Method exec

dnn/src/rocm/matrix_mul/blas.cpp:33–146

}

void MatrixMulForwardImpl::AlgoBlas::exec(const ExecArgs& args) const {
    auto m = args.layout_c.shape[0], n = args.layout_c.shape[1];
    auto k = args.layout_a.shape[args.opr->param().transposeA ? 0 : 1];
    auto&& handle = concrete_handle(args.opr->handle());
    auto rocblas_handle_ = handle->get_rocblas_handle();

    auto sgemm = [&]() {
        auto zero = handle->zero_device();
        auto one = handle->one_device();
        rocblas_check(rocblas_sgemm(
                rocblas_handle_,
                args.opr->param().transposeB ? rocblas_operation_transpose
                                             : rocblas_operation_none,
                args.opr->param().transposeA ? rocblas_operation_transpose
                                             : rocblas_operation_none,
                n, m, k, one, args.tensor_b.ptr<dt_float32>(), args.layout_b.stride[0],
                args.tensor_a.ptr<dt_float32>(), args.layout_a.stride[0], zero,
                args.tensor_c.ptr<dt_float32>(), args.layout_c.stride[0]));
    };

#if !MEGDNN_DISABLE_FLOAT16
    //! used for FLOAT_IO16xC32, not tested
    auto gemm_ex = [&]() {
        auto zero = handle->zero_device();
        auto one = handle->one_device();
        //! These two arguments for future use, see
        //! https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/library/src/blas_ex/rocblas_gemm_ex.cpp
        int32_t solution_index = 0;
        uint32_t flags = 1;
        size_t ws_size = 0;
        auto gemm_ex_err = rocblas_gemm_ex(
                rocblas_handle_,
                args.opr->param().transposeB ? rocblas_operation_transpose
                                             : rocblas_operation_none,
                args.opr->param().transposeA ? rocblas_operation_transpose
                                             : rocblas_operation_none,
                n, m, k, one, args.tensor_b.raw_ptr(), rocblas_datatype_f16_r,
                args.layout_b.stride[0], args.tensor_a.raw_ptr(),
                rocblas_datatype_f16_r, args.layout_a.stride[0], zero,
                args.tensor_c.raw_ptr(), rocblas_datatype_f16_r,
                args.layout_c.stride[0], args.tensor_c.raw_ptr(),
                rocblas_datatype_f16_r, args.layout_c.stride[0], rocblas_datatype_f32_r,
                rocblas_gemm_algo_standard, solution_index, flags, &ws_size, nullptr);
        rocblas_check(gemm_ex_err);
        MEGDNN_MARK_USED_VAR(ws_size);
    };

    auto hgemm = [&]() {
        auto one_half = handle->one_device_h();
        auto zero_half = handle->zero_device_h();
        auto hgemm_err = rocblas_hgemm(
                rocblas_handle_,
                args.opr->param().transposeB ? rocblas_operation_transpose
                                             : rocblas_operation_none,
                args.opr->param().transposeA ? rocblas_operation_transpose
                                             : rocblas_operation_none,
                n, m, k, reinterpret_cast<const rocblas_half*>(one_half),
                static_cast<const rocblas_half*>(args.tensor_b.raw_ptr()),