• 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_NEGEMMCONVOLUTIONLAYER_H
25 #define ARM_COMPUTE_NEGEMMCONVOLUTIONLAYER_H
26 
27 #include "arm_compute/runtime/IFunction.h"
28 
29 #include "arm_compute/core/Types.h"
30 #include "arm_compute/runtime/IWeightsManager.h"
31 #include "arm_compute/runtime/MemoryGroup.h"
32 #include "arm_compute/runtime/NEON/functions/NEGEMM.h"
33 #include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
34 #include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h"
35 #include "arm_compute/runtime/NEON/functions/NEReshapeLayer.h"
36 #include "arm_compute/runtime/Tensor.h"
37 
38 #include <memory>
39 
40 namespace arm_compute
41 {
42 class ITensor;
43 class NECol2ImKernel;
44 class NEIm2ColKernel;
45 class NEWeightsReshapeKernel;
46 
47 /** Function to reshape the weights. This function calls the following kernel:
48  * -# @ref NEWeightsReshapeKernel
49  */
50 class NEConvolutionLayerReshapeWeights : public IFunction
51 {
52 public:
53     /** Constructor */
54     NEConvolutionLayerReshapeWeights();
55     /** Prevent instances of this class from being copied (As this class contains pointers) */
56     NEConvolutionLayerReshapeWeights(const NEConvolutionLayerReshapeWeights &) = delete;
57     /** Prevent instances of this class from being moved (As this class contains non movable objects) */
58     NEConvolutionLayerReshapeWeights(NEConvolutionLayerReshapeWeights &&) = delete;
59     /** Prevent instances of this class from being copied (As this class contains pointers) */
60     NEConvolutionLayerReshapeWeights &operator=(const NEConvolutionLayerReshapeWeights &) = delete;
61     /** Prevent instances of this class from being moved (As this class contains non movable objects) */
62     NEConvolutionLayerReshapeWeights &operator=(NEConvolutionLayerReshapeWeights &&) = delete;
63     /** Default destructor */
64     ~NEConvolutionLayerReshapeWeights();
65     /** Set the input and output tensors.
66      *
67      * @param[in]  weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
68      *                     Data type supported: All.
69      * @param[in]  biases  Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
70      *                     Data type supported: same as @p weights.
71      *                     @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
72      * @param[out] output  Destination tensor. Data types supported: same as @p weights.
73      */
74     void configure(const ITensor *weights, const ITensor *biases, ITensor *output);
75     /** Static function to check if given info will lead to a valid configuration of @ref NEConvolutionLayerReshapeWeights
76      *
77      * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
78      *                    Data type supported: All.
79      * @param[in] biases  Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
80      *                    Data type supported: same as @p weights.
81      *                    @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
82      * @param[in] output  Destination tensor. Data types supported: same as @p weights.
83      *
84      * @return an error status
85      */
86     static Status validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output);
87 
88     // Inherited methods overridden:
89     void run() override;
90 
91 private:
92     std::unique_ptr<NEWeightsReshapeKernel> _weights_reshape_kernel;
93 };
94 
95 namespace weights_transformations
96 {
97 /** Basic function to manage the reshape weights generated from @ref NEConvolutionLayerReshapeWeights */
98 class NEConvolutionLayerReshapeWeightsTransform : public ITransformWeights
99 {
100 public:
101     /** Constructor */
102     NEConvolutionLayerReshapeWeightsTransform() = default;
103     /** Prevent instances of this class from being copied (As this class contains pointers) */
104     NEConvolutionLayerReshapeWeightsTransform(const NEConvolutionLayerReshapeWeightsTransform &) = delete;
105     /** Prevent instances of this class from being copied (As this class contains pointers) */
106     NEConvolutionLayerReshapeWeightsTransform &operator=(const NEConvolutionLayerReshapeWeightsTransform &) = delete;
107     /** Prevent instances of this class from being moved (As this class contains non movable objects) */
108     NEConvolutionLayerReshapeWeightsTransform(NEConvolutionLayerReshapeWeightsTransform &&) = delete;
109     /** Prevent instances of this class from being moved (As this class contains non movable objects) */
110     NEConvolutionLayerReshapeWeightsTransform &operator=(NEConvolutionLayerReshapeWeightsTransform &&) = delete;
111     /** Default destructor */
