nvbench
nvbench copied to clipboard
NVIDIA
CUDA Kernel Benchmarking Library
Fixes: #295 Can compare the files in #295: ``` ['base.json', 'base_bif.json'] # mul ## [0] NVIDIA H200 | T{ct} | OffsetT{ct} | Elements{io} | Ref Time | Ref Noise |...
When I run the same benchmark, once full and once a subset (to save execution time): ``` ./bin/cub.bench.transform.babelstream.base -d 0 --stopping-criterion entropy --json base.json ./bin/cub.bench.transform.babelstream.base -d 0 -a "T{ct}=I16" -a...
This PR builds on top of #290. Hence, it should be merged after #290 has been merged. This PR moves doc-strings for Python objects exposed in `cuda.bench` namespace from `__init__.pyi`...
This PR adds `try/catch` around invocation of Python function that defines benchmark function. If execution of benchmark for some input parameters resulted in Python exception being raised, it gets propagated...
Presently, all doc-strings are collected in `__init__.pyi` ``` In [1]: import cuda.bench as bench In [2]: bench.State.__doc__ is None Out[2]: True In [3]: bench.Launch.__doc__ is None Out[3]: True ``` Even...
`NVBench` library contains a number of CUDA driver and CUDA runtime calls. These calls are direct, such as, in [main.cuh:178](https://github.com/NVIDIA/nvbench/blob/main/nvbench/main.cuh#L178): ```cpp NVBENCH_DRIVER_API_CALL(cuInit(0)); ``` When `libnvbench.so` is compiled it would contain...
When running a pynvbench script with multiple benchmarks, if I hit `Ctrl-C`, it terminates the current benchmark and moves on to the next benchmark (instead of terminating execution altogether). This...
This issue pens down issues that should be addressed prior to release of Python API of NVBench. - [ ] Fix Ctrl-C interrupt behavior #284 - [ ] Make doc-strings...
This PR introduces "multi-CUDA" wheels builds, exactly similar to the ones we build for CCCL, as described [here](https://github.com/NVIDIA/cccl/pull/5613). Instead of shipping multiple packages (`pynvbench-cu12`, `pynvbench-cu13`); we ship a single one...
This pull request is based on the idea from the issue [#121](https://github.com/NVIDIA/nvbench/issues/121). The user would be able to provide Cupti metrics it wants to use for his benchmark. By filling...
