183657749412669e+04,
|
|
"cpu_time": 9.3781428904655186e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0757516909899685e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8142144219185968e+04,
|
|
"cpu_time": 9.3678368548934741e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0776137499493740e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4892283854347244e+02,
|
|
"cpu_time": 1.7421547494814325e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.7201043394261673e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.1841427030915663e-03,
|
|
"cpu_time": 1.8576756292044024e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1848656896172329e-03,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 5792,
|
|
"real_time": 1.2072106075157732e+05,
|
|
"cpu_time": 1.4748171754143783e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4743763630698532e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 5792,
|
|
"real_time": 1.2084446967688082e+05,
|
|
"cpu_time": 1.4771521667817334e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4708282565308220e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 5792,
|
|
"real_time": 1.2072299147943973e+05,
|
|
"cpu_time": 1.4754276968232269e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4743207972230621e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 5792,
|
|
"real_time": 1.2084516200836700e+05,
|
|
"cpu_time": 1.4776004178176296e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4708083718813644e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 5792,
|
|
"real_time": 1.2072385216040516e+05,
|
|
"cpu_time": 1.4761979247237407e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4742960276209961e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2077150721533402e+05,
|
|
"cpu_time": 1.4762390763121418e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4729259632652199e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2072385216040513e+05,
|
|
"cpu_time": 1.4761979247237407e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4742960276209961e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.6929395512893478e+01,
|
|
"cpu_time": 1.1585354108160281e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9242443004051227e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.5418200083865181e-04,
|
|
"cpu_time": 7.8478847322631287e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.5407006102599495e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 3015,
|
|
"real_time": 2.3212904277258393e+05,
|
|
"cpu_time": 2.6022232570478087e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6137692637703644e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 3015,
|
|
"real_time": 2.3215918917174931e+05,
|
|
"cpu_time": 2.6028630613599313e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6133000076056358e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 3015,
|
|
"real_time": 2.3213021463373210e+05,
|
|
"cpu_time": 2.6028078009950297e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6137510204072357e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 3015,
|
|
"real_time": 2.3215351011187062e+05,
|
|
"cpu_time": 2.6026375986731265e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6133883980292526e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 3015,
|
|
"real_time": 2.3212771798842348e+05,
|
|
"cpu_time": 2.6022625671640839e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6137898880384254e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3213993493567189e+05,
|
|
"cpu_time": 2.6025588570479961e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6135997155701828e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3213021463373210e+05,
|
|
"cpu_time": 2.6026375986731268e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6137510204072357e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5144208933000286e+01,
|
|
"cpu_time": 3.0046990064203175e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3573569948737081e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.5237413533336628e-05,
|
|
"cpu_time": 1.1545172161172395e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.5235697930691953e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1538,
|
|
"real_time": 4.5517874964205874e+05,
|
|
"cpu_time": 4.8925332964892109e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6858522093118767e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1538,
|
|
"real_time": 4.5520611214378831e+05,
|
|
"cpu_time": 4.8927714499350835e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6856306522308945e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1538,
|
|
"real_time": 4.5517771081697382e+05,
|
|
"cpu_time": 4.8920610858251480e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6858606213137024e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1538,
|
|
"real_time": 4.5519882965768158e+05,
|
|
"cpu_time": 4.8928543823144294e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6856896166927307e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1538,
|
|
"real_time": 4.5516146904403181e+05,
|
|
"cpu_time": 4.8933293237969477e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6859921458723015e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5518457426090690e+05,
|
|
"cpu_time": 4.8927099076721631e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6858050490843010e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5517874964205874e+05,
|
|
"cpu_time": 4.8927714499350835e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6858522093118767e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7902377917613968e+01,
|
|
"cpu_time": 4.6381516000840733e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4496247238509697e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.9329931043209125e-05,
|
|
"cpu_time": 9.4797191895866907e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.9329934832313321e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 776,
|
|
"real_time": 9.0165870132455218e+05,
|
|
"cpu_time": 9.4742139948441042e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7214116550650444e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 776,
|
|
"real_time": 9.0166447483442398e+05,
|
|
"cpu_time": 9.4743642783495155e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7213878262378838e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 776,
|
|
"real_time": 9.0166405837961298e+05,
|
|
"cpu_time": 9.4741972036072961e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7213895450486198e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 776,
|
|
"real_time": 9.0165725588431279e+05,
|
|
"cpu_time": 9.4742789304114971e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7214176208332100e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 776,
|
|
"real_time": 9.0167354480315419e+05,
|
|
"cpu_time": 9.4745992525775405e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7213503926551735e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.0166360704521113e+05,
|
|
"cpu_time": 9.4743307319579925e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7213914079679863e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.0166405837961286e+05,
|
|
"cpu_time": 9.4742789304114971e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7213895450486198e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.4054533835177025e+00,
|
|
"cpu_time": 1.6383635801129611e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6436756956934032e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.1040389491915258e-06,
|
|
"cpu_time": 1.7292657671180671e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.1039979563368352e-06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 390,
|
|
"real_time": 1.7947245159377465e+06,
|
|
"cpu_time": 1.8645698743590889e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7392292468315407e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 390,
|
|
"real_time": 1.7946711177221285e+06,
|
|
"cpu_time": 1.8644369051283041e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7393405029650986e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 390,
|
|
"real_time": 1.7946739925835759e+06,
|
|
"cpu_time": 1.8645012333335076e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7393345129714310e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 390,
|
|
"real_time": 1.7946776856357853e+06,
|
|
"cpu_time": 1.8644881307692137e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7393268182429039e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 390,
|
|
"real_time": 1.7947157859826125e+06,
|
|
"cpu_time": 1.8646701384615228e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7392474353958885e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7946926195723698e+06,
|
|
"cpu_time": 1.8645332564103275e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7392957032813728e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7946776856357851e+06,
|
|
"cpu_time": 1.8645012333335076e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7393268182429039e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5428213735452722e+01,
|
|
"cpu_time": 9.0031859627769705e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.2980126205965190e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4168562046859939e-05,
|
|
"cpu_time": 4.8286539978966412e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4168477277545377e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 195,
|
|
"real_time": 3.5809334725714647e+06,
|
|
"cpu_time": 3.7104005487184948e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7481212378854500e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 195,
|
|
"real_time": 3.5809064021286294e+06,
|
|
"cpu_time": 3.7102641897438215e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7481495724159615e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 195,
|
|
"real_time": 3.5808358257875228e+06,
|
|
"cpu_time": 3.7098207948713773e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7482234464206940e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 195,
|
|
"real_time": 3.5809506889050589e+06,
|
|
"cpu_time": 3.7103595999998362e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7481032178368118e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 195,
|
|
"real_time": 3.5808858819879019e+06,
|
|
"cpu_time": 3.7101362615384208e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7481710510553894e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5809024542761156e+06,
|
|
"cpu_time": 3.7101962789743901e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7481537051228615e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5809064021286294e+06,
|
|
"cpu_time": 3.7102641897438215e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7481495724159615e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.4772437669193451e+01,
|
|
"cpu_time": 2.3323923628735528e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.6863874905944342e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2503115692450288e-05,
|
|
"cpu_time": 6.2864392810999656e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2503189194693973e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 98,
|
|
"real_time": 7.1530135357942507e+06,
|
|
"cpu_time": 7.3861540612242334e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7527603527761765e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 98,
|
|
"real_time": 7.1529531233696910e+06,
|
|
"cpu_time": 7.3860815816320907e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7527920478464211e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 98,
|
|
"real_time": 7.1530951690689037e+06,
|
|
"cpu_time": 7.3863855102039063e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7527175251457115e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 98,
|
|
"real_time": 7.1528486346787941e+06,
|
|
"cpu_time": 7.3859355918371798e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7528468685686699e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 98,
|
|
"real_time": 7.1528763985451385e+06,
|
|
"cpu_time": 7.3858466428574435e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7528323019058250e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1529573722913563e+06,
|
|
"cpu_time": 7.3860806775509715e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7527898192485619e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1529531233696919e+06,
|
|
"cpu_time": 7.3860815816320907e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7527920478464211e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0067390775274653e+02,
|
|
"cpu_time": 2.0864502088968885e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.2818203892218822e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4074445367552494e-05,
|
|
"cpu_time": 2.8248408052708953e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4074383708170166e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4297099650970528e+07,
|
|
"cpu_time": 1.4720706857142616e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7551036581294006e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4296812347459549e+07,
|
|
"cpu_time": 1.4720467816325290e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7551791193188492e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4297027768073034e+07,
|
|
"cpu_time": 1.4720667795918232e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7551225381187042e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4296629542142762e+07,
|
|
"cpu_time": 1.4719948102041673e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7552271353009712e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4296616408593800e+07,
|
|
"cpu_time": 1.4720305285714969e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7552305850304756e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4296837143447936e+07,
|
|
"cpu_time": 1.4720419171428556e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7551726071796806e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4296812347459551e+07,
|
|
"cpu_time": 1.4720467816325292e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7551791193188492e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2230878368830494e+02,
|
|
"cpu_time": 3.0895786428595409e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.8391013007140066e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5549508010601237e-05,
|
|
"cpu_time": 2.0988387673472137e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5549488429772762e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 24,
|
|
"real_time": 2.8585107919449609e+07,
|
|
"cpu_time": 2.9388081958330762e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7562979542554565e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 24,
|
|
"real_time": 2.8585060965269804e+07,
|
|
"cpu_time": 2.9387933333334595e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7563041243976067e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 24,
|
|
"real_time": 2.8584861118967336e+07,
|
|
"cpu_time": 2.9388330291666877e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7563303859731689e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 24,
|
|
"real_time": 2.8585374355316162e+07,
|
|
"cpu_time": 2.9388557791667342e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7562629429070641e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 24,
|
|
"real_time": 2.8585894654194515e+07,
|
|
"cpu_time": 2.9389574333336551e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7561945742441399e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time_mean",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8585259802639484e+07,
|
|
"cpu_time": 2.9388495541667223e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7562779963554878e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time_median",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8585107919449609e+07,
|
|
"cpu_time": 2.9388330291666877e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7562979542554565e+10,
|
|
"own_gpu": 3.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time_stddev",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.9928460399800508e+02,
|
|
"cpu_time": 6.4847967017841904e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.2467854204264923e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time_cv",
|
|
"family_index": 3,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3968199231169354e-05,
|
|
"cpu_time": 2.2065766151894361e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3968043434264351e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57080,
|
|
"real_time": 1.2261260430753995e+04,
|
|
"cpu_time": 3.3466426769445090e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3406027244363165e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57080,
|
|
"real_time": 1.2264207094591759e+04,
|
|
"cpu_time": 3.3403966590749093e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3398000934004486e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57080,
|
|
"real_time": 1.2263321058495751e+04,
|
|
"cpu_time": 3.3506142466712598e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3400413969936669e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57080,
|
|
"real_time": 1.2261616701992434e+04,
|
|
"cpu_time": 3.3458255693763873e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3405056605092102e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57080,
|
|
"real_time": 1.2264576575842399e+04,
|
|
"cpu_time": 3.3634982568323830e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3396994789595205e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2262996372335268e+04,
|
|
"cpu_time": 3.3493954817798898e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3401298708598328e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2263321058495752e+04,
|
|
"cpu_time": 3.3466426769445083e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3400413969936669e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4987741705402411e+00,
|
|
"cpu_time": 8.6843921943942718e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.0823594403236959e+04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2221924601734399e-04,
|
|
"cpu_time": 2.5928237622686858e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2222157814698368e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57081,
|
|
"real_time": 1.2262265415366213e+04,
|
|
"cpu_time": 3.3627538603038796e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6806578739800870e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57081,
|
|
"real_time": 1.2260440714283041e+04,
|
|
"cpu_time": 3.3563467370928978e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6816521452255535e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57081,
|
|
"real_time": 1.2263144969048621e+04,
|
|
"cpu_time": 3.3563416478336221e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6801787149023151e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57081,
|
|
"real_time": 1.2260672591328328e+04,
|
|
"cpu_time": 3.3634224452970288e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6815257800734353e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57081,
