• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2017-2021 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_CLGEMMLOWPOUTPUTSTAGE_H
25 #define ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H
26 
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/runtime/CL/ICLSimpleFunction.h"
29 
30 #include <limits>
31 
32 /** This file contains all available output stages for GEMMLowp on OpenCL.
33  *
34  *  In gemmlowp, the "output stage" is the process that takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyCore),
35  *  and processes it to obtain the final QASYMM8/QASYMM8_SIGNED value.
36  *
37  *  More information about the GEMMLowp output stage can be found at https://github.com/google/gemmlowp/blob/master/doc/output.md
38  */
39 
40 namespace arm_compute
41 {
42 class CLCompileContext;
43 class ITensor;
44 class ICLTensor;
45 class ITensorInfo;
46 struct GEMMLowpOutputStageInfo;
47 
48 /** Basic function to execute GEMMLowpQuantizeDown kernels on CL.
49  *
50  *  This function calls the following CL kernels:
51  *
52  * -# @ref opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleKernel
53  * -# @ref opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFloatKernel
54  * -# @ref opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel
55 */
56 class CLGEMMLowpOutputStage : public IFunction
57 {
58 public:
59     CLGEMMLowpOutputStage();
60     /** Prevent instances of this class from being copied (As this class contains pointers) */
61     CLGEMMLowpOutputStage(const CLGEMMLowpOutputStage &) = delete;
62     /** Default move constructor */
63     CLGEMMLowpOutputStage(CLGEMMLowpOutputStage &&);
64     /** Prevent instances of this class from being copied (As this class contains pointers) */
65     CLGEMMLowpOutputStage &operator=(const CLGEMMLowpOutputStage &) = delete;
66     /** Default move assignment operator */
67     CLGEMMLowpOutputStage &operator=(CLGEMMLowpOutputStage &&);
68     /** Default destructor */
69     ~CLGEMMLowpOutputStage();
70     /** Initialise the kernel's inputs, output
71      *
72      * Valid data layouts:
73      * - All
74      *
75      * Valid data type configurations:
76      * |src0           |src1          |dst           |
77      * |:--------------|:-------------|:-------------|
78      * |S32            |S32           |QASYMM8       |
79      * |S32            |S32           |QASYMM8_SIGNED|
80      * |S32            |S32           |QSYMM16       |
81      *
82      * @param[in]  input  Input tensor. Data type supported: S32
83      * @param[in]  bias   Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
84      *                    Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
85      * @param[out] output Output tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM16
86      * @param[in]  info   GEMMLowp output stage metadata.
87      */
88     void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const GEMMLowpOutputStageInfo &info);
89     /** Initialise the kernel's inputs, output
90      *
91      * @param[in]  compile_context The compile context to be used.
92      * @param[in]  input           Input tensor. Data type supported: S32
93      * @param[in]  bias            Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
94      *                             Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
95      * @param[out] output          Output tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
96      * @param[in]  info            GEMMLowp output stage metadata.
97      */
98     void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const GEMMLowpOutputStageInfo &info);
99     /** Static function to check if given info will lead to a valid configuration of @ref opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel
100      *
101      * @param[in] input  Input tensor. It is the output of @ref CLGEMMLowpMatrixMultiplyCore function. Data type supported: S32
102      * @param[in] bias   Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
103      *                   Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
104      * @param[in] output Output tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
105      * @param[in] info   GEMMLowp output stage metadata.
106      *
107      * @return a status
108      */
109     static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo &info);
110 
111     // Inherited methods overridden:
112     void run() override;
113 
114 private:
115     struct Impl;
116     std::unique_ptr<Impl> _impl;
117 };
118 } // namespace arm_compute
119 #endif /*ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H */
120