1 /*
2 * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
3 *
4 * SPDX-License-Identifier: Apache-2.0
5 *
6 * Licensed under the Apache License, Version 2.0 (the License); you may
7 * not use this file except in compliance with the License.
8 * You may obtain a copy of the License at
9 *
10 * www.apache.org/licenses/LICENSE-2.0
11 *
12 * Unless required by applicable law or agreed to in writing, software
13 * distributed under the License is distributed on an AS IS BASIS, WITHOUT
14 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15 * See the License for the specific language governing permissions and
16 * limitations under the License.
17 */
18
19 /* ----------------------------------------------------------------------
20 * Project: CMSIS NN Library
21 * Title: arm_softmax_q15.c
22 * Description: Q15 softmax function
23 *
24 * $Date: 20. February 2018
25 * $Revision: V.1.0.0
26 *
27 * Target Processor: Cortex-M cores
28 *
29 * -------------------------------------------------------------------- */
30
31 #include "arm_math.h"
32 #include "arm_nnfunctions.h"
33
34 /**
35 * @ingroup groupNN
36 */
37
38 /**
39 * @addtogroup Softmax
40 * @{
41 */
42
43 /**
44 * @brief Q15 softmax function
45 * @param[in] vec_in pointer to input vector
46 * @param[in] dim_vec input vector dimention
47 * @param[out] p_out pointer to output vector
48 * @return none.
49 *
50 * @details
51 *
52 * Here, instead of typical e based softmax, we use
53 * 2-based softmax, i.e.,:
54 *
55 * y_i = 2^(x_i) / sum(2^x_j)
56 *
57 * The relative output will be different here.
58 * But mathematically, the gradient will be the same
59 * with a log(2) scaling factor.
60 *
61 */
62
arm_softmax_q15(const q15_t * vec_in,const uint16_t dim_vec,q15_t * p_out)63 void arm_softmax_q15(const q15_t * vec_in, const uint16_t dim_vec, q15_t * p_out)
64 {
65 q31_t sum;
66 int16_t i;
67 uint8_t shift;
68 q31_t base;
69 base = -1 * 0x100000;
70 for (i = 0; i < dim_vec; i++)
71 {
72 if (vec_in[i] > base)
73 {
74 base = vec_in[i];
75 }
76 }
77
78 /* we ignore really small values
79 * anyway, they will be 0 after shrinking
80 * to q15_t
81 */
82 base = base - 16;
83
84 sum = 0;
85
86 for (i = 0; i < dim_vec; i++)
87 {
88 if (vec_in[i] > base)
89 {
90 shift = (uint8_t)__USAT(vec_in[i] - base, 5);
91 sum += 0x1 << shift;
92 }
93 }
94
95 /* This is effectively (0x1 << 32) / sum */
96 int64_t div_base = 0x100000000LL;
97 int output_base = (int32_t)(div_base / sum);
98
99 /* Final confidence will be output_base >> ( 17 - (vec_in[i] - base) )
100 * so 32768 (0x1<<15) -> 100% confidence when sum = 0x1 << 16, output_base = 0x1 << 16
101 * and vec_in[i]-base = 16
102 */
103 for (i = 0; i < dim_vec; i++)
104 {
105 if (vec_in[i] > base)
106 {
107 /* Here minimum value of 17+base-vec[i] will be 1 */
108 shift = (uint8_t)__USAT(17+base-vec_in[i], 5);
109 p_out[i] = (q15_t) __SSAT((output_base >> shift), 16);
110 } else
111 {
112 p_out[i] = 0;
113 }
114 }
115
116 }
117
118 /**
119 * @} end of Softmax group
120 */
121