src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp

   1 /*
   2  * Copyright (c) 2018 ARM Limited.
   3  *
   4  * SPDX-License-Identifier: MIT
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to
   8  * deal in the Software without restriction, including without limitation the
   9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  10  * sell copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in all
  14  * copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  */
  24 #include "arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h"
  25
  26 #include "arm_compute/core/AccessWindowStatic.h"
  27 #include "arm_compute/core/CL/CLHelpers.h"
  28 #include "arm_compute/core/CL/CLHelpers.h"
  29 #include "arm_compute/core/CL/CLKernelLibrary.h"
  30 #include "arm_compute/core/CL/ICLTensor.h"
  31 #include "arm_compute/core/Helpers.h"
  32 #include "arm_compute/core/IAccessWindow.h"
  33 #include "arm_compute/core/TensorInfo.h"
  34 #include "arm_compute/core/Types.h"
  35 #include "arm_compute/core/Utils.h"
  36 #include "arm_compute/core/Validate.h"
  37 #include "arm_compute/core/Window.h"
  38 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
  39
  40 #include "support/ToolchainSupport.h"
  41
  42 using namespace arm_compute;
  43 using namespace arm_compute::misc::shape_calculator;
  44
  45 namespace
  46 {
  47 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info)
  48 {
  49     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
  50     ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() != DataLayout::NCHW);
  51
  52     const Size2D kernel_size      = winograd_info.kernel_size;
  53     const Size2D output_tile_size = winograd_info.output_tile_size;
  54
  55     const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
  56     const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
  57
  58     ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size != Size2D(3U, 3U) && kernel_size != Size2D(5U, 5U), "Winograd filter transform only supports 3x3 and 5x5 kernels");
  59     ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size == Size2D(3U, 3U) && output_tile_size != Size2D(2U, 2U)
  60                                     && output_tile_size != Size2D(4U, 4U),
  61                                     "Winograd filter transform only supports 2x2 or 4x4 output tile for 3x3 kernels");
  62     ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size == Size2D(5U, 5U) && output_tile_size != Size2D(4U, 4U), "Winograd filter transform only supports 4x4 output tile for 5x5 kernels");
  63     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_w) != kernel_size.width || input->dimension(idx_h) != kernel_size.height);
  64     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
  65
  66     // Checks performed when output is configured
  67     if(output->total_size() != 0)
  68     {
  69         const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(compute_winograd_filter_transform_shape(*input, winograd_info));
  70
  71         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
  72         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
  73     }
  74
  75     return Status{};
  76 }
  77
  78 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
  79 {
  80     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
  81
  82     const unsigned int num_elems_processed_per_iteration_x = input->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH));
  83     const unsigned int num_elems_processed_per_iteration_y = input->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT));
  84
  85     Window win            = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
  86     bool   window_changed = false;
  87
  88     AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
  89     AccessWindowStatic    output_access(output, 0, 0, output->dimension(0), output->dimension(1));
  90     window_changed = update_window_and_padding(win, input_access, output_access);
  91     output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
  92
  93     Window win_collapsed = win.collapse(win, Window::DimZ);
  94
  95     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
  96     return std::make_pair(err, win_collapsed);
  97 }
  98 } // namespace
  99
 100 CLWinogradFilterTransformKernel::CLWinogradFilterTransformKernel()
 101     : _input(nullptr), _output(nullptr)
 102 {
 103 }
 104
 105 void CLWinogradFilterTransformKernel::configure(const ICLTensor *input, ICLTensor *output, const WinogradInfo &winograd_info)
 106 {
 107     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 108
 109     // Output auto initialization if not yet initialized
 110     auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_winograd_filter_transform_shape(*input->info(), winograd_info)));
 111
 112     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), winograd_info));
 113
 114     const size_t idx_c = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::CHANNEL);
 115
 116     // Set build options
 117     CLBuildOptions build_opts;
 118     build_opts.add_option("-DNUM_CHANNELS=" + support::cpp11::to_string(input->info()->dimension(idx_c)));
 119
 120     const Size2D kernel_size      = winograd_info.kernel_size;
 121     const Size2D output_tile_size = winograd_info.output_tile_size;
 122
 123     // Create kernel
 124     std::string kernel_name = "winograd_filter_transform_" + output_tile_size.to_string() + "_" + kernel_size.to_string() + "_nchw";
 125     _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 126
 127     _input  = input;
 128     _output = output;
 129
 130     // Configure kernel window
 131     auto win_config = validate_and_configure_window(input->info(), output->info());
 132     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
 133     ICLKernel::configure(win_config.second);
 134 }
 135
 136 Status CLWinogradFilterTransformKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info)
 137 {
 138     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, winograd_info));
 139     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
 140
 141     return Status{};
 142 }
 143
 144 void CLWinogradFilterTransformKernel::run(const Window &window, cl::CommandQueue &queue)
 145 {
 146     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
 147     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
 148
 149     // Setup output window
 150     Window window_out;
 151     window_out.use_tensor_dimensions(_output->info()->tensor_shape(), 0);
 152
 153     unsigned int idx = 0;
 154     add_4D_tensor_argument(idx, _input, window);
 155     add_3D_tensor_argument(idx, _output, window_out);
 156     enqueue(queue, *this, window);
 157 }