112     ~NEConvolutionLayerReshapeWeightsTransform() = default;
configure(const ITensor * input,const ITensor * biases)113     void configure(const ITensor *input, const ITensor *biases)
114     {
115         _bias_bit = (biases != nullptr) ? 1 : 0;
116         _func.configure(input, biases, &_output);
117     }
118 
run()119     void run() override
120     {
121         _output.allocator()->allocate();
122         _func.run();
123         _reshape_run = true;
124     }
125 
get_weights()126     ITensor *get_weights() override
127     {
128         return &_output;
129     }
130 
release()131     void release() override
132     {
133         _output.allocator()->free();
134     }
135 
uid()136     uint32_t uid() override
137     {
138         return ((0x8) | (_bias_bit << 7));
139     }
140 
is_reshape_run()141     bool is_reshape_run()
142     {
143         return _reshape_run;
144     }
145 
146 private:
147     Tensor                           _output{};
148     NEConvolutionLayerReshapeWeights _func{};
149     int32_t                          _bias_bit{ 0 };
150 };
151 } // namespace weights_transformations
152 
153 /** Basic function to compute the convolution layer. This function calls the following NEON kernels/functions:
154  *
155  * -# @ref NEIm2ColKernel
156  * -# @ref NEGEMM (if the data type is BFLOAT16/FP16/FP32)
157  * -# @ref NEGEMMLowpMatrixMultiplyCore (if the data type is QASYMM8/QASYMM8_SIGNED)
158  * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if the data type is QASYMM8/QASYMM8_SIGNED)
159  * -# @ref NEArithmeticAdditionKernel (if biases != nullptr and we have a 1x1 convolution with the NHWC data layout)
160  * -# @ref NECol2ImKernel (if NCHW data layout)
161  *
162  */
163 class NEGEMMConvolutionLayer : public IFunction
164 {
165 public:
166     /** Constructor */
167     NEGEMMConvolutionLayer(const std::shared_ptr<IMemoryManager> &memory_manager = nullptr, IWeightsManager *weights_manager = nullptr);
168     /** Prevent instances of this class from being copied (As this class contains pointers) */
169     NEGEMMConvolutionLayer(const NEGEMMConvolutionLayer &) = delete;
170     /** Prevent instances of this class from being moved (As this class contains non movable objects) */
171     NEGEMMConvolutionLayer(NEGEMMConvolutionLayer &&) = delete;
172     /** Prevent instances of this class from being copied (As this class contains pointers) */
173     NEGEMMConvolutionLayer &operator=(const NEGEMMConvolutionLayer &) = delete;
174     /** Prevent instances of this class from being moved (As this class contains non movable objects) */
175     NEGEMMConvolutionLayer &operator=(NEGEMMConvolutionLayer &&) = delete;
176     /** Default destructor */
177     ~NEGEMMConvolutionLayer();
178     /** Set the input and output tensors.
179      *
180      * @param[in]  input        Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
181      *                          while every optional dimension from 4 and above represent a batch of inputs.
182      *                          Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
183      * @param[in]  weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
184      *                          Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
185      * @param[in]  biases       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
186      *                          Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
187      * @param[out] output       Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
188      *                          Data types supported: Same as @p input.
189      * @param[in]  conv_info    Contains padding and stride information described in @ref PadStrideInfo.
190      * @param[in]  weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
191      *                          tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input.
192      * @param[in]  dilation     (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
193      * @param[in]  act_info     (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
194      * @param[in]  num_groups   (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported
195      */
196     void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo(),
197                    const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), unsigned int num_groups = 1);
198     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMConvolutionLayer
199      *
200      * @param[in] input        Source tensor info. 3 lower dimensions represent a single input [width, height, IFM],
201      *                         while every optional dimension from 4 and above represent a batch of inputs.
