Method run

src/core/NEON/kernels/NEROIPoolingLayerKernel.cpp:183–256

}

void NEROIPoolingLayerKernel::run(const Window &window, const ThreadInfo &info)
{
    ARM_COMPUTE_UNUSED(info);
    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);

    const size_t values_per_roi = _rois->info()->dimension(0);

    const int   roi_list_start = window.x().start();
    const int   roi_list_end   = window.x().end();
    const int   width          = _input->info()->dimension(Window::DimX);
    const int   height         = _input->info()->dimension(Window::DimY);
    const int   fms            = _input->info()->dimension(Window::DimZ);
    const int   pooled_w       = _pool_info.pooled_width();
    const int   pooled_h       = _pool_info.pooled_height();
    const float spatial_scale  = _pool_info.spatial_scale();

    const auto *rois_ptr  = reinterpret_cast<const uint16_t *>(_rois->buffer());
    const auto  data_type = _input->info()->data_type();

    for (int roi_indx = roi_list_start; roi_indx < roi_list_end; ++roi_indx)
    {
        const unsigned int roi_batch = rois_ptr[values_per_roi * roi_indx];
        const auto         x1        = rois_ptr[values_per_roi * roi_indx + 1];
        const auto         y1        = rois_ptr[values_per_roi * roi_indx + 2];
        const auto         x2        = rois_ptr[values_per_roi * roi_indx + 3];
        const auto         y2        = rois_ptr[values_per_roi * roi_indx + 4];

        // Scale ROI
        const int roi_anchor_x = support::cpp11::round(x1 * spatial_scale);
        const int roi_anchor_y = support::cpp11::round(y1 * spatial_scale);
        const int roi_width    = std::max(support::cpp11::round((x2 - x1) * spatial_scale), 1.f);
        const int roi_height   = std::max(support::cpp11::round((y2 - y1) * spatial_scale), 1.f);

        // Iterate through all feature maps
        for (int fm = 0; fm < fms; ++fm)
        {
            // Iterate through all output pixels
            for (int py = 0; py < pooled_h; ++py)
            {
                for (int px = 0; px < pooled_w; ++px)
                {
                    auto region_start_x = static_cast<int>(std::floor((static_cast<float>(px) / pooled_w) * roi_width));
                    auto region_end_x =
                        static_cast<int>(std::floor((static_cast<float>(px + 1) / pooled_w) * roi_width));
                    auto region_start_y =
                        static_cast<int>(std::floor((static_cast<float>(py) / pooled_h) * roi_height));
                    auto region_end_y =
                        static_cast<int>(std::floor((static_cast<float>(py + 1) / pooled_h) * roi_height));

                    region_start_x = std::min(std::max(region_start_x + roi_anchor_x, 0), width);
                    region_end_x   = std::min(std::max(region_end_x + roi_anchor_x, 0), width);
                    region_start_y = std::min(std::max(region_start_y + roi_anchor_y, 0), height);
                    region_end_y   = std::min(std::max(region_end_y + roi_anchor_y, 0), height);

                    switch (data_type)
                    {
                        case DataType::F32: