src/gpu/cl/kernels/ClWinogradInputTransformKernel.cpp

/*
 * Copyright (c) 2018-2023 Arm Limited.
 *
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to
 * deal in the Software without restriction, including without limitation the
 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
 * sell copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */
#include "src/gpu/cl/kernels/ClWinogradInputTransformKernel.h"

#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibrary.h"
#include "arm_compute/core/CL/ICLTensor.h"
#include "arm_compute/core/CL/OpenCL.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/core/utils/StringUtils.h"
#include "src/core/AccessWindowStatic.h"
#include "src/core/CL/CLValidate.h"
#include "src/core/helpers/AutoConfiguration.h"
#include "src/core/helpers/WindowHelpers.h"
#include "support/Cast.h"
#include "support/StringSupport.h"

namespace arm_compute
{
namespace opencl
{
namespace kernels
{
namespace
{
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info)
{
    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);

    const PadStrideInfo conv_info        = winograd_info.convolution_info;
    const Size2D        output_tile_size = winograd_info.output_tile_size;
    const Size2D        kernel_size      = winograd_info.kernel_size;
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd input transform only supports unit strides");
    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!cl_winograd_convolution_layer_supported(output_tile_size, kernel_size, input->data_layout()), "Winograd input transform not supported");

    ARM_COMPUTE_UNUSED(conv_info);
    ARM_COMPUTE_UNUSED(output_tile_size);
    ARM_COMPUTE_UNUSED(kernel_size);

    // Validate configured output
    if(output->total_size() != 0)
    {
        const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info);

        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
    }

    return Status{};
}

std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const WinogradInfo &winograd_info)
{
    ARM_COMPUTE_UNUSED(output);
    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);

    bool window_changed                    = false;
    int  num_elems_processed_per_iteration = 1;

    if(input->data_layout() == DataLayout::NHWC)
    {
        // In the case of FP16 computation, we can perform more
        // output feature maps in a single work-item.
        // From experiments, num_elems_processed_per_iteration = 2 looks good for fp16 to
        // improve the performance. However, in order to make the implementation simpler,
        // we set num_elems_processed_per_iteration = 2 only when the OFMs are multiple of 2.
        // Note: At the moment, only Winograd Input Transform 3x3 can support N0 != 1
        const DataType dt   = input->data_type();
        const size_t   dim0 = input->dimension(0);
        const size_t   k_sz = winograd_info.kernel_size.area();
        const bool     cond = dt == DataType::F16 && ((dim0 % 2) == 0);
        if(cond)
        {
            if(k_sz == 3 || k_sz == 9)
            {
                num_elems_processed_per_iteration = 2;
            }
        }
    }
    Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));

    if(input->data_layout() == DataLayout::NCHW)
    {
        const PadStrideInfo conv_info        = winograd_info.convolution_info;
        const Size2D        output_tile_size = winograd_info.output_tile_size;
        const Size2D        kernel_size      = winograd_info.kernel_size;

        unsigned int num_elems_read_per_iteration_x = output_tile_size.width + kernel_size.width - 1;
        unsigned int num_elems_read_per_iteration_y = output_tile_size.height + kernel_size.height - 1;

        AccessWindowRectangle input_access(input, -conv_info.pad_left(), -conv_info.pad_top(), num_elems_read_per_iteration_x, num_elems_read_per_iteration_y);
        window_changed = update_window_and_padding(win, input_access);
    }

    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
    return std::make_pair(err, win);
}
} // namespace

ClWinogradInputTransformKernel::ClWinogradInputTransformKernel()
{
    _type = CLKernelType::WINOGRAD;
}

BorderSize ClWinogradInputTransformKernel::border_size() const
{
    return _border_size;
}

void ClWinogradInputTransformKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const WinogradInfo &winograd_info)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, winograd_info));

    auto padding_info = get_padding_info({ src, dst });

    const PadStrideInfo conv_info        = winograd_info.convolution_info;
    const Size2D        output_tile_size = winograd_info.output_tile_size;
    const Size2D        kernel_size      = winograd_info.kernel_size;

    _data_layout = src->data_layout();

    const size_t idx_w = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
    const size_t idx_h = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);

    // Compute the number of output tiles along the x and y direction of size "output_tile_size"
    const Size2D num_tiles = compute_winograd_convolution_tiles(Size2D(src->dimension(idx_w), src->dimension(idx_h)),
                                                                kernel_size,
                                                                output_tile_size,
                                                                conv_info);

    _num_tiles_x = num_tiles.width;
    _num_tiles_y = num_tiles.height;

    const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*src, winograd_info);

    // Output auto initialization if not yet initialized
    auto_init_if_empty(*dst, src->clone()->set_tensor_shape(output_shape));

    ARM_COMPUTE_ERROR_ON(_num_tiles_x * _num_tiles_y != static_cast<int>(dst->dimension(1)));
    const size_t total_batches = src->tensor_shape().total_size_upper(3);

    // Create window and update padding
    auto win_config = validate_and_configure_window(src, dst, winograd_info);
    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
    IClKernel::configure_internal(win_config.second, cl::NDRange(1, 1, 8));