202      *                         Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
203      * @param[in] weights      Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
204      *                         Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
205      * @param[in] biases       Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
206      *                         Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
207      * @param[in] output       Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
208      *                         Data types supported: Same as @p input.
209      * @param[in] conv_info    Contains padding and stride information described in @ref PadStrideInfo.
210      * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
211      *                         tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input.
212      * @param[in] dilation     (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
213      * @param[in] act_info     (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
214      * @param[in] num_groups   (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported
215      *
216      * @return a status
217      */
218     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
219                            const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), unsigned int num_groups = 1);
220 
221     // Inherited methods overridden:
222     void run() override;
223     void prepare() override;
224 
225 private:
226     /** Configures the appropriate matrix multiply routine
227      *
228      * @param[in]  input         Input tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
229      * @param[in]  weights       Weights tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
230      * @param[in]  biases        Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
231      *                           Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
232      * @param[out] output        Output tensor. Data types supported: Same as @p input,
233      *                           except for input of QASYMM8/QASYMM8_SIGNED type where output should be of S32 type.
234      * @param[in]  act_info      (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
235      * @param[in]  gemm_3d_depth (Optional) Depth of GEMM 3D (Defaults to 1)
236      */
237     void configure_mm(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const ActivationLayerInfo &act_info = ActivationLayerInfo(), int gemm_3d_depth = 1);
238     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMConvolutionLayer matrix multiply routines
239      *
240      * @param[in] input         Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
241      * @param[in] weights       Weights tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
242      * @param[in] biases        Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
243      *                          Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
244      * @param[in] output        Output tensor info. Data types supported: Same as @p input,
245      *                          except for input of QASYMM8/QASYMM8_SIGNED type where output should be of S32 type.
246      * @param[in] act_info      (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
247      * @param[in] gemm_3d_depth (Optional) Depth of GEMM 3D (Defaults to 1)
248      * @param[in] skip_im2col   (Optional) Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout. (Default to false)
249      *
250      * @return a status
251      */
252     static Status validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo(),
253                               int gemm_3d_depth = 1, bool skip_im2col = false);
254     /** Static function to check if GEMM3D is supported in @ref NEGEMM or in @ref NEGEMMLowpMatrixMultiplyCore
255      *
256      * @param[in] input_info    Input tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
257      * @param[in] weights_info  Weights tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
258      * @param[in] act_info      Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
259      * @param[in] gemm_3d_depth Depth of GEMM 3D
260      * @param[in] skip_im2col   Flag which specifies if im2col has to be skipped. i.e. 1x1 convolution with NHWC data layout
261      *
262      * @return a status
263      */
264     static Status validate_gemm3d(const ITensorInfo *input_info, const ITensorInfo *weights_info, const ActivationLayerInfo &act_info, int gemm_3d_depth, bool skip_im2col);
265 
266 private:
267     MemoryGroup                                                        _memory_group;
268     IWeightsManager                                                   *_weights_manager;
269     NEConvolutionLayerReshapeWeights                                   _reshape_weights;
270     weights_transformations::NEConvolutionLayerReshapeWeightsTransform _reshape_weights_managed;
271     std::unique_ptr<NEIm2ColKernel>                                    _im2col_kernel;
272     NEGEMM                                                             _mm_gemm;
273     NEGEMMLowpMatrixMultiplyCore                                       _mm_gemmlowp;
274     std::unique_ptr<NECol2ImKernel>                                    _col2im_kernel;
275     NEReshapeLayer                                                     _reshape_layer;
276 
277     const ITensor *_original_weights;
278 
279     Tensor _im2col_output;
280     Tensor _weights_reshaped;
281     Tensor _gemm_output;
282     Tensor _tmp_output;
283 
284     DataLayout _data_layout;
285 
286     bool _skip_im2col;
287     bool _skip_col2im;
288     bool _is_quantized;
289     bool _is_prepared;
290 };
291 } // namespace arm_compute
292 #endif /* ARM_COMPUTE_NECONVOLUTIONGEMMLAYER_H */
293