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_fully_connected_q15.c
22  * Description:  Q15 basic fully-connected layer function
23  *
24  * $Date:        17. January 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 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 
65 arm_status
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)66 arm_fully_connected_q15(const q15_t * pV,
67                         const q15_t * pM,
68                         const uint16_t dim_vec,
69                         const uint16_t num_of_rows,
70                         const uint16_t bias_shift,
71                         const uint16_t out_shift,
72                         const q15_t * bias,
73                         q15_t * pOut,
74                         q15_t * vec_buffer)
75 {
76 
77 #if defined (ARM_MATH_DSP)
78     /* Run the following code for Cortex-M4 and Cortex-M7 */
79 
80     const q15_t *pB = pM;
81     const q15_t *pB2 = pB + dim_vec;
82     q15_t    *pO = pOut;
83     const q15_t    *pA;
84     const q15_t    *pBias = bias;
85     uint16_t rowCnt = num_of_rows >> 1;
86 
87     /* this loop loops over different output */
88     while (rowCnt) {
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 = *__SIMD32(pA)++;
101             inM1 = *__SIMD32(pB)++;
102             sum = __SMLAD(inV1, inM1, sum);
103             inM2 = *__SIMD32(pB2)++;
104             sum2 = __SMLAD(inV1, inM2, sum2);
105 
106             inV1 = *__SIMD32(pA)++;
107             inM1 = *__SIMD32(pB)++;
108             sum = __SMLAD(inV1, inM1, sum);
109             inM2 = *__SIMD32(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         q31_t     sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
137 
138         uint16_t  colCnt = dim_vec >> 2;
139 
140         pA = pV;
141 
142         while (colCnt) {
143             q31_t     inV1, inM1;
144             inV1 = *__SIMD32(pA)++;
145             inM1 = *__SIMD32(pB)++;
146             sum = __SMLAD(inV1, inM1, sum);
147 
148             inV1 = *__SIMD32(pA)++;
149             inM1 = *__SIMD32(pB)++;
150             sum = __SMLAD(inV1, inM1, sum);
151 
152             colCnt--;
153 	}
154 
155 	/* left-over of the vector */
156 	colCnt = dim_vec & 0x3;
157 	while(colCnt) {
158             q15_t     inV = *pA++;
159             q15_t     inM = *pB++;
160 
161             sum += inV * inM;
162 
163             colCnt--;
164 	}
165 
166         *pO++ =  (q15_t) (__SSAT((sum >> out_shift), 16));
167 
168         rowCnt --;
169     }
170 
171 #else
172     int       i, j;
173     /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
174     for (i = 0; i < num_of_rows; i++)
175     {
176         int       ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
177         for (j = 0; j < dim_vec; j++)
178         {
179             ip_out += pV[j] * pM[i * dim_vec + j];
180         }
181         pOut[i] = (q15_t) __SSAT((ip_out >> out_shift), 16);
182     }
183 
184 #endif                          /* ARM_MATH_DSP */
185 
186     /* Return to application */
187     return (ARM_MATH_SUCCESS);
188 
189 }
190 
191 /**
192  * @} end of FC group
193  */
194