xref: /aosp_15_r20/external/ComputeLibrary/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3) 
1 /*
2  * Copyright (c) 2017-2022 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 "arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h"
25 
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/Utils.h"
29 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
30 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
31 #include "arm_compute/runtime/CL/CLScheduler.h"
32 #include "src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.h"
33 #include "src/runtime/heuristics/dwc_native/ClDWCNativeKernelConfig.h"
34 #include "src/runtime/heuristics/dwc_native/IClDWCNativeKernelConfig.h"
35 
36 #include "src/common/utils/Log.h"
37 
38 namespace arm_compute
39 {
40 using namespace arm_compute::misc;
41 using namespace arm_compute::misc::shape_calculator;
42 using namespace arm_compute::cl_dwc;
43 
CLDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)44 CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
45     : _memory_group(std::move(memory_manager)),
46       _dwc_native_kernel(std::make_unique<CLDepthwiseConvolutionLayerNativeKernel>()),
47       _permute_input_to_nhwc(),
48       _permute_weights_to_nhwc(),
49       _permute_output_to_nchw(),
50       _permuted_input(),
51       _permuted_weights(),
52       _permuted_output(),
53       _output_multipliers(),
54       _output_shifts(),
55       _original_weights(),
56       _input(),
57       _output(),
58       _needs_permute(false),
59       _is_prepared(false),
60       _is_quantized(false)
61 {
62 }
63 
64 CLDepthwiseConvolutionLayer::~CLDepthwiseConvolutionLayer() = default;
65 
configure(ICLTensor * input,const ICLTensor * weights,const ICLTensor * biases,ICLTensor * output,const PadStrideInfo & conv_info,unsigned int depth_multiplier,ActivationLayerInfo act_info,const Size2D & dilation)66 void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
67                                             unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
68 {
69     configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
70 }
71 
configure(const CLCompileContext & compile_context,ICLTensor * input,const ICLTensor * weights,const ICLTensor * biases,ICLTensor * output,const PadStrideInfo & conv_info,unsigned int depth_multiplier,ActivationLayerInfo act_info,const Size2D & dilation)72 void CLDepthwiseConvolutionLayer::configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases,
73                                             ICLTensor *output, const PadStrideInfo &conv_info,
74                                             unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
75 {
76     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights);
77     ARM_COMPUTE_ERROR_THROW_ON(CLDepthwiseConvolutionLayer::validate(input->info(),
78                                                                      weights->info(),
79                                                                      biases != nullptr ? biases->info() : nullptr,
80                                                                      output != nullptr ? output->info() : input->info(),
81                                                                      conv_info,
82                                                                      depth_multiplier,
83                                                                      act_info,
84                                                                      dilation));
85     ARM_COMPUTE_LOG_PARAMS(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
86 
87     _is_quantized     = is_data_type_quantized(input->info()->data_type());
88     _is_prepared      = false;
89     _original_weights = weights;
90     _input            = input;
91     _output           = output;
92     _needs_permute    = input->info()->data_layout() == DataLayout::NCHW;
93 
94     const GPUTarget gpu_target = CLScheduler::get().target();
95 
96     ICLTensor       *input_to_use   = input;
97     const ICLTensor *weights_to_use = weights;
98     ICLTensor       *output_to_use  = output;
99     if(_needs_permute)
100     {
101         _memory_group.manage(&_permuted_input);
102         _memory_group.manage(&_permuted_output);
103 
104         // Configure the function to transform the input tensor from NCHW -> NHWC
105         _permute_input_to_nhwc.configure(compile_context, input, &_permuted_input, PermutationVector(2U, 0U, 1U));
106         _permuted_input.info()->set_data_layout(DataLayout::NHWC);
107 
108         // Configure the function to transform the weights tensor from IHW -> HWI
109         _permute_weights_to_nhwc.configure(compile_context, weights, &_permuted_weights, PermutationVector(2U, 0U, 1U));
