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_convolve_HWC_q15_fast.c
22  * Description:  Fast Q15 version of convolution
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 NNConv
40  * @{
41  */
42 
43   /**
44    * @brief Fast Q15 convolution function
45    * @param[in]       Im_in       pointer to input tensor
46    * @param[in]       dim_im_in   input tensor dimention
47    * @param[in]       ch_im_in    number of input tensor channels
48    * @param[in]       wt          pointer to kernel weights
49    * @param[in]       ch_im_out   number of filters, i.e., output tensor channels
50    * @param[in]       dim_kernel  filter kernel size
51    * @param[in]       padding     padding sizes
52    * @param[in]       stride      convolution stride
53    * @param[in]       bias        pointer to bias
54    * @param[in]       bias_shift  amount of left-shift for bias
55    * @param[in]       out_shift   amount of right-shift for output
56    * @param[in,out]   Im_out      pointer to output tensor
57    * @param[in]       dim_im_out  output tensor dimension
58    * @param[in,out]   bufferA     pointer to buffer space for input
59    * @param[in,out]   bufferB     pointer to buffer space for output
60    * @return     The function returns either
61    * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
62    *
63    * @details
64    *
65    * <b>Buffer size:</b>
66    *
67    * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel
68    *
69    * bufferB size: 0
70    *
71    * <b>Input dimension constraints:</b>
72    *
73    * ch_im_in is multiple of 2
74    *
75    * ch_im_out is multipe of 2
76    *
77    */
78 
79 arm_status
arm_convolve_HWC_q15_fast(const q15_t * Im_in,const uint16_t dim_im_in,const uint16_t ch_im_in,const q15_t * wt,const uint16_t ch_im_out,const uint16_t dim_kernel,const uint16_t padding,const uint16_t stride,const q15_t * bias,const uint16_t bias_shift,const uint16_t out_shift,q15_t * Im_out,const uint16_t dim_im_out,q15_t * bufferA,q7_t * bufferB)80 arm_convolve_HWC_q15_fast(const q15_t * Im_in,
81                           const uint16_t dim_im_in,
82                           const uint16_t ch_im_in,
83                           const q15_t * wt,
84                           const uint16_t ch_im_out,
85                           const uint16_t dim_kernel,
86                           const uint16_t padding,
87                           const uint16_t stride,
88                           const q15_t * bias,
89                           const uint16_t bias_shift,
90                           const uint16_t out_shift,
91                           q15_t * Im_out,
92                           const uint16_t dim_im_out,
93                           q15_t * bufferA,
94                           q7_t * bufferB)
95 {
96 
97 #if defined (ARM_MATH_DSP)
98     int16_t   i_out_y, i_out_x, i_ker_y, i_ker_x;
99 
100     q15_t    *pBuffer = bufferA;
101     q15_t    *im_buffer = bufferA;
102     q15_t    *pOut = Im_out;
103 
104     if (ch_im_in % 2 != 0 || ch_im_out % 2 != 0)
105     {
106         /* check if the input dimension meets the constraints */
107         return ARM_MATH_SIZE_MISMATCH;
108     }
109 
110     /* Run the following code for Cortex-M4 and Cortex-M7 */
111 
112     /* This part implements the im2col function */
113     for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++)
114     {
115         for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++)
116         {
117             for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
118             {
119                 for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++)
120                 {
121                     if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in)
122                     {
123                         /* arm_fill_q15(0, pBuffer, ch_im_in); */
124                         memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
125                     } else
126                     {
127                         /* arm_copy_q15((q15_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */
128                         memcpy(pBuffer, (q15_t *) Im_in + (i_ker_y * dim_im_in + i_ker_x) * ch_im_in, sizeof(q15_t)*ch_im_in);
129                     }
130                     pBuffer += ch_im_in;
131                 }
132             }
133 
134             if (i_out_x & 0x1)
135             {
136                 int       i;
137                 /* initialize the matrix pointers for A */
138                 const q15_t *pA = wt;
139 
140                 /* set up the second output pointers */
141                 q15_t    *pOut2 = pOut + ch_im_out;
142 
143                 /* this loop over rows in A */
144                 for (i = 0; i < ch_im_out; i += 2)
145                 {
146                     /* setup pointers for B */
