1 /*
2 * Copyright (c) 2017-2020 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 "src/core/CL/kernels/CLPoolingLayerKernel.h"
25
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/CLKernelLibrary.h"
28 #include "arm_compute/core/CL/ICLTensor.h"
29 #include "arm_compute/core/CL/OpenCL.h"
30 #include "arm_compute/core/Helpers.h"
31 #include "arm_compute/core/TensorInfo.h"
32 #include "arm_compute/core/Utils.h"
33 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
34 #include "src/core/AccessWindowStatic.h"
35 #include "src/core/CL/CLValidate.h"
36 #include "src/core/CL/ICLKernel.h"
37 #include "src/core/helpers/AutoConfiguration.h"
38 #include "src/core/helpers/WindowHelpers.h"
39 #include "support/StringSupport.h"
40
41 #include <set>
42 #include <string>
43 #include <tuple>
44
45 namespace arm_compute
46 {
47 using namespace arm_compute::misc::shape_calculator;
48
49 namespace
50 {
51 // Internal window config info
52 using CLPoolingConfig = std::pair<unsigned int, BorderSize>; //num_elems_processed_per_iteration, border_size
53
auto_init(const ITensorInfo * input,ITensorInfo * output,ITensorInfo * indices,PoolingLayerInfo pool_info)54 void auto_init(const ITensorInfo *input, ITensorInfo *output, ITensorInfo *indices, PoolingLayerInfo pool_info)
55 {
56 TensorShape out_shape = compute_pool_shape(*input, pool_info);
57 auto_init_if_empty(*output, input->clone()->set_tensor_shape(out_shape));
58 if(indices)
59 {
60 auto_init_if_empty(*indices, input->clone()->set_tensor_shape(out_shape).set_data_type(DataType::U32));
61 }
62 }
63
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,const PoolingLayerInfo & pool_info,const ITensorInfo * indices)64 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices)
65 {
66 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
67 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
68 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
69 ARM_COMPUTE_RETURN_ERROR_ON_MSG((is_data_type_quantized_asymmetric(input->data_type()) && pool_info.pool_type == PoolingType::L2),
70 "Unsupported combination of parameters!");
71
72 // Check indices
73 if(indices)
74 {
75 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
76 ARM_COMPUTE_RETURN_ERROR_ON_MSG(pool_info.pool_type != PoolingType::MAX, "Pooling indices only supported for MAX pooling method");
77 ARM_COMPUTE_RETURN_ERROR_ON_MSG((pool_info.pool_size != Size2D(2, 2)), "Pooling indices only supported for pool size 2x2");
78
79 if(indices->total_size() != 0)
80 {
81 TensorInfo idx_info(TensorInfo(compute_pool_shape(*input, pool_info), 1, DataType::U32));
82 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(indices, &idx_info);
83 }
84 }
85
86 // Checks performed when output is configured
87 if(output->total_size() != 0)
88 {
89 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
90 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
91 TensorInfo out_info(TensorInfo(compute_pool_shape(*input, pool_info), 1, output->data_type()));
92 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &out_info);
93 }
94
95 return Status{};
96 }
97
validate_and_configure_window(ITensorInfo * input,ITensorInfo * output,const PoolingLayerInfo & pool_info,ITensorInfo * indices=nullptr)98 std::tuple<Status, Window, CLPoolingConfig> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const PoolingLayerInfo &pool_info, ITensorInfo *indices = nullptr)
99 {
100 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
101
102 // Get data layout
103 const DataLayout data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? input->data_layout() : pool_info.data_layout;
104 const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
105 const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
106
107 int pool_stride_x = 0;
108 int pool_stride_y = 0;
109 unsigned int pooled_w = 0;
110 unsigned int pooled_h = 0;
111 int pool_size_x = pool_info.is_global_pooling ? input->dimension(idx_width) : pool_info.pool_size.width;
112 int pool_size_y = pool_info.is_global_pooling ? input->dimension(idx_height) : pool_info.pool_size.height;
113 const PadStrideInfo pad_stride_info = pool_info.pad_stride_info;
114 std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride();
115 const int pool_pad_right = pad_stride_info.pad_right();
116 const int pool_pad_top = pad_stride_info.pad_top();
117 const int pool_pad_left = pad_stride_info.pad_left();
118 const int pool_pad_bottom = pad_stride_info.pad_bottom();
119 BorderSize border_size = BorderSize();
120
121 auto_init(input, output, indices, pool_info);
122 pooled_w = output->tensor_shape()[idx_width];
123 pooled_h = output->tensor_shape()[idx_height];
124
125 const DataType data_type = input->data_type();
126
127 const int input_width = input->dimension(idx_width);
128 const int input_height = input->dimension(idx_height);
129
130 unsigned int num_elems_processed_per_iteration = 0;
131 bool window_changed = false;
132 Window win{};
133 switch(data_layout)
134 {
135 case DataLayout::NCHW:
136 {
137 // Initialize border size
138 border_size = BorderSize(pool_pad_top, pool_pad_right, pool_pad_bottom, pool_pad_left);
139 // Change the number of elements processed per iteration
140 // for pooling 3x3 with stride less equal than 3
141 const bool can_optimize = (pool_size_x == 3) && (pool_size_y == 3) && (pool_stride_x <= 3) && !is_data_type_quantized(data_type);
142 num_elems_processed_per_iteration = can_optimize ? 4 : 1;
143 const unsigned int num_elems_read_per_iteration = (num_elems_processed_per_iteration - 1) * pool_stride_x + pool_size_x;
144
145 // Number of iterations in X dimension
146 const int num_iterations_x = (pooled_w + num_elems_processed_per_iteration - 1) / num_elems_processed_per_iteration;
147
148 // Upper limit for the number of right/bottom border elements that are accessed
149 const int upper_bound_w = ((num_iterations_x - 1) * num_elems_processed_per_iteration * pool_stride_x - pool_pad_left + num_elems_read_per_iteration) - input_width;
150 const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_top + pool_size_y) - input_height;
151
152 border_size.right = std::max(upper_bound_w, pool_pad_right);
153 border_size.bottom = std::max(upper_bound_h, pool_pad_bottom);
154
155 win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
156
157 AccessWindowRectangle input_access(input, -pool_pad_left, -pool_pad_top, num_elems_read_per_iteration, pool_size_y,
158 pool_stride_x, pool_stride_y);
159 AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
160
161 // Update indices window
162 if(indices)
163 {
164 AccessWindowHorizontal indices_access(indices, 0, num_elems_processed_per_iteration);
165 window_changed = update_window_and_padding(win, input_access, output_access, indices_access);
166 indices_access.set_valid_region(win, ValidRegion(Coordinates(), indices->tensor_shape()));
167 }
168 else
169 {
170 window_changed = update_window_and_padding(win, input_access, output_access);
171 }
172
173 output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
174 break;
175 }
176 case DataLayout::NHWC:
177 {
178 // Initialize border size
179 border_size = BorderSize();
180 num_elems_processed_per_iteration = adjust_vec_size(4, output->dimension(0));
181 win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
182
183 if(indices != nullptr)
184 {
185 indices->set_valid_region(ValidRegion(Coordinates(), indices->tensor_shape()));
186 }
187
188 output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
189 break;
190 }
191 default:
192 ARM_COMPUTE_ERROR("Not implemented");
193 }
194
195 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
196 return std::make_tuple(err, win, CLPoolingConfig(num_elems_processed_per_iteration, border_size));
197 }
198 } // namespace
199
CLPoolingLayerKernel()200 CLPoolingLayerKernel::CLPoolingLayerKernel()
201 : _input(nullptr), _output(nullptr), _indices(nullptr), _pool_info(), _data_layout(DataLayout::UNKNOWN), _border_size(0), _num_elems_processed_per_iteration(1)
202 {
203 }
204
border_size() const205 BorderSize CLPoolingLayerKernel::border_size() const
206 {
207 return _border_size;
208 }
209
configure(const ICLTensor * input,ICLTensor * output,const PoolingLayerInfo & pool_info,ICLTensor * indices)210 void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output, const PoolingLayerInfo &pool_info, ICLTensor *indices)
211 {
212 configure(CLKernelLibrary::get().get_compile_context(), input, output, pool_info, indices);
213 }
214
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,const PoolingLayerInfo & pool_info,ICLTensor * indices)215 void CLPoolingLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const PoolingLayerInfo &pool_info, ICLTensor *indices)
216 {
217 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
218
219 auto padding_info = get_padding_info({ input, output, indices });
220
221 // Set instance variables
222 _input = input;
223 _output = output;
224 _pool_info = pool_info;
