1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
|
/*
* Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
*
* SPDX-License-Identifier: Apache-2.0
*
* Licensed under the Apache License, Version 2.0 (the License); you may
* not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS, WITHOUT
* WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/* ----------------------------------------------------------------------
* Project: CMSIS NN Library
* Title: arm_softmax_q7.c
* Description: Q7 softmax function
*
* $Date: 20. February 2018
* $Revision: V.1.0.0
*
* Target Processor: Cortex-M cores
*
* -------------------------------------------------------------------- */
#include "arm_math.h"
#include "arm_nnfunctions.h"
/**
* @ingroup groupNN
*/
/**
* @addtogroup Softmax
* @{
*/
/**
* @brief Q7 softmax function
* @param[in] vec_in pointer to input vector
* @param[in] dim_vec input vector dimention
* @param[out] p_out pointer to output vector
* @return none.
*
* @details
*
* Here, instead of typical natural logarithm e based softmax, we use
* 2-based softmax here, i.e.,:
*
* y_i = 2^(x_i) / sum(2^x_j)
*
* The relative output will be different here.
* But mathematically, the gradient will be the same
* with a log(2) scaling factor.
*
*/
void arm_softmax_q7(const q7_t * vec_in, const uint16_t dim_vec, q7_t * p_out)
{
q31_t sum;
int16_t i;
uint8_t shift;
q15_t base;
base = -257;
/* We first search for the maximum */
for (i = 0; i < dim_vec; i++)
{
if (vec_in[i] > base)
{
base = vec_in[i];
}
}
/*
* So the base is set to max-8, meaning
* that we ignore really small values.
* anyway, they will be 0 after shrinking to q7_t.
*/
base = base - 8;
sum = 0;
for (i = 0; i < dim_vec; i++)
{
if (vec_in[i] > base)
{
shift = (uint8_t)__USAT(vec_in[i] - base, 5);
sum += 0x1 << shift;
}
}
/* This is effectively (0x1 << 20) / sum */
int output_base = 0x100000 / sum;
/*
* Final confidence will be output_base >> ( 13 - (vec_in[i] - base) )
* so 128 (0x1<<7) -> 100% confidence when sum = 0x1 << 8, output_base = 0x1 << 12
* and vec_in[i]-base = 8
*/
for (i = 0; i < dim_vec; i++)
{
if (vec_in[i] > base)
{
/* Here minimum value of 13+base-vec_in[i] will be 5 */
shift = (uint8_t)__USAT(13+base-vec_in[i], 5);
p_out[i] = (q7_t) __SSAT((output_base >> shift), 8);
} else {
p_out[i] = 0;
}
}
}
/**
* @} end of Softmax group
*/
|