110         _permuted_weights.info()->set_data_layout(DataLayout::NHWC);
111 
112         // Set output quantization info before dwc kernel configure
113         _permuted_output.info()->set_quantization_info(output->info()->quantization_info());
114 
115         input_to_use   = &_permuted_input;
116         weights_to_use = &_permuted_weights;
117         output_to_use  = &_permuted_output;
118     }
119 
120     CLTensor *output_multipliers_to_use = nullptr;
121     CLTensor *output_shifts_to_use      = nullptr;
122     if(_is_quantized)
123     {
124         const size_t idx_c       = get_data_layout_dimension_index(weights->info()->data_layout(), DataLayoutDimension::CHANNEL);
125         const size_t num_filters = (is_data_type_quantized_per_channel(weights->info()->data_type())) ? weights->info()->dimension(idx_c) : 1;
126 
127         _output_multipliers.allocator()->init(TensorInfo(TensorShape(num_filters), 1, DataType::S32));
128         _output_shifts.allocator()->init(TensorInfo(TensorShape(num_filters), 1, DataType::S32));
129 
130         output_multipliers_to_use = &_output_multipliers;
131         output_shifts_to_use      = &_output_shifts;
132     }
133 
134     // Get the depthwise convolution compute parameters
135     auto t = ClDWCNativeKernelConfigurationFactory::create(gpu_target);
136     const DWCComputeKernelInfo dwc_native_compute_info = t->configure(input_to_use->info(), weights_to_use->info(), conv_info, dilation, depth_multiplier);
137 
138     const ConvolutionInfo conv_kernel_info{ conv_info, depth_multiplier, act_info, dilation };
139 
140     _dwc_native_kernel->set_target(gpu_target);
141     _dwc_native_kernel->configure(compile_context, input_to_use, weights_to_use, biases, output_to_use,
142                                   dwc_native_compute_info, conv_kernel_info, output_multipliers_to_use, output_shifts_to_use);
143 
144     if(_needs_permute)
145     {
146         _permuted_input.allocator()->allocate();
147 
148         // Configure the function to transform the convoluted output to NCHW format
149         _permuted_output.info()->set_data_layout(DataLayout::NCHW);
150         _permute_output_to_nchw.configure(compile_context, &_permuted_output, output, PermutationVector(1U, 2U, 0U));
151         _permuted_output.allocator()->allocate();
152     }
153 
154     if(_is_quantized)
155     {
156         _output_multipliers.allocator()->allocate();
157         _output_shifts.allocator()->allocate();
158     }
159 }
160 
validate(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * output,const PadStrideInfo & conv_info,unsigned int depth_multiplier,ActivationLayerInfo act_info,const Size2D & dilation)161 Status CLDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
162                                              const PadStrideInfo &conv_info,
163                                              unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
164 {
165     const bool in_place = input == output || output == nullptr;
166     if(in_place)
167     {
168         output = input;
169     }
170     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
171     const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
172     const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
173 
174     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) + (weights->dimension(idx_w) - 1) * (dilation.x() - 1) > input->dimension(idx_w) + conv_info.pad_left() + conv_info.pad_right());
175     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_h) + (weights->dimension(idx_h) - 1) * (dilation.y() - 1) > input->dimension(idx_h) + conv_info.pad_top() + conv_info.pad_bottom());
176 
177     const GPUTarget gpu_target = CLScheduler::get().target();
178 
179     const ConvolutionInfo conv_kernel_info{ conv_info, depth_multiplier, act_info, dilation };
180 
181     const bool needs_permute = input->data_layout() == DataLayout::NCHW;
182 
183     const bool is_quantized = is_data_type_quantized(input->data_type());
184 
185     TensorInfo output_multipliers_shifts_info(TensorInfo(TensorShape(1U), 1, DataType::S32));
186     if(is_quantized)
187     {
188         if(is_data_type_quantized_per_channel(weights->data_type()))
189         {
190             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QSYMM8_PER_CHANNEL);
191 
192             const size_t idx_c = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::CHANNEL);
193             output_multipliers_shifts_info.set_tensor_shape(TensorShape(weights->dimension(idx_c)));
194         }
195         else
196         {
197             ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
