File size: 9,400 Bytes
25ade36
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
#include "edge-impulse-sdk/classifier/ei_classifier_config.h"
#if EI_CLASSIFIER_TFLITE_LOAD_CMSIS_NN_SOURCES
/*
 * SPDX-FileCopyrightText: Copyright 2010-2022 Arm Limited and/or its affiliates <open-source-office@arm.com>
 *
 * SPDX-License-Identifier: Apache-2.0
 *
 * Licensed under the Apache License, Version 2.0 (the License); you may
 * not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an AS IS BASIS, WITHOUT
 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

/* ----------------------------------------------------------------------
 * Project:      CMSIS NN Library
 * Title:        arm_convolve_HWC_q7_RGB.c
 * Description:  Q7 version of convolution for RGB image
 *
 * $Date:        4 Aug 2022
 * $Revision:    V.2.0.1
 *
 * Target Processor:  Cortex-M cores
 *
 * -------------------------------------------------------------------- */

#include "edge-impulse-sdk/CMSIS/NN/Include/arm_nnfunctions.h"
#include "edge-impulse-sdk/CMSIS/NN/Include/arm_nnsupportfunctions.h"

/**
 *  @ingroup groupNN
 */

/**
 * @addtogroup NNConv
 * @{
 */

/*
 * Q7 convolution function for RGB image
 * Refer function header for details
 */

arm_cmsis_nn_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)
{
    (void)bufferB;
#if defined(ARM_MATH_DSP) && !defined(ARM_MATH_MVEI)
    /* Run the following code for Cortex-M4 and Cortex-M7 */
    int16_t i_out_y, i_out_x, i_ker_y, i_ker_x;

    /*
     *  Here we use bufferA as q15_t internally as computation are done with q15_t level
     *  im2col are done to output in q15_t format from q7_t input
     */
    q15_t *pBuffer = bufferA;
    q7_t *pOut = Im_out;

    // check if number of input channels is 3
    if (ch_im_in != 3)
    {
        return ARM_CMSIS_NN_ARG_ERROR;
    }
    // This part implements the im2col function
    for (i_out_y = 0; i_out_y < dim_im_out; i_out_y++)
    {
        for (i_out_x = 0; i_out_x < dim_im_out; i_out_x++)
        {
            for (i_ker_y = i_out_y * stride - padding; i_ker_y < i_out_y * stride - padding + dim_kernel; i_ker_y++)
            {
                for (i_ker_x = i_out_x * stride - padding; i_ker_x < i_out_x * stride - padding + dim_kernel; i_ker_x++)
                {
                    if (i_ker_y < 0 || i_ker_y >= dim_im_in || i_ker_x < 0 || i_ker_x >= dim_im_in)
                    {
                        /* Equivalent to arm_fill_q15(0, pBuffer, ch_im_in) with assumption: ch_im_in = 3 */
                        arm_memset_q7((q7_t *)pBuffer, (q7_t)0, 3 * sizeof(q15_t));
                        pBuffer += 3;
                    }
                    else
                    {
                        /*
                         * Equivalent to:
                         *  arm_q7_to_q15_no_shift( (q7_t*)Im_in+(i_ker_y*dim_im_in+i_ker_x)*3, pBuffer, 3);
                         */

                        const q7_t *pPixel = Im_in + (i_ker_y * dim_im_in + i_ker_x) * 3;
                        q31_t buf = arm_nn_read_q7x4(pPixel);

                        union arm_nnword top;
                        union arm_nnword bottom;

                        top.word = __SXTB16(buf);
                        bottom.word = __SXTB16(__ROR(buf, 8));