|
|
"real_time": 1.2259723206693387e+04,
|
|
"cpu_time": 3.3590401271877461e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6820431928899097e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2261249379343919e+04,
|
|
"cpu_time": 3.3595809635430356e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6812115414142609e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2260672591328326e+04,
|
|
"cpu_time": 3.3590401271877468e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6815257800734353e+08,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4090259136127774e+00,
|
|
"cpu_time": 3.3937496544479316e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.6775647701598718e+04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1491699336827063e-04,
|
|
"cpu_time": 1.0101705216441223e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1491276279114350e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57071,
|
|
"real_time": 1.2263166340044489e+04,
|
|
"cpu_time": 3.3963914159556312e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3360334146736004e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57071,
|
|
"real_time": 1.2262319483218602e+04,
|
|
"cpu_time": 3.3922213383328257e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3361256834338770e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57071,
|
|
"real_time": 1.2261537826315061e+04,
|
|
"cpu_time": 3.3978376811340248e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3362108596882138e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57071,
|
|
"real_time": 1.2261290521540757e+04,
|
|
"cpu_time": 3.4039529673563251e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3362378104667227e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57071,
|
|
"real_time": 1.2261635351952569e+04,
|
|
"cpu_time": 3.4011537768743729e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3362002318386495e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2261989904614295e+04,
|
|
"cpu_time": 3.3983114359306361e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3361616000202129e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2261635351952567e+04,
|
|
"cpu_time": 3.3978376811340248e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3362002318386495e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.6023703650054464e-01,
|
|
"cpu_time": 4.4973877954651769e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.2837861874869748e+04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.1999483151952407e-05,
|
|
"cpu_time": 1.3234183741707461e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.1996888605103312e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57086,
|
|
"real_time": 1.2264711294775700e+04,
|
|
"cpu_time": 3.4484693742772804e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6717302358317981e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57086,
|
|
"real_time": 1.2263937395814692e+04,
|
|
"cpu_time": 3.4382277108222515e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6718988317065873e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57086,
|
|
"real_time": 1.2264811948735653e+04,
|
|
"cpu_time": 3.4418894685211497e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6717083096718798e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57086,
|
|
"real_time": 1.2263997236809302e+04,
|
|
"cpu_time": 3.4454998563572080e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6718857944332991e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57086,
|
|
"real_time": 1.2265588813265000e+04,
|
|
"cpu_time": 3.4458759065270744e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6715390919155903e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2264609337880069e+04,
|
|
"cpu_time": 3.4439924633009927e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6717524527118311e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2264711294775701e+04,
|
|
"cpu_time": 3.4454998563572080e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6717302358317981e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.7767842518035437e-01,
|
|
"cpu_time": 3.9848865192542419e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4762440123502619e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.5254790960793102e-05,
|
|
"cpu_time": 1.1570543669061381e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.5253772139401348e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57122,
|
|
"real_time": 1.2253964631494018e+04,
|
|
"cpu_time": 3.5281728826021135e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3481466587201805e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57122,
|
|
"real_time": 1.2252925193353705e+04,
|
|
"cpu_time": 3.5219073754419805e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3486003518203449e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57122,
|
|
"real_time": 1.2250869435134899e+04,
|
|
"cpu_time": 3.5227314589825481e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3494978741709490e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57122,
|
|
"real_time": 1.2251511444566187e+04,
|
|
"cpu_time": 3.5219100591716211e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3492175472820244e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57122,
|
|
"real_time": 1.2253808193994926e+04,
|
|
"cpu_time": 3.5239633836349014e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3482149355101233e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2252615779708749e+04,
|
|
"cpu_time": 3.5237370319666334e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3487354735007248e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2252925193353705e+04,
|
|
"cpu_time": 3.5227314589825481e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3486003518203449e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3789729168077893e+00,
|
|
"cpu_time": 2.6184277504215817e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.0199053281592391e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1254518558326721e-04,
|
|
"cpu_time": 7.4308262128181880e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1254819682116820e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 44488,
|
|
"real_time": 1.5733384244771998e+04,
|
|
"cpu_time": 4.0322709629565405e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3308204999540091e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 44488,
|
|
"real_time": 1.5731974807325549e+04,
|
|
"cpu_time": 4.0266409121561825e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3315668633645859e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 44488,
|
|
"real_time": 1.5731129476158418e+04,
|
|
"cpu_time": 4.0400497819637938e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3320145701329594e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 44488,
|
|
"real_time": 1.5733265740094212e+04,
|
|
"cpu_time": 4.0354951582451075e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3308832486048851e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 44488,
|
|
"real_time": 1.5733014659922552e+04,
|
|
"cpu_time": 4.0409858770905441e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3310161995771780e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5732553785654547e+04,
|
|
"cpu_time": 4.0350885384824338e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3312602763267231e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5733014659922552e+04,
|
|
"cpu_time": 4.0354951582451075e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.3310161995771780e+09,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.7118740433920669e-01,
|
|
"cpu_time": 5.8919224697651430e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.1431543441743997e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.1731071609287411e-05,
|
|
"cpu_time": 1.4601717938960144e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.1733209305543763e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 35977,
|
|
"real_time": 1.9455984883557852e+04,
|
|
"cpu_time": 4.5984581093476023e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3473694678984690e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 35977,
|
|
"real_time": 1.9454792625772730e+04,
|
|
"cpu_time": 4.5870060205132191e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3474520394153410e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 35977,
|
|
"real_time": 1.9454094593126858e+04,
|
|
"cpu_time": 4.5962629819048969e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3475003873612068e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 35977,
|
|
"real_time": 1.9456513242316149e+04,
|
|
"cpu_time": 4.6037705422908504e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3473328788935349e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 35977,
|
|
"real_time": 1.9454925463418487e+04,
|
|
"cpu_time": 4.5965871306668720e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3474428390533541e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9455262161638417e+04,
|
|
"cpu_time": 4.5964169569446887e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3474195225243813e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9454925463418487e+04,
|
|
"cpu_time": 4.5965871306668720e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3474428390533541e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.7269924831009558e-01,
|
|
"cpu_time": 6.0592885825319378e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.7366002962918917e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.9996717609287674e-05,
|
|
"cpu_time": 1.3182634733293742e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9996309120346698e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 26436,
|
|
"real_time": 2.6477830882206774e+04,
|
|
"cpu_time": 5.3052355651386948e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9801017777189747e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 26436,
|
|
"real_time": 2.6475434148644832e+04,
|
|
"cpu_time": 5.3037153124525619e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9802810297894062e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 26436,
|
|
"real_time": 2.6477728112677029e+04,
|
|
"cpu_time": 5.3004312036618081e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9801094632019466e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 26436,
|
|
"real_time": 2.6477073736881299e+04,
|
|
"cpu_time": 5.2977924534724276e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9801584012272923e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 26436,
|
|
"real_time": 2.6475554886509151e+04,
|
|
"cpu_time": 5.2988329172337886e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9802719990097565e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6476724353383819e+04,
|
|
"cpu_time": 5.3012014903918564e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9801845341894753e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.6477073736881302e+04,
|
|
"cpu_time": 5.3004312036618081e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9801584012272923e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1604189728995216e+00,
|
|
"cpu_time": 3.1787605125958020e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.6788129695252678e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.3827890392008258e-05,
|
|
"cpu_time": 5.9963020050400552e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.3828303977122317e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 17285,
|
|
"real_time": 4.0499767597695965e+04,
|
|
"cpu_time": 6.7159375990744375e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5890914002668339e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 17285,
|
|
"real_time": 4.0495428006946466e+04,
|
|
"cpu_time": 6.7049465490309609e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5893688537385761e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 17285,
|
|
"real_time": 4.0493825183149725e+04,
|
|
"cpu_time": 6.7088463754701530e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5894713459580326e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 17285,
|
|
"real_time": 4.0496593906015718e+04,
|
|
"cpu_time": 6.7113299971074928e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5892943056730392e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 17285,
|
|
"real_time": 4.0492557072135089e+04,
|
|
"cpu_time": 6.7105251374020518e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5895524407905975e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.0495634353188594e+04,
|
|
"cpu_time": 6.7103171316170206e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5893556692854160e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.0495428006946466e+04,
|
|
"cpu_time": 6.7105251374020518e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.5893688537385761e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7744272732519630e+00,
|
|
"cpu_time": 3.9895205939404136e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7739613697710556e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.8511762256997629e-05,
|
|
"cpu_time": 5.9453532756938989e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.8509760586911396e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10286,
|
|
"real_time": 6.8078314875118624e+04,
|
|
"cpu_time": 9.4875103344345247e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0804992806402004e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10286,
|
|
"real_time": 6.8191617501816931e+04,
|
|
"cpu_time": 9.5045485125409192e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0753809292529575e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10286,
|
|
"real_time": 6.8213138921454243e+04,
|
|
"cpu_time": 9.5196397141746420e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0744106387111420e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10286,
|
|
"real_time": 6.8234924239211236e+04,
|
|
"cpu_time": 9.5123244215434985e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0734290737218559e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10286,
|
|
"real_time": 6.8080320686723760e+04,
|
|
"cpu_time": 9.4931478806143146e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0804085216492855e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8159663244864976e+04,
|
|
"cpu_time": 9.5034341726615807e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0768256887950886e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8191617501816931e+04,
|
|
"cpu_time": 9.5045485125409177e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 3.0753809292529575e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4929531204256250e+01,
|
|
"cpu_time": 1.3250229075015477e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.3833695960518412e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0993236708795112e-03,
|
|
"cpu_time": 1.3942569427304772e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0996299232592529e-03,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 5794,
|
|
"real_time": 1.2073241482629217e+05,
|
|
"cpu_time": 1.4843050949258605e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4740496212510086e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 5794,
|
|
"real_time": 1.2081014825032407e+05,
|
|
"cpu_time": 1.4856975578183771e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4718142976773880e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 5794,
|
|
"real_time": 1.2072965329522437e+05,
|
|
"cpu_time": 1.4848015913012842e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4741290855391792e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 5794,
|
|
"real_time": 1.2081296236820599e+05,
|
|
"cpu_time": 1.4863603658958140e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4717334280876831e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 5794,
|
|
"real_time": 1.2073152597275631e+05,
|
|
"cpu_time": 1.4857569209525562e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4740751980112190e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2076334094256058e+05,
|
|
"cpu_time": 1.4853843061787786e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4731603261132957e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2073241482629217e+05,
|
|
"cpu_time": 1.4856975578183771e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.4740496212510086e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.4036020746693382e+01,
|
|
"cpu_time": 8.2065743221675888e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2663097744131962e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.6464725472971563e-04,
|
|
"cpu_time": 5.5248828791515836e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.6459870996807207e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 3016,
|
|
"real_time": 2.3213452258501176e+05,
|
|
"cpu_time": 2.6187314588859299e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6136839564343315e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 3016,
|
|
"real_time": 2.3208592761243417e+05,
|
|
"cpu_time": 2.6185974535811110e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6144406023653175e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 3016,
|
|
"real_time": 2.3214040961635634e+05,
|
|
"cpu_time": 2.6197133720158800e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6135923141788704e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 3016,
|
|
"real_time": 2.3208629211475109e+05,
|
|
"cpu_time": 2.6200249966840190e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6144349257182312e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 3016,
|
|
"real_time": 2.3213807281326110e+05,
|
|
"cpu_time": 2.6197573872679184e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6136286901752869e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3211704494836292e+05,
|
|
"cpu_time": 2.6193649336869718e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6139560977744072e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3213452258501179e+05,
|
|
"cpu_time": 2.6197133720158800e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.6136839564343315e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8317715095299160e+01,
|
|
"cpu_time": 6.5219404324282337e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.4091270991705377e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2199756851806707e-04,
|
|
"cpu_time": 2.4898937710248967e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2200278531014317e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1538,
|
|
"real_time": 4.5517738917990861e+05,
|
|
"cpu_time": 4.9009228933679679e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6858632258134453e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1538,
|
|
"real_time": 4.5516854742486199e+05,
|
|
"cpu_time": 4.9008389271775843e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6859348245650780e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1538,
|
|
"real_time": 4.5518248715672299e+05,
|
|
"cpu_time": 4.9018049739923287e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6858219446881912e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1538,
