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_q7_RGB.c
22  * Description:  Q7 version of convolution for RGB image
23  *
24  * $Date:        17. January 2018
25  * $Revision:    V.1.0.0
26  *
27  * Target Processor:  Cortex-M cores
28  *
29  * -------------------------------------------------------------------- */
30 #include "arm_math.h"
31 #include "arm_nnfunctions.h"
32 
33 /**
34  *  @ingroup groupNN
35  */
36 
37 /**
38  * @addtogroup NNConv
39  * @{
40  */
41 
42   /**
43    * @brief Q7 convolution function for RGB image
44    * @param[in]       Im_in       pointer to input tensor
45    * @param[in]       dim_im_in   input tensor dimention
46    * @param[in]       ch_im_in    number of input tensor channels
47    * @param[in]       wt          pointer to kernel weights
48    * @param[in]       ch_im_out   number of filters, i.e., output tensor channels
49    * @param[in]       dim_kernel  filter kernel size
50    * @param[in]       padding     padding sizes
51    * @param[in]       stride      convolution stride
52    * @param[in]       bias        pointer to bias
53    * @param[in]       bias_shift  amount of left-shift for bias
54    * @param[in]       out_shift   amount of right-shift for output
55    * @param[in,out]   Im_out      pointer to output tensor
56    * @param[in]       dim_im_out  output tensor dimension
57    * @param[in,out]   bufferA     pointer to buffer space for input
58    * @param[in,out]   bufferB     pointer to buffer space for output
59    * @return     The function returns either
60    * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
61    *
62    * @details
63    *
64    * <b>Buffer size:</b>
65    *
66    * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel
67    *
68    * bufferB size: 0
69    *
70    * <b>Input dimension constraints:</b>
71    *
72    * ch_im_in equals 3
73    *
74    * This kernel is written exclusively for convolution with ch_im_in
75    * equals 3. This applies on the first layer of CNNs which has input
76    * image with RGB format.
77    */
78 
79 arm_status
arm_convolve_HWC_q7_RGB(const q7_t * Im_in,const uint16_t dim_im_in,const uint16_t ch_im_in,const q7_t * wt,const uint16_t ch_im_out,const uint16_t dim_kernel,const uint16_t padding,const uint16_t stride,const q7_t * bias,const uint16_t bias_shift,const uint16_t out_shift,q7_t * Im_out,const uint16_t dim_im_out,q15_t * bufferA,q7_t * bufferB)80 arm_convolve_HWC_q7_RGB(const q7_t * Im_in,
81                         const uint16_t dim_im_in,
82                         const uint16_t ch_im_in,
83                         const q7_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 q7_t * bias,
89                         const uint16_t bias_shift,
90                         const uint16_t out_shift,
91                         q7_t * Im_out, const uint16_t dim_im_out, q15_t * bufferA, q7_t * bufferB)
92 {
93 
94 #if defined (ARM_MATH_DSP)
95     /* Run the following code for Cortex-M4 and Cortex-M7 */
96     int16_t   i_out_y, i_out_x, i_ker_y, i_ker_x;
97 
98     /*
99      *  Here we use bufferA as q15_t internally as computation are done with q15_t level
100      *  im2col are done to output in q15_t format from q7_t input
101      */
102     q15_t    *pBuffer = bufferA;
103     q7_t     *pOut = Im_out;
104 
105     // check if number of input channels is 3
106     if (ch_im_in != 3)
107     {
108         return ARM_MATH_SIZE_MISMATCH;
109     }
110     // This part implements the im2col function
111     for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++)
112     {
113         for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++)
114         {
115             for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
116             {
117                 for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++)
118                 {
119                     if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in)
120                     {
121                         /* Equivalent to arm_fill_q15(0, pBuffer, ch_im_in) with assumption: ch_im_in = 3 */
122                         *__SIMD32(pBuffer) = 0x0;
123                         *(pBuffer + 2) = 0;
124                         pBuffer += 3;
125                     } else
126                     {
127                         /*
128                          * Equivalent to:
129                          *  arm_q7_to_q15_no_shift( (q7_t*)Im_in+(i_ker_y*dim_im_in+i_ker_x)*3, pBuffer, 3);
130                          */
131 
132                         const q7_t *pPixel = Im_in + (i_ker_y * dim_im_in + i_ker_x) * 3;
133                         q31_t     buf = *__SIMD32(pPixel);
134 
135                         union arm_nnword top;
136                         union arm_nnword bottom;
137 
138                         top.word = __SXTB16(buf);
139                         bottom.word = __SXTB16(__ROR(buf, 8));
140 
141 #ifndef ARM_MATH_BIG_ENDIAN
142                         /*
143                          *  little-endian, | omit | 3rd  | 2nd  | 1st  |
144                          *                MSB                         LSB
145                          *   top | 3rd | 1st |; bottom | omit | 2nd |