#ifndef ARM_MATH_BIG_ENDIAN
                        /*
                         *  little-endian, | omit | 3rd  | 2nd  | 1st  |
                         *                MSB                         LSB
                         *   top | 3rd | 1st |; bottom | omit | 2nd |
                         *
                         *  version 1, need to swap 2nd and 3rd weight
                         * *__SIMD32(pBuffer) = top.word;
                         * *(pBuffer+2) = bottom.half_words[0];
                         *
                         *  version 2, no weight shuffling required
                         */
                        *pBuffer++ = top.half_words[0];
                        int32_t packed_word = __PKHBT(bottom.word, top.word, 0);
                        arm_memcpy_q7((q7_t *)pBuffer, (q7_t *)&packed_word, 4);
#else
                        /*
                         *  big-endian,    | 1st  | 2nd  | 3rd  | omit |
                         *                MSB                         LSB
                         *  top | 2nd | omit |; bottom | 1st | 3rd |
                         *
                         *  version 1, need to swap 2nd and 3rd weight
                         * *__SIMD32(pBuffer) = bottom.word;
                         * *(pBuffer+2) = top.half_words[1];
                         *
                         *  version 2, no weight shuffling required
                         */
                        *pBuffer++ = bottom.half_words[0];
                        int32_t packed_word = __PKHTB(top.word, bottom.word, 0);
                        arm_memcpy_q7((q7_t *)pBuffer, (q7_t *)&packed_word, 4);
#endif
                        pBuffer += 2;
                    }
                }
            }

            if (pBuffer == bufferA + 2 * 3 * dim_kernel * dim_kernel)
            {
                pOut = arm_nn_mat_mult_kernel_q7_q15(
                    wt, bufferA, ch_im_out, 3 * dim_kernel * dim_kernel, bias_shift, out_shift, bias, pOut);

                /* counter reset */
                pBuffer = bufferA;
            }
        }
    }

    /* left-over because odd number of output pixels */
    if (pBuffer != bufferA)
    {
        const q7_t *pA = wt;
        int i;

        for (i = 0; i < ch_im_out; i++)
        {
            q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
            q15_t *pB = bufferA;
            /* basically each time it process 4 entries */
            uint16_t colCnt = 3 * dim_kernel * dim_kernel >> 2;

            while (colCnt)
            {

                q31_t inA1, inA2;
                q31_t inB1, inB2;

                pA = read_and_pad(pA, &inA1, &inA2);

                inB1 = arm_nn_read_q15x2_ia((const q15_t **)&pB);
                sum = __SMLAD(inA1, inB1, sum);
                inB2 = arm_nn_read_q15x2_ia((const q15_t **)&pB);
                sum = __SMLAD(inA2, inB2, sum);

                colCnt--;
            }
            colCnt = 3 * dim_kernel * dim_kernel & 0x3;
            while (colCnt)
            {
                q7_t inA1 = *pA++;
                q15_t inB1 = *pB++;
                sum += inA1 * inB1;
                colCnt--;
            }
            *pOut++ = (q7_t)__SSAT((sum >> out_shift), 8);
        }
    }
#else
    (void)bufferA;
    /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
    int i, j, k, l, m, n;
    int conv_out;
    int in_row, in_col;

    // check if number of input channels is 3
    if (ch_im_in != 3)
    {
        return ARM_CMSIS_NN_ARG_ERROR;
    }

    for (i = 0; i < ch_im_out; i++)
    {
        for (j = 0; j < dim_im_out; j++)
        {
            for (k = 0; k < dim_im_out; k++)
            {
                conv_out = (bias[i] << bias_shift) + NN_ROUND(out_shift);
                for (m = 0; m < dim_kernel; m++)
                {
                    for (n = 0; n < dim_kernel; n++)
                    {
                        /* if-for implementation */
                        in_row = stride * j + m - padding;
                        in_col = stride * k + n - padding;
                        if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in && in_col < dim_im_in)
                        {
                            for (l = 0; l < ch_im_in; l++)
                            {
                                conv_out += Im_in[(in_row * dim_im_in + in_col) * ch_im_in + l] *
                                    wt[i * ch_im_in * dim_kernel * dim_kernel + (m * dim_kernel + n) * ch_im_in + l];
                            }
                        }
                    }
                }
                Im_out[i + (j * dim_im_out + k) * ch_im_out] = (q7_t)__SSAT((conv_out >> out_shift), 8);
            }
        }
    }

#endif /* ARM_MATH_DSP */

    /* Return to application */
    return (ARM_CMSIS_NN_SUCCESS);
}

/**
 * @} end of NNConv group
 */

#endif // EI_CLASSIFIER_TFLITE_LOAD_CMSIS_NN_SOURCES