提交 | 用户 | age
|
bfc108
|
1 |
/* ----------------------------------------------------------------------
|
Q |
2 |
* Copyright (C) 2010-2014 ARM Limited. All rights reserved.
|
|
3 |
*
|
|
4 |
* $Date: 19. March 2015
|
|
5 |
* $Revision: V.1.4.5
|
|
6 |
*
|
|
7 |
* Project: CMSIS DSP Library
|
|
8 |
* Title: arm_conv_q7.c
|
|
9 |
*
|
|
10 |
* Description: Convolution of Q7 sequences.
|
|
11 |
*
|
|
12 |
* Target Processor: Cortex-M4/Cortex-M3/Cortex-M0
|
|
13 |
*
|
|
14 |
* Redistribution and use in source and binary forms, with or without
|
|
15 |
* modification, are permitted provided that the following conditions
|
|
16 |
* are met:
|
|
17 |
* - Redistributions of source code must retain the above copyright
|
|
18 |
* notice, this list of conditions and the following disclaimer.
|
|
19 |
* - Redistributions in binary form must reproduce the above copyright
|
|
20 |
* notice, this list of conditions and the following disclaimer in
|
|
21 |
* the documentation and/or other materials provided with the
|
|
22 |
* distribution.
|
|
23 |
* - Neither the name of ARM LIMITED nor the names of its contributors
|
|
24 |
* may be used to endorse or promote products derived from this
|
|
25 |
* software without specific prior written permission.
|
|
26 |
*
|
|
27 |
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
|
|
28 |
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
|
|
29 |
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
|
|
30 |
* FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
|
|
31 |
* COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
|
|
32 |
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
|
|
33 |
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
|
|
34 |
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
|
35 |
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
|
|
36 |
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
|
|
37 |
* ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
|
|
38 |
* POSSIBILITY OF SUCH DAMAGE.
|
|
39 |
* -------------------------------------------------------------------- */
|
|
40 |
|
|
41 |
#include "arm_math.h"
|
|
42 |
|
|
43 |
/**
|
|
44 |
* @ingroup groupFilters
|
|
45 |
*/
|
|
46 |
|
|
47 |
/**
|
|
48 |
* @addtogroup Conv
|
|
49 |
* @{
|
|
50 |
*/
|
|
51 |
|
|
52 |
/**
|
|
53 |
* @brief Convolution of Q7 sequences.
|
|
54 |
* @param[in] *pSrcA points to the first input sequence.
|
|
55 |
* @param[in] srcALen length of the first input sequence.
|
|
56 |
* @param[in] *pSrcB points to the second input sequence.
|
|
57 |
* @param[in] srcBLen length of the second input sequence.
|
|
58 |
* @param[out] *pDst points to the location where the output result is written. Length srcALen+srcBLen-1.
|
|
59 |
* @return none.
|
|
60 |
*
|
|
61 |
* @details
|
|
62 |
* <b>Scaling and Overflow Behavior:</b>
|
|
63 |
*
|
|
64 |
* \par
|
|
65 |
* The function is implemented using a 32-bit internal accumulator.
|
|
66 |
* Both the inputs are represented in 1.7 format and multiplications yield a 2.14 result.
|
|
67 |
* The 2.14 intermediate results are accumulated in a 32-bit accumulator in 18.14 format.
|
|
68 |
* This approach provides 17 guard bits and there is no risk of overflow as long as <code>max(srcALen, srcBLen)<131072</code>.
|
|
69 |
* The 18.14 result is then truncated to 18.7 format by discarding the low 7 bits and then saturated to 1.7 format.
|
|
70 |
*
|
|
71 |
* \par
|
|
72 |
* Refer the function <code>arm_conv_opt_q7()</code> for a faster implementation of this function.
