xref: /aosp_15_r20/external/ComputeLibrary/tests/validation/reference/SoftmaxLayer.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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 #include "SoftmaxLayer.h"
25 
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/Types.h"
28 #include "utils/TypePrinter.h"
29 
30 namespace arm_compute
31 {
32 namespace test
33 {
34 namespace validation
35 {
36 namespace reference
37 {
38 template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type>
softmax_layer_generic(const SimpleTensor<T> & src,float beta,int32_t axis,bool is_log)39 SimpleTensor<T> softmax_layer_generic(const SimpleTensor<T> &src, float beta, int32_t axis, bool is_log)
40 {
41     // Create reference
42     SimpleTensor<T> dst{ src.shape(), src.data_type(), 1 };
43 
44     const int32_t n_dims = static_cast<int32_t>(src.shape().num_dimensions());
45     ARM_COMPUTE_ERROR_ON(axis < -n_dims || axis >= n_dims);
46 
47     const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, n_dims));
48     Window             window;
49     window.use_tensor_dimensions(src.shape());
50     const unsigned int axis_dimension = src.shape()[actual_axis];
51     window.set(actual_axis, Window::Dimension(0, 1, 1));
52 
53     execute_window_loop(window, [&](const Coordinates & id)
54     {
55         // Find max along axis
56         Coordinates offset(id);
57         offset.set(actual_axis, 0);
58         T max = *reinterpret_cast<const T *>(src(offset));
59         for(unsigned int axis_id = 1; axis_id < axis_dimension; ++axis_id)
60         {
61             offset.set(actual_axis, axis_id);
62             const T val = *reinterpret_cast<const T *>(src(offset));
63             if(val > max)
64             {
65                 max = val;
66             }
67         }
68 
69         // Regularize
70         T sum(0.f);
71         for(unsigned int axis_id = 0; axis_id < axis_dimension; ++axis_id)
72         {
73             offset.set(actual_axis, axis_id);
74             const T val = *reinterpret_cast<const T *>(src(offset));
75             T       res{ (val - max) *beta };
76             if(is_log)
77             {
78                 sum += std::exp(res);
79             }
80             else
81             {
82                 res = std::exp(res);
83                 sum += res;
84             }
85             *reinterpret_cast<T *>(dst(offset)) = res;
86         }
87 
88         // Normalize
89         for(unsigned int axis_id = 0; axis_id < axis_dimension; ++axis_id)
90         {
91             offset.set(actual_axis, axis_id);
92             const T val = *reinterpret_cast<const T *>(dst(offset));
93             if(is_log)
94             {
95                 *reinterpret_cast<T *>(dst(offset)) = val - static_cast<T>(std::log(sum));
96             }
97             else
98             {
99                 *reinterpret_cast<T *>(dst(offset)) = val / sum;
100             }
101         }
102     });
103     return dst;
104 }
105 
106 template SimpleTensor<float> softmax_layer_generic(const SimpleTensor<float> &src, float beta, int32_t axis, bool is_log);
107 template SimpleTensor<half> softmax_layer_generic(const SimpleTensor<half> &src, float beta, int32_t axis, bool is_log);
108 
109 template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type>
softmax_layer(const SimpleTensor<T> & src,float beta,int32_t axis,bool is_log)110 SimpleTensor<T> softmax_layer(const SimpleTensor<T> &src, float beta, int32_t axis, bool is_log)
111 {
112     return softmax_layer_generic<T>(src, beta, axis, is_log);
113 }
114 
115 template < typename T, typename std::enable_if < std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int >::type >
softmax_layer(const SimpleTensor<T> & src,float beta,int32_t axis,bool is_log)116 SimpleTensor<T> softmax_layer(const SimpleTensor<T> &src, float beta, int32_t axis, bool is_log)
117 {
118     const QuantizationInfo output_quantization_info = arm_compute::get_softmax_output_quantization_info(src.data_type(), is_log);
119 
120     SimpleTensor<float> src_tmp = convert_from_asymmetric(src);
121     SimpleTensor<float> dst_tmp = softmax_layer<float>(src_tmp, beta, axis, is_log);
122     SimpleTensor<T>     dst     = convert_to_asymmetric<T>(dst_tmp, output_quantization_info);
123     return dst;
124 }
125 
126 template SimpleTensor<float> softmax_layer(const SimpleTensor<float> &src, float beta, int32_t axis, bool is_log);
127 template SimpleTensor<half> softmax_layer(const SimpleTensor<half> &src, float beta, int32_t axis, bool is_log);
128 template SimpleTensor<uint8_t> softmax_layer(const SimpleTensor<uint8_t> &src, float beta, int32_t axis, bool is_log);
129 template SimpleTensor<int8_t> softmax_layer(const SimpleTensor<int8_t> &src, float beta, int32_t axis, bool is_log);
130 
131 } // namespace reference
132 } // namespace validation
133 } // namespace test
134 } // namespace arm_compute
135