Method configure

src/gpu/cl/kernels/ClDirectConv2dKernel.cpp:187–452

}

void ClDirectConv2dKernel::configure(const CLCompileContext            &compile_context,
                                     ITensorInfo                       *src,
                                     ITensorInfo                       *weights,
                                     ITensorInfo                       *biases,
                                     ITensorInfo                       *dst,
                                     const PadStrideInfo               &conv_info,
                                     const ActivationLayerInfo         &act_info,
                                     const DirectConvComputeKernelInfo &desc)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst);

    // Perform validation
    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, weights, biases, dst, conv_info, act_info, desc));

    const int conv_stride_x = std::get<0>(conv_info.stride());
    const int conv_stride_y = std::get<1>(conv_info.stride());

    _data_layout = src->data_layout();
    _conv_info   = conv_info;

    const unsigned int width_idx   = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
    const unsigned int height_idx  = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
    const unsigned int channel_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
    const unsigned int kernel_size = weights->dimension(width_idx);
    const DataType     data_type   = src->data_type();

    const GPUTarget gpu_target                         = get_target();
    unsigned int    _num_elems_processed_per_iteration = 0;

    // Get dst shape
    TensorShape output_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *weights, conv_info);

    // Output auto inizialitation if not yet initialized
    auto_init_if_empty(*dst, output_shape, 1, src->data_type(), src->quantization_info());

    // Configure kernel window
    Window win;
    if (_data_layout == DataLayout::NHWC)
    {
        output_shape.collapse(2U, 1U);
        const unsigned int n0 = adjust_vec_size(desc.n0, output_shape[0]);
        const unsigned int m0 = adjust_vec_size(desc.m0, output_shape[1]);

        // Create window and update padding
        win = calculate_max_window(output_shape, Steps(n0, m0));
    }
    else if (_data_layout == DataLayout::NCHW)
    {
        _num_elems_processed_per_iteration = 1u;
        win                                = calculate_max_window(*dst, Steps(_num_elems_processed_per_iteration));
    }

    ICLKernel::configure_internal(win);

    std::stringstream kernel_name;
    CLBuildOptions    build_options;

    if (_data_layout == DataLayout::NHWC)