|
|
73 |
*
|
|
74 |
*/
|
|
75 |
|
|
76 |
void arm_conv_q7(
|
|
77 |
q7_t * pSrcA,
|
|
78 |
uint32_t srcALen,
|
|
79 |
q7_t * pSrcB,
|
|
80 |
uint32_t srcBLen,
|
|
81 |
q7_t * pDst)
|
|
82 |
{
|
|
83 |
|
|
84 |
|
|
85 |
#ifndef ARM_MATH_CM0_FAMILY
|
|
86 |
|
|
87 |
/* Run the below code for Cortex-M4 and Cortex-M3 */
|
|
88 |
|
|
89 |
q7_t *pIn1; /* inputA pointer */
|
|
90 |
q7_t *pIn2; /* inputB pointer */
|
|
91 |
q7_t *pOut = pDst; /* output pointer */
|
|
92 |
q7_t *px; /* Intermediate inputA pointer */
|
|
93 |
q7_t *py; /* Intermediate inputB pointer */
|
|
94 |
q7_t *pSrc1, *pSrc2; /* Intermediate pointers */
|
|
95 |
q7_t x0, x1, x2, x3, c0, c1; /* Temporary variables to hold state and coefficient values */
|
|
96 |
q31_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
|
|
97 |
q31_t input1, input2; /* Temporary input variables */
|
|
98 |
q15_t in1, in2; /* Temporary input variables */
|
|
99 |
uint32_t j, k, count, blkCnt, blockSize1, blockSize2, blockSize3; /* loop counter */
|
|
100 |
|
|
101 |
/* The algorithm implementation is based on the lengths of the inputs. */
|
|
102 |
/* srcB is always made to slide across srcA. */
|
|
103 |
/* So srcBLen is always considered as shorter or equal to srcALen */
|
|
104 |
if(srcALen >= srcBLen)
|
|
105 |
{
|
|
106 |
/* Initialization of inputA pointer */
|
|
107 |
pIn1 = pSrcA;
|
|
108 |
|
|
109 |
/* Initialization of inputB pointer */
|
|
110 |
pIn2 = pSrcB;
|
|
111 |
}
|
|
112 |
else
|
|
113 |
{
|
|
114 |
/* Initialization of inputA pointer */
|
|
115 |
pIn1 = pSrcB;
|
|
116 |
|
|
117 |
/* Initialization of inputB pointer */
|
|
118 |
pIn2 = pSrcA;
|
|
119 |
|
|
120 |
/* srcBLen is always considered as shorter or equal to srcALen */
|
|
121 |
j = srcBLen;
|
|
122 |
srcBLen = srcALen;
|
|
123 |
srcALen = j;
|
|
124 |
}
|
|
125 |
|
|
126 |
/* conv(x,y) at n = x[n] * y[0] + x[n-1] * y[1] + x[n-2] * y[2] + ...+ x[n-N+1] * y[N -1] */
|
|
127 |
/* The function is internally
|
|
128 |
* divided into three stages according to the number of multiplications that has to be
|
|
129 |
* taken place between inputA samples and inputB samples. In the first stage of the
|
|
130 |
* algorithm, the multiplications increase by one for every iteration.
|
|
131 |
* In the second stage of the algorithm, srcBLen number of multiplications are done.
|
|
132 |
* In the third stage of the algorithm, the multiplications decrease by one
|
|
133 |
* for every iteration. */
|
|
134 |
|
|
135 |
/* The algorithm is implemented in three stages.
|
|
136 |
The loop counters of each stage is initiated here. */
|
|
137 |
blockSize1 = srcBLen - 1u;
|
|
138 |
blockSize2 = (srcALen - srcBLen) + 1u;
|
|
139 |
blockSize3 = blockSize1;
|
|
140 |
|
|
141 |
/* --------------------------
|
|
142 |
* Initializations of stage1
|
|
143 |
* -------------------------*/
|
|
144 |
|
|
145 |
/* sum = x[0] * y[0]
|
|
146 |
* sum = x[0] * y[1] + x[1] * y[0]
|
|
147 |
* ....
|
|
148 |
* sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
|
|
149 |
*/
|
|
150 |
|
|
151 |
/* In this stage the MAC operations are increased by 1 for every iteration.
|
|
152 |
The count variable holds the number of MAC operations performed */
|
|
153 |
count = 1u;
|
|
154 |
|
|
155 |
/* Working pointer of inputA */
|
|
156 |
px = pIn1;
|
|
157 |
|
|
158 |
/* Working pointer of inputB */
|
|
159 |
py = pIn2;
|
|
160 |
|
|
161 |
|
|
162 |
/* ------------------------
|
|
163 |
* Stage1 process
|
|
164 |
* ----------------------*/
|
|
165 |
|
|
166 |
/* The first stage starts here */
|
|
167 |
while(blockSize1 > 0u)
|
|
168 |
{
|
|
169 |
/* Accumulator is made zero for every iteration */
|
|
170 |
sum = 0;
|
|
171 |
|
|
172 |
/* Apply loop unrolling and compute 4 MACs simultaneously. */
|
|
173 |
k = count >> 2u;
|
|
174 |
|
|
175 |
/* First part of the processing with loop unrolling. Compute 4 MACs at a time.