    _src_width  = src->dimension(idx_w);
    _src_height = src->dimension(idx_h);

    CLBuildOptions build_opts;
    if(_data_layout == DataLayout::NHWC)
    {
        build_opts.add_option("-DNHWC");
        build_opts.add_option("-DN0=" + support::cpp11::to_string(win_config.second.x().step()));
        build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
        build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
        build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width));
        build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height));
        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type()));
        build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL");
        build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_INPUT_TRANSFORM_VERTICAL");
        build_opts.add_option_if(total_batches > 1, "-DIS_BATCHED");
    }
    else
    {
        build_opts.add_option("-DNUM_TILES_X=" + support::cpp11::to_string(_num_tiles_x));
        build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
        build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
        build_opts.add_option("-DOUTPUT_TILE_W=" + support::cpp11::to_string(output_tile_size.width));
        build_opts.add_option("-DOUTPUT_TILE_H=" + support::cpp11::to_string(output_tile_size.height));
        build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src->data_type()));
        build_opts.add_option_if(winograd_info.kernel_size.height == 1, "-DWINOGRAD_INPUT_TRANSFORM_HORIZONTAL");
        build_opts.add_option_if(winograd_info.kernel_size.width == 1, "-DWINOGRAD_INPUT_TRANSFORM_VERTICAL");
        build_opts.add_option_if(total_batches > 1, "-DSRC_DEPTH=" + support::cpp11::to_string(src->dimension(2)));
    }

    // Create kernel
    std::string kernel_name = "winograd_input_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string();

    // Get the maximum dimension from the tile size
    const unsigned int tile_max_dim = std::max(output_tile_size.width, output_tile_size.height);

    // Check optimized kernel if output_dims == 2x2
    if((tile_max_dim == 2) && (_data_layout == DataLayout::NCHW))
    {
        _step_z = (src->dimension(2) % 2) != 0 ? 1 : 2;
    }

    // Append stepz and data layout
    kernel_name += "_stepz";
    kernel_name += support::cpp11::to_string(_step_z);
    kernel_name += "_" + lower_string(string_from_data_layout(_data_layout));

    // A macro guard to compile ONLY the kernel of interest
    build_opts.add_option("-D" + upper_string(kernel_name));
    _kernel = create_kernel(compile_context, kernel_name, build_opts.options());

    _border_size = BorderSize(src->padding());

    ARM_COMPUTE_ERROR_ON((src->data_layout() == DataLayout::NHWC) && has_padding_changed(padding_info));

    _config_id = kernel_name;
    _config_id += support::cpp11::to_string(src->dimension(0));
    _config_id += "_";
    _config_id += support::cpp11::to_string(src->dimension(1));
    _config_id += "_";
    _config_id += support::cpp11::to_string(src->dimension(2));
    _config_id += "_";
    _config_id += support::cpp11::to_string(conv_info.pad_left());
    _config_id += "_";
    _config_id += support::cpp11::to_string(conv_info.pad_top());
    _config_id += "_";
    _config_id += lower_string(string_from_data_layout(_data_layout));
}

Status ClWinogradInputTransformKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const WinogradInfo &winograd_info)
{
    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, winograd_info));
    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), dst->clone().get(), winograd_info).first);
    return Status{};
}

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

    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));

    const size_t idx_w         = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
    const size_t idx_h         = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
    const size_t idx_c         = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
    const size_t total_batches = window.shape().total_size_upper(3);

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

    if(_data_layout == DataLayout::NHWC)
    {
        Window slice = window_collapsed.first_slice_window_3D();
        slice.set(1, Window::Dimension(0, _num_tiles_x * _num_tiles_y, 1));
        slice.set(2, Window::Dimension(0, total_batches, 1));

        unsigned int idx = 0;
        add_4D_tensor_argument(idx, src, slice);
        add_4D_tensor_argument(idx, dst, slice);
        _kernel.setArg<cl_uint>(idx++, _src_width);
        _kernel.setArg<cl_uint>(idx++, _src_height);
        _kernel.setArg<cl_uint>(idx++, _num_tiles_x);
        _kernel.setArg<cl_uint>(idx++, _num_tiles_y);
        enqueue(queue, *this, slice, lws_hint());
    }
    else
    {
        Window slice = window_collapsed.first_slice_window_3D();
        slice.set(idx_w, Window::Dimension(0, _num_tiles_x, 1));
        slice.set(idx_h, Window::Dimension(0, _num_tiles_y, 1));

        ARM_COMPUTE_ERROR_ON(((slice[idx_c].end() - slice[idx_c].start()) % _step_z) != 0);
        slice.set(idx_c, Window::Dimension(slice[idx_c].start(), slice[idx_c].end(), _step_z));

        unsigned int idx = 2 * 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()[3]));

        do
        {
            unsigned int idx = 0;
            add_3D_tensor_argument(idx, src, slice);
            add_3D_tensor_argument(idx, dst, slice);

            enqueue(queue, *this, slice, lws_hint());
        }
        while(window_collapsed.slide_window_slice_3D(slice));
    }
}
} // namespace kernels
} // namespace opencl
} // namespace arm_compute