1 /*
2 * Copyright (C) 2010-2020 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_fully_connected_q15.c
22 * Description: Q15 basic fully-connected layer function
23 *
24 * $Date: 09. October 2020
25 * $Revision: V.1.0.1
26 *
27 * Target Processor: Cortex-M cores
28 *
29 * -------------------------------------------------------------------- */
30
31 #include "arm_nnfunctions.h"
32 #include "arm_nnsupportfunctions.h"
33
34 /**
35 * @ingroup groupNN
36 */
37
38 /**
39 * @addtogroup FC
40 * @{
41 */
42
43 /**
44 * @brief Q15 opt fully-connected layer function
45 * @param[in] pV pointer to input vector
46 * @param[in] pM pointer to matrix weights
47 * @param[in] dim_vec length of the vector
48 * @param[in] num_of_rows number of rows in weight matrix
49 * @param[in] bias_shift amount of left-shift for bias
50 * @param[in] out_shift amount of right-shift for output
51 * @param[in] bias pointer to bias
52 * @param[in,out] pOut pointer to output vector
53 * @param[in,out] vec_buffer pointer to buffer space for input
54 * @return The function returns <code>ARM_MATH_SUCCESS</code>
55 *
56 *
57 * @details
58 *
59 * <b>Buffer size:</b>
60 *
61 * vec_buffer size: 0
62 *
63 */
64
arm_fully_connected_q15(const q15_t * pV,const q15_t * pM,const uint16_t dim_vec,const uint16_t num_of_rows,const uint16_t bias_shift,const uint16_t out_shift,const q15_t * bias,q15_t * pOut,q15_t * vec_buffer)65 arm_status arm_fully_connected_q15(const q15_t *pV,
66 const q15_t *pM,
67 const uint16_t dim_vec,
68 const uint16_t num_of_rows,
69 const uint16_t bias_shift,
70 const uint16_t out_shift,
71 const q15_t *bias,
72 q15_t *pOut,
73 q15_t *vec_buffer)
74 {
75 (void)vec_buffer;
76 #if defined(ARM_MATH_DSP)
77 /* Run the following code for Cortex-M4 and Cortex-M7 */
78
79 const q15_t *pB = pM;
80 const q15_t *pB2 = pB + dim_vec;
81 q15_t *pO = pOut;
82 const q15_t *pA;
83 const q15_t *pBias = bias;
84 uint16_t rowCnt = num_of_rows >> 1;
85
86 /* this loop loops over different output */
87 while (rowCnt)
88 {
89 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
90 q31_t sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
91
92 uint16_t colCnt = dim_vec >> 2;
93
94 pA = pV;
95 pB2 = pB + dim_vec;
96
97 while (colCnt)
98 {
99 q31_t inV1, inM1, inM2;
100 inV1 = arm_nn_read_q15x2_ia(&pA);
101 inM1 = arm_nn_read_q15x2_ia(&pB);
102 sum = __SMLAD(inV1, inM1, sum);
103 inM2 = arm_nn_read_q15x2_ia(&pB2);
104 sum2 = __SMLAD(inV1, inM2, sum2);
105
106 inV1 = arm_nn_read_q15x2_ia(&pA);
107 inM1 = arm_nn_read_q15x2_ia(&pB);
108 sum = __SMLAD(inV1, inM1, sum);
109 inM2 = arm_nn_read_q15x2_ia(&pB2);
110 sum2 = __SMLAD(inV1, inM2, sum2);
111
112 colCnt--;
113 }
114 colCnt = dim_vec & 0x3;
115 while (colCnt)
116 {
117 q15_t inV = *pA++;
118 q15_t inM = *pB++;
119 q15_t inM2 = *pB2++;
120
121 sum += inV * inM;
122 sum2 += inV * inM2;
123 colCnt--;
124 } /* while over colCnt */
125 *pO++ = (q15_t)(__SSAT((sum >> out_shift), 16));
126 *pO++ = (q15_t)(__SSAT((sum2 >> out_shift), 16));
127
128 /* adjust the pointers and counters */
129 pB = pB + dim_vec;
130 rowCnt--;
131 }
132
133 rowCnt = num_of_rows & 0x1;
134
135 while (rowCnt)
136 {
137 q31_t sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
138
139 uint16_t colCnt = dim_vec >> 2;
140
141 pA = pV;
142
143 while (colCnt)
144 {
145 q31_t inV1, inM1;
146 inV1 = arm_nn_read_q15x2_ia(&pA);
147 inM1 = arm_nn_read_q15x2_ia(&pB);
148 sum = __SMLAD(inV1, inM1, sum);
149
150 inV1 = arm_nn_read_q15x2_ia(&pA);
151 inM1 = arm_nn_read_q15x2_ia(&pB);
152 sum = __SMLAD(inV1, inM1, sum);
153
154 colCnt--;
155 }
156
157 /* left-over of the vector */
158 colCnt = dim_vec & 0x3;
159 while (colCnt)
160 {
161 q15_t inV = *pA++;
162 q15_t inM = *pB++;
163
164 sum += inV * inM;
165
166 colCnt--;
167 }
168
169 *pO++ = (q15_t)(__SSAT((sum >> out_shift), 16));
170
171 rowCnt--;
172 }
173
174 #else
175 int i, j;
176 /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
177 for (i = 0; i < num_of_rows; i++)
178 {
179 int ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
180 for (j = 0; j < dim_vec; j++)
181 {
182 ip_out += pV[j] * pM[i * dim_vec + j];
183 }
184 pOut[i] = (q15_t)__SSAT((ip_out >> out_shift), 16);
185 }
186
187 #endif /* ARM_MATH_DSP */
188
189 /* Return to application */
190 return (ARM_MATH_SUCCESS);
191 }
192
193 /**
194 * @} end of FC group
195 */
196
