1 /*
2 * Copyright (c) 2017-2021 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/gpu/cl/kernels/ClPool2dKernel.h"
25
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/TensorInfo.h"
28 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
29 #include "src/core/CL/CLValidate.h"
30 #include "src/core/helpers/AutoConfiguration.h"
31 #include "src/core/helpers/WindowHelpers.h"
32 #include "support/Cast.h"
33
34 namespace arm_compute
35 {
36 namespace opencl
37 {
38 namespace kernels
39 {
40 using namespace arm_compute::misc::shape_calculator;
41
42 namespace
43 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * dst,const PoolingLayerInfo & pool_info,const ITensorInfo * indices)44 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, const ITensorInfo *indices)
45 {
46 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
47 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
48 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
49 ARM_COMPUTE_RETURN_ERROR_ON_MSG((is_data_type_quantized_asymmetric(src->data_type()) && pool_info.pool_type == PoolingType::L2),
50 "Unsupported combination of parameters!");
51
52 const auto data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout;
53 const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
54 const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
55 const bool is_global_pooling = pool_info.is_global_pooling;
56 unsigned int pool_size_x = is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width;
57 unsigned int pool_size_y = is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height;
58 int output_width = 0;
59 int output_height = 0;
60
61 ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_pool_region_entirely_outside_input(pool_info), "Pooling region that is entirely outside input tensor is unsupported");
62
63 std::tie(output_width, output_height) = scaled_dimensions_signed(src->tensor_shape()[idx_width], src->tensor_shape()[idx_height],
64 pool_size_x, pool_size_y, pool_info.pad_stride_info);
65 ARM_COMPUTE_RETURN_ERROR_ON_MSG((output_width < 1 || output_height < 1), "Calculated output dimension size is invalid");
66
67 // Check indices
68 if(indices)
69 {
70 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32);
71 ARM_COMPUTE_RETURN_ERROR_ON_MSG(pool_info.pool_type != PoolingType::MAX, "Pooling indices only supported for MAX pooling method");
72 ARM_COMPUTE_RETURN_ERROR_ON_MSG((pool_info.pool_size != Size2D(2, 2)), "Pooling indices only supported for pool size 2x2");
73
74 if(indices->total_size() != 0)
75 {
76 TensorInfo idx_info(TensorInfo(compute_pool_shape(*src, pool_info), 1, DataType::U32));
77 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(indices, &idx_info);
78 }
79 }
80
81 // Checks performed when dst is configured
82 if(dst->total_size() != 0)
83 {
84 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
85 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, dst);
86 TensorInfo out_info(TensorInfo(compute_pool_shape(*src, pool_info), 1, dst->data_type()));
87 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &out_info);
88 }
89
90 return Status{};
91 }
92 } // namespace
93
ClPool2dKernel()94 ClPool2dKernel::ClPool2dKernel()
95 {
96 _type = CLKernelType::POOL;
97 }
98
configure(const ClCompileContext & compile_context,ITensorInfo * src,ITensorInfo * dst,const PoolingLayerInfo & pool_info,ITensorInfo * indices)99 void ClPool2dKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &pool_info, ITensorInfo *indices)
100 {
101 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
102 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, pool_info, indices));
103
104 auto padding_info = get_padding_info({ src, dst, indices });
105
106 // Auto init if empty
107 TensorShape out_shape = compute_pool_shape(*src, pool_info);
108 auto_init_if_empty(*dst, src->clone()->set_tensor_shape(out_shape));
109 if(indices)
110 {
111 auto_init_if_empty(*indices, src->clone()->set_tensor_shape(out_shape).set_data_type(DataType::U32));
112 }
113
114 // Set instance variables
115 _pool_info = pool_info;
116 _data_layout = pool_info.data_layout == DataLayout::UNKNOWN ? src->data_layout() : pool_info.data_layout;
117 _num_elems_processed_per_iteration = (_data_layout == DataLayout::NCHW) ? 1 : ((dst->data_type() == DataType::F32) ? 2 : 4);
118 _num_elems_processed_per_iteration = adjust_vec_size(_num_elems_processed_per_iteration, dst->dimension(0));
119
120 int pool_stride_x = 0;
121 int pool_stride_y = 0;
122 const PoolingType pool_type = pool_info.pool_type;
123 const int idx_width = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
124 const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
125 const int idx_channel = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
126 const int idx_batch_size = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::BATCHES);
127 const int pool_size_x = pool_info.is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width;
128 const int pool_size_y = pool_info.is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height;
129 const PadStrideInfo pad_stride_info = pool_info.pad_stride_info;
