gemmBench

int8_t gemm benchmark between Eigen, kpu's intgemm, dnnl, fbgemm, and mkl.

Compilation

mkdir build
cd build
cmake -DWITH_MKL=OFF ..
make -j

If you have Intel MKL installed on your system, you can set -DWITH_MKL=ON during the CMake configuration.

Usage

./benchmark [iterations=1000] [arch=any] [use_eigen=0]

Changing parameters

Some paramters are hardcoded

• The memory alignment can be changed here: https://github.com/XapaJIaMnu/gemmBench/blob/master/bench.cpp#L338
• Matrix sizes can be changed here: https://github.com/XapaJIaMnu/gemmBench/blob/master/bench.cpp#L372
• For intgemm to work, you need M and N to be a multiple of 8 and K to be a multiple of 32
• The number of iterations of the loop can be varied through command one.
• You can limit arch for intgemm and dnnl. Supported values: ssse3, avx2, avx512, avx512vnni and any
• Since Eigen is a lot slower than the other two, its execution is disabled by default. To enable it, provide the argument.

Caveats

• Fbgemm only supports AVX2 processors or newer, so the test is skipped on older architectures.
• Fbgemm doesn't allow for limiting the arch type, so the test is skipped in case explicit arch is requested