Function gemm_finalize_impl

src/targets/gpu/gemm_impl.cpp:612–643

 */
template <class T>
static int32_t gemm_finalize_impl(context& ctx,
                                  const shape& output_shape,
                                  const std::vector<shape>& input_shapes,
                                  T alpha,
                                  T beta,
                                  bool compute_fp32,
                                  int32_t solution_idx)
{
#ifdef MIGRAPHX_USE_ROCBLAS_TUNING_API

    // This code should be called only if either the environment var.
    // MIGRAPHX_ENABLE_GEMM_TUNING, or option --exhaustive-tune, is set

    if(solution_idx == 0)
    {
        auto gemm_item = gemm_impl<T>(output_shape, input_shapes, alpha, beta, compute_fp32);
        solution_idx   = gemm_item.tune(ctx, input_shapes);
        gemm_save_solution(ctx, output_shape, input_shapes, solution_idx);
    }
    else
    {
        // If a tuned solution index is already given, don't tune again but validate
        // in case the data was tuned with a different rocBLAS version
        auto gemm_item = gemm_impl<T>(output_shape, input_shapes, alpha, beta, compute_fp32);
        solution_idx   = gemm_item.validate(ctx, input_shapes, solution_idx);
    }
#else
    (void)ctx, (void)output_shape, (void)input_shapes;
    (void)alpha, (void)beta, (void)compute_fp32;
#endif
    return solution_idx;
}

int32_t gemm_finalize(context& ctx,
                      const shape& output_shape,