IntMatrix.h with SSE extension

IntMatrix.h is simple std::vector<std::vector> wrapper to represent matrixes. IntMatrix.h uses SSE extension to speed up addition and substraction operations.

Build

Mac M1

• brew for x86 is required
• cmake installed under x86 brew is required

To build under Rosetta 2 with SSE extension enabled. Script adds required defintions automatically.

./scripts/build_x86.sh

To build without SSE support use:

./scripts/build_arm64.sh

x86

• cmake is required

cmake . -Bbuild/x86 -DBUILD_X86_WITH_SSE=1
cmake --build build/x86

Used SSE intrinsics

See this guide

1. __m128i _mm_lddqu_si128 (__m128i const* mem_addr)
2. __m128i _mm_sub_epi32 (__m128i a, __m128i b)
3. __m128i _mm_add_epi32 (__m128i a, __m128i b)
4. void _mm_storeu_si32 (void* mem_addr, __m128i a)
5. __m128i _mm_bsrli_si128 (__m128i a, int imm8)

Testing

You could generate your own tests using script generateTestcases.py. To see usage run without parameters. Paste generated tests into tests/test.cpp and build.

Benchmarking

The following script build two versions of static library IntMatrix.h with support of custom SSE and without it. As second step, it builds three executables:

1. With -O3 and with custom SSE support
2. With -O3 and without custom SSE support
3. Without -O3 and without custom SSE support

As third step all executables run with time recording functions.

./scripts/runBenchmark.sh

Сonclusions

1. Custom SSE twices perfomance comparing to code without compiler optimizations
2. -O3 flag still gives more perfomance because compiler seems to be using AVX