summaryrefslogtreecommitdiff
path: root/Drivers/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7.c
blob: 75e924f4be8b1328d42bb0c5dcd3036e6f94370a (plain) (blame)
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
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
/*
 * 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_fully_connected_q7.c
 * Description:  Q7 basic fully-connected layer function
 *
 * $Date:        17. January 2018
 * $Revision:    V.1.0.0
 *
 * Target Processor:  Cortex-M cores
 *
 * -------------------------------------------------------------------- */

#include "arm_math.h"
#include "arm_nnfunctions.h"

/**
 *  @ingroup groupNN
 */

/**
 * @addtogroup FC
 * @{
 */

  /**
   * @brief Q7 basic fully-connected layer function
   * @param[in]       pV          pointer to input vector
   * @param[in]       pM          pointer to matrix weights
   * @param[in]       dim_vec     length of the vector
   * @param[in]       num_of_rows number of rows in weight matrix
   * @param[in]       bias_shift  amount of left-shift for bias
   * @param[in]       out_shift   amount of right-shift for output
   * @param[in]       bias        pointer to bias
   * @param[in,out]   pOut        pointer to output vector
   * @param[in,out]   vec_buffer  pointer to buffer space for input
   * @return     The function returns <code>ARM_MATH_SUCCESS</code>
   *
   * @details
   *
   * <b>Buffer size:</b>
   *
   * vec_buffer size: dim_vec
   *
   * This basic function is designed to work with regular weight
   * matrix without interleaving.
   *
   */

arm_status
arm_fully_connected_q7(const q7_t * pV,
                       const q7_t * pM,
                       const uint16_t dim_vec,
                       const uint16_t num_of_rows,
                       const uint16_t bias_shift,
                       const uint16_t out_shift, const q7_t * bias, q7_t * pOut, q15_t * vec_buffer)
{

#if defined (ARM_MATH_DSP)
    /* Run the following code for Cortex-M4 and Cortex-M7 */

    const q7_t *pB = pM;
    const q7_t *pB2;
    q7_t     *pO = pOut;
    const q7_t *pBias = bias;
    q15_t    *pA;
    uint16_t  rowCnt = num_of_rows >> 1;

    /* expand the vector into the buffer */
    arm_q7_to_q15_reordered_no_shift(pV, vec_buffer, dim_vec);

    while (rowCnt)
    {
        q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
        q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
        uint16_t  colCnt = dim_vec >> 2;

        pA = vec_buffer;
        pB2 = pB + dim_vec;

        while (colCnt)
        {
            q31_t     inV, inM11, inM12, inM21, inM22;
            pB = (q7_t *) read_and_pad_reordered((void *)pB, &inM11, &inM12);
            pB2 = (q7_t *) read_and_pad_reordered((void *)pB2, &inM21, &inM22);

            inV = *__SIMD32(pA)++;

            sum = __SMLAD(inV, inM11, sum);
            sum2 = __SMLAD(inV, inM21, sum2);

            inV = *__SIMD32(pA)++;

            sum = __SMLAD(inV, inM12, sum);
            sum2 = __SMLAD(inV, inM22, sum2);

            colCnt--;
        }
        colCnt = dim_vec & 0x3;
        while (colCnt)
        {
            q7_t      inV = *pA++;
            q15_t     inM = *pB++;
            q15_t     inM2 = *pB2++;

            sum += inV * inM;
            sum2 += inV * inM2;
            colCnt--;
        }                       /* while over colCnt */
        *pO++ = (q7_t) (__SSAT((sum >> out_shift), 8));
        *pO++ = (q7_t) (__SSAT((sum2 >> out_shift), 8));

        /* adjust the pointers and counters */
        pB += dim_vec;
        rowCnt--;
    }

    /* left-over part of the rows */
    rowCnt = num_of_rows & 0x1;

    while (rowCnt)
    {
        uint16_t  colCnt = dim_vec >> 2;
        q31_t     sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);

        pA = vec_buffer;

        while (colCnt)
        {
            q31_t     inV1, inV2, inM11, inM12;

            pB = (q7_t *) read_and_pad_reordered((void *)pB, &inM11, &inM12);

            inV1 = *__SIMD32(pA)++;
            sum = __SMLAD(inV1, inM11, sum);

            inV2 = *__SIMD32(pA)++;
            sum = __SMLAD(inV2, inM12, sum);

            colCnt--;
        }

        /* left-over of the vector */
        colCnt = dim_vec & 0x3;
        while (colCnt)
        {
            q7_t      inV = *pA++;
            q15_t     inM = *pB++;
            sum += inV * inM;
            colCnt--;
        }

        *pO++ = (q7_t) (__SSAT((sum >> out_shift), 8));

        rowCnt--;
    }

#else
    int       i, j;

    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
    for (i = 0; i < num_of_rows; i++)
    {
        int       ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
        for (j = 0; j < dim_vec; j++)
        {
            ip_out += pV[j] * pM[i * dim_vec + j];
        }
        pOut[i] = (q7_t) __SSAT((ip_out >> out_shift), 8);
    }

#endif                          /* ARM_MATH_DSP */

    /* Return to ARM_MATH_SUCCESS */
    return (ARM_MATH_SUCCESS);

}

/**
 * @} end of FC group
 */