Function reference_gpu_gemm

include/reference.h:79–90

template<class AType, class BType, class CType, class AccumType>
void reference_gpu_gemm(GemmParams<AType, BType, CType, AccumType> params) {
    /// don't know what tilings are good, but not important here
    const int mTile = 4;
    const int nTile = 4;
    const int kTile = 4;
    const int tx = 128;
    const int ty = 1;
    dim3 grid(ceil_div(params.N, nTile * tx), ceil_div(params.M, mTile * ty));
    dim3 block(tx, ty);
    reference_gpu_gemm_kernel<AType, BType, CType, AccumType, mTile, nTile, kTile><<<grid, block>>>(params);
    CUDA_CHECK(cudaGetLastError());
}