• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2020-2021 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_NEGEMMCONV2D_H
25 #define ARM_COMPUTE_NEGEMMCONV2D_H
26 
27 #include "arm_compute/runtime/FunctionDescriptors.h"
28 #include "arm_compute/runtime/IFunction.h"
29 #include "arm_compute/runtime/IMemoryManager.h"
30 
31 #include <memory>
32 
33 namespace arm_compute
34 {
35 // Forward declarations
36 class ITensor;
37 class ITensorInfo;
38 
39 /** Basic function to compute the convolution layer. This function calls the following kernels/functions:
40  *
41  * Supports only NHWC data layout
42  *
43  * -# @ref cpu::CpuGemmAssemblyDispatch
44  * -# @ref NEActivationLayer, in case activation cannot be fused in the assembly dispatch
45  *
46  * Weights are transformed from OHWI to HWIO format using the following kernels:
47  * -# @ref NEPermute
48  */
49 class NEGEMMConv2d : public IFunction
50 {
51 public:
52     /** Constructor */
53     NEGEMMConv2d(const std::shared_ptr<IMemoryManager> &memory_manager = nullptr);
54     /** Prevent instances of this class from being copied (As this class contains pointers) */
55     NEGEMMConv2d(const NEGEMMConv2d &) = delete;
56     /** Default move constructor */
57     NEGEMMConv2d(NEGEMMConv2d &&) = default;
58     /** Prevent instances of this class from being copied (As this class contains pointers) */
59     NEGEMMConv2d &operator=(const NEGEMMConv2d &) = delete;
60     /** Default move assignment operator */
61     NEGEMMConv2d &operator=(NEGEMMConv2d &&) = default;
62     /** Destructor */
63     ~NEGEMMConv2d();
64     /** Set the input and output tensors.
65      *
66      * Valid data layouts:
67      * - All
68      *
69      * Valid data type configurations:
70      * |src0           |src1           |src2           |dst            |
71      * |:--------------|:--------------|:--------------|:--------------|
72      * |QASYMM8        |QASYMM8        |S32            |QASYMM8        |
73      * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32            |QASYMM8_SIGNED |
74      * |F16            |F16            |F16            |F16            |
75      * |F32            |F32            |F32            |F32            |
76      * |BFLOAT16       |BFLOAT16       |BFLOAT16       |BFLOAT16       |
77      *
78      * @param[in]  input   Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
79      *                     while every optional dimension from 4 and above represent a batch of inputs.
80      *                     Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
81      * @param[in]  weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
82      *                     Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
83      * @param[in]  biases  Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
84      *                     Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
85      * @param[out] output  Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
86      *                     Data types supported: Same as @p input.
87      * @param[in]  info    Convolution layer descriptor
88      */
89     void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const Conv2dInfo &info);
90     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMConv2d
91      *
92      * @param[in] input   Source tensor info. 3 lower dimensions represent a single input [width, height, IFM],
93      *                    while every optional dimension from 4 and above represent a batch of inputs.
94      *                    Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
95      * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
96      *                    Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
97      * @param[in] biases  Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
98      *                    Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
99      * @param[in] output  Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
100      *                    Data types supported: Same as @p input.
101      * @param[in] info    Contains padding and stride information described in @ref PadStrideInfo.
102      *
103      * @return a status
104      */
105     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const Conv2dInfo &info);
106 
107     // Inherited methods overridden:
108     void run() override;
109     void prepare() override;
110 
111 private:
112     struct Impl;
113     std::unique_ptr<Impl> _impl;
114 };
115 } // namespace arm_compute
116 #endif /* ARM_COMPUTE_NEGEMMCONV2D_H */
117