|
|
"real_time": 4.5516928221142682e+05,
|
|
"cpu_time": 4.9010239271785005e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6859288743054855e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1538,
|
|
"real_time": 4.5518748591937392e+05,
|
|
"cpu_time": 4.9018367620288976e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6857814678525017e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5517703837845894e+05,
|
|
"cpu_time": 4.9012854967490555e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6858660674449402e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5517738917990867e+05,
|
|
"cpu_time": 4.9010239271785005e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.6858632258134453e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2343433709861049e+00,
|
|
"cpu_time": 4.9322262714033982e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.6678761740152317e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.8090419060505475e-05,
|
|
"cpu_time": 1.0063127876706764e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8090391924190114e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 776,
|
|
"real_time": 9.0167600587706466e+05,
|
|
"cpu_time": 9.4810044329899934e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7213402354387192e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 776,
|
|
"real_time": 9.0157517595648684e+05,
|
|
"cpu_time": 9.4791530283514143e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7217564208555199e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 776,
|
|
"real_time": 9.0167518616918009e+05,
|
|
"cpu_time": 9.4819675773191871e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7213436184883797e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 776,
|
|
"real_time": 9.0153187763284333e+05,
|
|
"cpu_time": 9.4801565721656953e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7219351675177628e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 776,
|
|
"real_time": 9.0167930736158462e+05,
|
|
"cpu_time": 9.4810102448442671e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7213266098103165e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.0162751059943193e+05,
|
|
"cpu_time": 9.4806583711341117e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7215404104221390e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.0167518616918009e+05,
|
|
"cpu_time": 9.4810044329899922e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.7213436184883797e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9268021200381739e+01,
|
|
"cpu_time": 1.0577634214086231e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.8592022168989726e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.6825540909161087e-05,
|
|
"cpu_time": 1.1157067157162943e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.6828460840887386e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 390,
|
|
"real_time": 1.7946201666162754e+06,
|
|
"cpu_time": 1.8652801282049343e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7394466666744621e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 390,
|
|
"real_time": 1.7946710424999206e+06,
|
|
"cpu_time": 1.8651841025641074e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7393406596965790e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 390,
|
|
"real_time": 1.7946000581081861e+06,
|
|
"cpu_time": 1.8653105179489616e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7394885672044479e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 390,
|
|
"real_time": 1.7946025204414930e+06,
|
|
"cpu_time": 1.8652506615382708e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7394834363372253e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 390,
|
|
"real_time": 1.7946295956603417e+06,
|
|
"cpu_time": 1.8652857538462251e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7394270194962990e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7946246766652435e+06,
|
|
"cpu_time": 1.8652622328204997e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7394372698818024e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7946201666162754e+06,
|
|
"cpu_time": 1.8652801282049343e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.7394466666744621e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8684919785585965e+01,
|
|
"cpu_time": 4.8587976406616477e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.9769695599024091e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5983798817972371e-05,
|
|
"cpu_time": 2.6048871601901055e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5983606966861438e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 195,
|
|
"real_time": 3.5809293198279846e+06,
|
|
"cpu_time": 3.7112841538461135e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7481255845185837e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 195,
|
|
"real_time": 3.5809227301237676e+06,
|
|
"cpu_time": 3.7111114717950281e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7481324819137054e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 195,
|
|
"real_time": 3.5809711469576145e+06,
|
|
"cpu_time": 3.7115209076925265e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7480818049603973e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 195,
|
|
"real_time": 3.5810384457596601e+06,
|
|
"cpu_time": 3.7112968461536719e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7480113668963379e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 195,
|
|
"real_time": 3.5808997622762737e+06,
|
|
"cpu_time": 3.7115032820513360e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7481565223898285e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5809522809890606e+06,
|
|
"cpu_time": 3.7113433323077350e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7481015521357712e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5809293198279841e+06,
|
|
"cpu_time": 3.7112968461536719e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.7481255845185837e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.4632798121335497e+01,
|
|
"cpu_time": 1.7068925425636652e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.7182257611955132e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5256499901262549e-05,
|
|
"cpu_time": 4.5991232546580625e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5256325586848390e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 98,
|
|
"real_time": 7.1529610871280339e+06,
|
|
"cpu_time": 7.3871267959185494e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7527878696706123e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 98,
|
|
"real_time": 7.1530623304448565e+06,
|
|
"cpu_time": 7.3867885204080902e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7527347533025864e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 98,
|
|
"real_time": 7.1530019180202968e+06,
|
|
"cpu_time": 7.3868247040817253e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7527664479404144e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 98,
|
|
"real_time": 7.1528761799694328e+06,
|
|
"cpu_time": 7.3870431734687779e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7528324165838860e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 98,
|
|
"real_time": 7.1527537538156826e+06,
|
|
"cpu_time": 7.3865333469389407e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7528966498644157e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1529310538756605e+06,
|
|
"cpu_time": 7.3868633081632182e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7528036274723839e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1529610871280329e+06,
|
|
"cpu_time": 7.3868247040817244e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.7527878696706123e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2000386942569791e+02,
|
|
"cpu_time": 2.3331052692737177e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.2960752625742962e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6776880487429894e-05,
|
|
"cpu_time": 3.1584519327647564e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6776990984777094e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4296756221019492e+07,
|
|
"cpu_time": 1.4720904367346074e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7551938614626259e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4296965160388120e+07,
|
|
"cpu_time": 1.4721432306122996e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7551389821350418e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4297036967259280e+07,
|
|
"cpu_time": 1.4721083285714522e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7551201219487183e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4296870488597423e+07,
|
|
"cpu_time": 1.4721003734693186e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7551638481175690e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 49,
|
|
"real_time": 1.4296811682229139e+07,
|
|
"cpu_time": 1.4721384448978137e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7551792940472710e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4296888103898693e+07,
|
|
"cpu_time": 1.4721161628570985e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7551592215422455e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4296870488597419e+07,
|
|
"cpu_time": 1.4721083285714522e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.7551638481175690e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1362376407920799e+02,
|
|
"cpu_time": 2.3460955131261366e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9843880069454777e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.9474472524005646e-06,
|
|
"cpu_time": 1.5936891206824397e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.9474339991362288e-06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 24,
|
|
"real_time": 2.8586001368239522e+07,
|
|
"cpu_time": 2.9390131999998439e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7561805520410454e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 24,
|
|
"real_time": 2.8586894972249866e+07,
|
|
"cpu_time": 2.9391024958333865e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7560631367705818e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 24,
|
|
"real_time": 2.8584967677791912e+07,
|
|
"cpu_time": 2.9388645083334569e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7563163831533943e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 24,
|
|
"real_time": 2.8585627830276888e+07,
|
|
"cpu_time": 2.9389113041664436e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7562296353090088e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 24,
|
|
"real_time": 2.8585280912617844e+07,
|
|
"cpu_time": 2.9389290458330680e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7562752217909431e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time_mean",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8585754552235208e+07,
|
|
"cpu_time": 2.9389641108332403e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7562129858129944e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time_median",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8585627830276888e+07,
|
|
"cpu_time": 2.9389290458330683e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.7562296353090088e+10,
|
|
"own_gpu": 4.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time_stddev",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4515773364450092e+02,
|
|
"cpu_time": 9.4205757977892677e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.7913524491767737e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time_cv",
|
|
"family_index": 4,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.6067450214856572e-05,
|
|
"cpu_time": 3.2054068857337606e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6067085349414856e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57030,
|
|
"real_time": 1.2250648481064340e+04,
|
|
"cpu_time": 3.1200915360337596e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3434964739467716e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57030,
|
|
"real_time": 1.2261579275654864e+04,
|
|
"cpu_time": 3.1172345204277972e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3405158568215847e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57030,
|
|
"real_time": 1.2264264368075752e+04,
|
|
"cpu_time": 3.1213434543221858e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3397844967057389e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57030,
|
|
"real_time": 1.2252449458469930e+04,
|
|
"cpu_time": 3.1194847518851140e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3430050161671942e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57030,
|
|
"real_time": 1.2257444071804395e+04,
|
|
"cpu_time": 3.1201317482025388e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3416428221132690e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2257277131013858e+04,
|
|
"cpu_time": 3.1196572021742795e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3416889331509119e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2257444071804395e+04,
|
|
"cpu_time": 3.1200915360337600e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3416428221132690e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.8009314339842728e+00,
|
|
"cpu_time": 1.5129781954491600e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5814859076197929e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.7326427982170047e-04,
|
|
"cpu_time": 4.8498219432400232e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.7325946228292230e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57241,
|
|
"real_time": 1.2252486989997618e+04,
|
|
"cpu_time": 3.1204193008506416e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6859895519069576e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57241,
|
|
"real_time": 1.2255875393931577e+04,
|
|
"cpu_time": 3.1136875229294037e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6841410643390024e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57241,
|
|
"real_time": 1.2254429806069651e+04,
|
|
"cpu_time": 3.1198086266836173e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6849295557941687e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57241,
|
|
"real_time": 1.2251700876519912e+04,
|
|
"cpu_time": 3.1190809996330456e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6864185491989684e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57241,
|
|
"real_time": 1.2246667323272946e+04,
|
|
"cpu_time": 3.1205516919690137e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6891667616644883e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2252232077958342e+04,
|
|
"cpu_time": 3.1187096284131450e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6861290965807176e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2252486989997618e+04,
|
|
"cpu_time": 3.1198086266836173e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6859895519069576e+08,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.5148758868507555e+00,
|
|
"cpu_time": 2.8672040060627353e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9184521740194620e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.8687637195299188e-04,
|
|
"cpu_time": 9.1935587075530961e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.8693017234748835e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57244,
|
|
"real_time": 1.2228349285219010e+04,
|
|
"cpu_time": 3.1467156732583404e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3398374235027883e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57244,
|
|
"real_time": 1.2227971308188162e+04,
|
|
"cpu_time": 3.1280754803997843e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3398788390212247e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57244,
|
|
"real_time": 1.2227322963409306e+04,
|
|
"cpu_time": 3.1466993781009103e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3399498851081054e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57244,
|
|
"real_time": 1.2224486685783482e+04,
|
|
"cpu_time": 3.1345997344700536e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3402607750437357e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57244,
|
|
"real_time": 1.2226823309924344e+04,
|
|
"cpu_time": 3.1454189574452579e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3400046426369252e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2226990710504862e+04,
|
|
"cpu_time": 3.1403018447348695e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3399863130625558e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2227322963409308e+04,
|
|
"cpu_time": 3.1454189574452583e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3399498851081054e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5178810640734641e+00,
|
|
"cpu_time": 8.5183407760805665e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6636640333913575e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2414183506080293e-04,
|
|
"cpu_time": 2.7125866229587734e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2415530048131852e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57186,
|
|
"real_time": 1.2243312071067616e+04,
|
|
"cpu_time": 3.1912615028153199e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6763999651233778e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57186,
|
|
"real_time": 1.2237201269873654e+04,
|
|
"cpu_time": 3.1757756706187021e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6777364592891364e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57186,
|
|
"real_time": 1.2243586316959874e+04,
|
|
"cpu_time": 3.1870052950021924e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6763400160465736e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57186,
|
|
"real_time": 1.2241778816141485e+04,
|
|
"cpu_time": 3.1821885286610108e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6767351781257086e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57186,
|
|
"real_time": 1.2244803484545742e+04,
|
|
"cpu_time": 3.1935120256705453e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6760739803914971e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2242136391717677e+04,
|
|
"cpu_time": 3.1859486045535545e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6766571197952590e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2243312071067619e+04,
|
|
"cpu_time": 3.1870052950021927e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6763999651233778e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.9614682632210427e+00,
|
|
"cpu_time": 7.1445009832680668e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4764992716744752e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.4190779848071303e-04,
|
|
"cpu_time": 2.2425035272247351e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4196223056653109e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57027,
|
|
"real_time": 1.2273278884209927e+04,
|
|
"cpu_time": 3.2718596629665095e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3397303702040644e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57027,
|
|
"real_time": 1.2276864389541019e+04,
|
|
