host_gemm.hpp

#pragma once
#include "host_tensor.hpp"
#include "gemm_common.hpp"

template <typename AType, typename BType, typename CType>
void host_gemm(const Tensor<AType>& a,
               const Tensor<BType>& b,
               Tensor<CType>& c,
               const GemmMatrixLayout layout)
{
    auto f_mk_nk_mn = [&](auto m, auto n) {
        const int K = a.mDesc.GetLengths()[1];

        double v = 0;

        for(int k = 0; k < K; ++k)
        {
            v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(n, k));
        }

        c(m, n) = v;
    };

    if(layout == GemmMatrixLayout::MK_NK_MN)
    {
        make_ParallelTensorFunctor(f_mk_nk_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
            std::thread::hardware_concurrency());
    }
    else
    {
        throw std::runtime_error("wrong! not supported layout");
    }
}