DGEMM (Double-precision GEneral Matrix Multiplication) kernels with a bunch of basic optimizations such as SIMD vectorization, loop unrolling, blocking and multi-threading.
- CBLAS (For details, see: http://www.netlib.org/blas/)
main: A main function which executes all the kernels and show the performance of each.dgemm...: Performs DGEMM formatrices in column major order (for easier SIMD vectorization).
...avx2 (or avx512)...: A SIMD vectorized version of the kernel using AVX2 (or AVX-512) instruction intrinsics....unroll...: A loop-unrolled version of the kernel. Unrolling factor is statically defined indgemm.h....block...: A blocked (tiled) version of the kernel. Block size is statically defined indgemm.h....omp...: A multi-threaded version of the kernel using OpenMP directives.
Use make run to build and run the kernels. To enable AVX-512 versions, build with make avx512=1 and run.
You can also specify the size of the matrices with
make run size=(an exponent of 2).