1 /*
2 * Copyright (c) 2022 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/ClPool3dKernel.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 #include "utils/TypePrinter.h"
34
35 namespace arm_compute
36 {
37 namespace opencl
38 {
39 namespace kernels
40 {
41 using namespace arm_compute::misc::shape_calculator;
42
43 namespace
44 {
validate_arguments(const ITensorInfo * src,const ITensorInfo * dst,const Pooling3dLayerInfo & pool_info)45 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const Pooling3dLayerInfo &pool_info)
46 {
47 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
48 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->data_layout() != DataLayout::NDHWC, "Only NDHWC layout supported");
49
50 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
51 ARM_COMPUTE_RETURN_ERROR_ON_MSG((pool_info.stride.x() == 0 || pool_info.stride.y() == 0 || pool_info.stride.z() == 0), "Strides cannot be zero.");
52 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32, DataType::QASYMM8_SIGNED, DataType::QASYMM8);
53 ARM_COMPUTE_RETURN_ERROR_ON_MSG((!is_data_type_float(src->data_type())) && (!pool_info.exclude_padding
54 && (pool_info.pool_type == PoolingType::AVG)),
55 "Exclude padding is unsupported for non-float types for Avg op");
56
57 const auto data_layout = src->data_layout();
58 const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
59 const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
60 const int idx_depth = get_data_layout_dimension_index(data_layout, DataLayoutDimension::DEPTH);
61 const bool is_global_pooling = pool_info.is_global_pooling;
62 const unsigned int pool_size_x = is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width;
63 const unsigned int pool_size_y = is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height;
64 const unsigned int pool_size_z = is_global_pooling ? src->dimension(idx_depth) : pool_info.pool_size.depth;
65 int output_width = 0;
66 int output_height = 0;
67 int output_depth = 0;
68
69 bool round_type_ceil_with_asymm_padding = (pool_info.round_type == DimensionRoundingType::CEIL) && (!is_symmetric(pool_info.padding));
70 ARM_COMPUTE_RETURN_ERROR_ON_MSG(round_type_ceil_with_asymm_padding, "Cannot use dimension round type CEIL when padding is asymmetric.");
71
72 ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_pool_3d_region_entirely_outside_input(pool_info), "Pooling region that is entirely outside input tensor is unsupported");
73 std::tie(output_width, output_height, output_depth) = scaled_3d_dimensions_signed(src->tensor_shape()[idx_width], src->tensor_shape()[idx_height],
74 src->tensor_shape()[idx_depth], pool_size_x, pool_size_y,
75 pool_size_z, pool_info);
76
77 ARM_COMPUTE_RETURN_ERROR_ON_MSG((output_width < 1 || output_height < 1 || output_depth < 1), "Calculated output dimension size is invalid");
78 // Checks performed when dst is configured
79 if(dst->total_size() != 0)
80 {
81 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
82 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(src, dst);
83 TensorInfo out_info(TensorInfo(compute_pool3d_shape(src->tensor_shape(), pool_info), 1, dst->data_type()));
84 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &out_info);
85 }
86
87 return Status{};
88 }
89 } // namespace
90
ClPool3dKernel()91 ClPool3dKernel::ClPool3dKernel()
92 {
93 _type = CLKernelType::POOL;
94 }
95
configure(const ClCompileContext & compile_context,const ITensorInfo * src,ITensorInfo * dst,const Pooling3dLayerInfo & pool_info)96 void ClPool3dKernel::configure(const ClCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *dst, const Pooling3dLayerInfo &pool_info)
97 {
98 ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
99 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, pool_info));
100 auto padding_info = get_padding_info({ src, dst });
101
102 // Auto init if empty
103 TensorShape out_shape = compute_pool3d_shape(src->tensor_shape(), pool_info);
104 auto_init_if_empty(*dst, src->clone()->set_tensor_shape(out_shape));
105
106 // Set instance variables
107 _pool_info = pool_info;
108 _data_layout = src->data_layout();
109
110 _num_elems_processed_per_iteration = (dst->data_type() == DataType::F32) ? 2 : 4;
111 _num_elems_processed_per_iteration = adjust_vec_size(_num_elems_processed_per_iteration, dst->dimension(0));
112
113 const PoolingType pool_type = pool_info.pool_type;
114 const int idx_width = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
115 const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
116 const int idx_depth = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::DEPTH);
117 const int idx_channel = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
118 const int idx_batch_size = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::BATCHES);
119 const int pool_size_x = pool_info.is_global_pooling ? src->dimension(idx_width) : pool_info.pool_size.width;
120 const int pool_size_y = pool_info.is_global_pooling ? src->dimension(idx_height) : pool_info.pool_size.height;
121 const int pool_size_z = pool_info.is_global_pooling ? src->dimension(idx_depth) : pool_info.pool_size.depth;
122 const bool exclude_padding = pool_info.exclude_padding;
123 const int pool_stride_x = pool_info.stride.x();
124 const int pool_stride_y = pool_info.stride.y();
125 const int pool_stride_z = pool_info.stride.z();
126 const int pool_pad_top = pool_info.padding.top;
127 const int pool_pad_left = pool_info.padding.left;
128 const int pool_pad_front = pool_info.padding.front;
129 const DataType data_type = src->data_type();
130
131 // Set build options
132 CLBuildOptions build_opts;
133 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(_num_elems_processed_per_iteration));
134 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
135 build_opts.add_option("-DPOOL_" + string_from_pooling_type(pool_type));
136 build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(pool_stride_x));