"cpu_time": 3.2609850228838324e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3381708814696875e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57027,
|
|
"real_time": 1.2272978130593869e+04,
|
|
"cpu_time": 3.2684937485753333e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3398612221619616e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57027,
|
|
"real_time": 1.2276555678472298e+04,
|
|
"cpu_time": 3.2672538692199083e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3383051172016792e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57027,
|
|
"real_time": 1.2272771280635161e+04,
|
|
"cpu_time": 3.2737367334772855e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3399512222155800e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2274489672690455e+04,
|
|
"cpu_time": 3.2684658074245737e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3392037626505947e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2273278884209927e+04,
|
|
"cpu_time": 3.2684937485753337e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3397303702040644e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0378480904035934e+00,
|
|
"cpu_time": 4.9176373686243238e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.8637907845345722e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.6602304004032092e-04,
|
|
"cpu_time": 1.5045705411552812e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6601334540815933e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 43925,
|
|
"real_time": 1.5935099115920548e+04,
|
|
"cpu_time": 3.8032167125781612e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.2253645896088390e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 43925,
|
|
"real_time": 1.5932053307819679e+04,
|
|
"cpu_time": 3.7915072373363277e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.2269370725534792e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 43925,
|
|
"real_time": 1.5936510336413105e+04,
|
|
"cpu_time": 3.7991421627773969e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.2246362116375914e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 43925,
|
|
"real_time": 1.5935299164917311e+04,
|
|
"cpu_time": 3.7998262219692886e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.2252613298007164e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 43925,
|
|
"real_time": 1.5935536383021381e+04,
|
|
"cpu_time": 3.8024403665338978e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.2251388876782017e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5934899661618405e+04,
|
|
"cpu_time": 3.7992265402390149e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.2254676182557669e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5935299164917309e+04,
|
|
"cpu_time": 3.7998262219692886e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 8.2252613298007164e+09,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6808332975488018e+00,
|
|
"cpu_time": 4.6423197267075025e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.6772207458379201e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0548126020506679e-04,
|
|
"cpu_time": 1.2219117963982867e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0549212699566798e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 35258,
|
|
"real_time": 1.9850857906383819e+04,
|
|
"cpu_time": 4.4163758551250925e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3205676109126617e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 35258,
|
|
"real_time": 1.9849641616455287e+04,
|
|
"cpu_time": 4.3862553831755642e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3206485289018190e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 35258,
|
|
"real_time": 1.9853580822497464e+04,
|
|
"cpu_time": 4.4096904787564265e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3203864952308578e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 35258,
|
|
"real_time": 1.9854274513592365e+04,
|
|
"cpu_time": 4.3988119717511734e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3203403620743460e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 35258,
|
|
"real_time": 1.9852283023240001e+04,
|
|
"cpu_time": 4.4124716036076228e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3204728125884670e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9852127576433788e+04,
|
|
"cpu_time": 4.4047210584831759e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3204831619416304e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9852283023240001e+04,
|
|
"cpu_time": 4.4096904787564265e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.3204728125884670e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9064027664361280e+00,
|
|
"cpu_time": 1.2211662897961469e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2680814711997018e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.6030148864205102e-05,
|
|
"cpu_time": 2.7724032318556663e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.6031627494221327e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 25675,
|
|
"real_time": 2.7264763213224356e+04,
|
|
"cpu_time": 5.1517422122690325e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9229508648206493e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 25675,
|
|
"real_time": 2.7262780662914971e+04,
|
|
"cpu_time": 5.1341946368059340e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9230907018709900e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 25675,
|
|
"real_time": 2.7270093180905496e+04,
|
|
"cpu_time": 5.1513697487830243e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9225750221018902e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 25675,
|
|
"real_time": 2.7268451392384108e+04,
|
|
"cpu_time": 5.1415033768258771e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9226907771756706e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 25675,
|
|
"real_time": 2.7266596589966211e+04,
|
|
"cpu_time": 5.1488185589092624e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9228215676647076e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7266537007879029e+04,
|
|
"cpu_time": 5.1455257067186256e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9228257867267818e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7266596589966204e+04,
|
|
"cpu_time": 5.1488185589092624e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 1.9228215676647076e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8970355501924621e+00,
|
|
"cpu_time": 7.5533589812838187e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.0429931397241645e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0624875279744270e-04,
|
|
"cpu_time": 1.4679469915816828e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0624951848612053e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 16637,
|
|
"real_time": 4.2075368417441423e+04,
|
|
"cpu_time": 6.7560925226906984e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.4921374177803650e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 16637,
|
|
"real_time": 4.2075544173780196e+04,
|
|
"cpu_time": 6.7425998857966595e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.4921270077201542e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 16637,
|
|
"real_time": 4.2079650145093685e+04,
|
|
"cpu_time": 6.7506499068341276e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.4918838354987125e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 16637,
|
|
"real_time": 4.2078764538335578e+04,
|
|
"cpu_time": 6.7446936707340967e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.4919362806973618e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 16637,
|
|
"real_time": 4.2077139885268465e+04,
|
|
"cpu_time": 6.7551634489397242e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.4920324975964314e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2077293431983875e+04,
|
|
"cpu_time": 6.7498398869990633e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.4920234078586048e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.2077139885268472e+04,
|
|
"cpu_time": 6.7506499068341276e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 2.4920324975964314e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9044298339085466e+00,
|
|
"cpu_time": 6.0620346463602097e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1278897058134719e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 4.5260274095028830e-05,
|
|
"cpu_time": 8.9810051021156211e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.5259996445325017e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 9826,
|
|
"real_time": 7.1267099252624219e+04,
|
|
"cpu_time": 9.7091479442302239e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.9426650193325752e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 9826,
|
|
"real_time": 7.1179719086906669e+04,
|
|
"cpu_time": 9.6884420008145884e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.9462774325359283e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 9826,
|
|
"real_time": 7.1247687342773308e+04,
|
|
"cpu_time": 9.7110829737439097e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.9434667681360401e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 9826,
|
|
"real_time": 7.1276736262186605e+04,
|
|
"cpu_time": 9.7200397313258247e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.9422671547218010e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 9826,
|
|
"real_time": 7.1184832537505994e+04,
|
|
"cpu_time": 9.6831355790758986e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.9460657913257694e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1231214896399368e+04,
|
|
"cpu_time": 9.7023696458380902e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.9441484332104233e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1247687342773308e+04,
|
|
"cpu_time": 9.7091479442302254e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 2.9434667681360401e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5919399125050106e+01,
|
|
"cpu_time": 1.5795790210731144e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8982421487243403e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.4465275780901021e-04,
|
|
"cpu_time": 1.6280342624862656e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4475083094041460e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 5510,
|
|
"real_time": 1.2704977420130796e+05,
|
|
"cpu_time": 1.5385468874773654e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.3013077168907085e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 5510,
|
|
"real_time": 1.2704495616517487e+05,
|
|
"cpu_time": 1.5372612105263007e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.3014329152484119e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 5510,
|
|
"real_time": 1.2704783926139875e+05,
|
|
"cpu_time": 1.5379821796734206e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.3013579958414654e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 5510,
|
|
"real_time": 1.2705129047654173e+05,
|
|
"cpu_time": 1.5382861270418635e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.3012683179116707e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 5510,
|
|
"real_time": 1.2703029450579308e+05,
|
|
"cpu_time": 1.5384224627950561e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.3018139620299183e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2704483092204327e+05,
|
|
"cpu_time": 1.5380997735028013e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.3014361815844357e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2704783926139874e+05,
|
|
"cpu_time": 1.5382861270418635e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 3.3013579958414654e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.4636729506510271e+00,
|
|
"cpu_time": 5.1375653319661367e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1995631508097239e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.6619577429674964e-05,
|
|
"cpu_time": 3.3402029052160051e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.6624433423217121e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2863,
|
|
"real_time": 2.4447221017562677e+05,
|
|
"cpu_time": 2.7253116032133502e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.4313135198367516e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2863,
|
|
"real_time": 2.4451144421284474e+05,
|
|
"cpu_time": 2.7253745930838917e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.4307629350460182e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2863,
|
|
"real_time": 2.4450384201379470e+05,
|
|
"cpu_time": 2.7256295214809489e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.4308696055282116e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2863,
|
|
"real_time": 2.4451809669484539e+05,
|
|
"cpu_time": 2.7269972651064763e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.4306695959885723e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2863,
|
|
"real_time": 2.4446685173653829e+05,
|
|
"cpu_time": 2.7257093782746146e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.4313887303790356e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4449448896673002e+05,
|
|
"cpu_time": 2.7258044722318568e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.4310008773557178e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4450384201379470e+05,
|
|
"cpu_time": 2.7256295214809489e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 3.4308696055282116e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.3411779538857211e+01,
|
|
"cpu_time": 6.8740843339152292e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.2854660156757063e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.5755857883745644e-05,
|
|
"cpu_time": 2.5218552555557328e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.5758238867249264e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1458,
|
|
"real_time": 4.8018269090266799e+05,
|
|
"cpu_time": 5.1399065637854621e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.4939235249112099e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1458,
|
|
"real_time": 4.8012451386265713e+05,
|
|
"cpu_time": 5.1385185665296519e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.4943468861911179e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1458,
|
|
"real_time": 4.8014417513697164e+05,
|
|
"cpu_time": 5.1401618381345132e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.4942037972685875e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1458,
|
|
"real_time": 4.8006965360136010e+05,
|
|
"cpu_time": 5.1389469478739740e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.4947462048770638e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1458,
|
|
"real_time": 4.8014844094408833e+05,
|
|
"cpu_time": 5.1401153909465432e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.4941727535367859e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.8013389488954906e+05,
|
|
"cpu_time": 5.1395298614540289e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.4942786333569534e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.8014417513697158e+05,
|
|
"cpu_time": 5.1399065637854626e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 3.4942037972685875e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.1573033897529783e+01,
|
|
"cpu_time": 7.4944139545005456e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.0257005402650144e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.6586334228899467e-05,
|
|
"cpu_time": 1.4581905653876859e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.6590133694010082e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 736,
|
|
"real_time": 9.5107386218644585e+05,
|
|
"cpu_time": 9.9687406657607097e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.5280574237274208e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 736,
|
|
"real_time": 9.5102841829509044e+05,
|
|
"cpu_time": 9.9673068614120316e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.5282260082357018e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 736,
|
|
"real_time": 9.5117604091090325e+05,
|
|
"cpu_time": 9.9699780434786808e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.5276784272095695e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 736,
|
|
"real_time": 9.5109646423699579e+05,
|
|
"cpu_time": 9.9694280978252192e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.5279735822505226e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 736,
|
|
"real_time": 9.5116298641938635e+05,
|
|
"cpu_time": 9.9699648505428527e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.5277268437782959e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.5110755440976447e+05,
|
|
"cpu_time": 9.9690837038038985e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.5279324570403023e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.5109646423699579e+05,
|
|
"cpu_time": 9.9694280978252203e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 3.5279735822505226e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.1813813177313321e+01,
|
|
"cpu_time": 1.1146427981695331e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.2928596804514662e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.4991401751270464e-05,
|
|
"cpu_time": 1.1180995478493369e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.4991598007378549e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8867470419430989e+06,
|
|
"cpu_time": 1.9565940673855017e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5568553975781929e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8871821751165383e+06,
|
|
"cpu_time": 1.9568567169810778e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5560352829135780e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8868794406806882e+06,
|
|
"cpu_time": 1.9567112398923112e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5566058198074699e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8873145729127638e+06,
|
|
"cpu_time": 1.9570511212937052e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5557858219908913e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8866478366340508e+06,
|
|
"cpu_time": 1.9565170754714750e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5570424271510170e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8869542134574284e+06,
|
|
"cpu_time": 1.9567460442048144e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5564649498882294e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8868794406806887e+06,
|
|
"cpu_time": 1.9567112398923114e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 3.5566058198074699e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8469422957786486e+02,
|
|
"cpu_time": 2.1333979419330618e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.3656363955229018e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5087500668933845e-04,
|
|
"cpu_time": 1.0902783977774875e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5086993604960258e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 187,
|
|
"real_time": 3.7477242581068194e+06,
|
|
"cpu_time": 3.8791317807484018e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5813127849432739e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 187,
|
|
"real_time": 3.7486371696294150e+06,
|
|
