search

NVIDIA / nvbench

CUDA Kernel Benchmarking Library
Apache License 2.0
525 stars 66 forks source link

No batch tag #146

Closed gevtushenko closed 1 year ago

gevtushenko commented 1 year ago

NVBench always collects batch measurements:

NumElements DataSize Samples CPU Time Noise GPU Time Noise Elem/s GlobalMem BW BWUtil Samples Batch GPU
16777216 64.000 MiB 2624x 200.347 us 16.22% 190.839 us 2.91% 87.913G 703.303 GB/s 73.25% 3059x 174.476 us

In some cases, batch measurements are not required. By providing a way to disable batch measurements, we can speedup benchmarking. This PR introduces a new tag:

state.exec(nvbench::exec_tag::no_batch, [&](nvbench::launch &launch) {
/// ...

leads to absence of Samples and Batch GPU columns:

NumElements DataSize Samples CPU Time Noise GPU Time Noise Elem/s GlobalMem BW BWUtil
16777216 64.000 MiB 2656x 197.683 us 16.99% 188.337 us 2.55% 89.081G 712.645 GB/s 74.23%
alliepiper commented 1 year ago

The commit needs to be signed, but feel free to merge once that's done.