This repository contains implementations of matrix multiplication algorithms for both GPU and CPU with different optimizations.
CUDA implementation (float_cuda_2d_tiled_register_cache_matrix_multiplication.cu) is 17.3 % slower than cublasSgemm() from cublas_v2.h and 567 % faster than algorithm without any optimizations on large matrices.
- Refactor the code to use single main function and just include the kernels at compile time
- Add CPU performance comparison to some library implementation