|
|
176 |
** a second loop below computes MACs for the remaining 1 to 3 samples. */
|
|
177 |
while(k > 0u)
|
|
178 |
{
|
|
179 |
/* x[0] , x[1] */
|
|
180 |
in1 = (q15_t) * px++;
|
|
181 |
in2 = (q15_t) * px++;
|
|
182 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
183 |
|
|
184 |
/* y[srcBLen - 1] , y[srcBLen - 2] */
|
|
185 |
in1 = (q15_t) * py--;
|
|
186 |
in2 = (q15_t) * py--;
|
|
187 |
input2 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
188 |
|
|
189 |
/* x[0] * y[srcBLen - 1] */
|
|
190 |
/* x[1] * y[srcBLen - 2] */
|
|
191 |
sum = __SMLAD(input1, input2, sum);
|
|
192 |
|
|
193 |
/* x[2] , x[3] */
|
|
194 |
in1 = (q15_t) * px++;
|
|
195 |
in2 = (q15_t) * px++;
|
|
196 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
197 |
|
|
198 |
/* y[srcBLen - 3] , y[srcBLen - 4] */
|
|
199 |
in1 = (q15_t) * py--;
|
|
200 |
in2 = (q15_t) * py--;
|
|
201 |
input2 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
202 |
|
|
203 |
/* x[2] * y[srcBLen - 3] */
|
|
204 |
/* x[3] * y[srcBLen - 4] */
|
|
205 |
sum = __SMLAD(input1, input2, sum);
|
|
206 |
|
|
207 |
/* Decrement the loop counter */
|
|
208 |
k--;
|
|
209 |
}
|
|
210 |
|
|
211 |
/* If the count is not a multiple of 4, compute any remaining MACs here.
|
|
212 |
** No loop unrolling is used. */
|
|
213 |
k = count % 0x4u;
|
|
214 |
|
|
215 |
while(k > 0u)
|
|
216 |
{
|
|
217 |
/* Perform the multiply-accumulates */
|
|
218 |
sum += ((q15_t) * px++ * *py--);
|
|
219 |
|
|
220 |
/* Decrement the loop counter */
|
|
221 |
k--;
|
|
222 |
}
|
|
223 |
|
|
224 |
/* Store the result in the accumulator in the destination buffer. */
|
|
225 |
*pOut++ = (q7_t) (__SSAT(sum >> 7u, 8));
|
|
226 |
|
|
227 |
/* Update the inputA and inputB pointers for next MAC calculation */
|
|
228 |
py = pIn2 + count;
|
|
229 |
px = pIn1;
|
|
230 |
|
|
231 |
/* Increment the MAC count */
|
|
232 |
count++;
|
|
233 |
|
|
234 |
/* Decrement the loop counter */
|
|
235 |
blockSize1--;
|
|
236 |
}
|
|
237 |
|
|
238 |
/* --------------------------
|
|
239 |
* Initializations of stage2
|
|
240 |
* ------------------------*/
|
|
241 |
|
|
242 |
/* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
|
|
243 |
* sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
|
|
244 |
* ....
|
|
245 |
* sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
|
|
246 |
*/
|
|
247 |
|
|
248 |
/* Working pointer of inputA */
|
|
249 |
px = pIn1;
|
|
250 |
|
|
251 |
/* Working pointer of inputB */
|
|
252 |
pSrc2 = pIn2 + (srcBLen - 1u);
|
|
253 |
py = pSrc2;
|
|
254 |
|
|
255 |
/* count is index by which the pointer pIn1 to be incremented */
|
|
256 |
count = 0u;
|
|
257 |
|
|
258 |
/* -------------------
|
|
259 |
* Stage2 process
|
|
260 |
* ------------------*/
|
|
261 |
|
|
262 |
/* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
|
|
263 |
* So, to loop unroll over blockSize2,
|
|
264 |
* srcBLen should be greater than or equal to 4 */
|
|
265 |
if(srcBLen >= 4u)
|
|
266 |
{
|
|
267 |
/* Loop unroll over blockSize2, by 4 */
|
|
268 |
blkCnt = blockSize2 >> 2u;
|
|
269 |
|
|
270 |
while(blkCnt > 0u)
|
|
271 |
{
|
|
272 |
/* Set all accumulators to zero */
|
|
273 |
acc0 = 0;
|
|
274 |
acc1 = 0;
|
|
275 |
acc2 = 0;
|
|
276 |
acc3 = 0;
|
|
277 |
|
|
278 |
/* read x[0], x[1], x[2] samples */
|
|
279 |
x0 = *(px++);
|
|
280 |
x1 = *(px++);
|
|
281 |
x2 = *(px++);
|
|
282 |
|
|
283 |
/* Apply loop unrolling and compute 4 MACs simultaneously. */
|
|
284 |
k = srcBLen >> 2u;
|
|
285 |
|
|
286 |
/* First part of the processing with loop unrolling. Compute 4 MACs at a time.
