Method configure

src/gpu/cl/kernels/ClTransposedConvolutionKernel.cpp:117–259

} // namespace

void ClTransposedConvolutionKernel::configure(const CLCompileContext &compile_context,
                                              const ITensorInfo      *input,
                                              const ITensorInfo      *weights,
                                              const ITensorInfo      *biases,
                                              ITensorInfo            *output,
                                              const PadStrideInfo    &deconv_info)
{
    ARM_COMPUTE_UNUSED(biases, deconv_info);
    ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);

    // Perform validation
    ARM_COMPUTE_ERROR_THROW_ON(validate(input, weights, biases, output, deconv_info));

    constexpr unsigned int channel_idx = 0;
    constexpr unsigned int width_idx   = 1;
    constexpr unsigned int height_idx  = 2;

    const size_t input_channels  = input->dimension(channel_idx); // same as weight channels
    const size_t input_width     = input->dimension(width_idx);
    const size_t input_height    = input->dimension(height_idx);
    const size_t weights_width   = weights->dimension(width_idx);
    const size_t weights_height  = weights->dimension(height_idx);
    const size_t output_width    = output->dimension(width_idx);
    const size_t output_height   = output->dimension(height_idx);
    const size_t output_channels = output->dimension(channel_idx);

    // Calculate output shape
    auto out_dims =
        deconvolution_output_dimensions(input_width, input_height, weights_width, weights_height, deconv_info);
    TensorShape output_shape = misc::shape_calculator::compute_deconvolution_output_shape(out_dims, *input, *weights);
    auto_init_if_empty(*output, output_shape, 1, input->data_type(), input->quantization_info());

    // Calculate updated paddings
    // p' = k - p - 1 (k: kernel dimensions)
    const uint32_t pad_left = weights_width - deconv_info.pad_left() - 1;
    const uint32_t pad_top  = weights_height - deconv_info.pad_top() - 1;

    // Configure kernel window
    Window win;
    output_shape.collapse(2U, 1U); // Collapse width and height into single dimension

    const unsigned int n0               = adjust_vec_size(16 / output->element_size(), output_channels);
    const unsigned int m0               = 1;
    const unsigned int k0               = adjust_vec_size(16 / input->element_size(), input_channels);
    const unsigned int partial_store_n0 = output_channels % n0;

    // Create window and update padding
    win = calculate_max_window(output_shape, Steps(n0, m0));
    ICLKernel::configure_internal(win);

    const std::string kernel_name = "transposed_convolution_nhwc";
    CLBuildOptions    build_options;

    const DataType    input_data_type = input->data_type();
    const PaddingInfo strides         = deconv_info.stride();

    if (biases != nullptr)
    {