compute/ARMComputeEx/src/core/CL/kernels/CLArgMinMaxLayerKernelEx.cpp

   1 /*
   2  * Copyright (c) 2020 Samsung Electronics Co., Ltd. All Rights Reserved
   3  *
   4  * Licensed under the Apache License, Version 2.0 (the "License");
   5  * you may not use this file except in compliance with the License.
   6  * You may obtain a copy of the License at
   7  *
   8  *      http://www.apache.org/licenses/LICENSE-2.0
   9  *
  10  * Unless required by applicable law or agreed to in writing, software
  11  * distributed under the License is distributed on an "AS IS" BASIS,
  12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13  * See the License for the specific language governing permissions and
  14  * limitations under the License.
  15  */
  16
  17 /*
  18  * Copyright (c) 2019-2020 ARM Limited.
  19  *
  20  * SPDX-License-Identifier: MIT
  21  *
  22  * Permission is hereby granted, free of charge, to any person obtaining a copy
  23  * of this software and associated documentation files (the "Software"), to
  24  * deal in the Software without restriction, including without limitation the
  25  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  26  * sell copies of the Software, and to permit persons to whom the Software is
  27  * furnished to do so, subject to the following conditions:
  28  *
  29  * The above copyright notice and this permission notice shall be included in all
  30  * copies or substantial portions of the Software.
  31  *
  32  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  33  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  34  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  35  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  36  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  37  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  38  * SOFTWARE.
  39  */
  40 #include "arm_compute/core/CL/kernels/CLArgMinMaxLayerKernelEx.h"
  41
  42 #include "arm_compute/core/AccessWindowStatic.h"
  43 #include "arm_compute/core/CL/CLHelpers.h"
  44 #include "arm_compute/core/CL/CLKernelLibraryEx.h"
  45 #include "arm_compute/core/CL/CLValidate.h"
  46 #include "arm_compute/core/CL/ICLTensor.h"
  47 #include "arm_compute/core/Helpers.h"
  48 #include "arm_compute/core/TensorInfo.h"
  49 #include "arm_compute/core/Utils.h"
  50 #include "arm_compute/core/Validate.h"
  51 #include "arm_compute/core/Window.h"
  52
  53 #include "support/StringSupport.h"
  54
  55 namespace arm_compute
  56 {
  57 namespace
  58 {
  59 constexpr unsigned int vector_size = 16;
  60
  61 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *prev_output,
  62                           const ITensorInfo *output, unsigned int axis, ReductionOperation op)
  63 {
  64   ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
  65   ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
  66   ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::S32,
  67                                                        DataType::F16, DataType::F32);
  68   ARM_COMPUTE_RETURN_ERROR_ON_MSG(op != ReductionOperation::ARG_IDX_MAX &&
  69                                       op != ReductionOperation::ARG_IDX_MIN,
  70                                   "Only ARG_IDX_MAX and ARG_IDX_MIN are supported");
  71   ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions,
  72                                   "Reduction axis greater than max number of dimensions");
  73   ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
  74
  75   if (output->total_size() != 0)
  76   {
  77     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U32, DataType::S32,
  78                                                          DataType::S64);
  79   }
  80   if (prev_output != nullptr && prev_output->total_size() != 0)
  81   {
  82     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(prev_output, 1, DataType::U32,
  83                                                          DataType::S32, DataType::S64);
  84     if (output->total_size() != 0)
  85     {
  86       ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(prev_output, output);
  87     }
  88   }
  89
  90   return Status{};
  91 }
  92
  93 std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input,
  94                                                          ITensorInfo *prev_output,
  95                                                          ITensorInfo *output, unsigned int axis,
  96                                                          ReductionOperation op)
  97 {
  98   ARM_COMPUTE_UNUSED(op);
  99   // Output tensor auto initialization if not yet initialized
 100   TensorShape output_shape{input->tensor_shape()};
 101   output_shape.set(axis, 1);
 102   DataType output_data_type = (prev_output != nullptr) ? (prev_output->data_type()) : DataType::S32;
 103   auto_init_if_empty(*output, input->clone()
 104                                   ->set_tensor_shape(output_shape)
 105                                   .set_data_type(output_data_type)
 106                                   .reset_padding()
 107                                   .set_is_resizable(true));
 108
 109   Window win = calculate_max_window((prev_output != nullptr) ? (*prev_output) : (*input),
 110                                     Steps(vector_size));
 111   bool window_changed = false;
 112
 113   switch (axis)
 114   {
 115     case 0:
 116     {
 117       ITensorInfo *input_tensor_access = prev_output != nullptr ? prev_output : input;
 118       AccessWindowStatic input_access(input_tensor_access, 0, 0,
 119                                       static_cast<int>(input_tensor_access->dimension(0)), 1);
 120       AccessWindowHorizontal output_access(output, 0, 1);
 121       window_changed = update_window_and_padding(win, input_access, output_access);
 122       output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
 123     }
 124     break;
 125     case 1:
 126     case 2:
 127     case 3:
 128     {
 129       AccessWindowHorizontal input_access(input, 0, vector_size);
 130       AccessWindowHorizontal output_access(output, 0, vector_size);
 131       window_changed = update_window_and_padding(win, input_access, output_access);
 132       output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
 133     }
 134     break;
 135     default:
 136       ARM_COMPUTE_ERROR("Not supported");
 137   }
 138
 139   Status err = (window_changed)
 140                    ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!")