"cpu_time": 3.8797838074863059e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5804406221919998e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 187,
|
|
"real_time": 3.7486808309779768e+06,
|
|
"cpu_time": 3.8804396363637708e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5803989203579262e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 187,
|
|
"real_time": 3.7489007538393699e+06,
|
|
"cpu_time": 3.8806970213909638e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5801888823688736e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 187,
|
|
"real_time": 3.7485909284675186e+06,
|
|
"cpu_time": 3.8810409251335147e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5804847891170204e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7485067882042201e+06,
|
|
"cpu_time": 3.8802186342245913e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5805651997958183e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7486371696294150e+06,
|
|
"cpu_time": 3.8804396363637708e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 3.5804406221919998e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5330234666635971e+02,
|
|
"cpu_time": 7.6226362575001860e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.3305100666595846e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2092877838525169e-04,
|
|
"cpu_time": 1.9644862766923609e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2094487392399758e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 94,
|
|
"real_time": 7.4779554616976930e+06,
|
|
"cpu_time": 7.7114002978719128e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.5896904892645363e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 94,
|
|
"real_time": 7.4783486631163890e+06,
|
|
"cpu_time": 7.7114855425522961e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.5895017482126480e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 94,
|
|
"real_time": 7.4793716070262045e+06,
|
|
"cpu_time": 7.7128112234035181e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.5890108167353096e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 94,
|
|
"real_time": 7.4770856410899060e+06,
|
|
"cpu_time": 7.7115020744689358e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.5901080833530640e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 94,
|
|
"real_time": 7.4781954655067082e+06,
|
|
"cpu_time": 7.7115905319151636e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.5895752824082047e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4781913676873818e+06,
|
|
"cpu_time": 7.7117579340423662e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.5895772839947525e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4781954655067082e+06,
|
|
"cpu_time": 7.7115020744689358e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 3.5895752824082047e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.2097600362140292e+02,
|
|
"cpu_time": 5.9266490867090329e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.9406795622684169e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0978269520739590e-04,
|
|
"cpu_time": 7.6852115138971805e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0978115946518726e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.4941837598985814e+07,
|
|
"cpu_time": 1.5365463659575054e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.5930715244585472e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.4938668169556780e+07,
|
|
"cpu_time": 1.5362947276595691e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.5938338405165108e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.4938920181482397e+07,
|
|
"cpu_time": 1.5362803148936439e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.5937732143818581e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.4939182180356473e+07,
|
|
"cpu_time": 1.5362689531914528e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.5937101878704674e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.4940292039450178e+07,
|
|
"cpu_time": 1.5364400191488806e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.5934432244187744e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4939780033966329e+07,
|
|
"cpu_time": 1.5363660761702105e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.5935663983292320e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.4939182180356473e+07,
|
|
"cpu_time": 1.5362947276595693e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 3.5937101878704674e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3065930132538626e+03,
|
|
"cpu_time": 1.2232176027102794e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1426325290749473e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.7457312643376181e-05,
|
|
"cpu_time": 7.9617587349980129e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.7451633856996811e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 23,
|
|
"real_time": 2.9843611804687459e+07,
|
|
"cpu_time": 3.0646972304349102e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.5978950236557831e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 23,
|
|
"real_time": 2.9850422121260475e+07,
|
|
"cpu_time": 3.0653481043476976e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.5970741708045891e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 23,
|
|
"real_time": 2.9844042966547221e+07,
|
|
"cpu_time": 3.0647140434780337e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.5978430442670876e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 23,
|
|
"real_time": 2.9848202982026599e+07,
|
|
"cpu_time": 3.0651950782607462e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.5973416042720047e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 23,
|
|
"real_time": 2.9846303731850956e+07,
|
|
"cpu_time": 3.0648937739131369e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.5975705187712723e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time_mean",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.9846516721274544e+07,
|
|
"cpu_time": 3.0649696460869052e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.5975448723541466e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time_median",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.9846303731850952e+07,
|
|
"cpu_time": 3.0648937739131369e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 3.5975705187712723e+10,
|
|
"own_gpu": 5.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time_stddev",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.8591516504861993e+03,
|
|
"cpu_time": 2.9126872600641914e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.4461864817795334e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time_cv",
|
|
"family_index": 5,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.5795153491000224e-05,
|
|
"cpu_time": 9.5031520582360902e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.5792730988910008e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57084,
|
|
"real_time": 1.2261890215419478e+04,
|
|
"cpu_time": 3.3540234093616593e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3404311472706139e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57084,
|
|
"real_time": 1.2258462008974715e+04,
|
|
"cpu_time": 3.3527390302012354e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3413653335966778e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57084,
|
|
"real_time": 1.2259773622601197e+04,
|
|
"cpu_time": 3.3369967293811365e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3410078571507406e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57084,
|
|
"real_time": 1.2255505202027158e+04,
|
|
"cpu_time": 3.3371788907575705e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3421714833285612e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57084,
|
|
"real_time": 1.2253279289312737e+04,
|
|
"cpu_time": 3.3403986686286240e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3427786172902429e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2257782067667056e+04,
|
|
"cpu_time": 3.3442673456660457e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3415508877273679e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2258462008974715e+04,
|
|
"cpu_time": 3.3403986686286240e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3413653335966778e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4192330747800037e+00,
|
|
"cpu_time": 8.4413407386074752e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.3214671699255588e+04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.7894386242998065e-04,
|
|
"cpu_time": 2.5241225853390300e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7895631349385822e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2254569777059996e+04,
|
|
"cpu_time": 3.3441265133468609e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6848532009137166e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2250127466719588e+04,
|
|
"cpu_time": 3.3382765974216345e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6872773546687853e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2257251306693242e+04,
|
|
"cpu_time": 3.3404604568065020e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6833907497080064e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2263895922021316e+04,
|
|
"cpu_time": 3.3470145221747058e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6797696686990535e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2265241481948511e+04,
|
|
"cpu_time": 3.3446781493029783e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6790368636905003e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2258217190888532e+04,
|
|
"cpu_time": 3.3429112478105366e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6828655675360131e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2257251306693244e+04,
|
|
"cpu_time": 3.3441265133468609e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6833907497080064e+08,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.3495836306828499e+00,
|
|
"cpu_time": 3.4972143373203060e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.4617167405782931e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.1798589728059812e-04,
|
|
"cpu_time": 1.0461582967872184e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.1799885926100348e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2260011389988596e+04,
|
|
"cpu_time": 3.3758588698942724e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3363772250145714e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2260731841870811e+04,
|
|
"cpu_time": 3.3726189781406014e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3362986982594373e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2261822009944455e+04,
|
|
"cpu_time": 3.3762720503748431e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3361798912683954e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2261020251877153e+04,
|
|
"cpu_time": 3.3857088961675472e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3362672651561458e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2259284651565964e+04,
|
|
"cpu_time": 3.3875007409094229e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3364564463316510e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2260574029049396e+04,
|
|
"cpu_time": 3.3795919070973374e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3363159052060404e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2260731841870813e+04,
|
|
"cpu_time": 3.3762720503748438e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3362986982594373e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.6960466286221147e-01,
|
|
"cpu_time": 6.5868422102360000e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0568071006574473e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.9083137589226567e-05,
|
|
"cpu_time": 1.9490052027889085e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.9083628095746056e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57074,
|
|
"real_time": 1.2260231231472846e+04,
|
|
"cpu_time": 3.4166200879559277e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6727065241544805e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57074,
|
|
"real_time": 1.2260901204050464e+04,
|
|
"cpu_time": 3.4188575428392964e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6725604794185023e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57074,
|
|
"real_time": 1.2260004825392425e+04,
|
|
"cpu_time": 3.4024525931247517e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6727558811504087e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57074,
|
|
"real_time": 1.2259959234488568e+04,
|
|
"cpu_time": 3.4024580527035730e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6727658202826757e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57074,
|
|
"real_time": 1.2262709748558957e+04,
|
|
"cpu_time": 3.4071553544522256e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6721663214650178e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2260761248792651e+04,
|
|
"cpu_time": 3.4095087262151545e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6725910052942171e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2260231231472846e+04,
|
|
"cpu_time": 3.4071553544522249e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6727065241544805e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1524496196347007e+00,
|
|
"cpu_time": 7.7944083785239314e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.5118574991428156e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.3994948294763130e-05,
|
|
"cpu_time": 2.2860796098258967e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.3985854706799519e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57057,
|
|
"real_time": 1.2275330744436977e+04,
|
|
"cpu_time": 3.4871695217063861e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3388378174413004e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57057,
|
|
"real_time": 1.2271480318828813e+04,
|
|
"cpu_time": 3.4830350000875645e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3405129859878826e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57057,
|
|
"real_time": 1.2265687225050648e+04,
|
|
"cpu_time": 3.4753257794836340e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3430353144953432e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57057,
|
|
"real_time": 1.2269444882980295e+04,
|
|
"cpu_time": 3.4791396182764016e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3413989487746944e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57057,
|
|
"real_time": 1.2273103994496210e+04,
|
|
"cpu_time": 3.4768654117811544e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3398064604837675e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2271009433158590e+04,
|
|
"cpu_time": 3.4803070662670281e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3407183054365978e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2271480318828813e+04,
|
|
"cpu_time": 3.4791396182764016e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3405129859878826e+09,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6761517913194481e+00,
|
|
"cpu_time": 4.8103029015986756e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.6001295819526617e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 2.9958022698489581e-04,
|
|
"cpu_time": 1.3821489914561474e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.9960943274686585e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2262225427540754e+04,
|
|
"cpu_time": 3.6185585896447075e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0689087455986126e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2260647363350616e+04,
|
|
"cpu_time": 3.6201989000211426e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0690463245178953e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2260233032232083e+04,
|
|
"cpu_time": 3.5951123169620703e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0690824526370131e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2259557543402589e+04,
|
|
"cpu_time": 3.6142917764309350e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0691413579647142e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57092,
|
|
"real_time": 1.2260022705512840e+04,
|
|
"cpu_time": 3.6178363816296129e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0691007932722847e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2260537214407776e+04,
|
|
"cpu_time": 3.6131995929376935e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0690559347981041e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2260233032232081e+04,
|
|
"cpu_time": 3.6178363816296129e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0690824526370131e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.0221543268822102e+00,
|
|
"cpu_time": 1.0338657694858223e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.9120163159635209e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.3369456738081744e-05,
|
|
"cpu_time": 2.8613580370887926e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.3363423988161984e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 44705,
|
|
"real_time": 1.5657522508725173e+04,
|
|
"cpu_time": 4.1370846818029087e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6742367756707355e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 44705,
|
|
"real_time": 1.5658348501274942e+04,
|
|
"cpu_time": 4.1296303254668564e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6741484581126520e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 44705,
|
|
"real_time": 1.5654268575138494e+04,
|
|
"cpu_time": 4.1216390716920876e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6745847865184004e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 44705,
|
|
"real_time": 1.5658413058940181e+04,
|
|
"cpu_time": 4.1289556470192285e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6741415558093784e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 44705,
|
|
"real_time": 1.5654712485068107e+04,
|
|
"cpu_time": 4.1252827692652514e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6745373014677858e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5656653025829379e+04,
|
|
"cpu_time": 4.1285184990492671e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6743297755157906e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5657522508725171e+04,
|
|
"cpu_time": 4.1289556470192285e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6742367756707355e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0111972625057866e+00,
|
|
"cpu_time": 5.7576733862937807e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.1508676253456417e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2845639864330122e-04,
|
|
"cpu_time": 1.3946100490090289e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2846140926347976e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 36875,
|
|
"real_time": 1.8984367859735295e+04,
|
|
"cpu_time": 4.4748152949151990e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7616826847944901e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 36875,
|
|
"real_time": 1.8979591554103783e+04,
|
|
