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_1x1_HWC_q7_fast_nonsquare.c
22  * Description:  Fast Q7 version of 1x1 convolution (non-square shape)
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 Q7 version of 1x1 convolution (non-sqaure shape)
45  * @param[in]       Im_in        pointer to input tensor
46  * @param[in]       dim_im_in_x  input tensor dimention x
47  * @param[in]       dim_im_in_y  input tensor dimention y
48  * @param[in]       ch_im_in     number of input tensor channels
49  * @param[in]       wt           pointer to kernel weights
50  * @param[in]       ch_im_out    number of filters, i.e., output tensor channels
51  * @param[in]       dim_kernel_x filter kernel size x
52  * @param[in]       dim_kernel_y filter kernel size y
53  * @param[in]       padding_x    padding size x
54  * @param[in]       padding_y    padding size y
55  * @param[in]       stride_x     convolution stride x
56  * @param[in]       stride_y     convolution stride y
57  * @param[in]       bias         pointer to bias
58  * @param[in]       bias_shift   amount of left-shift for bias
59  * @param[in]       out_shift    amount of right-shift for output
60  * @param[in,out]   Im_out       pointer to output tensor
61  * @param[in]       dim_im_out_x output tensor dimension x
62  * @param[in]       dim_im_out_y output tensor dimension y
63  * @param[in,out]   bufferA      pointer to buffer space for input
64  * @param[in,out]   bufferB      pointer to buffer space for output
65  * @return     The function returns either
66  * <code>ARM_MATH_SIZE_MISMATCH</code> or <code>ARM_MATH_SUCCESS</code> based on the outcome of size checking.
67  *
68  * This function is optimized for convolution with 1x1 kernel size (i.e., dim_kernel_x=1
69  * and dim_kernel_y=1). It can be used for the second half of MobileNets [1] after depthwise
70  * separable convolution.
71  *
72  * This function is the version with full list of optimization tricks, but with
73  * some contraints:
74  *   ch_im_in is multiple of 4
75  *   ch_im_out is multiple of 2
76  *
77  * [1] MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
78  * https://arxiv.org/abs/1704.04861
79  */
80 
arm_convolve_1x1_HWC_q7_fast_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)81 arm_status arm_convolve_1x1_HWC_q7_fast_nonsquare(const q7_t * Im_in,
82                                                   const uint16_t dim_im_in_x,
83                                                   const uint16_t dim_im_in_y,
84                                                   const uint16_t ch_im_in,
85                                                   const q7_t * wt,
86                                                   const uint16_t ch_im_out,
87                                                   const uint16_t dim_kernel_x,
88                                                   const uint16_t dim_kernel_y,
89                                                   const uint16_t padding_x,
90                                                   const uint16_t padding_y,
91                                                   const uint16_t stride_x,
92                                                   const uint16_t stride_y,
93                                                   const q7_t * bias,
94                                                   const uint16_t bias_shift,
95                                                   const uint16_t out_shift,
96                                                   q7_t * Im_out,
97                                                   const uint16_t dim_im_out_x,
98                                                   const uint16_t dim_im_out_y,
99                                                   q15_t * bufferA,
100                                                   q7_t * bufferB)
101 {
102 
103 #if defined (ARM_MATH_DSP)
104     /* Run the following code for Cortex-M4 and Cortex-M7 */
105 
106     int16_t   i_out_y, i_out_x;
107     int16_t   i_ch_out;
108 
109     /* -----------------------
110      *  Here we use bufferA as q15_t internally as computation are done with q15_t level
111      *  im2col are done to output in q15_t format from q7_t input
112      */
113 
114     q15_t    *pBuffer = bufferA;
115     q7_t     *pOut = Im_out;
116 
117     if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0 || dim_kernel_x != 1 || dim_kernel_y != 1
118         || padding_x != 0 || padding_y != 0 || stride_x != 1 || stride_y != 1)
