• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2017-2020 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/CL/CLTensor.h"
28 #include "arm_compute/runtime/CL/functions/CLPermute.h"
29 #include "arm_compute/runtime/IFunction.h"
30 #include "arm_compute/runtime/IMemoryManager.h"
31 #include "arm_compute/runtime/MemoryGroup.h"
32 
33 #include <memory>
34 
35 namespace arm_compute
36 {
37 class CLCompileContext;
38 class CLLogits1DMaxShiftExpSumKernel;
39 class CLLogits1DNormKernel;
40 class ICLTensor;
41 class ITensorInfo;
42 
43 /** Basic function to compute a SoftmaxLayer.
44  *
45  * Softmax is calculated by :
46  * @f[ out = exp((x - max(x)) * beta) / sum(exp((x - max(x)) * beta)) @f]
47  *
48  * Log Softmax is calculated by :
49  * @f[ out = (x - max(x) * beta) - log(\sum{e^{x - max(x) * beta}}) @f]
50  *
51  * This function runs the following kernels:
52  * -# If axis is not 0:
53  * -#   @ref CLPermute
54  * -# @ref CLLogits1DNormKernel
55  * -# @ref CLLogits1DMaxShiftExpSumKernel
56  */
57 template <bool IS_LOG = false>
58 class CLSoftmaxLayerGeneric : public IFunction
59 {
60 public:
61     /** Constructor */
62     CLSoftmaxLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
63     /** Prevent instances of this class from being copied */
64     CLSoftmaxLayerGeneric(const CLSoftmaxLayerGeneric &) = delete;
65     /** Prevent instances of this class from being copied */
66     CLSoftmaxLayerGeneric &operator=(const CLSoftmaxLayerGeneric &) = delete;
67     /** Prevent instances of this class to be moved */
68     CLSoftmaxLayerGeneric(CLSoftmaxLayerGeneric &&) = delete;
69     /** Prevent instances of this class to be moved */
70     CLSoftmaxLayerGeneric &operator=(CLSoftmaxLayerGeneric &&) = delete;
71     /** Default destructor */
72     ~CLSoftmaxLayerGeneric();
73     /** Set the input and output tensors.
74      *
75      * @param[in]  input  Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32 for Softmax and F16/F32 for Log Softmax
76      * @param[out] output Destination tensor. Data types supported: same as @p input
77      * @param[in]  beta   (Optional) A scaling factor for the exponent. Defaults to 1.f
78      * @param[in]  axis   (Optional) The dimension in which to apply the function. E.g. for input of shape 4x5x6 and
79      *                       axis=1, softmax will be applied to 4x6=24 vectors of size 5. Defaults to 0
80      */
81     void configure(const ICLTensor *input, ICLTensor *output, float beta = 1.0f, int32_t axis = 0);
82     /** Set the input and output tensors.
83      *
84      * @param[in]  compile_context The compile context to be used.
85      * @param[in]  input           Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32 for Softmax and F16/F32 for Log Softmax
86      * @param[out] output          Destination tensor. Data types supported: same as @p input
87      * @param[in]  beta            (Optional) A scaling factor for the exponent. Defaults to 1.f
88      * @param[in]  axis            (Optional) The dimension in which to apply the function. E.g. for input of shape 4x5x6 and
89      *                       axis=1, softmax will be applied to 4x6=24 vectors of size 5. Defaults to 0
90      */
91     void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, float beta = 1.0f, int32_t axis = 0);
92     /** Static function to check if given info will lead to a valid configuration of @ref CLSoftmaxLayer
93      *
94      * @param[in] input  Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32 for Softmax and F16/F32 for Log Softmax
95      * @param[in] output Destination tensor. Data types supported: same as @p input
96      * @param[in] beta   (Optional) A scaling factor for the exponent. Defaults to 1.f
97      * @param[in] axis   (Optional) The dimension in which to apply the function. E.g. for input of shape 4x5x6 and
98      *                       axis=1, softmax will be applied to 4x6=24 vectors of size 5. Defaults to 0
99      *
100      * @return a status
101      */
102     static Status validate(const ITensorInfo *input, const ITensorInfo *output, float beta = 1.0f, int32_t axis = 0);
103 
104     // Inherited methods overridden:
105     void run() override;
106 
107 private:
108     MemoryGroup                                     _memory_group;
109     CLPermute                                       _permute_input;
110     CLPermute                                       _permute_output;
111     std::unique_ptr<CLLogits1DMaxShiftExpSumKernel> _max_shift_exp_sum_kernel;
112     std::unique_ptr<CLLogits1DNormKernel>           _norm_kernel;
113     CLTensor                                        _max;
114     CLTensor                                        _sum;
115     CLTensor                                        _tmp;
116     CLTensor                                        _input_permuted;
117     CLTensor                                        _output_permuted;
118     bool                                            _needs_permute;
119 };
120 
121 using CLSoftmaxLayer    = CLSoftmaxLayerGeneric<false>;
122 using CLLogSoftmaxLayer = CLSoftmaxLayerGeneric<true>;
123 } // namespace arm_compute
124 #endif /* ARM_COMPUTE_CLSOFTMAXLAYER_H */
125