Method run_op

src/gpu/cl/kernels/ClIm2ColKernel.cpp:421–494

}

void ClIm2ColKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
{
    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
    ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(IClKernel::window(), window);
    ARM_COMPUTE_ERROR_ON(tensors.empty());

    // Get initial windows
    // Collapse in order to have (SRC_DEPTH * BATCH_SIZE) on the 3rd dimension
    Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
    window_collapsed.set_dimension_step(Window::DimZ, 1);

    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));
    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);

    Window window_output;
    window_output.use_tensor_dimensions(dst->info()->tensor_shape());

    const Window first_slice_3d = window_collapsed.first_slice_window_3D();

    Window slice     = first_slice_3d;
    Window slice_in  = first_slice_3d;
    Window slice_out = window_output.first_slice_window_2D();

    if (_data_layout == DataLayout::NHWC)
    {
        const Window tmp_win     = window.collapse_if_possible(ICLKernel::window(), 3);
        const int    num_batches = tmp_win[3].end();

        slice.set(1, Window::Dimension(0, static_cast<int>(dst->info()->tensor_shape()[1]), 1));
        slice.set(2, Window::Dimension(0, static_cast<int>(num_batches), 1));
    }
    else
    {
        slice.set(0,
                  Window::Dimension(
                      0, static_cast<int>(ceil_to_multiple(_convolved_dims.first, _num_elems_processed_per_iteration)),
                      _num_elems_processed_per_iteration));
        slice.set(1, Window::Dimension(0, static_cast<int>(_convolved_dims.second), 1));
        // Note: In case of NCHW the 3rd dimension is already set collapsing the input window
    }

    // Setup input slice
    // The dimensions of the input are increased within the OpenCL kernel
    slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
    slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
    slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));

    // Setup output slice
    // The dimensions of the output are increased within the OpenCL kernel
    slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
    slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));

    unsigned int idx = num_arguments_per_3D_tensor() +
                       (_num_groups == 1 ? num_arguments_per_2D_tensor() : num_arguments_per_3D_tensor());
    _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src->info()->strides_in_bytes()[3]));
    _kernel.setArg<cl_uint>(idx++,
                            static_cast<unsigned int>(dst->info()->strides_in_bytes()[((_num_groups == 1) ? 2 : 3)]));