Method run

src/core/CL/kernels/CLReductionOperationKernel.cpp:237–379

}

void CLReductionOperationKernel::run(const Window &window, cl::CommandQueue &queue)
{
    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);

    const bool is_serial_op = needs_serialized_reduction(_op, _input->info()->data_type(), _reduction_axis);
    switch (_reduction_axis)
    {
        case 0:
        {
            // We use parallel reduction only in non quantized types
            if (is_serial_op)
            {
                // Get first input and output slices
                Window window_in{window};
                window_in.set(Window::DimX,
                              Window::Dimension(0, _input->info()->dimension(0), _input->info()->dimension(0)));

                Window out_window{window};
                out_window.set(Window::DimX, Window::Dimension(0, 0, 0));

                Window in_slice  = window_in.first_slice_window_1D();
                Window out_slice = out_window.first_slice_window_1D();

                do
                {
                    unsigned int idx = 0;
                    add_1D_tensor_argument(idx, _input, in_slice);
                    add_1D_tensor_argument(idx, _output, out_slice);
                    enqueue(queue, *this, in_slice);
                } while (window_in.slide_window_slice_1D(in_slice) && out_window.slide_window_slice_1D(out_slice));
            }
            else
            {
                // Set out window
                bool   has_collapsed = true;
                Window window_in     = window.collapse_if_possible(window, 2, &has_collapsed);
                ARM_COMPUTE_ERROR_ON(!has_collapsed);

                Window window_out = window_in;
                window_out.set(0, Window::Dimension());

                unsigned int idx = 0;
                add_3D_tensor_argument(idx, _input, window_in);
                add_3D_tensor_argument(idx, _output, window_out);
                enqueue(queue, *this, window_in);
            }
        }
        break;
        case 1:
        {
            bool   has_collapsed = true;
            Window actual_window = window.collapse_if_possible(window, 2, &has_collapsed);
            ARM_COMPUTE_ERROR_ON(!has_collapsed);

            actual_window = actual_window.shift_dimensions(1, Window::DimY);

            const ITensorInfo *input_info    = _input->info();