 141                    : Status{};
 142   return std::make_tuple(err, win);
 143 }
 144 } // namespace
 145
 146 CLArgMinMaxLayerKernelEx::CLArgMinMaxLayerKernelEx()
 147     : _input(nullptr), _prev_output(nullptr), _output(nullptr), _reduction_axis(0),
 148       _op(ReductionOperation::ARG_IDX_MAX)
 149 {
 150 }
 151
 152 void CLArgMinMaxLayerKernelEx::configure(const ICLTensor *input, const ICLTensor *prev_output,
 153                                          ICLTensor *output, unsigned int axis,
 154                                          ReductionOperation op)
 155 {
 156   ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 157   ARM_COMPUTE_ERROR_THROW_ON(
 158       validate_arguments(input->info(), (prev_output != nullptr) ? prev_output->info() : nullptr,
 159                          output->info(), axis, op));
 160   auto win_config = validate_and_configure_window(
 161       input->info(), (prev_output != nullptr) ? prev_output->info() : nullptr, output->info(), axis,
 162       op);
 163   ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
 164
 165   _input = input;
 166   _prev_output = prev_output;
 167   _output = output;
 168   _reduction_axis = axis;
 169   _op = op;
 170
 171   // Set build options
 172   CLBuildOptions build_opts;
 173
 174   build_opts.add_option_if(_prev_output != nullptr, "-DPREV_OUTPUT");
 175   build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
 176   build_opts.add_option_if(is_data_type_float(input->info()->data_type()), "-DFLOAT_DATA_TYPE");
 177   build_opts.add_option_if_else(op == ReductionOperation::ARG_IDX_MAX, "-DARG_MAX", "-DARG_MIN");
 178   build_opts.add_option("-DDATA_TYPE_OUTPUT=" +
 179                         get_cl_type_from_data_type(output->info()->data_type()));
 180   build_opts.add_option("-DDATA_TYPE_SELECT=" +
 181                         get_cl_signed_type_from_element_size(input->info()->element_size()));
 182
 183   // Create kernel
 184   cl::NDRange lws_hint = CLKernelLibrary::get().default_ndrange();
 185   std::string kernel_axis_name;
 186   switch (axis)
 187   {
 188     case 0:
 189     {
 190       const ICLTensor *input_for_width = prev_output != nullptr ? _prev_output : _input;
 191       build_opts.add_option("-DWIDTH=" +
 192                             support::cpp11::to_string(input_for_width->info()->dimension(0)));
 193
 194       kernel_axis_name = "x";
 195       lws_hint = create_lws_hint_parallel_implementations(input_for_width->info()->dimension(0),
 196                                                           vector_size);
 197     }
 198     break;
 199     case 1:
 200       build_opts.add_option("-DHEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
 201       kernel_axis_name = "y";
 202       break;
 203     case 2:
 204       build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
 205       kernel_axis_name = "z";
 206       break;
 207     case 3:
 208       build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
 209       build_opts.add_option("-DBATCH=" + support::cpp11::to_string(input->info()->dimension(3)));
 210       kernel_axis_name = "w";
 211       break;
 212     default:
 213       ARM_COMPUTE_ERROR("Not supported");
 214   }
 215   _kernel = static_cast<cl::Kernel>(CLKernelLibraryEx::get().create_kernel(
 216       "arg_min_max_ex_" + kernel_axis_name, build_opts.options()));
 217
 218   // Configure kernel window
 219   ICLKernel::configure_internal(std::get<1>(win_config), lws_hint);
 220 }
 221
 222 Status CLArgMinMaxLayerKernelEx::validate(const ITensorInfo *input, const ITensorInfo *prev_output,
 223                                           const ITensorInfo *output, unsigned int axis,
 224                                           ReductionOperation op)
 225 {
 226   ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, prev_output, output, axis, op));
 227   ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(
 228       input->clone().get(), (prev_output != nullptr) ? prev_output->clone().get() : nullptr,
