xref: /aosp_15_r20/external/ComputeLibrary/arm_compute/runtime/CL/functions/CLSoftmaxLayer.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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_CLSOFTMAXLAYER_H
25 #define ARM_COMPUTE_CLSOFTMAXLAYER_H
26 
27 #include "arm_compute/runtime/IFunction.h"
28 #include "arm_compute/runtime/IMemoryManager.h"
29 #include "arm_compute/runtime/MemoryGroup.h"
30 
31 #include <memory>
32 
33 namespace arm_compute
34 {
35 class ICLTensor;
36 class ITensorInfo;
37 class CLCompileContext;
38 
39 /** Basic function to compute a SoftmaxLayer.
40  *
41  * Softmax is calculated by :
42  * @f[ out = exp((x - max(x)) * beta) / sum(exp((x - max(x)) * beta)) @f]
43  *
44  * Log Softmax is calculated by :
45  * @f[ out = (x - max(x) * beta) - log(\sum{e^{x - max(x) * beta}}) @f]
46  *
47  * This function runs the following operators/kernels:
48  * -# If axis is not 0:
49  * -# @ref opencl::ClPermute
50  * -# @ref opencl::kernels::ClLogits1DNormKernel
51  * -# @ref opencl::kernels::ClLogits1DMaxShiftExpSumKernel
52  */
53 template <bool IS_LOG = false>
54 class CLSoftmaxLayerGeneric : public IFunction
55 {
56 public:
57     /** Constructor */
58     CLSoftmaxLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
59     /** Default destructor */
60     ~CLSoftmaxLayerGeneric();
61     /** Set the input and output tensors.
62      *
63      * Valid data layouts:
64      * - All
65      *
66      * Valid data type configurations:
67      * |src            |dst            |
68      * |:--------------|:--------------|
69      * |QASYMM8        |QASYMM8        |
70      * |QASYMM8_SIGNED |QASYMM8_SIGNED |
71      * |F16            |F16            |
72      * |F32            |F32            |
73      *
74      * @param[in]  input  Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32 for Softmax and F16/F32 for Log Softmax
75      * @param[out] output Destination tensor. Data types supported: same as @p input
76      * @param[in]  beta   (Optional) A scaling factor for the exponent. Defaults to 1.f
77      * @param[in]  axis   (Optional) The dimension in which to apply the function. E.g. for input of shape 4x5x6 and
78      *                       axis=1, softmax will be applied to 4x6=24 vectors of size 5. Defaults to 0
79      */
80     void configure(const ICLTensor *input, ICLTensor *output, float beta = 1.0f, int32_t axis = 0);
81     /** Set the input and output tensors.
82      *
83      * @param[in]  compile_context The compile context to be used.
84      * @param[in]  input           Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32 for Softmax and F16/F32 for Log Softmax
85      * @param[out] output          Destination tensor. Data types supported: same as @p input
86      * @param[in]  beta            (Optional) A scaling factor for the exponent. Defaults to 1.f
87      * @param[in]  axis            (Optional) The dimension in which to apply the function. E.g. for input of shape 4x5x6 and
88      *                       axis=1, softmax will be applied to 4x6=24 vectors of size 5. Defaults to 0
89      */
90     void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, float beta = 1.0f, int32_t axis = 0);
91     /** Static function to check if given info will lead to a valid configuration of @ref CLSoftmaxLayer
92      *
93      * @param[in] input  Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32 for Softmax and F16/F32 for Log Softmax
94      * @param[in] output Destination tensor. Data types supported: same as @p input
95      * @param[in] beta   (Optional) A scaling factor for the exponent. Defaults to 1.f
96      * @param[in] axis   (Optional) The dimension in which to apply the function. E.g. for input of shape 4x5x6 and
97      *                       axis=1, softmax will be applied to 4x6=24 vectors of size 5. Defaults to 0
98      *
99      * @return a status
100      */
101     static Status validate(const ITensorInfo *input, const ITensorInfo *output, float beta = 1.0f, int32_t axis = 0);
102 
103     // Inherited methods overridden:
104     void run() override;
105 
106 private:
107     struct Impl;
108     std::unique_ptr<Impl> _impl;
109 };
110 
111 using CLSoftmaxLayer    = CLSoftmaxLayerGeneric<false>;
112 using CLLogSoftmaxLayer = CLSoftmaxLayerGeneric<true>;
113 } // namespace arm_compute
114 #endif /* ARM_COMPUTE_CLSOFTMAXLAYER_H */
115