147                     q15_t    *pB = im_buffer;
148                     const q15_t *pB2 = pB + ch_im_in * dim_kernel * dim_kernel;
149 
150                     /* aling the second pointer for A */
151                     const q15_t *pA2 = pA + ch_im_in * dim_kernel * dim_kernel;
152 
153                     /* init the sum with bias */
154                     q31_t     sum =  ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
155                     q31_t     sum2 = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
156                     q31_t     sum3 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift);
157                     q31_t     sum4 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift);
158 
159                     uint16_t  colCnt = ch_im_in * dim_kernel * dim_kernel >> 1;
160                     /* accumulate over the vector */
161                     while (colCnt)
162                     {
163                         q31_t     inA1 = *__SIMD32(pA)++;
164                         q31_t     inB1 = *__SIMD32(pB)++;
165                         q31_t     inA2 = *__SIMD32(pA2)++;
166                         q31_t     inB2 = *__SIMD32(pB2)++;
167 
168                         sum = __SMLAD(inA1, inB1, sum);
169                         sum2 = __SMLAD(inA1, inB2, sum2);
170                         sum3 = __SMLAD(inA2, inB1, sum3);
171                         sum4 = __SMLAD(inA2, inB2, sum4);
172 
173                         colCnt--;
174                     }           /* while over colCnt */
175                     colCnt = ch_im_in * dim_kernel * dim_kernel & 0x1;
176                     while (colCnt)
177                     {
178                         q15_t     inA1 = *pA++;
179                         q15_t     inB1 = *pB++;
180                         q15_t     inA2 = *pA2++;
181                         q15_t     inB2 = *pB2++;
182 
183                         sum += inA1 * inB1;
184                         sum2 += inA1 * inB2;
185                         sum3 += inA2 * inB1;
186                         sum4 += inA2 * inB2;
187                         colCnt--;
188                     }           /* while over colCnt */
189                     *pOut++ = (q15_t) __SSAT(sum >> out_shift, 16);
190                     *pOut++ = (q15_t) __SSAT(sum3 >> out_shift, 16);
191                     *pOut2++ = (q15_t) __SSAT(sum2 >> out_shift, 16);
192                     *pOut2++ = (q15_t) __SSAT(sum4 >> out_shift, 16);
193 
194                     /* skip the row computed with A2 */
195                     pA += ch_im_in * dim_kernel * dim_kernel;
196                 }               /* for over ch_im_out */
197 
198                 pOut += ch_im_out;
199                 /* counter reset */
200                 pBuffer = im_buffer;
201             }
202         }
203     }
204 
205 #else
206     /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
207     uint16_t  i, j, k, l, m, n;
208     int       conv_out;
209     signed char in_row, in_col;
210 
211     if (ch_im_in % 2 != 0 || ch_im_out % 2 != 0)
212     {
213         /* check if the input dimension meets the constraints */
214         return ARM_MATH_SIZE_MISMATCH;
215     }
216 
217     for (i = 0; i < ch_im_out; i++)
218     {
219         for (j = 0; j < dim_im_out; j++)
220         {
221             for (k = 0; k < dim_im_out; k++)
222             {
223                 conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
224                 for (m = 0; m < dim_kernel; m++)
225                 {
226                     for (n = 0; n < dim_kernel; n++)
227                     {
228                         in_row = stride * j + m - padding;
229                         in_col = stride * k + n - padding;
230                         if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in)
231                         {
232                             for (l = 0; l < ch_im_in; l++)
233                             {
234                                 conv_out +=
235                                     Im_in[(in_row * dim_im_in + in_col) * ch_im_in +
236                                           l] * wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel +
237                                                                                             n) * ch_im_in + l];
238                             }
239                         }
240                     }
241                 }
242                 Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q15_t) __SSAT((conv_out >> out_shift), 16);
243             }
244         }
245     }
246 
247 #endif                          /* ARM_MATH_DSP */
248 
249     /* Return to application */
250     return ARM_MATH_SUCCESS;
251 }
252 
253 /**
254  * @} end of NNConv group
255  */
256