|
|
287 |
** a second loop below computes MACs for the remaining 1 to 3 samples. */
|
|
288 |
do
|
|
289 |
{
|
|
290 |
/* Read y[srcBLen - 1] sample */
|
|
291 |
c0 = *(py--);
|
|
292 |
/* Read y[srcBLen - 2] sample */
|
|
293 |
c1 = *(py--);
|
|
294 |
|
|
295 |
/* Read x[3] sample */
|
|
296 |
x3 = *(px++);
|
|
297 |
|
|
298 |
/* x[0] and x[1] are packed */
|
|
299 |
in1 = (q15_t) x0;
|
|
300 |
in2 = (q15_t) x1;
|
|
301 |
|
|
302 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
303 |
|
|
304 |
/* y[srcBLen - 1] and y[srcBLen - 2] are packed */
|
|
305 |
in1 = (q15_t) c0;
|
|
306 |
in2 = (q15_t) c1;
|
|
307 |
|
|
308 |
input2 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
309 |
|
|
310 |
/* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
|
|
311 |
acc0 = __SMLAD(input1, input2, acc0);
|
|
312 |
|
|
313 |
/* x[1] and x[2] are packed */
|
|
314 |
in1 = (q15_t) x1;
|
|
315 |
in2 = (q15_t) x2;
|
|
316 |
|
|
317 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
318 |
|
|
319 |
/* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
|
|
320 |
acc1 = __SMLAD(input1, input2, acc1);
|
|
321 |
|
|
322 |
/* x[2] and x[3] are packed */
|
|
323 |
in1 = (q15_t) x2;
|
|
324 |
in2 = (q15_t) x3;
|
|
325 |
|
|
326 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
327 |
|
|
328 |
/* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
|
|
329 |
acc2 = __SMLAD(input1, input2, acc2);
|
|
330 |
|
|
331 |
/* Read x[4] sample */
|
|
332 |
x0 = *(px++);
|
|
333 |
|
|
334 |
/* x[3] and x[4] are packed */
|
|
335 |
in1 = (q15_t) x3;
|
|
336 |
in2 = (q15_t) x0;
|
|
337 |
|
|
338 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
339 |
|
|
340 |
/* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
|
|
341 |
acc3 = __SMLAD(input1, input2, acc3);
|
|
342 |
|
|
343 |
/* Read y[srcBLen - 3] sample */
|
|
344 |
c0 = *(py--);
|
|
345 |
/* Read y[srcBLen - 4] sample */
|
|
346 |
c1 = *(py--);
|
|
347 |
|
|
348 |
/* Read x[5] sample */
|
|
349 |
x1 = *(px++);
|
|
350 |
|
|
351 |
/* x[2] and x[3] are packed */
|
|
352 |
in1 = (q15_t) x2;
|
|
353 |
in2 = (q15_t) x3;
|
|
354 |
|
|
355 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
356 |
|
|
357 |
/* y[srcBLen - 3] and y[srcBLen - 4] are packed */
|
|
358 |
in1 = (q15_t) c0;
|
|
359 |
in2 = (q15_t) c1;
|
|
360 |
|
|
361 |
input2 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
362 |
|
|
363 |
/* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
|
|
364 |
acc0 = __SMLAD(input1, input2, acc0);
|
|
365 |
|
|
366 |
/* x[3] and x[4] are packed */
|
|
367 |
in1 = (q15_t) x3;
|
|
368 |
in2 = (q15_t) x0;
|
|
369 |
|
|
370 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
371 |
|
|
372 |
/* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
|
|
373 |