"cpu_time": 4.4703122630510159e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7623776755440136e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 36875,
|
|
"real_time": 1.8982403192637583e+04,
|
|
"cpu_time": 4.4701080542373827e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7619685172599625e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 36875,
|
|
"real_time": 1.8977204543860011e+04,
|
|
"cpu_time": 4.4598541396611588e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7627251357715431e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 36875,
|
|
"real_time": 1.8981127369551035e+04,
|
|
"cpu_time": 4.4658698413556936e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7621541639357384e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8980938903977541e+04,
|
|
"cpu_time": 4.4681919186440908e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7621816354611500e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8981127369551039e+04,
|
|
"cpu_time": 4.4701080542373827e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.7621541639357384e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7248025094715791e+00,
|
|
"cpu_time": 5.6335413048184407e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.9653187534724115e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.4355467467947985e-04,
|
|
"cpu_time": 1.2608100563701807e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.4355749464717573e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 27325,
|
|
"real_time": 2.5608575346724476e+04,
|
|
"cpu_time": 5.1476323915828805e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.0946284039737511e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 27325,
|
|
"real_time": 2.5614729927826545e+04,
|
|
"cpu_time": 5.1434225178408487e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.0936445668352722e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 27325,
|
|
"real_time": 2.5610038652371277e+04,
|
|
"cpu_time": 5.1375129515097498e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.0943944452145935e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 27325,
|
|
"real_time": 2.5612539984208612e+04,
|
|
"cpu_time": 5.1329873559009284e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.0939945848654549e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 27325,
|
|
"real_time": 2.5610074945819764e+04,
|
|
"cpu_time": 5.1348215407138319e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.0943886428225983e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5611191771390142e+04,
|
|
"cpu_time": 5.1392753515096483e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.0942101287423340e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5610074945819764e+04,
|
|
"cpu_time": 5.1375129515097491e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 4.0943886428225983e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4371545532109087e+00,
|
|
"cpu_time": 6.1139283707053309e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.8958742451983741e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.5159747932285443e-05,
|
|
"cpu_time": 1.1896479469443841e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.5155698478893532e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 18064,
|
|
"real_time": 3.8702415077767546e+04,
|
|
"cpu_time": 6.4854706875553689e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 5.4186592639917732e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 18064,
|
|
"real_time": 3.8686304345614808e+04,
|
|
"cpu_time": 6.4784892548714051e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 5.4209158395294426e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 18064,
|
|
"real_time": 3.8690288854792474e+04,
|
|
"cpu_time": 6.4699785595656605e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 5.4203575679436432e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 18064,
|
|
"real_time": 3.8745940424951878e+04,
|
|
"cpu_time": 6.4719088795395961e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 5.4125721998206085e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 18064,
|
|
"real_time": 3.8686507697231231e+04,
|
|
"cpu_time": 6.4585273306026604e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 5.4208873450474106e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8702291280071593e+04,
|
|
"cpu_time": 6.4728749424269387e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 5.4186784432665764e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.8690288854792481e+04,
|
|
"cpu_time": 6.4719088795395961e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 5.4203575679436432e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.5268680949928498e+01,
|
|
"cpu_time": 1.0066647602876718e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.5352236722737640e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.5289883658489582e-04,
|
|
"cpu_time": 1.5552050197809525e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.5241436805809109e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 11347,
|
|
"real_time": 6.1801748240772526e+04,
|
|
"cpu_time": 8.8604384066265338e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.7867076893350533e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 11347,
|
|
"real_time": 6.1697005059267947e+04,
|
|
"cpu_time": 8.8363602009333976e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.7982295023410439e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 11347,
|
|
"real_time": 6.1804976276084235e+04,
|
|
"cpu_time": 8.8339598307919689e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.7863532238309555e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 11347,
|
|
"real_time": 6.1711713765708082e+04,
|
|
"cpu_time": 8.8158485679036443e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.7966091752433037e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 11347,
|
|
"real_time": 6.1801704443957780e+04,
|
|
"cpu_time": 8.8278549572574586e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.7867124988493225e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.1763429557158110e+04,
|
|
"cpu_time": 8.8348923927026015e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.7909224179199371e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.1801704443957773e+04,
|
|
"cpu_time": 8.8339598307919689e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.7867124988493225e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.4189840681876738e+01,
|
|
"cpu_time": 1.6338724862714452e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.9602345116600908e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.7737745572770540e-04,
|
|
"cpu_time": 1.8493405619982226e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.7767671972399197e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 6085,
|
|
"real_time": 1.1494432716210844e+05,
|
|
"cpu_time": 1.4311416088742725e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 7.2979747736218124e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 6085,
|
|
"real_time": 1.1500420274991418e+05,
|
|
"cpu_time": 1.4316839244044069e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 7.2941751687472656e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 6085,
|
|
"real_time": 1.1494913915165141e+05,
|
|
"cpu_time": 1.4304901495481000e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 7.2976692665205444e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 6085,
|
|
"real_time": 1.1502234437152815e+05,
|
|
"cpu_time": 1.4313837715695403e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 7.2930247125761597e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 6085,
|
|
"real_time": 1.1495129568331079e+05,
|
|
"cpu_time": 1.4308933541495795e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 7.2975323593659149e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1497426182370260e+05,
|
|
"cpu_time": 1.4311185617091798e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 7.2960752561663406e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1495129568331079e+05,
|
|
"cpu_time": 1.4311416088742722e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 7.2975323593659149e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.6273498153800688e+01,
|
|
"cpu_time": 4.5719872759802051e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.3015393537619989e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.1549233348783024e-04,
|
|
"cpu_time": 3.1946949737832322e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1544896029092259e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 3146,
|
|
"real_time": 2.2249025823723839e+05,
|
|
"cpu_time": 2.5632302256835319e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 7.5406519516511505e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 3146,
|
|
"real_time": 2.2250480420866155e+05,
|
|
"cpu_time": 2.5642352542909456e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 7.5401589910241165e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 3146,
|
|
"real_time": 2.2251365121805691e+05,
|
|
"cpu_time": 2.5651400381437072e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 7.5398591988222855e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 3146,
|
|
"real_time": 2.2251497365226282e+05,
|
|
"cpu_time": 2.5644752479338963e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 7.5398143884998672e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 3146,
|
|
"real_time": 2.2246131996372691e+05,
|
|
"cpu_time": 2.5641163127782015e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 7.5416328567750946e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2249700145598935e+05,
|
|
"cpu_time": 2.5642394157660566e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 7.5404234773545029e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2250480420866158e+05,
|
|
"cpu_time": 2.5642352542909453e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 7.5401589910241165e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.2247119085073525e+01,
|
|
"cpu_time": 6.8927022381456851e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.5401451848939881e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.9988399571641323e-05,
|
|
"cpu_time": 2.6880104079854482e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.9996309325844218e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1596,
|
|
"real_time": 4.3857730154451658e+05,
|
|
"cpu_time": 4.8439061779446172e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 7.6507452350664230e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1596,
|
|
"real_time": 4.3864095836885413e+05,
|
|
"cpu_time": 4.8440543859651324e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 7.6496349371423737e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1596,
|
|
"real_time": 4.3853429691497830e+05,
|
|
"cpu_time": 4.8411639348370518e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 7.6514955012755661e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1596,
|
|
"real_time": 4.3860697687500244e+05,
|
|
"cpu_time": 4.8420137218043057e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 7.6502275999049133e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1596,
|
|
"real_time": 4.3851775357562152e+05,
|
|
"cpu_time": 4.8415920614033862e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 7.6517841584294266e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3857545745579462e+05,
|
|
"cpu_time": 4.8425460563908983e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 7.6507774863637405e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.3857730154451664e+05,
|
|
"cpu_time": 4.8420137218043057e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 7.6507452350664230e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.0771970341922426e+01,
|
|
"cpu_time": 1.3443153744804277e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.8568731077463217e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.1576564415267102e-04,
|
|
"cpu_time": 2.7760507774753779e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1576435366905193e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 802,
|
|
"real_time": 8.7244182709253801e+05,
|
|
"cpu_time": 9.4279990523694071e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.6920732037394516e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 802,
|
|
"real_time": 8.7239293831846502e+05,
|
|
"cpu_time": 9.4288081670821935e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.6925042664091431e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 802,
|
|
"real_time": 8.7232790773262992e+05,
|
|
"cpu_time": 9.4261525685781357e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.6930777297301590e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 802,
|
|
"real_time": 8.7247872624971310e+05,
|
|
"cpu_time": 9.4283343391519284e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.6917478880502472e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 802,
|
|
"real_time": 8.7242866025835811e+05,
|
|
"cpu_time": 9.4271304613469914e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.6921892937500015e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.7241401193034090e+05,
|
|
"cpu_time": 9.4276849177057331e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.6923184763358017e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 8.7242866025835811e+05,
|
|
"cpu_time": 9.4279990523694060e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.6921892937500015e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.7083775766897006e+01,
|
|
"cpu_time": 1.0532845769756265e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.0333919000848718e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 6.5431979526086453e-05,
|
|
"cpu_time": 1.1172250517170954e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.5434002967626779e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 402,
|
|
"real_time": 1.7401005575243400e+06,
|
|
"cpu_time": 1.8575510995025134e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.7132167689752945e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 402,
|
|
"real_time": 1.7401097114753926e+06,
|
|
"cpu_time": 1.8575667014924400e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.7131761931378677e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 402,
|
|
"real_time": 1.7400718866314609e+06,
|
|
"cpu_time": 1.8576384800993609e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.7133438584440903e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 402,
|
|
"real_time": 1.7400886069992511e+06,
|
|
"cpu_time": 1.8575105522389058e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.7132697415596466e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 402,
|
|
"real_time": 1.7401272134475447e+06,
|
|
"cpu_time": 1.8575589502489013e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.7130986150195007e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7400995952155977e+06,
|
|
"cpu_time": 1.8575651567164245e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.7132210354272797e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7401005575243402e+06,
|
|
"cpu_time": 1.8575589502489013e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.7132167689752945e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0943604851720256e+01,
|
|
"cpu_time": 4.6358590415079554e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.2835356574960169e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2035865596029489e-05,
|
|
"cpu_time": 2.4956642972904692e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2035874007572439e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 201,
|
|
"real_time": 3.4771777645439208e+06,
|
|
"cpu_time": 3.6997205223877388e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.7199232877071213e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 201,
|
|
"real_time": 3.4771546946180547e+06,
|
|
"cpu_time": 3.6999876865672437e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.7199745071878677e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 201,
|
|
"real_time": 3.4772916129368604e+06,
|
|
"cpu_time": 3.7001623383082286e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.7196705332770203e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 201,
|
|
"real_time": 3.4772263496965910e+06,
|
|
"cpu_time": 3.7000458059704066e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.7198154219503891e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 201,
|
|
"real_time": 3.4772016742818095e+06,
|
|
"cpu_time": 3.7000759601988057e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.7198702044063461e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4772104192154473e+06,
|
|
"cpu_time": 3.6999984626864851e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.7198507909057495e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4772016742818095e+06,
|
|
"cpu_time": 3.7000458059704066e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.7198702044063461e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.2664567478282869e+01,
|
|
"cpu_time": 1.6768228151973415e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.1692106033046399e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.5145637200225982e-05,
|
|
"cpu_time": 4.5319554375702053e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.5145507795073065e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.9536661664669467e+06,
|
|
"cpu_time": 7.3782772673257114e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.7206886144316589e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.9533002028002003e+06,
|
|
"cpu_time": 7.3779065742572509e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.7210949670171555e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.9539053826638963e+06,
|
|
"cpu_time": 7.3782628910892904e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.7204230206873459e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.9536820542768091e+06,
|
|
"cpu_time": 7.3776275346528906e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.7206709741611160e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 101,
|
|
"real_time": 6.9536803991045104e+06,
|
|
"cpu_time": 7.3779166930690547e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.7206728118988297e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9536468410624731e+06,
|
|
"cpu_time": 7.3779981920788391e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.7207100776392212e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.9536803991045104e+06,
|
|
"cpu_time": 7.3779166930690538e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.7206728118988297e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1779657800046374e+02,
|
|
"cpu_time": 2.7403076125755689e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.4182641322072325e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.1321202094178517e-05,
|
|
"cpu_time": 3.7141614042649344e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.1321783979572389e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.3912737444043159e+07,
|
|
"cpu_time": 1.4716493779999383e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.7176891199059494e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.3911940604448318e+07,