225 _data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? input->info()->data_layout() : pool_info.data_layout;
226 _indices = indices;
227 int pool_stride_x = 0;
228 int pool_stride_y = 0;
229 const PoolingType pool_type = pool_info.pool_type;
230 const int idx_width = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
231 const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
232 const int idx_channel = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
233 const int idx_batch_size = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::BATCHES);
234 const int pool_size_x = pool_info.is_global_pooling ? input->info()->dimension(idx_width) : pool_info.pool_size.width;
235 const int pool_size_y = pool_info.is_global_pooling ? input->info()->dimension(idx_height) : pool_info.pool_size.height;
236 const PadStrideInfo pad_stride_info = pool_info.pad_stride_info;
237 const bool exclude_padding = pool_info.exclude_padding;
238 std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride();
239 const int pool_pad_top = pad_stride_info.pad_top();
240 const int pool_pad_left = pad_stride_info.pad_left();
241
242 // Set build options
243 CLBuildOptions build_opts;
244 const DataType data_type = input->info()->data_type();
245
246 // Configure kernel window
247 auto win_config = validate_and_configure_window(input->info(), output->info(), pool_info, (indices ? indices->info() : nullptr));
248
249 ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
250 ICLKernel::configure_internal(std::get<1>(win_config));
251
252 CLPoolingConfig pooling_config = std::get<2>(win_config);
253 _num_elems_processed_per_iteration = pooling_config.first;
254 _border_size = pooling_config.second;
255
256 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(_num_elems_processed_per_iteration));
257
258 // Tensor paddings are used to calculate the indicies for MAX pooling
259 if(pool_info.pool_size == Size2D(2, 2) && pool_type == PoolingType::MAX && _indices && is_data_type_float(data_type))
260 {
261 build_opts.add_option("-DPAD_TENSOR_LEFT=" + support::cpp11::to_string(input->info()->padding().left));
262 build_opts.add_option("-DPAD_TENSOR_RIGHT=" + support::cpp11::to_string(input->info()->padding().right));
263 build_opts.add_option("-DPAD_TENSOR_TOP=" + support::cpp11::to_string(input->info()->padding().top));
264 build_opts.add_option("-DPAD_TENSOR_BOTTOM=" + support::cpp11::to_string(input->info()->padding().bottom));
265 build_opts.add_option("-DTENSOR_CHANNEL=" + support::cpp11::to_string(input->info()->dimension(idx_channel)));
266 build_opts.add_option("-DTENSOR_WIDTH=" + support::cpp11::to_string(input->info()->dimension(idx_width)));
267 build_opts.add_option("-DTENSOR_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(idx_height)));
268 }
269
270 if(is_data_type_quantized_asymmetric(data_type) && input->info()->quantization_info() != output->info()->quantization_info())
271 {
272 const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform();
273 const UniformQuantizationInfo oq_info = output->info()->quantization_info().uniform();
274
275 build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset));
276 build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
277 build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale));
278 build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
279 }
280
281 // Check output dimensions
282 auto_init(input->info(), output->info(), indices ? indices->info() : nullptr, pool_info);
283
284 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), pool_info, (indices) ? indices->info() : nullptr));
285
286 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
287 build_opts.add_option("-DPOOL_" + string_from_pooling_type(pool_type));
288 build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(pool_stride_x));
289 build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(pool_stride_y));
290 build_opts.add_option("-DPAD_X=" + support::cpp11::to_string(pool_pad_left));
291 build_opts.add_option("-DPAD_Y=" + support::cpp11::to_string(pool_pad_top));
292 build_opts.add_option("-DPOOL_SIZE_X=" + support::cpp11::to_string(pool_size_x));
293 build_opts.add_option("-DPOOL_SIZE_Y=" + support::cpp11::to_string(pool_size_y));
294
295 // Set the initial value for the pooling operation accordingly with the data type
296 if(pool_type == PoolingType::MAX)
297 {
298 if(is_data_type_quantized(data_type))
299 {
300 PixelValue type_min{};
301 std::tie(type_min, std::ignore) = get_min_max(data_type);