137 build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(pool_stride_y));
138 build_opts.add_option("-DSTRIDE_Z=" + support::cpp11::to_string(pool_stride_z));
139 build_opts.add_option("-DPAD_X=" + support::cpp11::to_string(pool_pad_left));
140 build_opts.add_option("-DPAD_Y=" + support::cpp11::to_string(pool_pad_top));
141 build_opts.add_option("-DPAD_Z=" + support::cpp11::to_string(pool_pad_front));
142 build_opts.add_option("-DPOOL_SIZE_X=" + support::cpp11::to_string(pool_size_x));
143 build_opts.add_option("-DPOOL_SIZE_Y=" + support::cpp11::to_string(pool_size_y));
144 build_opts.add_option("-DPOOL_SIZE_Z=" + support::cpp11::to_string(pool_size_z));
145 build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src->dimension(idx_width)));
146 build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src->dimension(idx_height)));
147 build_opts.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(src->dimension(idx_depth)));
148
149 // If datatype is quantized add relevant parameters
150 if(is_data_type_quantized_asymmetric(data_type) && src->quantization_info() != dst->quantization_info())
151 {
152 const UniformQuantizationInfo iq_info = src->quantization_info().uniform();
153 const UniformQuantizationInfo oq_info = dst->quantization_info().uniform();
154
155 build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset));
156 build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
157 build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale));
158 build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
159 }
160
161 // Set the initial value for the pooling operation accordingly with the data type
162 if(pool_type == PoolingType::MAX)
163 {
164 if(is_data_type_quantized(data_type))
165 {
166 PixelValue type_min{};
167 std::tie(type_min, std::ignore) = get_min_max(data_type);
168 build_opts.add_option("-DINITIAL_VALUE=" + support::cpp11::to_string(type_min.get<int32_t>()));
169 }
170 else
171 {
172 build_opts.add_option("-DINITIAL_VALUE=" + float_to_string_with_full_precision(std::numeric_limits<float>::lowest()));
173 }
174 }
175 else
176 {
177 // Pool AVG and Pool L2 initial value
178 build_opts.add_option("-DINITIAL_VALUE=0");
179 }
180 // Create kernel
181 // Floating point mixed precision is support on F16 only
182 const auto use_fp_mixed_precision = (data_type == DataType::F16) && pool_info.fp_mixed_precision && pool_type != PoolingType::MAX;
183
184 // Wider accumulation is required to avoid accuracy loss
185 // Case 1: Floating point mixed precision (fp16 src data and fp32 accumulation)
186 DataType acc_data_type = data_type;
187 if(use_fp_mixed_precision)
188 {
189 acc_data_type = DataType::F32;
190 }
191 else if(is_data_type_quantized(data_type) && pool_type != PoolingType::MAX) // Use S32 for avg pooling to allow for integer division
192 {
193 acc_data_type = DataType::S32;
194 }
195
196 build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(acc_data_type));
197 build_opts.add_option_if(use_fp_mixed_precision, "-DFP_MIXED_PRECISION");
198 build_opts.add_option_if(exclude_padding, "-DEXCLUDE_PADDING");
199 build_opts.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst->dimension(idx_height)));
200 build_opts.add_option("-DDST_DEPTH=" + support::cpp11::to_string(dst->dimension(idx_depth)));
201 build_opts.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(dst->dimension(idx_channel)));
202 build_opts.add_option("-DDST_BATCH_SIZE=" + support::cpp11::to_string(dst->dimension(idx_batch_size)));
203 build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(src->dimension(0) % _num_elems_processed_per_iteration));
204
205 // if datatype is quantized use quantized kernel function
206 std::string kernel_name = (is_data_type_quantized_asymmetric(data_type) ? "pooling_3d_layer_MxN_ndhwc_quantized" : "pooling_3d_layer_MxN_ndhwc");
207 _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
208
209 // Configure kernel window
210 Window win = calculate_max_window(*dst, Steps(_num_elems_processed_per_iteration));
211 ICLKernel::configure_internal(win);
212
213 // Set config_id for enabling LWS tuning
214 _config_id = "pooling_layer_3d";
215 _config_id += lower_string(string_from_data_type(data_type));
216 _config_id += "_";
217 _config_id += lower_string(string_from_data_layout(_data_layout));
218 _config_id += "_";
219 _config_id += support::cpp11::to_string(dst->dimension(idx_width));
220 _config_id += "_";
221 _config_id += support::cpp11::to_string(dst->dimension(idx_height));
222 _config_id += "_";
223 _config_id += support::cpp11::to_string(dst->dimension(idx_channel));
224 _config_id += "_";
225 _config_id += lower_string(string_from_data_layout(src->data_layout()));
226
227 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
228 }
229
validate(const ITensorInfo * src,const ITensorInfo * dst,const Pooling3dLayerInfo & pool_info)230 Status ClPool3dKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const Pooling3dLayerInfo &pool_info)
231 {
232 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, pool_info));
233 return Status{};
234 }
235
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)236 void ClPool3dKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
237 {
238 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
239 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
240
241 const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
242 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST_0));
243
244 // Collapse 3D window
245 Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
246
247 // Set CL kernel arguments
248 unsigned int idx = 0;
249 // Passing of the window not needed, as the steps are not used for the pool3d kernel
250 add_5D_tensor_argument(idx, src, window);
251 add_5D_tensor_argument(idx, dst, window);
252 enqueue(queue, *this, window_collapsed, lws_hint());
253 }
254 } // namespace kernels
255 } // namespace opencl
256 } // namespace arm_compute
257