nvbench
nvbench copied to clipboard
NVIDIA
CUDA Kernel Benchmarking Library
## Description In an attempt to gather more accurate timings, [nvbench will "flush" the L2 cache](https://github.com/NVIDIA/nvbench/blob/1a13a2e724b8aa8aee27649ac6878babb63862a6/nvbench/detail/l2flush.cuh#L28) by querying the device's L2 cache size, allocating device memory of that size, memset...
Similar to #47 This PR updates the code thus numbers of bytes are displayed in base-ten format. i.e. B, MB, KB and GB.
Nvbench currently does not support benchmarking CPU-only code natively. Although adding `nvbench::exec_tag::sync` gives plausible measurements for cold runs, there is no mechanism for batch measurements. We could enable this feature...
Hello, I am getting a lot of errors when compiling the library with tests and examples with MSVC. MSVC version = 19.29.30154.0 NVCC version = 12.3.107 I am compiling with...
The steps to reproduce: ```bash git clone https://github.com/NVIDIA/nvbench_demo git submodule update --init --recursive ``` Modify CMakeLists.txt and use c++20. ```cmake cmake_minimum_required(VERSION 3.20.1) project(NVBenchDemo CUDA CXX) # CXX added add_subdirectory(nvbench) add_executable(example_bench...
The custom stopping criterion introduced in https://github.com/NVIDIA/nvbench/pull/151 currently triggers linear regression and R2 computation every second `is_finished()` call. Per @julie-is-late [comment](https://github.com/NVIDIA/nvbench/pull/151#discussion_r1445601451): > depending on the expected amount of time a...
The custom stopping criterion introduced in #151 stores all parameters for all criterion implementations in a single map. This requires that all parameters must be uniquely named across implementations along...
Currently, NVBench relies on a fixed [criteria](https://github.com/NVIDIA/nvbench/blob/57c4d42ba505d525a1a3815d2b2560450a4f83dd/nvbench/detail/measure_cold.cu#L102-L154) to understand when enough samples are collected. We can consider an API that would allow users to modify this criterion, accepting the set...
Adds an `NVBENCH_ENVIRONMENT` macro that can be custom defined to a global, fixture-like class which will be created and destroyed in the nvbench `main()` runtime. The definition should occur before...