119     {
120         /* check if the input dimension meets the constraints */
121         return ARM_MATH_SIZE_MISMATCH;
122     }
123 
124     for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++)
125     {
126         for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++)
127         {
128             /* This part implements the im2col function */
129             arm_q7_to_q15_reordered_no_shift((q7_t *) Im_in + (i_out_y * dim_im_in_x + i_out_x) * ch_im_in, pBuffer,
130                                              ch_im_in);
131             pBuffer += ch_im_in;
132 
133             if (pBuffer == bufferA + 2 * ch_im_in * dim_kernel_x * dim_kernel_y)
134             {
135                 pOut =
136                     arm_nn_mat_mult_kernel_q7_q15_reordered(wt, bufferA, ch_im_out, ch_im_in, bias_shift, out_shift, bias, pOut);
137                 /* counter reset */
138                 pBuffer = bufferA;
139             }
140         }
141     }
142 
143     /* check if there is left-over for compute */
144     if (pBuffer != bufferA)
145     {
146         const q7_t *pA = wt;
147         for (i_ch_out = 0; i_ch_out < ch_im_out; i_ch_out++)
148         {
149             q31_t     sum = ((q31_t)(bias[i_ch_out]) << bias_shift) + NN_ROUND(out_shift);
150             q15_t    *pB = bufferA;
151             /* basically each time it process 4 entries */
152             uint16_t  colCnt = ch_im_in * dim_kernel_x * dim_kernel_y >> 2;
153 
154             while (colCnt)
155             {
156 
157                 q31_t     inA1, inA2;
158                 q31_t     inB1, inB2;
159 
160                 pA = (const q7_t *)read_and_pad_reordered((void *)pA, &inA1, &inA2);
161 
162                 inB1 = *__SIMD32(pB)++;
163                 sum = __SMLAD(inA1, inB1, sum);
164                 inB2 = *__SIMD32(pB)++;
165                 sum = __SMLAD(inA2, inB2, sum);
166 
167                 colCnt--;
168             }
169             colCnt = ch_im_in * dim_kernel_y * dim_kernel_x & 0x3;
170             while (colCnt)
171             {
172                 q7_t      inA1 = *pA++;
173                 q15_t     inB1 = *pB++;
174                 sum += inA1 * inB1;
175                 colCnt--;
176             }
177             *pOut = (q7_t) __SSAT((sum >> out_shift), 8);
178             pOut++;
179 
180         }
181 
182     }
183 
184 #else
185     /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
186 
187     int       i, j, k, l, m, n;
188     int       conv_out;
189     int       in_row, in_col;
190 
191     if (ch_im_in % 4 != 0 || ch_im_out % 2 != 0 || dim_kernel_x != 1 || dim_kernel_y != 1
192         || padding_x != 0 || padding_y != 0 || stride_x != 1 || stride_y != 1)
193     {
194         /* check if the input dimension meets the constraints */
195         return ARM_MATH_SIZE_MISMATCH;
196     }
197 
198     for (i = 0; i < ch_im_out; i++)
199     {
200         for (j = 0; j < dim_im_out_y; j++)
201         {
202             for (k = 0; k < dim_im_out_x; k++)
203             {
204                 conv_out = ((q31_t)(bias[i]) << bias_shift) + NN_ROUND(out_shift);
205                 for (m = 0; m < dim_kernel_y; m++)
206                 {
207                     for (n = 0; n < dim_kernel_x; n++)
208                     {
209                         // if-for implementation
210                         in_row = stride_y * j + m - padding_y;
211                         in_col = stride_x * k + n - padding_x;
212                         if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x)
213                         {
214                             for (l = 0; l < ch_im_in; l++)
215                             {
216                                 conv_out += Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in + l] *
217                                     wt[i * ch_im_in * dim_kernel_y * dim_kernel_x + (m * dim_kernel_y + n) * ch_im_in + l];
218                             }
219                         }
220                     }
221                 }
222                 Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q7_t) __SSAT((conv_out >> out_shift), 8);
223             }
224         }
225     }
226 
227 #endif                          /* ARM_MATH_DSP */
228 
229     /* Return to application */
230     return ARM_MATH_SUCCESS;
231 }
232 
233 /**
234  * @} end of NNConv group
235  */
236