xref: /aosp_15_r20/external/ComputeLibrary/arm_compute/runtime/NEON/functions/NEGenerateProposalsLayer.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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_NEGENERATEPROPOSALSLAYER_H
25 #define ARM_COMPUTE_NEGENERATEPROPOSALSLAYER_H
26 
27 #include "arm_compute/core/Types.h"
28 #include "arm_compute/runtime/CPP/CPPScheduler.h"
29 #include "arm_compute/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.h"
30 #include "arm_compute/runtime/IFunction.h"
31 #include "arm_compute/runtime/MemoryGroup.h"
32 #include "arm_compute/runtime/NEON/functions/NEBoundingBoxTransform.h"
33 #include "arm_compute/runtime/NEON/functions/NEDequantizationLayer.h"
34 #include "arm_compute/runtime/NEON/functions/NEPadLayer.h"
35 #include "arm_compute/runtime/NEON/functions/NEPermute.h"
36 #include "arm_compute/runtime/NEON/functions/NEQuantizationLayer.h"
37 #include "arm_compute/runtime/NEON/functions/NEReshapeLayer.h"
38 #include "arm_compute/runtime/Tensor.h"
39 
40 namespace arm_compute
41 {
42 class ITensor;
43 class NEComputeAllAnchorsKernel;
44 
45 /** Basic function to generate proposals for a RPN (Region Proposal Network)
46  *
47  * This function calls the following Arm(R) Neon(TM) layers/kernels:
48  * -# @ref NEComputeAllAnchorsKernel
49  * -# @ref NEPermute x 2
50  * -# @ref NEReshapeLayer x 2
51  * -# @ref NEBoundingBoxTransform
52  * -# @ref NEPadLayerKernel
53  * -# @ref NEDequantizationLayer x 2
54  * -# @ref NEQuantizationLayer
55  * And the following CPP kernels:
56  * -# @ref CPPBoxWithNonMaximaSuppressionLimit
57  */
58 class NEGenerateProposalsLayer : public IFunction
59 {
60 public:
61     /** Default constructor
62      *
63      * @param[in] memory_manager (Optional) Memory manager.
64      */
65     NEGenerateProposalsLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
66     /** Prevent instances of this class from being copied (As this class contains pointers) */
67     NEGenerateProposalsLayer(const NEGenerateProposalsLayer &) = delete;
68     /** Prevent instances of this class from being copied (As this class contains pointers) */
69     NEGenerateProposalsLayer &operator=(const NEGenerateProposalsLayer &) = delete;
70     /** Default destructor */
71     ~NEGenerateProposalsLayer();
72 
73     /** Set the input and output tensors.
74      *
75      * Valid data layouts:
76      * - All
77      *
78      * Valid data type configurations:
79      * |src0           |src1               |src2     |dst            |
80      * |:--------------|:------------------|:--------|:--------------|
81      * |F16            |F16                |F16      |F16            |
82      * |F32            |F32                |F32      |F32            |
83      * |QASYMM8        |QSYMM8             |QSYMM16  |QASYMM8        |
84      *
85      * @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.
86      *                                 Data types supported: QASYMM8/F16/F32
87      * @param[in]  deltas              Bounding box deltas from convolution layer of size (W, H, 4*A). Data types supported: Same as @p scores
88      * @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
89      * @param[out] proposals           Box proposals output tensor of size (5, W*H*A).
90      *                                 Data types supported: QASYMM16 with scale of 0.125 and 0 offset if @p scores is QASYMM8, otherwise same as @p scores
91      * @param[out] scores_out          Box scores output tensor of size (W*H*A). Data types supported: Same as @p input
92      * @param[out] num_valid_proposals Scalar output tensor which says which of the first proposals are valid. Data types supported: U32
93      * @param[in]  info                Contains GenerateProposals operation information described in @ref GenerateProposalsInfo
94      *
95      * @note Only single image prediction is supported. Height and Width (and scale) of the image will be contained in the @ref GenerateProposalsInfo struct.
96      * @note Proposals contains all the proposals. Of those, only the first num_valid_proposals are valid.
97      */
98     void configure(const ITensor *scores, const ITensor *deltas, const ITensor *anchors, ITensor *proposals, ITensor *scores_out, ITensor *num_valid_proposals,
99                    const GenerateProposalsInfo &info);
100 
101     /** Static function to check if given info will lead to a valid configuration of @ref NEGenerateProposalsLayer
102      *
103      * @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.
104      *                                Data types supported: QASYMM8/F16/F32
105      * @param[in] deltas              Bounding box deltas info from convolution layer of size (W, H, 4*A). Data types supported: Same as @p scores
106      * @param[in] anchors             Anchors tensor info of size (4, A). Data types supported: QSYMM16 with scale of 0.125 if @p scores is QASYMM8, otherwise same as @p scores
107      * @param[in] proposals           Box proposals info  output tensor of size (5, W*H*A).
108      *                                Data types supported: QASYMM16 with scale of 0.125 and 0 offset if @p scores is QASYMM8, otherwise same as @p scores
109      * @param[in] scores_out          Box scores output tensor info of size (W*H*A). Data types supported: Same as @p input
110      * @param[in] num_valid_proposals Scalar output tensor info which says which of the first proposals are valid. Data types supported: U32
111      * @param[in] info                Contains GenerateProposals operation information described in @ref GenerateProposalsInfo
112      *
113      * @return a Status
114      */
115     static Status validate(const ITensorInfo *scores, const ITensorInfo *deltas, const ITensorInfo *anchors, const ITensorInfo *proposals, const ITensorInfo *scores_out,
116                            const ITensorInfo           *num_valid_proposals,
117                            const GenerateProposalsInfo &info);
118 
119     // Inherited methods overridden:
120     void run() override;
121 
122 private:
123     // Memory group manager
124     MemoryGroup _memory_group;
125 
126     // kernels/layers
127     NEPermute                                  _permute_deltas;
128     NEReshapeLayer                             _flatten_deltas;
129     NEPermute                                  _permute_scores;
130     NEReshapeLayer                             _flatten_scores;
131     std::unique_ptr<NEComputeAllAnchorsKernel> _compute_anchors;
132     NEBoundingBoxTransform                     _bounding_box;
133     NEPadLayer                                 _pad;
134     NEDequantizationLayer                      _dequantize_anchors;
135     NEDequantizationLayer                      _dequantize_deltas;
136     NEQuantizationLayer                        _quantize_all_proposals;
137 
138     // CPP functions
139     CPPBoxWithNonMaximaSuppressionLimit _cpp_nms;
140 
141     bool _is_nhwc;
142     bool _is_qasymm8;
143 
144     // Temporary tensors
145     Tensor _deltas_permuted;
146     Tensor _deltas_flattened;
147     Tensor _deltas_flattened_f32;
148     Tensor _scores_permuted;
149     Tensor _scores_flattened;
150     Tensor _all_anchors;
151     Tensor _all_anchors_f32;
152     Tensor _all_proposals;
153     Tensor _all_proposals_quantized;
154     Tensor _keeps_nms_unused;
155     Tensor _classes_nms_unused;
156     Tensor _proposals_4_roi_values;
157 
158     // Temporary tensor pointers
159     Tensor *_all_proposals_to_use;
160 
161     // Output tensor pointers
162     ITensor *_num_valid_proposals;
163     ITensor *_scores_out;
164 };
165 } // namespace arm_compute
166 #endif /* ARM_COMPUTE_NEGENERATEPROPOSALSLAYER_H */
167