302 build_opts.add_option("-DINITIAL_VALUE=" + support::cpp11::to_string(type_min.get<int32_t>()));
303 }
304 else
305 {
306 build_opts.add_option("-DINITIAL_VALUE=" + float_to_string_with_full_precision(std::numeric_limits<float>::lowest()));
307 }
308 }
309 else
310 {
311 // Pool AVG and Pool L2 initial value
312 build_opts.add_option("-DINITIAL_VALUE=0");
313 }
314
315 build_opts.add_option("-DMAX_WIDTH=" + support::cpp11::to_string(input->info()->dimension(idx_width) + (exclude_padding ? 0 : pool_pad_left)));
316 build_opts.add_option("-DMAX_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(idx_height) + (exclude_padding ? 0 : pool_pad_top)));
317
318 // Create kernel
319 switch(_data_layout)
320 {
321 case DataLayout::NCHW:
322 {
323 const auto use_fp_mixed_precision = (data_type == DataType::F16) && pool_info.fp_mixed_precision;
324 const auto use_wider_accumulator = use_fp_mixed_precision && (pool_type != PoolingType::MAX);
325 const auto acc_data_type = get_cl_type_from_data_type(use_wider_accumulator ? DataType::F32 : data_type);
326 build_opts.add_option("-DACC_DATA_TYPE=" + acc_data_type);
327 build_opts.add_option_if(use_wider_accumulator, "-DFP_MIXED_PRECISION");
328
329 if(pool_type != PoolingType::MAX)
330 {
331 build_opts.add_option_if(exclude_padding, "-DEXCLUDE_PADDING");
332 }
333
334 if((pool_size_x == 3) && (pool_size_y == 3) && !is_data_type_quantized_asymmetric(data_type))
335 {
336 // Check if we have pool3x3 with stride_x less equal than 3. In these cases, run an optimized OpenCL kernel where
337 // each thread computes 4 output elements
338 const bool is_pool3x3_stride_le3 = (pool_size_x == 3) && (pool_size_y == 3) && (pool_stride_x <= 3);
339
340 std::string kernel_name = ((is_pool3x3_stride_le3) ? "pooling_layer_optimized_" : "pooling_layer_")
341 + support::cpp11::to_string(pool_size_x);
342 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
343 }
344 else if(pool_info.pool_size == Size2D(2, 2) && pool_type == PoolingType::MAX && _indices && is_data_type_float(data_type))
345 {
346 // For max pooling with pool2x2, store indicies which will be used in max unpooling
347 if(data_type == DataType::F32)
348 {
349 std::string kernel_name = "pooling_layer_2_nchw_indices_fp32";
350 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
351 }
352 else if(data_type == DataType::F16)
353 {
354 std::string kernel_name = "pooling_layer_2_nchw_indices_fp16";
355 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
356 }
357 }
358 else // Run general case
359 {
360 std::string kernel_name = is_data_type_quantized_asymmetric(data_type) ? "pooling_layer_MxN_quantized_nchw" : "pooling_layer_MxN_nchw";
361 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
362 }
363 break;
364 }
365 case DataLayout::NHWC:
366 {
367 // Floating point mixed precision is support on F16 only
368 const auto use_fp_mixed_precision = (data_type == DataType::F16) && pool_info.fp_mixed_precision && pool_type != PoolingType::MAX;
369
370 // Wider accumulation is required to avoid accuracy loss
371 // Case 1: Floating point mixed precision (fp16 input data and fp32 accumulation)
372 // Cast 2: Quantized (int8/uint8 input data and int32 accumulation )
373 DataType acc_data_type = data_type;
374
375 if(use_fp_mixed_precision)
376 {
377 acc_data_type = DataType::F32;
378 }
379 else if(is_data_type_quantized(data_type) && pool_type != PoolingType::MAX)
380 {
381 acc_data_type = DataType::S32;
382 }
383
384 build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(acc_data_type));
385 build_opts.add_option_if(use_fp_mixed_precision, "-DFP_MIXED_PRECISION");
386 build_opts.add_option_if(exclude_padding, "-DEXCLUDE_PADDING");
387 build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(idx_width)));
388 build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(idx_height)));
389 build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(output->info()->dimension(idx_height)));
390 build_opts.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(output->info()->dimension(idx_channel)));
391 build_opts.add_option("-DDST_BATCH_SIZE=" + support::cpp11::to_string(output->info()->dimension(idx_batch_size)));
392 build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(input->info()->dimension(0) % _num_elems_processed_per_iteration));
393 if(pool_info.pool_size == Size2D(2, 2) && is_data_type_float(data_type))
394 {
395 build_opts.add_option_if(_indices != nullptr && pool_type == PoolingType::MAX, "-DEXTRACT_MAX_INDEX");
