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_basic.c
22  * Description:	 Q7 version of convolution
23  *
24  * $Date:        13. July 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 Basic Q7 convolution function (non-sqaure shape)
44    * @param[in]       Im_in        pointer to input tensor
45    * @param[in]       dim_im_in_x  input tensor dimention x
46    * @param[in]       dim_im_in_y  input tensor dimention y
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_x filter kernel size x
51    * @param[in]       dim_kernel_y filter kernel size y
52    * @param[in]       padding_x    padding size x
53    * @param[in]       padding_y    padding size y
54    * @param[in]       stride_x     convolution stride x
55    * @param[in]       stride_y     convolution stride y
56    * @param[in]       bias         pointer to bias
57    * @param[in]       bias_shift   amount of left-shift for bias
58    * @param[in]       out_shift    amount of right-shift for output
59    * @param[in,out]   Im_out       pointer to output tensor
60    * @param[in]       dim_im_out_x output tensor dimension x
61    * @param[in]       dim_im_out_y output tensor dimension y
62    * @param[in,out]   bufferA      pointer to buffer space for input
63    * @param[in,out]   bufferB      pointer to buffer space for output
64    * @return     The function returns <code>ARM_MATH_SUCCESS</code>
65    */
66 
arm_convolve_HWC_q7_basic_nonsquare(const q7_t * Im_in,const uint16_t dim_im_in_x,const uint16_t dim_im_in_y,const uint16_t ch_im_in,const q7_t * wt,const uint16_t ch_im_out,const uint16_t dim_kernel_x,const uint16_t dim_kernel_y,const uint16_t padding_x,const uint16_t padding_y,const uint16_t stride_x,const uint16_t stride_y,const q7_t * bias,const uint16_t bias_shift,const uint16_t out_shift,q7_t * Im_out,const uint16_t dim_im_out_x,const uint16_t dim_im_out_y,q15_t * bufferA,q7_t * bufferB)67 arm_status arm_convolve_HWC_q7_basic_nonsquare(const q7_t * Im_in,
68                                                const uint16_t dim_im_in_x,
69                                                const uint16_t dim_im_in_y,
70                                                const uint16_t ch_im_in,
71                                                const q7_t * wt,
72                                                const uint16_t ch_im_out,
73                                                const uint16_t dim_kernel_x,
74                                                const uint16_t dim_kernel_y,
75                                                const uint16_t padding_x,
76                                                const uint16_t padding_y,
77                                                const uint16_t stride_x,
78                                                const uint16_t stride_y,
79                                                const q7_t * bias,
80                                                const uint16_t bias_shift,
81                                                const uint16_t out_shift,
82                                                q7_t * Im_out,
83                                                const uint16_t dim_im_out_x,
84                                                const uint16_t dim_im_out_y,
85                                                q15_t * bufferA,
86                                                q7_t * bufferB)
87 {
88 
89 #if defined (ARM_MATH_DSP)
90     /* Run the following code for Cortex-M4 and Cortex-M7 */
91 
92     int16_t   i_out_y, i_out_x, i_ker_y, i_ker_x;
93 
94     /*
95      *  Here we use bufferA as q15_t internally as computation are done with q15_t level
96      *  im2col are done to output in q15_t format from q7_t input
97      */
98     q15_t    *pBuffer = bufferA;
99     q7_t     *pOut = Im_out;
100 
101     /* This part implements the im2col function */
102     for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++)
103     {
104         for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++)
105         {
106             for (i_ker_y = i_out_y * stride_y - padding_y; i_ker_y < i_out_y * stride_y - padding_y + dim_kernel_y; i_ker_y++)
107             {
108                 for (i_ker_x = i_out_x * stride_x - padding_x; i_ker_x < i_out_x * stride_x - padding_x + dim_kernel_x; i_ker_x++)
109                 {
110                     if (i_ker_y < 0 || i_ker_y >= dim_im_in_y || i_ker_x < 0 || i_ker_x >= dim_im_in_x)
111                     {
112                         /* Filling 0 for out-of-bound paddings */
113                         /* arm_fill_q15(0, pBuffer, ch_im_in); */
114                         memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
115                     } else
116                     {
117                         /* Copying the pixel data to column */
118                         arm_q7_to_q15_no_shift((q7_t *)
119                                                Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, pBuffer, ch_im_in);
120                     }
121                     pBuffer += ch_im_in;
122                 }
123             }
124 
125             /* Computation is filed for every 2 columns */
126             if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_y * dim_kernel_x)
127             {
128                 pOut =
129                     arm_nn_mat_mult_kernel_q7_q15(wt, bufferA,
130                                                   ch_im_out,
131                                                   ch_im_in *
132                                                   dim_kernel_y * dim_kernel_x, bias_shift, out_shift, bias, pOut);
133 
134                 /* counter reset */
135                 pBuffer = bufferA;
136             }
137         }
138     }
139 
140     /* left-over because odd number of output pixels */
141     if (pBuffer != bufferA)
142     {
143         const q7_t *pA = wt;
144         int       i;
145 
146         for (i = 0; i < ch_im_out; i++)
147         {
148             /* Load the accumulator with bias first */
149             q31_t     sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
150 
151             /* Point to the beging of the im2col buffer */
152             q15_t    *pB = bufferA;
153 
154             /* Each time it process 4 entries */
155             uint16_t  colCnt = ch_im_in * dim_kernel_y * dim_kernel_x >> 2;
156 
157             while (colCnt)
158             {
159                 q31_t     inA1, inA2;
160                 q31_t     inB1, inB2;
161 
162                 pA = (q7_t *) read_and_pad((void *)pA, &inA1, &inA2);
163 
164                 inB1 = *__SIMD32(pB)++;
165                 sum = __SMLAD(inA1, inB1, sum);
166                 inB2 = *__SIMD32(pB)++;
167                 sum = __SMLAD(inA2, inB2, sum);
168 
169                 colCnt--;
170             }
171             colCnt = ch_im_in * dim_kernel_y * dim_kernel_x & 0x3;
172             while (colCnt)
173             {
174                 q7_t      inA1 = *pA++;
175                 q15_t     inB1 = *pB++;
176                 sum += inA1 * inB1;
177                 colCnt--;
178             }
179             *pOut++ = (q7_t) __SSAT((sum >> out_shift), 8);
180         }
181     }
182 #else
183     /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
184 
185     uint16_t  i, j, k, l, m, n;
186     int       conv_out;
187     signed char in_row, in_col;
188 
189     for (i = 0; i < ch_im_out; i++)
190     {
191         for (j = 0; j < dim_im_out_y; j++)
192         {
193             for (k = 0; k < dim_im_out_x; k++)
194             {
195                 conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
196                 for (m = 0; m < dim_kernel_y; m++)
197                 {
198                     for (n = 0; n < dim_kernel_x; n++)
199                     {
200                         // if-for implementation
201                         in_row = stride_y * j + m - padding_y;
202                         in_col = stride_x * k + n - padding_x;
203                         if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x)
204                         {
205                             for (l = 0; l < ch_im_in; l++)
206                             {
207                                 conv_out +=
208                                     Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in + l] *
209                                          wt[i * ch_im_in * dim_kernel_y * dim_kernel_x +
210                                          (m * dim_kernel_x + n) * ch_im_in + l];
211                             }
212                         }
213                     }
214                 }
215                 Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8);
216             }
217         }
218     }
219 
220 #endif                          /* ARM_MATH_DSP */
221 
222     /* Return to application */
223     return ARM_MATH_SUCCESS;
224 }
225 
226 /**
227  * @} end of NNConv group
228  */
229