130 const bool exclude_padding = pool_info.exclude_padding;
131 std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride();
132 const int pool_pad_top = pad_stride_info.pad_top();
133 const int pool_pad_left = pad_stride_info.pad_left();
134 const DataType data_type = src->data_type();
135
136 // Set build options
137 CLBuildOptions build_opts;
138 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(_num_elems_processed_per_iteration));
139 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
140 build_opts.add_option("-DPOOL_" + string_from_pooling_type(pool_type));
141 build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(pool_stride_x));
142 build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(pool_stride_y));
143 build_opts.add_option("-DPAD_X=" + support::cpp11::to_string(pool_pad_left));
144 build_opts.add_option("-DPAD_Y=" + support::cpp11::to_string(pool_pad_top));
145 build_opts.add_option("-DPOOL_SIZE_X=" + support::cpp11::to_string(pool_size_x));
146 build_opts.add_option("-DPOOL_SIZE_Y=" + support::cpp11::to_string(pool_size_y));
147 build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src->dimension(idx_width)));
148 build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height)));
149 build_opts.add_option("-DMAX_WIDTH=" + support::cpp11::to_string(src->dimension(idx_width) + (exclude_padding ? 0 : pool_pad_left)));
150 build_opts.add_option("-DMAX_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height) + (exclude_padding ? 0 : pool_pad_top)));
151
152 // Tensor paddings are used to calculate the indicies for MAX pooling
153 if(pool_info.pool_size == Size2D(2, 2) && pool_type == PoolingType::MAX && indices && is_data_type_float(data_type))
154 {
155 build_opts.add_option("-DSRC_BATCH=" + support::cpp11::to_string(src->tensor_shape().total_size_lower(3)));
156 }
157
158 if(is_data_type_quantized_asymmetric(data_type))
159 {
160 build_opts.add_option("-DQUANTIZED");
161
162 if(src->quantization_info() != dst->quantization_info())
163 {
164 const UniformQuantizationInfo iq_info = src->quantization_info().uniform();
165 const UniformQuantizationInfo oq_info = dst->quantization_info().uniform();
166
167 build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset));
168 build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
169 build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale));
170 build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
171 }
172 }
173
174 // Set the initial value for the pooling operation accordingly with the data type
175 if(pool_type == PoolingType::MAX)
176 {
177 if(is_data_type_quantized(data_type))
178 {
179 PixelValue type_min{};
180 std::tie(type_min, std::ignore) = get_min_max(data_type);
181 build_opts.add_option("-DINITIAL_VALUE=" + support::cpp11::to_string(type_min.get<int32_t>()));
182 }
183 else
184 {
185 build_opts.add_option("-DINITIAL_VALUE=" + float_to_string_with_full_precision(std::numeric_limits<float>::lowest()));
186 }
187 }
188 else
189 {
190 // Pool AVG and Pool L2 initial value
191 build_opts.add_option("-DINITIAL_VALUE=0");
192 }
193
194 // Create kernel
195 switch(_data_layout)
196 {
197 case DataLayout::NCHW:
198 {
199 const auto use_fp_mixed_precision = (data_type == DataType::F16) && pool_info.fp_mixed_precision;
200 const auto use_wider_accumulator = use_fp_mixed_precision && (pool_type != PoolingType::MAX);
201 const auto acc_data_type = get_cl_type_from_data_type(use_wider_accumulator ? DataType::F32 : (is_data_type_quantized(data_type) ? DataType::S32 : data_type));
202 build_opts.add_option("-DACC_DATA_TYPE=" + acc_data_type);
203 build_opts.add_option_if(use_wider_accumulator, "-DFP_MIXED_PRECISION");
204
205 if(pool_type != PoolingType::MAX)
206 {
207 build_opts.add_option_if(exclude_padding, "-DEXCLUDE_PADDING");
208 }
209
210 if(pool_info.pool_size == Size2D(2, 2) && pool_type == PoolingType::MAX && indices && is_data_type_float(data_type))
211 {
212 // For max pooling with pool2x2, store indicies which will be used in max unpooling
213 std::string kernel_name = "pooling_layer_2_nchw_indices";
214 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
215 }
216 else // Run general case
217 {
218 std::string kernel_name = "pooling_layer_MxN_nchw";
219 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
220 }
221 break;
222 }
223 case DataLayout::NHWC:
224 {
225 // Floating point mixed precision is support on F16 only
226 const auto use_fp_mixed_precision = (data_type == DataType::F16) && pool_info.fp_mixed_precision && pool_type != PoolingType::MAX;
227
228 // Wider accumulation is required to avoid accuracy loss
229 // Case 1: Floating point mixed precision (fp16 src data and fp32 accumulation)
230 // Cast 2: Quantized (int8/uint8 src data and int32 accumulation )
231 DataType acc_data_type = data_type;
232
233 if(use_fp_mixed_precision)
234 {
235 acc_data_type = DataType::F32;
236 }
237 else if(is_data_type_quantized(data_type) && pool_type != PoolingType::MAX)
238 {
239 acc_data_type = DataType::S32;
240 }
241
242 build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(acc_data_type));