acc1 = __SMLAD(input1, input2, acc1);
|
|
374 |
|
|
375 |
/* x[4] and x[5] are packed */
|
|
376 |
in1 = (q15_t) x0;
|
|
377 |
in2 = (q15_t) x1;
|
|
378 |
|
|
379 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
380 |
|
|
381 |
/* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
|
|
382 |
acc2 = __SMLAD(input1, input2, acc2);
|
|
383 |
|
|
384 |
/* Read x[6] sample */
|
|
385 |
x2 = *(px++);
|
|
386 |
|
|
387 |
/* x[5] and x[6] are packed */
|
|
388 |
in1 = (q15_t) x1;
|
|
389 |
in2 = (q15_t) x2;
|
|
390 |
|
|
391 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
392 |
|
|
393 |
/* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
|
|
394 |
acc3 = __SMLAD(input1, input2, acc3);
|
|
395 |
|
|
396 |
} while(--k);
|
|
397 |
|
|
398 |
/* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
|
|
399 |
** No loop unrolling is used. */
|
|
400 |
k = srcBLen % 0x4u;
|
|
401 |
|
|
402 |
while(k > 0u)
|
|
403 |
{
|
|
404 |
/* Read y[srcBLen - 5] sample */
|
|
405 |
c0 = *(py--);
|
|
406 |
|
|
407 |
/* Read x[7] sample */
|
|
408 |
x3 = *(px++);
|
|
409 |
|
|
410 |
/* Perform the multiply-accumulates */
|
|
411 |
/* acc0 += x[4] * y[srcBLen - 5] */
|
|
412 |
acc0 += ((q15_t) x0 * c0);
|
|
413 |
/* acc1 += x[5] * y[srcBLen - 5] */
|
|
414 |
acc1 += ((q15_t) x1 * c0);
|
|
415 |
/* acc2 += x[6] * y[srcBLen - 5] */
|
|
416 |
acc2 += ((q15_t) x2 * c0);
|
|
417 |
/* acc3 += x[7] * y[srcBLen - 5] */
|
|
418 |
acc3 += ((q15_t) x3 * c0);
|
|
419 |
|
|
420 |
/* Reuse the present samples for the next MAC */
|
|
421 |
x0 = x1;
|
|
422 |
x1 = x2;
|
|
423 |
x2 = x3;
|
|
424 |
|
|
425 |
/* Decrement the loop counter */
|
|
426 |
k--;
|
|
427 |
}
|
|
428 |
|
|
429 |
|
|
430 |
/* Store the result in the accumulator in the destination buffer. */
|
|
431 |
*pOut++ = (q7_t) (__SSAT(acc0 >> 7u, 8));
|
|
432 |
*pOut++ = (q7_t) (__SSAT(acc1 >> 7u, 8));
|
|
433 |
*pOut++ = (q7_t) (__SSAT(acc2 >> 7u, 8));
|
|
434 |
*pOut++ = (q7_t) (__SSAT(acc3 >> 7u, 8));
|
|
435 |
|
|
436 |
/* Increment the pointer pIn1 index, count by 4 */
|
|
437 |
count += 4u;
|
|
438 |
|
|
439 |
/* Update the inputA and inputB pointers for next MAC calculation */
|
|
440 |
px = pIn1 + count;
|
|
441 |
py = pSrc2;
|
|
442 |
|
|
443 |
/* Decrement the loop counter */
|
|
444 |
blkCnt--;
|
|
445 |
}
|
|
446 |
|
|
447 |
/* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
|
|
448 |
** No loop unrolling is used. */
|
|
449 |
blkCnt = blockSize2 % 0x4u;
|
|
450 |
|
|
451 |
while(blkCnt > 0u)
|
|
452 |
{
|
|
453 |
/* Accumulator is made zero for every iteration */
|
|
454 |
sum = 0;
|
|
455 |
|
|
456 |
/* Apply loop unrolling and compute 4 MACs simultaneously. */
|
|
457 |
k = srcBLen >> 2u;
|
|
458 |
|
|
459 |
/* First part of the processing with loop unrolling. Compute 4 MACs at a time.
