Function brute_force_gemm

src/opr/test/blas.cpp:18–30

namespace {
template <typename dt_src, typename dt_dst>
void brute_force_gemm(
        size_t M, size_t N, size_t K, bool transa, bool transb, const dt_src* x,
        const dt_src* y, dt_dst* z) {
    for (size_t m = 0; m < M; ++m)
        for (size_t n = 0; n < N; ++n) {
            dt_dst cur = dt_dst(0);
            for (size_t k = 0; k < K; ++k) {
                cur += x[transa ? (k * M + m) : (m * K + k)] *
                       y[transb ? (n * K + k) : (k * N + n)];
            }
            z[m * N + n] = cur;
        }
}

float brute_force_dot(const HostTensorND& a, const HostTensorND& b) {
    auto sz = std::max(a.shape(0), b.shape(0));