|
|
"cpu_time": 1.4716060180001022e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.7181311689662689e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.3913163095712662e+07,
|
|
"cpu_time": 1.4717757559999429e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.7174530091642014e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.3912791814655066e+07,
|
|
"cpu_time": 1.4716280580000784e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.7176589594977768e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 50,
|
|
"real_time": 1.3913214355707169e+07,
|
|
"cpu_time": 1.4716734219998669e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.7174245760078690e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time_mean",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3912769462913275e+07,
|
|
"cpu_time": 1.4716665263999859e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.7176713667084137e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time_median",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3912791814655066e+07,
|
|
"cpu_time": 1.4716493779999381e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.7176589594977768e+10,
|
|
"own_gpu": 6.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time_stddev",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.1024690758509394e+02,
|
|
"cpu_time": 6.5980694474248867e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.8305159537582542e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time_cv",
|
|
"family_index": 6,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.6674718785877906e-05,
|
|
"cpu_time": 4.4833998253430346e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.6675777177662710e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57207,
|
|
"real_time": 1.2260957072471429e+04,
|
|
"cpu_time": 3.1164959934971252e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3406853769975501e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57207,
|
|
"real_time": 1.2255022700369378e+04,
|
|
"cpu_time": 3.1172651913227590e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3423030704598719e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57207,
|
|
"real_time": 1.2260906625475851e+04,
|
|
"cpu_time": 3.1179390476689292e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3406991221100122e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57207,
|
|
"real_time": 1.2258261285819086e+04,
|
|
"cpu_time": 3.1150892600555366e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3414200468531692e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57207,
|
|
"real_time": 1.2244752590520920e+04,
|
|
"cpu_time": 3.1150114846086424e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3451063790140224e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2255980054931333e+04,
|
|
"cpu_time": 3.1163601954305981e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3420427990869254e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2258261285819086e+04,
|
|
"cpu_time": 3.1164959934971248e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.0960000000000000e+03,
|
|
"bytes_per_second": 3.3414200468531692e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.7308673302959265e+00,
|
|
"cpu_time": 1.3004355684168184e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8363786406947777e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 0,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.4919046050402854e-04,
|
|
"cpu_time": 4.1729308772573787e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.4947789453698554e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 56959,
|
|
"real_time": 1.2243345222691767e+04,
|
|
"cpu_time": 3.1130226074896156e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6909817954140341e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 56959,
|
|
"real_time": 1.2242921934574038e+04,
|
|
"cpu_time": 3.1201710054599997e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6912131301481020e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 56959,
|
|
"real_time": 1.2252241255526656e+04,
|
|
"cpu_time": 3.1193639828649400e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6861236480344427e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 56959,
|
|
"real_time": 1.2251778468554932e+04,
|
|
"cpu_time": 3.1197908442915712e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6863762032796752e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 56959,
|
|
"real_time": 1.2243329173708564e+04,
|
|
"cpu_time": 3.1213481802701419e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6909905661865032e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2246723211011195e+04,
|
|
"cpu_time": 3.1187393240752543e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6891370686125517e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2243345222691767e+04,
|
|
"cpu_time": 3.1197908442915716e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 8.1920000000000000e+03,
|
|
"bytes_per_second": 6.6909817954140341e+08,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.8317812678714889e+00,
|
|
"cpu_time": 3.2799612846368333e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.6387299354045308e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 1,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.9453665969417593e-04,
|
|
"cpu_time": 1.0516945931700733e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.9447987211777246e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57266,
|
|
"real_time": 1.2230684025568908e+04,
|
|
"cpu_time": 3.1231239828171314e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3395816591899815e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57266,
|
|
"real_time": 1.2218672308094327e+04,
|
|
"cpu_time": 3.1314887612195624e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3408985515673687e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57266,
|
|
"real_time": 1.2223481086739876e+04,
|
|
"cpu_time": 3.1361764886669731e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3403710353651619e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57266,
|
|
"real_time": 1.2223416525988636e+04,
|
|
"cpu_time": 3.1349175025321561e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3403781148392842e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57266,
|
|
"real_time": 1.2221156569128731e+04,
|
|
"cpu_time": 3.1344852844621564e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3406259798182135e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2223482103104096e+04,
|
|
"cpu_time": 3.1320384039395962e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3403710681560020e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2223416525988634e+04,
|
|
"cpu_time": 3.1344852844621561e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6384000000000000e+04,
|
|
"bytes_per_second": 1.3403781148392842e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.4836989446221720e+00,
|
|
"cpu_time": 5.2718400315464308e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.9153709833541559e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 2,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.6681028423836431e-04,
|
|
"cpu_time": 1.6831977618522529e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.6671718005047759e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57234,
|
|
"real_time": 1.2230801818380267e+04,
|
|
"cpu_time": 3.1589970507041045e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6791375158051157e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57234,
|
|
"real_time": 1.2233723384759289e+04,
|
|
"cpu_time": 3.1713429167976868e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6784977042085333e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57234,
|
|
"real_time": 1.2236110629675159e+04,
|
|
"cpu_time": 3.1736003756508449e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6779751337431245e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57234,
|
|
"real_time": 1.2235084474805937e+04,
|
|
"cpu_time": 3.1682911888038387e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6781997351530132e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57234,
|
|
"real_time": 1.2235948664359361e+04,
|
|
"cpu_time": 3.1720462137890328e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6780105816761074e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2234333794396003e+04,
|
|
"cpu_time": 3.1688555491491017e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6783641341171789e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2235084474805935e+04,
|
|
"cpu_time": 3.1713429167976865e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 3.2768000000000000e+04,
|
|
"bytes_per_second": 2.6781997351530132e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1895161642187189e+00,
|
|
"cpu_time": 5.8392895839434118e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.7940271238281496e+05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 3,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7896488693333165e-04,
|
|
"cpu_time": 1.8427124535579963e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7899086471333438e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 57081,
|
|
"real_time": 1.2258094574876677e+04,
|
|
"cpu_time": 3.2292769678177454e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3463447846387110e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 57081,
|
|
"real_time": 1.2263402816699901e+04,
|
|
"cpu_time": 3.2445474588741879e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3440306071293058e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 57081,
|
|
"real_time": 1.2262501097420194e+04,
|
|
"cpu_time": 3.2408643892013981e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3444235787907553e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 57081,
|
|
"real_time": 1.2264002663501244e+04,
|
|
"cpu_time": 3.2412546363939149e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3437692243039808e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 57081,
|
|
"real_time": 1.2263764633484740e+04,
|
|
"cpu_time": 3.2417731189011516e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3438729426575747e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2262353157196552e+04,
|
|
"cpu_time": 3.2395433142376794e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3444882275040665e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2263402816699900e+04,
|
|
"cpu_time": 3.2412546363939146e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 6.5536000000000000e+04,
|
|
"bytes_per_second": 5.3440306071293058e+09,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4480773578193307e+00,
|
|
"cpu_time": 5.9178801901024016e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0672325488289795e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 4,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.9964172670908589e-04,
|
|
"cpu_time": 1.8267637182356924e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.9968844600250688e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 56956,
|
|
"real_time": 1.2284831303140652e+04,
|
|
"cpu_time": 3.3656538099584941e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0669417981058565e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 56956,
|
|
"real_time": 1.2285809616455834e+04,
|
|
"cpu_time": 3.3805630153100981e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0668568380258783e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 56956,
|
|
"real_time": 1.2289175769498994e+04,
|
|
"cpu_time": 3.3802966974506940e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0665646131070314e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 56956,
|
|
"real_time": 1.2288946945927140e+04,
|
|
"cpu_time": 3.3830181596320726e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0665844728334555e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 56956,
|
|
"real_time": 1.2289420256560546e+04,
|
|
"cpu_time": 3.3808280497226304e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0665433947547600e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2287636778316633e+04,
|
|
"cpu_time": 3.3780719464147973e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0666982233653965e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2288946945927139e+04,
|
|
"cpu_time": 3.3805630153100981e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3107200000000000e+05,
|
|
"bytes_per_second": 1.0665844728334555e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1491299956036198e+00,
|
|
"cpu_time": 7.0254214303027965e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.8658111276761107e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 5,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.7490181671027903e-04,
|
|
"cpu_time": 2.0797133814035515e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.7491461847471166e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 43266,
|
|
"real_time": 1.6182874302440559e+04,
|
|
"cpu_time": 3.9329253871400491e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6198852879952587e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 43266,
|
|
"real_time": 1.6182919232563512e+04,
|
|
"cpu_time": 3.9380457703508830e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6198807905591589e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 43266,
|
|
"real_time": 1.6185555949103271e+04,
|
|
"cpu_time": 3.9367459853003798e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6196169030234859e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 43266,
|
|
"real_time": 1.6184863558497318e+04,
|
|
"cpu_time": 3.9394522581239566e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6196861904490393e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 43266,
|
|
"real_time": 1.6185671191116813e+04,
|
|
"cpu_time": 3.9369702537787962e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6196053713476681e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6184376846744297e+04,
|
|
"cpu_time": 3.9368279309388134e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6197349086749222e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.6184863558497314e+04,
|
|
"cpu_time": 3.9369702537787962e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6214400000000000e+05,
|
|
"bytes_per_second": 1.6196861904490393e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.3860714890316514e+00,
|
|
"cpu_time": 2.4303829427154774e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3872110652384518e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 6,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.5642561474986802e-05,
|
|
"cpu_time": 6.1734548355937542e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.5644327217304083e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 35168,
|
|
"real_time": 1.9902199883447687e+04,
|
|
"cpu_time": 4.3178608308691051e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.6343218491943756e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 35168,
|
|
"real_time": 1.9907245755604959e+04,
|
|
"cpu_time": 4.3340142487487632e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.6336541299410282e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 35168,
|
|
"real_time": 1.9902290423065951e+04,
|
|
"cpu_time": 4.3215523544130148e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.6343098651217113e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 35168,
|
|
"real_time": 1.9904620585680630e+04,
|
|
"cpu_time": 4.3253008530479521e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.6340014759044060e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 35168,
|
|
"real_time": 1.9904716432425990e+04,
|
|
"cpu_time": 4.3232532813923412e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.6339887924547523e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9904214616045043e+04,
|
|
"cpu_time": 4.3243963136942351e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.6340552225232544e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.9904620585680626e+04,
|
|
"cpu_time": 4.3232532813923419e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 5.2428800000000000e+05,
|
|
"bytes_per_second": 2.6340014759044060e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.0836328942975122e+00,
|
|
"cpu_time": 6.0297186735590813e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7573064672321067e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 7,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0468299978125583e-04,
|
|
"cpu_time": 1.3943492307734457e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0467914429640491e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 25606,
|
|
"real_time": 2.7337470635266047e+04,
|
|
"cpu_time": 5.1089833671796659e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 3.8356730730139671e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 25606,
|
|
"real_time": 2.7338390026290272e+04,
|
|
"cpu_time": 5.1175977036632234e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 3.8355440791927574e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 25606,
|
|
"real_time": 2.7344070128935098e+04,
|
|
"cpu_time": 5.1157565570568542e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 3.8347473329890717e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 25606,
|
|
"real_time": 2.7345274184910333e+04,
|
|
"cpu_time": 5.1233575490118405e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 3.8345784829563904e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 25606,
|
|
"real_time": 2.7342227246594975e+04,
|
|
"cpu_time": 5.1192590252282767e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 3.8350057972346893e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7341486444399350e+04,
|
|
"cpu_time": 5.1169908404279726e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 3.8351097530773750e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.7342227246594975e+04,
|
|
"cpu_time": 5.1175977036632234e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0485760000000000e+06,
|
|
"bytes_per_second": 3.8350057972346893e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.4382306371433473e+00,
|
|
"cpu_time": 5.2841983008146599e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.8227886473781951e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 8,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2575141604444980e-04,
|
|
"cpu_time": 1.0326769121933200e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2575360179740058e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 16671,
|
|
"real_time": 4.1931743350214230e+04,
|
|
"cpu_time": 6.6641947573635873e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 5.0013470283946243e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 16671,
|
|
"real_time": 4.1991267478931040e+04,
|
|
"cpu_time": 6.6756989082834669e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9942574394836685e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 16671,
|
|
"real_time": 4.1919234626099111e+04,
|
|