 229       output->clone().get(), axis, op)));
 230   return Status{};
 231 }
 232
 233 void CLArgMinMaxLayerKernelEx::run(const Window &window, cl::CommandQueue &queue)
 234 {
 235   ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
 236   ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
 237
 238   switch (_reduction_axis)
 239   {
 240     case 0:
 241     {
 242       // Set out window
 243       Window out_window(window);
 244       out_window.set(Window::DimX, Window::Dimension(0, 0, 0));
 245
 246       // Get first input and output slices
 247       Window in_slice = window.first_slice_window_2D();
 248       Window out_slice = out_window.first_slice_window_2D();
 249
 250       // Reshape window
 251       const unsigned int num_tensors = _prev_output != nullptr ? 3 : 2;
 252
 253       // Set local sums buffer
 254       unsigned int local_res_size = lws_hint()[0] * _output->info()->element_size();
 255       _kernel.setArg(num_arguments_per_2D_tensor() * num_tensors, local_res_size, nullptr);
 256       do
 257       {
 258         unsigned int idx = 0;
 259         add_2D_tensor_argument(idx, _input, in_slice);
 260         if (_prev_output != nullptr)
 261         {
 262           add_2D_tensor_argument(idx, _prev_output, in_slice);
 263         }
 264         add_2D_tensor_argument(idx, _output, out_slice);
 265         enqueue(queue, *this, in_slice, lws_hint());
 266       } while (window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(out_slice));
 267     }
 268     break;
 269     case 1:
 270     {
 271       // Get first input and output slices
 272       Window window_in{window};
 273       window_in.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1),
 274                                                     _input->info()->dimension(1)));
 275       Window in_slice = window_in.first_slice_window_2D();
 276       Window out_slice = window.first_slice_window_2D();
 277
 278       do
 279       {
 280         unsigned int idx = 0;
 281         add_2D_tensor_argument(idx, _input, in_slice);
 282         add_2D_tensor_argument(idx, _output, out_slice);
 283         enqueue(queue, *this, in_slice, lws_hint());
 284       } while (window_in.slide_window_slice_2D(in_slice) &&
 285                window.slide_window_slice_2D(out_slice));
 286     }
 287     break;
 288     case 2:
 289     {
 290       // Get first input and output slices
 291       Window window_in{window};
 292       window_in.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2),
 293                                                     _input->info()->dimension(2)));
 294       Window in_slice = window_in.first_slice_window_3D();
 295       Window out_slice = window.first_slice_window_3D();
 296
 297       do
 298       {
 299         unsigned int idx = 0;
 300         add_3D_tensor_argument(idx, _input, in_slice);
 301         add_3D_tensor_argument(idx, _output, out_slice);
 302         enqueue(queue, *this, in_slice, lws_hint());
 303       } while (window_in.slide_window_slice_3D(in_slice) &&
 304                window.slide_window_slice_3D(out_slice));
 305     }
 306     break;
 307     case 3:
 308     {
 309       // Get first input and output slices
 310       Window window_in{window};
 311       window_in.set(3, Window::Dimension(0, 1, 1));
 312       Window in_slice = window_in.first_slice_window_4D();
 313       Window out_slice = window.first_slice_window_4D();
 314
 315       do
 316       {
 317         unsigned int idx = 0;
 318         add_4D_tensor_argument(idx, _input, in_slice);
 319         add_4D_tensor_argument(idx, _output, out_slice);
 320         enqueue(queue, *this, in_slice, lws_hint());
 321       } while (window_in.slide_window_slice_4D(in_slice) &&
 322                window.slide_window_slice_4D(out_slice));
 323     }
 324     break;
 325     default:
 326       ARM_COMPUTE_ERROR("Not supported");
 327   }
 328 }
 329 } // namespace arm_compute