146                          *
147                          *  version 1, need to swap 2nd and 3rd weight
148                          * *__SIMD32(pBuffer) = top.word;
149                          * *(pBuffer+2) = bottom.half_words[0];
150                          *
151                          *  version 2, no weight shuffling required
152                          */
153                         *pBuffer++ = top.half_words[0];
154                         *__SIMD32(pBuffer) = __PKHBT(bottom.word, top.word, 0);
155 #else
156                         /*
157                          *  big-endian,    | 1st  | 2nd  | 3rd  | omit |
158                          *                MSB                         LSB
159                          *  top | 2nd | omit |; bottom | 1st | 3rd |
160                          *
161                          *  version 1, need to swap 2nd and 3rd weight
162                          * *__SIMD32(pBuffer) = bottom.word;
163                          * *(pBuffer+2) = top.half_words[1];
164                          *
165                          *  version 2, no weight shuffling required
166                          */
167                         *pBuffer++ = bottom.half_words[0];
168                         *__SIMD32(pBuffer) = __PKHTB(top.word, bottom.word, 0);
169 #endif
170                         pBuffer += 2;
171                     }
172                 }
173             }
174 
175             if (pBuffer == bufferA + 2 * 3 * dim_kernel * dim_kernel)
176             {
177                 pOut =
178                     arm_nn_mat_mult_kernel_q7_q15(wt, bufferA,
179                                                   ch_im_out,
180                                                   3 * dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut);
181 
182                 /* counter reset */
183                 pBuffer = bufferA;
184             }
185         }
186     }
187 
188     /* left-over because odd number of output pixels */
189     if (pBuffer != bufferA)
190     {
191         const q7_t *pA = wt;
192         int       i;
193 
194         for (i = 0; i < ch_im_out; i++)
195         {
196             q31_t     sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
197             q15_t    *pB = bufferA;
198             /* basically each time it process 4 entries */
199             uint16_t  colCnt = 3 * dim_kernel * dim_kernel >> 2;
200 
201             while (colCnt)
202             {
203 
204                 q31_t     inA1, inA2;
205                 q31_t     inB1, inB2;
206 
207                 pA = (q7_t *) read_and_pad((void *)pA, &inA1, &inA2);
208 
209                 inB1 = *__SIMD32(pB)++;
210                 sum = __SMLAD(inA1, inB1, sum);
211                 inB2 = *__SIMD32(pB)++;
212                 sum = __SMLAD(inA2, inB2, sum);
213 
214                 colCnt--;
215             }
216             colCnt = 3 * dim_kernel * dim_kernel & 0x3;
217             while (colCnt)
218             {
219                 q7_t      inA1 = *pA++;
220                 q15_t     inB1 = *pB++;
221                 sum += inA1 * inB1;
222                 colCnt--;
223             }
224             *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8);
225         }
226     }
227 #else
228     /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
229 
230     uint16_t  i, j, k, l, m, n;
231     int       conv_out;
232     signed char in_row, in_col;
233 
234     // check if number of input channels is 3
235     if (ch_im_in != 3)
236     {
237         return ARM_MATH_SIZE_MISMATCH;
238     }
239 
240     for (i = 0; i < ch_im_out; i++)
241     {
242         for (j = 0; j < dim_im_out; j++)
243         {
244             for (k = 0; k < dim_im_out; k++)
245             {
246                 conv_out = (bias[i] << bias_shift) + NN_ROUND(out_shift);
247                 for (m = 0; m < dim_kernel; m++)
248                 {
249                     for (n = 0; n < dim_kernel; n++)
250                     {
251                         /* if-for implementation */
252                         in_row = stride * j + m - padding;
253                         in_col = stride * k + n - padding;
254                         if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in)
255                         {
256                             for (l = 0; l < ch_im_in; l++)
257                             {
258                                 conv_out +=
259                                     Im_in[(in_row * dim_im_in + in_col) * ch_im_in +
260                                           l] * wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel +
261                                                                                             n) * ch_im_in + l];
262                             }
263                         }
264                     }
265                 }
266                 Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8);
267             }
268         }
269     }
270 
271 #endif                          /* ARM_MATH_DSP */
272 
273     /* Return to application */
274     return (ARM_MATH_SUCCESS);
275 }
276 
277 /**
278  * @} end of NNConv group
279  */
280