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,
  67                                                        DataType::QASYMM8_SIGNED, DataType::S32,
  68                                                        DataType::F16, DataType::F32);
  69   ARM_COMPUTE_RETURN_ERROR_ON_MSG(op != ReductionOperation::ARG_IDX_MAX &&
  70                                     op != ReductionOperation::ARG_IDX_MIN,
  71                                   "Only ARG_IDX_MAX and ARG_IDX_MIN are supported");
  72   ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions,
  73                                   "Reduction axis greater than max number of dimensions");
  74   ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
  75
  76   if (output->total_size() != 0)
  77   {
  78     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U32, DataType::S32,
  79                                                          DataType::S64);
  80   }
  81   if (prev_output != nullptr && prev_output->total_size() != 0)
  82   {
  83     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(prev_output, 1, DataType::U32,
  84                                                          DataType::S32, DataType::S64);
  85     if (output->total_size() != 0)
  86     {
  87       ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(prev_output, output);
  88     }
  89   }
  90
  91   return Status{};
  92 }
  93
  94 std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input,
  95                                                          ITensorInfo *prev_output,
  96                                                          ITensorInfo *output, unsigned int axis,
  97                                                          ReductionOperation op)
  98 {
  99   ARM_COMPUTE_UNUSED(op);
 100   // Output tensor auto initialization if not yet initialized
 101   TensorShape output_shape{input->tensor_shape()};
 102   output_shape.set(axis, 1);
 103   DataType output_data_type = (prev_output != nullptr) ? (prev_output->data_type()) : DataType::S32;
 104   auto_init_if_empty(*output, input->clone()
 105                                 ->set_tensor_shape(output_shape)
 106                                 .set_data_type(output_data_type)
 107                                 .reset_padding()
 108                                 .set_is_resizable(true));
 109
 110   Window win =
 111     calculate_max_window((prev_output != nullptr) ? (*prev_output) : (*input), Steps(vector_size));
 112   bool window_changed = false;
 113
 114   switch (axis)
 115   {
 116     case 0:
 117     {
 118       ITensorInfo *input_tensor_access = prev_output != nullptr ? prev_output : input;
 119       AccessWindowStatic input_access(input_tensor_access, 0, 0,
 120                                       static_cast<int>(input_tensor_access->dimension(0)), 1);
 121       AccessWindowHorizontal output_access(output, 0, 1);
 122       window_changed = update_window_and_padding(win, input_access, output_access);
 123       output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
 124     }
 125     break;
 126     case 1:
 127     case 2:
 128     case 3:
 129     {
 130       AccessWindowHorizontal input_access(input, 0, vector_size);
 131       AccessWindowHorizontal output_access(output, 0, vector_size);
 132       window_changed = update_window_and_padding(win, input_access, output_access);
 133       output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
 134     }
 135     break;
 136     default:
 137       ARM_COMPUTE_ERROR("Not supported");
 138   }
 139
 140   Status err = (window_changed)
 141                  ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!")
 142                  : Status{};
 143   return std::make_tuple(err, win);
 144 }
 145 } // namespace
 146
 147 CLArgMinMaxLayerKernelEx::CLArgMinMaxLayerKernelEx()
 148   : _input(nullptr), _prev_output(nullptr), _output(nullptr), _reduction_axis(0),
 149     _op(ReductionOperation::ARG_IDX_MAX)
 150 {
 151 }
 152
 153 void CLArgMinMaxLayerKernelEx::configure(const ICLTensor *input, const ICLTensor *prev_output,
 154                                          ICLTensor *output, unsigned int axis,
 155                                          ReductionOperation op)
 156 {
 157   ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 158   ARM_COMPUTE_ERROR_THROW_ON(
 159     validate_arguments(input->info(), (prev_output != nullptr) ? prev_output->info() : nullptr,
 160                        output->info(), axis, op));
 161   auto win_config = validate_and_configure_window(
 162     input->info(), (prev_output != nullptr) ? prev_output->info() : nullptr, output->info(), axis,
 163     op);
 164   ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
 165
 166   _input = input;
 167   _prev_output = prev_output;
 168   _output = output;
 169   _reduction_axis = axis;
 170   _op = op;
 171
 172   // Set build options
 173   CLBuildOptions build_opts;
 174
 175   build_opts.add_option_if(_prev_output != nullptr, "-DPREV_OUTPUT");
 176   build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
 177   build_opts.add_option_if(is_data_type_float(input->info()->data_type()), "-DFLOAT_DATA_TYPE");
 178   build_opts.add_option_if_else(op == ReductionOperation::ARG_IDX_MAX, "-DARG_MAX", "-DARG_MIN");
 179   build_opts.add_option("-DDATA_TYPE_OUTPUT=" +
 180                         get_cl_type_from_data_type(output->info()->data_type()));
 181   build_opts.add_option("-DDATA_TYPE_SELECT=" +
 182                         get_cl_signed_type_from_element_size(input->info()->element_size()));
 183
 184   // Create kernel
 185   cl::NDRange lws_hint = CLKernelLibrary::get().default_ndrange();
 186   std::string kernel_axis_name;
 187   switch (axis)
 188   {
 189     case 0:
 190     {
