Method run_op

src/gpu/cl/kernels/ClPool2dKernel.cpp:339–404

}

void ClPool2dKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
{
    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);

    unsigned int pool_stride_x             = 0;
    unsigned int pool_stride_y             = 0;
    std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride();

    const auto src =
        utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
    auto dst     = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST_0));
    auto indices = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST_1));

    // Collapse window
    Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);

    switch (_data_layout)
    {
        case DataLayout::NCHW:
        {
            Window slice = window_collapsed.first_slice_window_3D();
            do
            {
                // Set srcs
                unsigned int idx = 0;
                add_3D_tensor_argument(idx, src, slice);
                add_3D_tensor_argument(idx, dst, slice);
                if (indices && is_data_type_float(src->info()->data_type()) && (_pool_info.pool_size == Size2D(2, 2)))
                {
                    add_3D_tensor_argument(idx, indices, slice);
                }
                enqueue(queue, *this, slice, lws_hint());
            } while (window_collapsed.slide_window_slice_3D(slice));
            break;
        }
        case DataLayout::NHWC:
        {
            const size_t batch_size = dst->info()->tensor_shape().total_size_upper(3);

            Window slice    = window_collapsed.first_slice_window_4D();
            Window in_slice = window_collapsed.first_slice_window_4D();
            in_slice.set(Window::DimX,
                         Window::Dimension(0, src->info()->dimension(0), _num_elems_processed_per_iteration));
            in_slice.set(Window::DimY, Window::Dimension(0, src->info()->dimension(1), pool_stride_x));
            in_slice.set(Window::DimZ, Window::Dimension(0, src->info()->dimension(2), pool_stride_y));
            in_slice.set(3, Window::Dimension(0, batch_size, 1));
            do
            {
                // Set srcs
                unsigned int idx = 0;
                add_4D_tensor_argument(idx, src, in_slice);
                add_4D_tensor_argument(idx, dst, slice);
                if (indices && is_data_type_float(src->info()->data_type()) &&
                    (_pool_info.pool_type == PoolingType::MAX) && (_pool_info.pool_size == Size2D(2, 2)))
                {
                    add_4D_tensor_argument(idx, indices, slice);
                }