• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #ifndef ARM_COMPUTE_CLFULLYCONNECTEDLAYER_H
25 #define ARM_COMPUTE_CLFULLYCONNECTEDLAYER_H
26 
27 #include "arm_compute/runtime/CL/ICLSimpleFunction.h"
28 
29 #include "arm_compute/runtime/CL/CLTensor.h"
30 #include "arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h"
31 #include "arm_compute/runtime/CL/functions/CLFlattenLayer.h"
32 #include "arm_compute/runtime/CL/functions/CLGEMM.h"
33 #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
34 #include "arm_compute/runtime/IWeightsManager.h"
35 #include "arm_compute/runtime/MemoryGroup.h"
36 
37 namespace arm_compute
38 {
39 /** Basic function to reshape the weights of Fully Connected layer with OpenCL. This function calls the following kernels:
40  *
41  *  -# @ref CLTransposeKernel
42  *
43  * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
44  */
45 class CLFullyConnectedLayerReshapeWeights : public ICLSimpleFunction
46 {
47 public:
48     /** Set the input and output tensors.
49      *
50      * @param[in]  input  Weights tensor. The weights must be 2 dimensional. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
51      * @param[out] output Destination tensor which stores the transposed input tensor. Data type supported: Same as @p input.
52      */
53     void configure(const ICLTensor *input, ICLTensor *output);
54     /** Set the input and output tensors.
55      *
56      * @param[in]  compile_context The compile context to be used.
57      * @param[in]  input           Weights tensor. The weights must be 2 dimensional. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
58      * @param[out] output          Destination tensor which stores the transposed input tensor. Data type supported: Same as @p input.
59      */
60     void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output);
61     /** Static function to check if given info will lead to a valid configuration of @ref CLFullyConnectedLayerReshapeWeights
62      *
63      * @param[in] input  Weights tensor. The weights must be 2 dimensional. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
64      * @param[in] output Destination tensor which stores the transposed input tensor. Data type supported: Same as @p input.
65      *
66      * @return a status
67      */
68     static Status validate(const ITensorInfo *input, const ITensorInfo *output);
69 };
70 
71 namespace weights_transformations
72 {
73 /** Basic function to manage the reshape weights generated from @ref CLFullyConnectedLayerReshapeWeights */
74 class CLFullyConnectedLayerReshapeWeightsManaged : public ITransformWeights
75 {
76 public:
77     //Inherited method override
run()78     void run() override
79     {
80         _output.allocator()->allocate();
81         _func.run();
82         _reshape_run = true;
83     }
84 
85     //Inherited method override
release()86     void release() override
87     {
88         _output.allocator()->free();
89     }
90 
91     //Inherited method override
get_weights()92     ICLTensor *get_weights() override
93     {
94         return &_output;
95     }
96 
97     //Inherited method override
uid()98     uint32_t uid() override
99     {
100         return _uid;
101     }
102 
103     /** Configures the @ref CLFullyConnectedLayerReshapeWeights function
104      *
105      * @param[in] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
106      */
configure(const ICLTensor * input)107     void configure(const ICLTensor *input)
108     {
109         configure(CLKernelLibrary::get().get_compile_context(), input);
110     }
111     /** Configures the @ref CLFullyConnectedLayerReshapeWeights function
112      *
113      * @param[in] compile_context The compile context to be used.
114      * @param[in] input           Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
115      */
configure(const CLCompileContext & compile_context,const ICLTensor * input)116     void configure(const CLCompileContext &compile_context, const ICLTensor *input)
117     {
118         _func.configure(compile_context, input, &_output);
119     }
120 
121 private:
122     static constexpr uint32_t           _uid = 0x0;
123     CLTensor                            _output{};
124     CLFullyConnectedLayerReshapeWeights _func{};
125 };
126 } // namespace weights_transformations
127 
128 /** Basic function to compute a Fully Connected layer on OpenCL. This function calls the following OpenCL kernels:
129  *
130  *  -# @ref CLIm2ColKernel (called when the input comes from a convolutional layer)
131  *  -# @ref CLFullyConnectedLayerReshapeWeights (if @p are_weights_reshaped is set to false and transpose_weights is set to true ) (called once)
132  *  -# @ref CLGEMMMatrixMultiplyKernel or @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
133  *
134  * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
135  */
136 class CLFullyConnectedLayer : public IFunction
137 {
138 public:
139     /** Constructor */
140     CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr, IWeightsManager *weights_manager = nullptr);
141     /** Prevent instances of this class from being copied (As this class contains pointers) */
142     CLFullyConnectedLayer(const CLFullyConnectedLayer &) = delete;
143     /** Default move constructor */
144     CLFullyConnectedLayer(CLFullyConnectedLayer &&) = default;
145     /** Prevent instances of this class from being copied (As this class contains pointers) */
146     CLFullyConnectedLayer &operator=(const CLFullyConnectedLayer &) = delete;
147     /** Default move assignment operator */
148     CLFullyConnectedLayer &operator=(CLFullyConnectedLayer &&) = default;
149     /** Set the input and output tensors.
150      *
151      * @param[in]  input   Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
152      * @param[in]  weights Weights tensor. The weights must be 2 dimensional.
153      *                     If this function is called after a Convolution Layer, the (transposed) weights will have as many rows as the product of the first 3 input's dimensions.
154      *                     If it is called after another FullyConnected Layer, the (transposed) weights will have as many rows as the input's first dimension.
155      *                     Data type supported: Same as @p input.
156      * @param[in]  biases  Bias tensor. Can be nullptr. Data type supported:Same as @p input.
157      * @param[out] output  Destination tensor. Its shape should be equal to the output of a matrix multiplication between:
158      *                     - The output of im2col on the input and the (transposed) 2D weights, if the function is called after a Convolution Layer
159      *                     - The input tensor and the (transposed) 2D weights, if the function is called after another FullyConnected Layer.
160      *                     Data type supported: Same as @p input.
161      * @param[in]  fc_info (Optional) Fully connected layer additional info
162      */
163     void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
164                    FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
165     /** Set the input and output tensors.
166      *
167      * @param[in]  compile_context The compile context to be used.
168      * @param[in]  input           Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
169      * @param[in]  weights         Weights tensor. The weights must be 2 dimensional.
170      *                             If this function is called after a Convolution Layer, the (transposed) weights will have as many rows as the product of the first 3 input's dimensions.
171      *                             If it is called after another FullyConnected Layer, the (transposed) weights will have as many rows as the input's first dimension.
172      *                             Data type supported: Same as @p input.
173      * @param[in]  biases          Bias tensor. Can be nullptr. Data type supported:Same as @p input.
174      * @param[out] output          Destination tensor. Its shape should be equal to the output of a matrix multiplication between:
175      *                             - The output of im2col on the input and the (transposed) 2D weights, if the function is called after a Convolution Layer
176      *                             - The input tensor and the (transposed) 2D weights, if the function is called after another FullyConnected Layer.
177      *                             Data type supported: Same as @p input.
178      * @param[in]  fc_info         (Optional) Fully connected layer additional info
179      */
180     void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
181                    FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
182     /** Static function to check if given info will lead to a valid configuration of @ref CLFullyConnectedLayer
183      *
184      * @param[in]  input   Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
185      * @param[in]  weights Weights tensor info. The weights must be 2 dimensional.
186      *                     If this function is called after a Convolution Layer, the (transposed) weights will have as many rows as the product of the first 3 input's dimensions.
187      *                     If it is called after another FullyConnected Layer, the (transposed) weights will have as many rows as the input's first dimension.
188      *                     Data type supported: Same as @p input.
189      * @param[in]  biases  Bias tensor info. Can be nullptr. Data type supported:Same as @p input.
190      * @param[out] output  Destination tensor info. Its shape should be equal to the output of a matrix multiplication between:
191      *                     - The output of im2col on the input and the (transposed) 2D weights, if the function is called after a Convolution Layer
192      *                     - The input tensor and the (transposed) 2D weights, if the function is called after another FullyConnected Layer.
193      *                     Data type supported: Same as @p input.
194      * @param[in]  fc_info (Optional) Fully connected layer additional info
195      *
196      * @return a status
197      */
198     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
199                            FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
200 
201     //Inherited methods override
202     void run() override;
203     void prepare() override;
204 
205 private:
206     void configure_fc_fc(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const FullyConnectedLayerInfo &fc_info);
207     void configure_conv_fc(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const FullyConnectedLayerInfo &fc_info);
208     void configure_mm(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const FullyConnectedLayerInfo &fc_info);
209 
210     MemoryGroup                                                         _memory_group;
211     IWeightsManager                                                    *_weights_manager;
212     CLConvertFullyConnectedWeights                                      _convert_weights;
213     weights_transformations::CLConvertFullyConnectedWeightsManaged      _convert_weights_managed;
214     weights_transformations::CLFullyConnectedLayerReshapeWeightsManaged _reshape_weights_managed_function;
215     CLFlattenLayer                                                      _flatten_layer;
216     CLFullyConnectedLayerReshapeWeights                                 _reshape_weights_function;
217     CLGEMM                                                              _mm_gemm;
218     CLGEMMLowpMatrixMultiplyCore                                        _mm_gemmlowp;
219     CLTensor                                                            _flatten_output;
220     CLTensor                                                            _converted_weights_output;
221     CLTensor                                                            _reshape_weights_output;
222     bool                                                                _are_weights_converted;
223     bool                                                                _are_weights_reshaped;
224     bool                                                                _is_fc_after_conv;
225     bool                                                                _is_quantized;
226     bool                                                                _is_prepared;
227     const ICLTensor                                                    *_original_weights;
228 };
229 } // namespace arm_compute
230 #endif /* ARM_COMPUTE_CLFULLYCONNECTEDLAYER_H */
231