"cpu_time": 6.6718153140184804e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 5.0028394332712921e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 16671,
|
|
"real_time": 4.1996690055792395e+04,
|
|
"cpu_time": 6.6876581188885481e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9936125852155113e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 16671,
|
|
"real_time": 4.2009340165280628e+04,
|
|
"cpu_time": 6.6840008397818703e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9921088780471466e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.1969655135263485e+04,
|
|
"cpu_time": 6.6766735876671912e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9968330728824493e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.1991267478931040e+04,
|
|
"cpu_time": 6.6756989082834654e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 2.0971520000000000e+06,
|
|
"bytes_per_second": 4.9942574394836685e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.1086534057036204e+01,
|
|
"cpu_time": 9.4157126419410744e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 4.8932505683867857e+07,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 9,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 9.7895810495985552e-04,
|
|
"cpu_time": 1.4102400721421064e-03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 9.7927036925491849e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 10264,
|
|
"real_time": 6.8199322885535250e+04,
|
|
"cpu_time": 9.4195721063910707e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.1500669252093002e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 10264,
|
|
"real_time": 6.8191607933330321e+04,
|
|
"cpu_time": 9.4266060502729466e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.1507627215664040e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 10264,
|
|
"real_time": 6.8186806416613923e+04,
|
|
"cpu_time": 9.4134987431801273e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.1511958403994781e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 10264,
|
|
"real_time": 6.8185746505807867e+04,
|
|
"cpu_time": 9.4205118959468862e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.1512914574349350e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 10264,
|
|
"real_time": 6.8194756660698360e+04,
|
|
"cpu_time": 9.4216452650035848e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.1504787250267273e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8191648080397150e+04,
|
|
"cpu_time": 9.4203668121589231e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.1507591339273697e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 6.8191607933330321e+04,
|
|
"cpu_time": 9.4205118959468862e+04,
|
|
"time_unit": "ns",
|
|
"bytes": 4.1943040000000000e+06,
|
|
"bytes_per_second": 6.1507627215664040e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.6311145290639590e+00,
|
|
"cpu_time": 4.7006784609475588e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.0790643225224074e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 10,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.2577774369449781e-05,
|
|
"cpu_time": 4.9899102175940380e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.2576218836248497e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 5522,
|
|
"real_time": 1.2677683264737017e+05,
|
|
"cpu_time": 1.5441367928287905e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 6.6168303978163879e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 5522,
|
|
"real_time": 1.2674266900372636e+05,
|
|
"cpu_time": 1.5446572546178967e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 6.6186139726577522e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 5522,
|
|
"real_time": 1.2678244130124550e+05,
|
|
"cpu_time": 1.5435044223108396e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 6.6165376797469749e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 5522,
|
|
"real_time": 1.2675655222392807e+05,
|
|
"cpu_time": 1.5430065483519694e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 6.6178890580588593e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 5522,
|
|
"real_time": 1.2677820796411020e+05,
|
|
"cpu_time": 1.5436003839189018e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 6.6167586170446114e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2676734062807607e+05,
|
|
"cpu_time": 1.5437810804056798e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 6.6173259450649170e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.2677683264737017e+05,
|
|
"cpu_time": 1.5436003839189018e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 8.3886080000000000e+06,
|
|
"bytes_per_second": 6.6168303978163879e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.7038973440097781e+01,
|
|
"cpu_time": 6.3310808737298984e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.8950853515410405e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 11,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.3441138195119663e-04,
|
|
"cpu_time": 4.1010224533041931e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.3442114572238708e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 2865,
|
|
"real_time": 2.4434878964450074e+05,
|
|
"cpu_time": 2.7799670924956398e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 6.8660933513969566e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 2865,
|
|
"real_time": 2.4431075356912069e+05,
|
|
"cpu_time": 2.7794436300173466e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 6.8671623147580238e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 2865,
|
|
"real_time": 2.4436649677049476e+05,
|
|
"cpu_time": 2.7805136684119818e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 6.8655958250107010e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 2865,
|
|
"real_time": 2.4434376223460588e+05,
|
|
"cpu_time": 2.7807827574170375e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 6.8662346223069977e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 2865,
|
|
"real_time": 2.4436158002118531e+05,
|
|
"cpu_time": 2.7809285479931737e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 6.8657339662583099e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4434627644798151e+05,
|
|
"cpu_time": 2.7803271392670355e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 6.8661640159461975e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.4434878964450076e+05,
|
|
"cpu_time": 2.7805136684119818e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 1.6777216000000000e+07,
|
|
"bytes_per_second": 6.8660933513969566e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 2.1895051306403349e+01,
|
|
"cpu_time": 6.1528362354925413e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 6.1529727547194613e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 12,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 8.9606650138843229e-05,
|
|
"cpu_time": 2.2129900286174897e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.9612959149091133e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 1460,
|
|
"real_time": 4.7949376222613739e+05,
|
|
"cpu_time": 5.2465781095891120e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 6.9978870724443665e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 1460,
|
|
"real_time": 4.7959699522710266e+05,
|
|
"cpu_time": 5.2479366849314678e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 6.9963807809327560e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 1460,
|
|
"real_time": 4.7952828122293320e+05,
|
|
"cpu_time": 5.2477668424654589e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 6.9973833273037979e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 1460,
|
|
"real_time": 4.7958504232380236e+05,
|
|
"cpu_time": 5.2467276301365311e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 6.9965551547258209e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 1460,
|
|
"real_time": 4.7948246796644124e+05,
|
|
"cpu_time": 5.2457814726030594e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 6.9980519084064728e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.7953730979328335e+05,
|
|
"cpu_time": 5.2469581479451270e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 6.9972516487626419e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.7952828122293326e+05,
|
|
"cpu_time": 5.2467276301365311e+05,
|
|
"time_unit": "ns",
|
|
"bytes": 3.3554432000000000e+07,
|
|
"bytes_per_second": 6.9973833273037979e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 5.2024698425985640e+01,
|
|
"cpu_time": 8.9353985661167002e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.5911672704953616e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 13,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.0848936540185412e-04,
|
|
"cpu_time": 1.7029673792263967e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.0848784139180910e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 738,
|
|
"real_time": 9.4906721269802703e+05,
|
|
"cpu_time": 1.0180919525746652e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.0710338637894348e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 738,
|
|
"real_time": 9.4905680780258623e+05,
|
|
"cpu_time": 1.0181658983740603e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.0711113864070557e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 738,
|
|
"real_time": 9.4918190292776795e+05,
|
|
"cpu_time": 1.0181512899728026e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.0701794664438446e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 738,
|
|
"real_time": 9.4918883705340070e+05,
|
|
"cpu_time": 1.0181699308942569e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.0701278165394714e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 738,
|
|
"real_time": 9.4921506047223613e+05,
|
|
"cpu_time": 1.0181914607045505e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.0699324941824280e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.4914196419080370e+05,
|
|
"cpu_time": 1.0181541065040672e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.0704770054724472e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 9.4918190292776807e+05,
|
|
"cpu_time": 1.0181658983740604e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 6.7108864000000000e+07,
|
|
"bytes_per_second": 7.0701794664438446e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.4119211705820817e+01,
|
|
"cpu_time": 3.7602655677559220e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.5215153936143294e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 14,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.8090754072823626e-05,
|
|
"cpu_time": 3.6932184860179618e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.8092544383367009e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8865249221366290e+06,
|
|
"cpu_time": 2.0029889514822485e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.1145483648309555e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8864783726386586e+06,
|
|
"cpu_time": 2.0030472264151266e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.1147239187410736e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8866043035794462e+06,
|
|
"cpu_time": 2.0031703692723266e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.1142490105290909e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8864947489169086e+06,
|
|
"cpu_time": 2.0031719083555427e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.1146621572659180e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 371,
|
|
"real_time": 1.8864075720405604e+06,
|
|
"cpu_time": 2.0030024285714694e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.1149909483672348e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8865019838624406e+06,
|
|
"cpu_time": 2.0030761768193431e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.1146348799468552e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.8864947489169086e+06,
|
|
"cpu_time": 2.0030472264151264e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 1.3421772800000000e+08,
|
|
"bytes_per_second": 7.1146621572659180e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.1622127163071326e+01,
|
|
"cpu_time": 8.9333547292740917e+01,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 2.7010957790644891e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 15,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 3.7965572141319226e-05,
|
|
"cpu_time": 4.4598177706148166e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.7965346425263997e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 186,
|
|
"real_time": 3.7616185746305892e+06,
|
|
"cpu_time": 3.9970415053760437e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.1361689303217499e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 186,
|
|
"real_time": 3.7612530104916100e+06,
|
|
"cpu_time": 3.9990545268816883e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.1368625096803696e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 186,
|
|
"real_time": 3.7613493272714238e+06,
|
|
"cpu_time": 4.0001872473115041e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.1366797562174255e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 186,
|
|
"real_time": 3.7620891003258605e+06,
|
|
"cpu_time": 4.0002375537628671e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.1352764073755966e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 186,
|
|
"real_time": 3.7622843751363377e+06,
|
|
"cpu_time": 3.9977109892473188e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.1349060632949219e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7617188775711642e+06,
|
|
"cpu_time": 3.9988463645158848e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.1359787333780136e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 3.7616185746305897e+06,
|
|
"cpu_time": 3.9990545268816883e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 2.6843545600000000e+08,
|
|
"bytes_per_second": 7.1361689303217499e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.5291075811271349e+02,
|
|
"cpu_time": 1.4425064344960381e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 8.5914947879772354e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 16,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 1.2039994822929062e-04,
|
|
"cpu_time": 3.6073064654252435e-04,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 1.2039686648435698e-04,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 93,
|
|
"real_time": 7.5121331779706860e+06,
|
|
"cpu_time": 7.9356560107518369e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.1467171744821136e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 93,
|
|
"real_time": 7.5120281938823964e+06,
|
|
"cpu_time": 7.9357932580644768e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.1468170531789795e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 93,
|
|
"real_time": 7.5127008541296888e+06,
|
|
"cpu_time": 7.9361916344080986e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.1461771528528931e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 93,
|
|
"real_time": 7.5124531623817263e+06,
|
|
"cpu_time": 7.9363012043031445e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.1464127681801346e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 93,
|
|
"real_time": 7.5116514238298582e+06,
|
|
"cpu_time": 7.9350650107527897e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.1471755238380493e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5121933624388725e+06,
|
|
"cpu_time": 7.9358014236560687e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.1466599345064346e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 7.5121331779706869e+06,
|
|
"cpu_time": 7.9357932580644768e+06,
|
|
"time_unit": "ns",
|
|
"bytes": 5.3687091200000000e+08,
|
|
"bytes_per_second": 7.1467171744821136e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 4.0300250301165721e+02,
|
|
"cpu_time": 4.9131473346485274e+02,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 3.8339415033200495e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 17,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 5.3646449654328435e-05,
|
|
"cpu_time": 6.1911167787071615e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.3646620077842426e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 0,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.5014225220743647e+07,
|
|
"cpu_time": 1.5817491446810851e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.1514967187019333e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 1,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.5012614765858397e+07,
|
|
"cpu_time": 1.5816773510637853e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.1522638843827362e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 2,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.5012999283189470e+07,
|
|
"cpu_time": 1.5816390446809934e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.1520806985070786e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 3,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.5013213645904623e+07,
|
|
"cpu_time": 1.5816717872338176e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.1519785791691605e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"run_type": "iteration",
|
|
"repetitions": 5,
|
|
"repetition_index": 4,
|
|
"threads": 1,
|
|
"iterations": 47,
|
|
"real_time": 1.5011160911873300e+07,
|
|
"cpu_time": 1.5814757382979728e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.1529565921227844e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time_mean",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "mean",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5012842765513888e+07,
|
|
"cpu_time": 1.5816426131915310e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.1521552945767380e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time_median",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "median",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.5012999283189470e+07,
|
|
"cpu_time": 1.5816717872338172e+07,
|
|
"time_unit": "ns",
|
|
"bytes": 1.0737418240000000e+09,
|
|
"bytes_per_second": 7.1520806985070786e+10,
|
|
"own_gpu": 7.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time_stddev",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "stddev",
|
|
"aggregate_unit": "time",
|
|
"iterations": 5,
|
|
"real_time": 1.1128620680396107e+03,
|
|
"cpu_time": 1.0157111569292431e+03,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 5.3018502358327704e+06,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": 0.0000000000000000e+00
|
|
},
|
|
{
|
|
"name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time_cv",
|
|
"family_index": 7,
|
|
"per_family_instance_index": 18,
|
|
"run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time",
|
|
"run_type": "aggregate",
|
|
"repetitions": 5,
|
|
"threads": 1,
|
|
"aggregate_name": "cv",
|
|
"aggregate_unit": "percentage",
|
|
"iterations": 5,
|
|
"real_time": 7.4127337868080143e-05,
|
|
"cpu_time": 6.4218752609331993e-05,
|
|
"time_unit": "ns",
|
|
"bytes": 0.0000000000000000e+00,
|
|
"bytes_per_second": 7.4129405996721046e-05,
|
|
"own_gpu": 0.0000000000000000e+00,
|
|
"wr_gpu": NaN
|
|
}
|
|
]
|
|
}
|