396
397 std::string kernel_name = "pooling_layer_2x2_nhwc";
398 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
399 }
400 else
401 {
402 std::string kernel_name = is_data_type_quantized_asymmetric(data_type) ? "pooling_layer_MxN_quantized_nhwc" : "pooling_layer_MxN_nhwc";
403 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
404 }
405 break;
406 }
407 default:
408 ARM_COMPUTE_ERROR("Not implemented");
409 }
410
411 // Set config_id for enabling LWS tuning
412 _config_id = "pooling_layer_";
413 _config_id += lower_string(string_from_data_type(data_type));
414 _config_id += "_";
415 _config_id += lower_string(string_from_data_layout(_data_layout));
416 _config_id += "_";
417 _config_id += support::cpp11::to_string(output->info()->dimension(idx_width));
418 _config_id += "_";
419 _config_id += support::cpp11::to_string(output->info()->dimension(idx_height));
420 _config_id += "_";
421 _config_id += support::cpp11::to_string(output->info()->dimension(idx_channel));
422 _config_id += "_";
423 _config_id += lower_string(string_from_data_layout(input->info()->data_layout()));
424
425 ARM_COMPUTE_ERROR_ON(input->info()->data_layout() == DataLayout::NHWC && has_padding_changed(padding_info));
426 }
427
validate(const ITensorInfo * input,const ITensorInfo * output,const PoolingLayerInfo & pool_info,const ITensorInfo * indices)428 Status CLPoolingLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices)
429 {
430 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, pool_info, indices));
431 ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get(), pool_info)));
432
433 return Status{};
434 }
435
run(const Window & window,cl::CommandQueue & queue)436 void CLPoolingLayerKernel::run(const Window &window, cl::CommandQueue &queue)
437 {
438 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
439 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
440
441 unsigned int pool_stride_x = 0;
442 unsigned int pool_stride_y = 0;
443 std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride();
444
445 // Collapse window
446 Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
447
448 switch(_data_layout)
449 {
450 case DataLayout::NCHW:
451 {
452 Window slice = window_collapsed.first_slice_window_3D();
453 do
454 {
455 // Upsample input by pool size
456 Window in_slice(slice);
457 in_slice.set(Window::DimX, Window::Dimension(in_slice.x().start() - _pool_info.pad_stride_info.pad_left(),
458 (in_slice.x().end() - _pool_info.pad_stride_info.pad_left()) * pool_stride_x,
459 pool_stride_x * _num_elems_processed_per_iteration));
460 in_slice.set(Window::DimY, Window::Dimension(in_slice.y().start() - _pool_info.pad_stride_info.pad_top(),
461 (in_slice.y().end() - _pool_info.pad_stride_info.pad_top()) * pool_stride_y,
462 pool_stride_y));
463
464 // Set inputs
465 unsigned int idx = 0;
466 add_3D_tensor_argument(idx, _input, in_slice);
467 add_3D_tensor_argument(idx, _output, slice);
468 if(_indices && is_data_type_float(_input->info()->data_type()) && (_pool_info.pool_size == Size2D(2, 2)))
469 {
470 add_3D_tensor_argument(idx, _indices, slice);
471 }
472 enqueue(queue, *this, slice, lws_hint());
473 }
474 while(window_collapsed.slide_window_slice_3D(slice));
475 break;
476 }
477 case DataLayout::NHWC:
478 {
479 const size_t batch_size = _output->info()->tensor_shape().total_size_upper(3);
480
481 Window slice = window_collapsed.first_slice_window_4D();
482 Window in_slice = window_collapsed.first_slice_window_4D();
483 in_slice.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _num_elems_processed_per_iteration));
484 in_slice.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(1), pool_stride_x));
485 in_slice.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(2), pool_stride_y));
486 in_slice.set(3, Window::Dimension(0, batch_size, 1));
487 do
488 {
489 // Set inputs
490 unsigned int idx = 0;
491 add_4D_tensor_argument(idx, _input, in_slice);
492 add_4D_tensor_argument(idx, _output, slice);
493 if(_indices && is_data_type_float(_input->info()->data_type()) && (_pool_info.pool_type == PoolingType::MAX) && (_pool_info.pool_size == Size2D(2, 2)))
494 {
495 add_4D_tensor_argument(idx, _indices, slice);
496 }
497 enqueue(queue, *this, slice, lws_hint());
498 }
499 while(window.slide_window_slice_4D(slice) && window.slide_window_slice_4D(in_slice));
500 break;
501 }
502 default:
503 ARM_COMPUTE_ERROR("Not implemented");
504 }
505 }
506 } // namespace arm_compute
507