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_q7.c
22  * Description:  Q7 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 Q7 basic 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    * @details
57    *
58    * <b>Buffer size:</b>
59    *
60    * vec_buffer size: dim_vec
61    *
62    * This basic function is designed to work with regular weight
63    * matrix without interleaving.
64    *
65    */
66 
67 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)68 arm_fully_connected_q7(const q7_t * pV,
69                        const q7_t * pM,
70                        const uint16_t dim_vec,
71                        const uint16_t num_of_rows,
72                        const uint16_t bias_shift,
73                        const uint16_t out_shift, const q7_t * bias, q7_t * pOut, q15_t * vec_buffer)
74 {
75 
76 #if defined (ARM_MATH_DSP)
77     /* Run the following code for Cortex-M4 and Cortex-M7 */
78 
79     const q7_t *pB = pM;
80     const q7_t *pB2;
81     q7_t     *pO = pOut;
82     const q7_t *pBias = bias;
83     q15_t    *pA;
84     uint16_t  rowCnt = num_of_rows >> 1;
85 
86     /* expand the vector into the buffer */
87     arm_q7_to_q15_reordered_no_shift(pV, vec_buffer, dim_vec);
88 
89     while (rowCnt)
90     {
91         q31_t     sum =  ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
92         q31_t     sum2 = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
93         uint16_t  colCnt = dim_vec >> 2;
94 
95         pA = vec_buffer;
96         pB2 = pB + dim_vec;
97 
98         while (colCnt)
99         {
100             q31_t     inV, inM11, inM12, inM21, inM22;
101             pB = (q7_t *) read_and_pad_reordered((void *)pB, &inM11, &inM12);
102             pB2 = (q7_t *) read_and_pad_reordered((void *)pB2, &inM21, &inM22);
103 
104             inV = *__SIMD32(pA)++;
105 
106             sum = __SMLAD(inV, inM11, sum);
107             sum2 = __SMLAD(inV, inM21, sum2);
108 
109             inV = *__SIMD32(pA)++;
110 
111             sum = __SMLAD(inV, inM12, sum);
112             sum2 = __SMLAD(inV, inM22, sum2);
113 
114             colCnt--;
115         }
116         colCnt = dim_vec & 0x3;
117         while (colCnt)
118         {
119             q7_t      inV = *pA++;
120             q15_t     inM = *pB++;
121             q15_t     inM2 = *pB2++;
122 
123             sum += inV * inM;
124             sum2 += inV * inM2;
125             colCnt--;
126         }                       /* while over colCnt */
127         *pO++ = (q7_t) (__SSAT((sum >> out_shift), 8));
128         *pO++ = (q7_t) (__SSAT((sum2 >> out_shift), 8));
129 
130         /* adjust the pointers and counters */
131         pB += dim_vec;
132         rowCnt--;
133     }
134 
135     /* left-over part of the rows */
136     rowCnt = num_of_rows & 0x1;
137 
138     while (rowCnt)
139     {
140         uint16_t  colCnt = dim_vec >> 2;
141         q31_t     sum = ((q31_t)(*pBias++) << bias_shift) + NN_ROUND(out_shift);
142 
143         pA = vec_buffer;
144 
145         while (colCnt)
146         {
147             q31_t     inV1, inV2, inM11, inM12;
148 
149             pB = (q7_t *) read_and_pad_reordered((void *)pB, &inM11, &inM12);
150 
151             inV1 = *__SIMD32(pA)++;
152             sum = __SMLAD(inV1, inM11, sum);
153 
154             inV2 = *__SIMD32(pA)++;
155             sum = __SMLAD(inV2, inM12, sum);
156 
157             colCnt--;
158         }
159 
160         /* left-over of the vector */
161         colCnt = dim_vec & 0x3;
162         while (colCnt)
163         {
164             q7_t      inV = *pA++;
165             q15_t     inM = *pB++;
166             sum += inV * inM;
167             colCnt--;
168         }
169 
170         *pO++ = (q7_t) (__SSAT((sum >> out_shift), 8));
171 
172         rowCnt--;
173     }
174 
175 #else
176     int       i, j;
177 
178     /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
179     for (i = 0; i < num_of_rows; i++)
180     {
181         int       ip_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
182         for (j = 0; j < dim_vec; j++)
183         {
184             ip_out += pV[j] * pM[i * dim_vec + j];
185         }
186         pOut[i] = (q7_t) __SSAT((ip_out >> out_shift), 8);
187     }
188 
189 #endif                          /* ARM_MATH_DSP */
190 
191     /* Return to ARM_MATH_SUCCESS */
192     return (ARM_MATH_SUCCESS);
193 
194 }
195 
196 /**
197  * @} end of FC group
198  */
199