198         }
199     }
200 
201     if(needs_permute)
202     {
203         ARM_COMPUTE_RETURN_ERROR_ON_MSG(in_place, "In-place is supported only with NHWC data layout");
204         TensorShape           permuted_input_shape   = input->tensor_shape();
205         TensorShape           permuted_weights_shape = weights->tensor_shape();
206         const ConvolutionInfo info{ conv_info, depth_multiplier, ActivationLayerInfo(), dilation };
207         TensorShape           permuted_output_shape = shape_calculator::compute_depthwise_convolution_shape(*input, *weights, info);
208 
209         permute(permuted_input_shape, PermutationVector(2U, 0U, 1U));
210         permute(permuted_weights_shape, PermutationVector(2U, 0U, 1U));
211         permute(permuted_output_shape, PermutationVector(2U, 0U, 1U));
212 
213         const TensorInfo permuted_input   = input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_input_shape).set_data_layout(DataLayout::NHWC);
214         const TensorInfo permuted_weights = weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_weights_shape).set_data_layout(DataLayout::NHWC);
215         const TensorInfo permuted_output  = output->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_output_shape).set_data_layout(DataLayout::NHWC);
216 
217         ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(input, &permuted_input, PermutationVector(2U, 0U, 1U)));
218         ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(weights, &permuted_weights, PermutationVector(2U, 0U, 1U)));
219 
220         // Get the depthwise convolution compute parameters
221         auto t = ClDWCNativeKernelConfigurationFactory::create(gpu_target);
222         const DWCComputeKernelInfo dwc_native_compute_info = t->configure(&permuted_input, &permuted_weights, conv_info, dilation, depth_multiplier);
223 
224         ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerNativeKernel::validate(&permuted_input, &permuted_weights, biases, &permuted_output,
225                                                                                       dwc_native_compute_info, conv_kernel_info, &output_multipliers_shifts_info, &output_multipliers_shifts_info));
226         ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(&permuted_output, output, PermutationVector(1U, 2U, 0U)));
227     }
228     else
229     {
230         // Get the depthwise convolution compute parameters
231         auto t = ClDWCNativeKernelConfigurationFactory::create(gpu_target);
232         const DWCComputeKernelInfo dwc_native_compute_info = t->configure(input, weights, conv_info, dilation, depth_multiplier);
233         ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayerNativeKernel::validate(input, weights, biases, output, dwc_native_compute_info, conv_kernel_info, &output_multipliers_shifts_info,
234                                                                                       &output_multipliers_shifts_info));
235     }
236     return Status{};
237 }
238 
run()239 void CLDepthwiseConvolutionLayer::run()
240 {
241     prepare();
242 
243     MemoryGroupResourceScope scope_mg(_memory_group);
244 
245     if(_needs_permute)
246     {
247         _permute_input_to_nhwc.run();
248     }
249     CLScheduler::get().enqueue(*_dwc_native_kernel);
250     if(_needs_permute)
251     {
252         _permute_output_to_nchw.run();
253     }
254 }
255 
prepare()256 void CLDepthwiseConvolutionLayer::prepare()
257 {
258     if(!_is_prepared)
259     {
260         if(_is_quantized)
261         {
262             _output_multipliers.map();
263             _output_shifts.map();
264             quantization::compute_quantized_multipliers_and_shifts(_input->info(),
265                                                                    _original_weights->info(),
266                                                                    _output != nullptr ? _output->info() : _input->info(),
267                                                                    reinterpret_cast<int32_t *>(_output_multipliers.ptr_to_element(Coordinates(0))),
268                                                                    reinterpret_cast<int32_t *>(_output_shifts.ptr_to_element(Coordinates(0))));
269             _output_multipliers.unmap();
270             _output_shifts.unmap();
271         }
272 
273         if(_needs_permute)
274         {
275             ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
276 
277             _permuted_weights.allocator()->allocate();
278             _permute_weights_to_nhwc.run();
279             _original_weights->mark_as_unused();
280         }
281         _is_prepared = true;
282     }
283 }
284 } // namespace arm_compute
285