243 build_opts.add_option_if(use_fp_mixed_precision, "-DFP_MIXED_PRECISION");
244 build_opts.add_option_if(exclude_padding, "-DEXCLUDE_PADDING");
245 build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src->dimension(idx_width)));
246 build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height)));
247 build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst->dimension(idx_height)));
248 build_opts.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(dst->dimension(idx_channel)));
249 build_opts.add_option("-DDST_BATCH_SIZE=" + support::cpp11::to_string(dst->dimension(idx_batch_size)));
250 build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % _num_elems_processed_per_iteration));
251 if(pool_info.pool_size == Size2D(2, 2) && is_data_type_float(data_type))
252 {
253 build_opts.add_option_if(indices != nullptr && pool_type == PoolingType::MAX, "-DEXTRACT_MAX_INDEX");
254
255 std::string kernel_name = "pooling_layer_2x2_nhwc";
256 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
257 }
258 else
259 {
260 std::string kernel_name = is_data_type_quantized_asymmetric(data_type) ? "pooling_layer_MxN_quantized_nhwc" : "pooling_layer_MxN_nhwc";
261 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
262 }
263 break;
264 }
265 default:
266 ARM_COMPUTE_ERROR("Not implemented");
267 }
268
269 // Configure kernel window
270 Window win = calculate_max_window(*dst, Steps(_num_elems_processed_per_iteration));
271 ICLKernel::configure_internal(win);
272
273 // Set config_id for enabling LWS tuning
274 _config_id = "pooling_layer_";
275 _config_id += lower_string(string_from_data_type(data_type));
276 _config_id += "_";
277 _config_id += lower_string(string_from_data_layout(_data_layout));
278 _config_id += "_";
279 _config_id += support::cpp11::to_string(dst->dimension(idx_width));
280 _config_id += "_";
281 _config_id += support::cpp11::to_string(dst->dimension(idx_height));
282 _config_id += "_";
283 _config_id += support::cpp11::to_string(dst->dimension(idx_channel));
284 _config_id += "_";
285 _config_id += lower_string(string_from_data_layout(src->data_layout()));
286
287 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
288 }
289
validate(const ITensorInfo * src,const ITensorInfo * dst,const PoolingLayerInfo & pool_info,const ITensorInfo * indices)290 Status ClPool2dKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const PoolingLayerInfo &pool_info, const ITensorInfo *indices)
291 {
292 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, pool_info, indices));
293 return Status{};
294 }
295
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)296 void ClPool2dKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
297 {
298 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
299 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
300
301 unsigned int pool_stride_x = 0;
302 unsigned int pool_stride_y = 0;
303 std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info.stride();
304
305 const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
306 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST_0));
307 auto indices = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST_1));
308
309 // Collapse window
310 Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
311
312 switch(_data_layout)
313 {
314 case DataLayout::NCHW:
315 {
316 Window slice = window_collapsed.first_slice_window_3D();
317 do
318 {
319 // Set srcs
320 unsigned int idx = 0;
321 add_3D_tensor_argument(idx, src, slice);
322 add_3D_tensor_argument(idx, dst, slice);
323 if(indices && is_data_type_float(src->info()->data_type()) && (_pool_info.pool_size == Size2D(2, 2)))
324 {
325 add_3D_tensor_argument(idx, indices, slice);
326 }
327 enqueue(queue, *this, slice, lws_hint());
328 }
329 while(window_collapsed.slide_window_slice_3D(slice));
330 break;
331 }
332 case DataLayout::NHWC:
333 {
334 const size_t batch_size = dst->info()->tensor_shape().total_size_upper(3);
335
336 Window slice = window_collapsed.first_slice_window_4D();
337 Window in_slice = window_collapsed.first_slice_window_4D();
338 in_slice.set(Window::DimX, Window::Dimension(0, src->info()->dimension(0), _num_elems_processed_per_iteration));
339 in_slice.set(Window::DimY, Window::Dimension(0, src->info()->dimension(1), pool_stride_x));
340 in_slice.set(Window::DimZ, Window::Dimension(0, src->info()->dimension(2), pool_stride_y));
341 in_slice.set(3, Window::Dimension(0, batch_size, 1));
342 do
343 {
344 // Set srcs
345 unsigned int idx = 0;
346 add_4D_tensor_argument(idx, src, in_slice);
347 add_4D_tensor_argument(idx, dst, slice);
348 if(indices && is_data_type_float(src->info()->data_type()) && (_pool_info.pool_type == PoolingType::MAX) && (_pool_info.pool_size == Size2D(2, 2)))
349 {
350 add_4D_tensor_argument(idx, indices, slice);
351 }
352 enqueue(queue, *this, slice, lws_hint());
353 }
354 while(window.slide_window_slice_4D(slice) && window.slide_window_slice_4D(in_slice));
355 break;
356 }
357 default:
358 ARM_COMPUTE_ERROR("Not implemented");
359 }
360 }
361 } // namespace kernels
362 } // namespace opencl
363 } // namespace arm_compute
364