kh4ster / cuda_bench_memory_space_transfer

Benchmark template project

The goal of this project is to be forked to serve as a base for anyone needing to benchmark several Cuda functions quickly.

Requirements

• Cuda Toolkit
• C++ compiler (g++ for linux, MSVC for Windows)
• GPU supported by CUDA
• CMake
• Conan

Additional libraries

These libraries are included in the conan file. Do not install them yourself. Conan will do the job for you.

• GoogleTest
• GoogleBenchmark

Build

• To build, execute the following commands :

mkdir build && cd build
conan install ..
cmake ..
make

Run :

cd build
./bin/Bench

• Additionnaly you can use --build=missing to build missing libraries:

conan install .. --build missing

---

Additional infos

• By default the program will run in release when it's inside a build or build_release folder. To build in debug, build the projet inside a build_debug folder.

• You can specify the "--no-check" option when running the bench binary to disable result checking :

./bin/Bench --no-check

To bench your own code

Compile your code

There are two ways to get your code compiled:

• Directly copy the function in the file src/to_bench.cu
• Create a new cmake target

Create a new cmake target

In CMakeLists.txt:

• Create a new library as follows:

add_library(LIB_NAME
	SOURCE_FILE1
	SOURCE_FILE2
	...
)

• Link this library to the Bench target (add this library among the others) as follows:

target_link_libraries(Bench LIB_NAME async_memcpy GTest::GTest benchmark::benchmark TestHelpers)

• Note: be careful not to have the same functions name for exported functions between libraries (avoid multiple definition compilation error)

Create new benches

In bench/main.cc:

• Include the header file containing the function(s) to bench #include header_file.cuh if not already in the included files
• Define a new bench as follows

BENCHMARK_DEFINE_F(Fixture, BENCH_NAME)
(benchmark::State &st)
{
    this->bench(st, NAME_OF_THE_FUNCTION_TO_BENCH, BUFFER_SIZE, FUNCTION_ARGS...);
}

• Register the new bench as follows

BENCHMARK_REGISTER_F(Fixture, BENCH_NAME)
    ->UseRealTime()
    ->Unit(benchmark::kMillisecond);

• Note: a template of this steps can directly be found in src/main.cc
• Note: the first function argument must be a cuda_tools::host_shared_ptr<int>

Extra ressources

• You can use premade host_shared_ptr to allocate data
• You can use premade test_helper to test your result