• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2019-2021 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 #ifndef ARM_COMPUTE_CLGENERATEPROPOSALSLAYER_H
25 #define ARM_COMPUTE_CLGENERATEPROPOSALSLAYER_H
26 
27 #include "arm_compute/core/Types.h"
28 #include "arm_compute/runtime/CL/CLScheduler.h"
29 #include "arm_compute/runtime/CL/CLTensor.h"
30 #include "arm_compute/runtime/CL/functions/CLPermute.h"
31 #include "arm_compute/runtime/CL/functions/CLReshapeLayer.h"
32 #include "arm_compute/runtime/CPP/CPPScheduler.h"
33 #include "arm_compute/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.h"
34 #include "arm_compute/runtime/IFunction.h"
35 #include "arm_compute/runtime/MemoryGroup.h"
36 
37 #include <memory>
38 
39 namespace arm_compute
40 {
41 class CLCompileContext;
42 class CLBoundingBoxTransformKernel;
43 class CLDequantizationLayer;
44 class CLComputeAllAnchorsKernel;
45 class CLPadLayerKernel;
46 class CLQuantizationLayer;
47 class ICLTensor;
48 class ITensorInfo;
49 
50 /** Basic function to generate proposals for a RPN (Region Proposal Network)
51  *
52  * This function calls the following OpenCL kernels:
53  * -# @ref CLComputeAllAnchorsKernel
54  * -# @ref CLPermute x 2
55  * -# @ref CLReshapeLayer x 2
56  * -# @ref CLBoundingBoxTransform
57  * -# @ref CLPadLayerKernel
58  * -# @ref CLDequantizationLayer x 2
59  * -# @ref CLQuantizationLayer
60  * And the following CPP functions:
61  * -# @ref CPPBoxWithNonMaximaSuppressionLimit
62  */
63 class CLGenerateProposalsLayer : public IFunction
64 {
65 public:
66     /** Default constructor
67      *
68      * @param[in] memory_manager (Optional) Memory manager.
69      */
70     CLGenerateProposalsLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
71     /** Prevent instances of this class from being copied (As this class contains pointers) */
72     CLGenerateProposalsLayer(const CLGenerateProposalsLayer &) = delete;
73     /** Prevent instances of this class from being copied (As this class contains pointers) */
74     CLGenerateProposalsLayer &operator=(const CLGenerateProposalsLayer &) = delete;
75     /** Default destructor */
76     ~CLGenerateProposalsLayer();
77 
78     /** Set the input and output tensors.
79      *
80      * Valid data layouts:
81      * - All
82      *
83      * Valid data type configurations:
84      * |src0           |src1               |src2     |dst            |
85      * |:--------------|:------------------|:--------|:--------------|
86      * |F16            |F16                |F16      |F16            |
87      * |F32            |F32                |F32      |F32            |
88      * |QASYMM8        |QSYMM8             |QSYMM16  |QASYMM8        |
89      *
90      * @param[in]  scores              Scores from convolution layer of size (W, H, A), where H and W are the height and width of the feature map, and A is the number of anchors.
91      *                                 Data types supported: QASYMM8/F16/F32
92      * @param[in]  deltas              Bounding box deltas from convolution layer of size (W, H, 4*A). Data types supported: Same as @p scores
93      * @param[in]  anchors             Anchors tensor of size (4, A). Data types supported: QSYMM16 with scale of 0.125 if @p scores is QASYMM8, otherwise same as @p scores
94      * @param[out] proposals           Box proposals output tensor of size (5, W*H*A).
95      *                                 Data types supported: QASYMM16 with scale of 0.125 and 0 offset if @p scores is QASYMM8, otherwise same as @p scores
96      * @param[out] scores_out          Box scores output tensor of size (W*H*A). Data types supported: Same as @p scores
97      * @param[out] num_valid_proposals Scalar output tensor which says which of the first proposals are valid. Data types supported: U32
98      * @param[in]  info                Contains GenerateProposals operation information described in @ref GenerateProposalsInfo
99      *
100      * @note Only single image prediction is supported. Height and Width (and scale) of the image will be contained in the @ref GenerateProposalsInfo struct.
101      * @note Proposals contains all the proposals. Of those, only the first num_valid_proposals are valid.
102      */
103     void configure(const ICLTensor *scores, const ICLTensor *deltas, const ICLTensor *anchors, ICLTensor *proposals, ICLTensor *scores_out, ICLTensor *num_valid_proposals,
104                    const GenerateProposalsInfo &info);
105     /** Set the input and output tensors.
106      *
107      * @param[in]  compile_context     The compile context to be used.
108      * @param[in]  scores              Scores from convolution layer of size (W, H, A), where H and W are the height and width of the feature map, and A is the number of anchors.
109      *                                 Data types supported: QASYMM8/F16/F32
110      * @param[in]  deltas              Bounding box deltas from convolution layer of size (W, H, 4*A). Data types supported: Same as @p scores
111      * @param[in]  anchors             Anchors tensor of size (4, A). Data types supported: QSYMM16 with scale of 0.125 if @p scores is QASYMM8, otherwise same as @p scores
112      * @param[out] proposals           Box proposals output tensor of size (5, W*H*A).
113      *                                 Data types supported: QASYMM16 with scale of 0.125 and 0 offset if @p scores is QASYMM8, otherwise same as @p scores
114      * @param[out] scores_out          Box scores output tensor of size (W*H*A). Data types supported: Same as @p scores
115      * @param[out] num_valid_proposals Scalar output tensor which says which of the first proposals are valid. Data types supported: U32
116      * @param[in]  info                Contains GenerateProposals operation information described in @ref GenerateProposalsInfo
117      *
118      * @note Only single image prediction is supported. Height and Width (and scale) of the image will be contained in the @ref GenerateProposalsInfo struct.
119      * @note Proposals contains all the proposals. Of those, only the first num_valid_proposals are valid.
120      */
121     void configure(const CLCompileContext &compile_context, const ICLTensor *scores, const ICLTensor *deltas, const ICLTensor *anchors, ICLTensor *proposals, ICLTensor *scores_out,
122                    ICLTensor *num_valid_proposals, const GenerateProposalsInfo &info);
123 
124     /** Static function to check if given info will lead to a valid configuration of @ref CLGenerateProposalsLayer
125      *
126      * @param[in] scores              Scores info from convolution layer of size (W, H, A), where H and W are the height and width of the feature map, and A is the number of anchors.
127      *                                Data types supported: QASYMM8/F16/F32
128      * @param[in] deltas              Bounding box deltas info from convolution layer of size (W, H, 4*A). Data types supported: Same as @p scores
129      * @param[in] anchors             Anchors tensor of size (4, A). Data types supported: QSYMM16 with scale of 0.125 if @p scores is QASYMM8, otherwise same as @p scores
130      * @param[in] proposals           Box proposals info  output tensor of size (5, W*H*A).
131      *                                Data types supported: QASYMM16 with scale of 0.125 and 0 offset if @p scores is QASYMM8, otherwise same as @p scores
132      * @param[in] scores_out          Box scores output tensor info of size (W*H*A). Data types supported: Same as @p scores
133      * @param[in] num_valid_proposals Scalar output tensor info which says which of the first proposals are valid. Data types supported: U32
134      * @param[in] info                Contains GenerateProposals operation information described in @ref GenerateProposalsInfo
135      *
136      * @return a Status
137      */
138     static Status validate(const ITensorInfo *scores, const ITensorInfo *deltas, const ITensorInfo *anchors, const ITensorInfo *proposals, const ITensorInfo *scores_out,
139                            const ITensorInfo           *num_valid_proposals,
140                            const GenerateProposalsInfo &info);
141 
142     // Inherited methods overridden:
143     void run() override;
144 
145 private:
146     // Memory group manager
147     MemoryGroup _memory_group;
148 
149     // OpenCL kernels
150     CLPermute                                     _permute_deltas;
151     CLReshapeLayer                                _flatten_deltas;
152     CLPermute                                     _permute_scores;
153     CLReshapeLayer                                _flatten_scores;
154     std::unique_ptr<CLComputeAllAnchorsKernel>    _compute_anchors_kernel;
155     std::unique_ptr<CLBoundingBoxTransformKernel> _bounding_box_kernel;
156     std::unique_ptr<CLPadLayerKernel>             _pad_kernel;
157     std::unique_ptr<CLDequantizationLayer>        _dequantize_anchors;
158     std::unique_ptr<CLDequantizationLayer>        _dequantize_deltas;
159     std::unique_ptr<CLQuantizationLayer>          _quantize_all_proposals;
160 
161     // CPP functions
162     CPPBoxWithNonMaximaSuppressionLimit _cpp_nms;
163 
164     bool _is_nhwc;
165     bool _is_qasymm8;
166 
167     // Temporary tensors
168     CLTensor _deltas_permuted;
169     CLTensor _deltas_flattened;
170     CLTensor _deltas_flattened_f32;
171     CLTensor _scores_permuted;
172     CLTensor _scores_flattened;
173     CLTensor _all_anchors;
174     CLTensor _all_anchors_f32;
175     CLTensor _all_proposals;
176     CLTensor _all_proposals_quantized;
177     CLTensor _keeps_nms_unused;
178     CLTensor _classes_nms_unused;
179     CLTensor _proposals_4_roi_values;
180 
181     // Temporary tensor pointers
182     CLTensor *_all_proposals_to_use;
183 
184     // Output tensor pointers
185     ICLTensor *_num_valid_proposals;
186     ICLTensor *_scores_out;
187 
188     /** Internal function to run the CPP BoxWithNMS kernel */
189     void run_cpp_nms_kernel();
190 };
191 } // namespace arm_compute
192 #endif /* ARM_COMPUTE_CLGENERATEPROPOSALSLAYER_H */
193