 191       const ICLTensor *input_for_width = prev_output != nullptr ? _prev_output : _input;
 192       build_opts.add_option("-DWIDTH=" +
 193                             support::cpp11::to_string(input_for_width->info()->dimension(0)));
 194
 195       kernel_axis_name = "x";
 196       lws_hint = create_lws_hint_parallel_implementations(input_for_width->info()->dimension(0),
 197                                                           vector_size);
 198     }
 199     break;
 200     case 1:
 201       build_opts.add_option("-DHEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
 202       kernel_axis_name = "y";
 203       break;
 204     case 2:
 205       build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
 206       kernel_axis_name = "z";
 207       break;
 208     case 3:
 209       build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
 210       build_opts.add_option("-DBATCH=" + support::cpp11::to_string(input->info()->dimension(3)));
 211       kernel_axis_name = "w";
 212       break;
 213     default:
 214       ARM_COMPUTE_ERROR("Not supported");
 215   }
 216   _kernel = static_cast<cl::Kernel>(CLKernelLibraryEx::get().create_kernel(
 217     "arg_min_max_ex_" + kernel_axis_name, build_opts.options()));
 218
 219   // Configure kernel window
 220   ICLKernel::configure_internal(std::get<1>(win_config), lws_hint);
 221 }
 222
 223 Status CLArgMinMaxLayerKernelEx::validate(const ITensorInfo *input, const ITensorInfo *prev_output,
 224                                           const ITensorInfo *output, unsigned int axis,
 225                                           ReductionOperation op)
 226 {
 227   ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, prev_output, output, axis, op));
 228   ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(
 229     input->clone().get(), (prev_output != nullptr) ? prev_output->clone().get() : nullptr,
 230     output->clone().get(), axis, op)));
 231   return Status{};
 232 }
 233
 234 void CLArgMinMaxLayerKernelEx::run(const Window &window, cl::CommandQueue &queue)
 235 {
 236   ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
 237   ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
 238
 239   switch (_reduction_axis)
 240   {
 241     case 0:
 242     {
 243       // Set out window
 244       Window out_window(window);
 245       out_window.set(Window::DimX, Window::Dimension(0, 0, 0));
 246
 247       // Get first input and output slices
 248       Window in_slice = window.first_slice_window_2D();
 249       Window out_slice = out_window.first_slice_window_2D();
 250
 251       // Reshape window
 252       const unsigned int num_tensors = _prev_output != nullptr ? 3 : 2;
 253
 254       // Set local sums buffer
 255       unsigned int local_res_size = lws_hint()[0] * _output->info()->element_size();
 256       _kernel.setArg(num_arguments_per_2D_tensor() * num_tensors, local_res_size, nullptr);
 257       do
 258       {
 259         unsigned int idx = 0;
 260         add_2D_tensor_argument(idx, _input, in_slice);
 261         if (_prev_output != nullptr)
 262         {
 263           add_2D_tensor_argument(idx, _prev_output, in_slice);
 264         }
 265         add_2D_tensor_argument(idx, _output, out_slice);
 266         enqueue(queue, *this, in_slice, lws_hint());
 267       } while (window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(out_slice));
 268     }
 269     break;
 270     case 1:
 271     {
 272       // Get first input and output slices
 273       Window window_in{window};
 274       window_in.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1),
 275                                                     _input->info()->dimension(1)));
 276       Window in_slice = window_in.first_slice_window_2D();
 277       Window out_slice = window.first_slice_window_2D();
 278
 279       do
 280       {
 281         unsigned int idx = 0;
 282         add_2D_tensor_argument(idx, _input, in_slice);
 283         add_2D_tensor_argument(idx, _output, out_slice);
 284         enqueue(queue, *this, in_slice, lws_hint());
 285       } while (window_in.slide_window_slice_2D(in_slice) &&
 286                window.slide_window_slice_2D(out_slice));
 287     }
 288     break;
 289     case 2:
 290     {
 291       // Get first input and output slices
 292       Window window_in{window};
 293       window_in.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2),
 294                                                     _input->info()->dimension(2)));
 295       Window in_slice = window_in.first_slice_window_3D();
 296       Window out_slice = window.first_slice_window_3D();
 297
 298       do
 299       {
 300         unsigned int idx = 0;
 301         add_3D_tensor_argument(idx, _input, in_slice);
 302         add_3D_tensor_argument(idx, _output, out_slice);
 303         enqueue(queue, *this, in_slice, lws_hint());
 304       } while (window_in.slide_window_slice_3D(in_slice) &&
 305                window.slide_window_slice_3D(out_slice));
 306     }
 307     break;
 308     case 3:
 309     {
 310       // Get first input and output slices
 311       Window window_in{window};
 312       window_in.set(3, Window::Dimension(0, 1, 1));
 313       Window in_slice = window_in.first_slice_window_4D();
 314       Window out_slice = window.first_slice_window_4D();
 315
 316       do
 317       {
 318         unsigned int idx = 0;
 319         add_4D_tensor_argument(idx, _input, in_slice);
 320         add_4D_tensor_argument(idx, _output, out_slice);
 321         enqueue(queue, *this, in_slice, lws_hint());
 322       } while (window_in.slide_window_slice_4D(in_slice) &&
 323                window.slide_window_slice_4D(out_slice));
 324     }
 325     break;
 326     default:
 327       ARM_COMPUTE_ERROR("Not supported");
 328   }
 329 }
 330 } // namespace arm_compute