xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/operators/CpuSoftmax.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2021-2022 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_CPU_SOFTMAX_H
25 #define ARM_COMPUTE_CPU_SOFTMAX_H
26 
27 #include "arm_compute/core/TensorInfo.h"
28 #include "arm_compute/core/experimental/Types.h"
29 #include "src/cpu/ICpuKernel.h"
30 #include "src/cpu/ICpuOperator.h"
31 #include "src/cpu/operators/CpuPermute.h"
32 #include <memory>
33 
34 namespace arm_compute
35 {
36 namespace cpu
37 {
38 class CpuLogits1DMaxKernel;
39 template <bool IS_LOG>
40 class CpuLogits1DSoftmaxKernel;
41 
42 /** Basic function to compute a SoftmaxLayer and a Log SoftmaxLayer.
43  *
44  * Softmax is calculated by :
45  * @f[ out = exp((x - max(x)) * beta) / sum(exp((x - max(x)) * beta)) @f]
46  *
47  * Log Softmax is calculated by :
48  * @f[ out = (x - max(x) * beta) - log(\sum{e^{x - max(x) * beta}}) @f]
49  *
50  * This function runs the following function/kernels:
51  * -# If axis is not 0:
52  * -# @ref CpuPermute
53  * -# @ref kernels::CpuLogits1DMaxKernel
54  * -# @ref kernels::CpuLogits1DSoftmaxKernel
55  */
56 template <bool IS_LOG = false>
57 class CpuSoftmaxGeneric : public ICpuOperator
58 {
59 public:
60     CpuSoftmaxGeneric();
61     /** Set the input and output tensors.
62      *
63      * @param[in,out] src  Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
64      *                     last value of each row to the nearest multiple.
65      * @param[out]    dst  Destination tensor ifo. Data types supported: same as @p input.
66      * @param[in]     beta (Optional) A scaling factor for the exponent.
67      * @param[in]     axis (Optional) The dimension in which to apply the function. E.g. for input of shape 4x5x6 and
68      *                       axis=1, softmax will be applied to 4x6=24 vectors of size 5. Defaults to 0
69      */
70     void configure(const ITensorInfo *src, ITensorInfo *dst, float beta = 1.0f, int32_t axis = 0);
71     /** Static function to check if given info will lead to a valid configuration
72      *
73      * Similar to @ref CpuSoftmaxGeneric::configure()
74      *
75      * @return a status
76      */
77     static Status validate(const ITensorInfo *src, const ITensorInfo *dst, float beta = 1.0f, int32_t axis = 0);
78 
79     // Inherited methods overridden:
80     void run(ITensorPack &tensors) override;
81     experimental::MemoryRequirements workspace() const override;
82 
83 private:
84     enum InternalTensorIdx
85     {
86         MAX = 0,
87         TMP,
88         PERMUTED_SRC,
89         PERMUTED_DST,
90         COUNT
91     };
92 
93     CpuPermute                  _permute_input;
94     CpuPermute                  _permute_output;
95     std::unique_ptr<ICPPKernel> _max_kernel;
96     std::unique_ptr<ICPPKernel> _softmax_kernel;
97 
98     TensorInfo _max;
99     TensorInfo _tmp;
100     TensorInfo _input_permuted;
101     TensorInfo _output_permuted;
102 
103     bool                             _needs_permute;
104     experimental::MemoryRequirements _aux_mem{};
105 };
106 using CpuSoftmax    = CpuSoftmaxGeneric<false>;
107 using CpuLogSoftmax = CpuSoftmaxGeneric<true>;
108 
109 } // namespace cpu
110 } // namespace arm_compute
111 #endif /* ARM_COMPUTE_CPU_SOFTMAX_H */
112