C++ Coordinates::set_num_dimensions方法代码示例

//.........这里部分代码省略.........
    if(is_interleaved_transposed)
    {
        build_opts.add_option("-DCOLS_B=" + support::cpp11::to_string(input1->info()->dimension(0)));
        if(data_type == DataType::F32)
        {
            kernel_name = "gemm_mm_interleaved_transposed_f32_" + string_from_target(arch_target);
        }
        else
        {
            kernel_name = "gemm_mm_interleaved_transposed_" + lower_string(string_from_data_type(data_type));
        }

        // Configure kernel window
        const unsigned int     num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(data_type);
        constexpr unsigned int num_elems_processed_per_iteration_y = 4;

        Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));

        AccessWindowRectangle input0_access(input0->info(), 0, 0, num_elems_processed_per_iteration_y, 1, 1.f, 0.25f);
        AccessWindowTranspose input1_access(input1->info(), 0, 0, num_elems_processed_per_iteration_x, 1, 0.f, 0.25f);
        AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);

        update_window_and_padding(win, input0_access, input1_access, output_access);

        output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));

        ICLKernel::configure(win);
    }
    else // The input tensors have not been reshaped
    {
        build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(input0->info()->dimension(0)));

        // Special case for 1xN, 2xN, 3xN and 4xN input0 tensor. num_elems_processed_per_iteration_x is set up for the default case.
        unsigned int       num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(data_type);
        const unsigned int num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->info()->dimension(1)), 4);

        // Create kernels according to the architecture, data type and input size.
        if(arch_target == GPUTarget::BIFROST && data_type == DataType::F32)
        {
            // The first kernel is optimized for the case of 1000 or less output elements (e.g. FC8 of AlexNet and VGG-16, and
            // FC1 of Inception v3). The second kernel is optimized for the case of greater than 1000 output elements (e.g.
            // FC6 and FC7 of AlexNet and VGG-16).
            if(input1->info()->dimension(0) <= 1000)
            {
                // Each work-item processes 2 elements in the X dimension.
                num_elems_processed_per_iteration_x = 2;
                kernel_name                         = "gemm_mm_floating_point_f32_bifrost_1000";
            }
            else
            {
                // Each work-item processes 4 elements in the X dimension (as in the default case).
                num_elems_processed_per_iteration_x = 4;
                kernel_name                         = "gemm_mm_floating_point_f32_bifrost";
            }
            // The work-group size equal to the Bifrost quad size has been proved to be optimal for these kernels
            // via exhaustive autotuning over a range of representative layer configurations.
            _lws_hint = cl::NDRange(4);
        }
        else if(is_data_type_fixed_point(data_type))
        {
            kernel_name = "gemm_mm_" + lower_string(string_from_data_type(data_type));
        }
        else // (MIDGARD and F32) or (F16)
        {
            build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
            kernel_name = "gemm_mm_floating_point";
        }
        build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_Y=" + support::cpp11::to_string(num_elems_processed_per_iteration_y));
        build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_X=" + support::cpp11::to_string(num_elems_processed_per_iteration_x));

        // Configure window
        Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));

        AccessWindowStatic    input0_access(input0->info(), 0, 0, input0->info()->dimension(0), ceil_to_multiple(input0->info()->dimension(1), num_elems_processed_per_iteration_y));
        AccessWindowStatic    input1_access(input1->info(), 0, 0, ceil_to_multiple(input1->info()->dimension(0), num_elems_processed_per_iteration_x), input1->info()->dimension(1));
        AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);

        update_window_and_padding(win, input0_access, input1_access, output_access);

        Coordinates coord;
        coord.set_num_dimensions(output->info()->num_dimensions());
        output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape()));

        ICLKernel::configure(win);
    }

    // Create kernel
    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));

    // Set config_id for enabling LWS tuning
    _config_id = "gemm_";
    _config_id += (is_interleaved_transposed ? "reshaped_" : "");
    _config_id += lower_string(string_from_data_type(input0->info()->data_type()));
    _config_id += "_";
    _config_id += support::cpp11::to_string(output->info()->dimension(1));
    _config_id += "_";
    _config_id += support::cpp11::to_string(output->info()->dimension(0));
    _config_id += "_";
    _config_id += (is_interleaved_transposed ? support::cpp11::to_string(input1->info()->dimension(0)) : support::cpp11::to_string(input1->info()->dimension(1)));
}