|
|
460 |
** a second loop below computes MACs for the remaining 1 to 3 samples. */
|
|
461 |
while(k > 0u)
|
|
462 |
{
|
|
463 |
|
|
464 |
/* Reading two inputs of SrcA buffer and packing */
|
|
465 |
in1 = (q15_t) * px++;
|
|
466 |
in2 = (q15_t) * px++;
|
|
467 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
468 |
|
|
469 |
/* Reading two inputs of SrcB buffer and packing */
|
|
470 |
in1 = (q15_t) * py--;
|
|
471 |
in2 = (q15_t) * py--;
|
|
472 |
input2 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
473 |
|
|
474 |
/* Perform the multiply-accumulates */
|
|
475 |
sum = __SMLAD(input1, input2, sum);
|
|
476 |
|
|
477 |
/* Reading two inputs of SrcA buffer and packing */
|
|
478 |
in1 = (q15_t) * px++;
|
|
479 |
in2 = (q15_t) * px++;
|
|
480 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
481 |
|
|
482 |
/* Reading two inputs of SrcB buffer and packing */
|
|
483 |
in1 = (q15_t) * py--;
|
|
484 |
in2 = (q15_t) * py--;
|
|
485 |
input2 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
486 |
|
|
487 |
/* Perform the multiply-accumulates */
|
|
488 |
sum = __SMLAD(input1, input2, sum);
|
|
489 |
|
|
490 |
/* Decrement the loop counter */
|
|
491 |
k--;
|
|
492 |
}
|
|
493 |
|
|
494 |
/* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
|
|
495 |
** No loop unrolling is used. */
|
|
496 |
k = srcBLen % 0x4u;
|
|
497 |
|
|
498 |
while(k > 0u)
|
|
499 |
{
|
|
500 |
/* Perform the multiply-accumulates */
|
|
501 |
sum += ((q15_t) * px++ * *py--);
|
|
502 |
|
|
503 |
/* Decrement the loop counter */
|
|
504 |
k--;
|
|
505 |
}
|
|
506 |
|
|
507 |
/* Store the result in the accumulator in the destination buffer. */
|
|
508 |
*pOut++ = (q7_t) (__SSAT(sum >> 7u, 8));
|
|
509 |
|
|
510 |
/* Increment the pointer pIn1 index, count by 1 */
|
|
511 |
count++;
|
|
512 |
|
|
513 |
/* Update the inputA and inputB pointers for next MAC calculation */
|
|
514 |
px = pIn1 + count;
|
|
515 |
py = pSrc2;
|
|
516 |
|
|
517 |
/* Decrement the loop counter */
|
|
518 |
blkCnt--;
|
|
519 |
}
|
|
520 |
}
|
|
521 |
else
|
|
522 |
{
|
|
523 |
/* If the srcBLen is not a multiple of 4,
|
|
524 |
* the blockSize2 loop cannot be unrolled by 4 */
|
|
525 |
blkCnt = blockSize2;
|
|
526 |
|
|
527 |
while(blkCnt > 0u)
|
|
528 |
{
|
|
529 |
/* Accumulator is made zero for every iteration */
|
|
530 |
sum = 0;
|
|
531 |
|
|
532 |
/* srcBLen number of MACS should be performed */
|
|
533 |
k = srcBLen;
|
|
534 |
|
|
535 |
while(k > 0u)
|
|
536 |
{
|
|
537 |
/* Perform the multiply-accumulate */
|
|
538 |
sum += ((q15_t) * px++ * *py--);
|
|
539 |
|
|
540 |
/* Decrement the loop counter */
|
|
541 |
k--;
|
|
542 |
}
|
|
543 |
|
|
544 |
/* Store the result in the accumulator in the destination buffer. */
|
|
545 |
*pOut++ = (q7_t) (__SSAT(sum >> 7u, 8));
|
|
546 |
|
|
547 |
/* Increment the MAC count */
|
|
548 |
count++;
|
|
549 |
|
|
550 |
/* Update the inputA and inputB pointers for next MAC calculation */
|
|
551 |
px = pIn1 + count;
|
|
552 |
py = pSrc2;
|
|
553 |
|
|
554 |
/* Decrement the loop counter */
|
|
555 |
blkCnt--;
|
|
556 |
}
|
|
557 |
}
|
|
558 |
|
|
559 |
|
|
560 |
/* --------------------------
|
|
561 |
* Initializations of stage3
|
|
562 |
* -------------------------*/
|
|
563 |
|
|
564 |
/* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
|
|
565 |
* sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
|
|
566 |
* ....
|
|
567 |
* sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
|
|
568 |
* sum += x[srcALen-1] * y[srcBLen-1]
|
|
569 |
*/
|
|
570 |
|
|
571 |
/* In this stage the MAC operations are decreased by 1 for every iteration.
|
|
572 |
The blockSize3 variable holds the number of MAC operations performed */
|
|
573 |
|
|
574 |
/* Working pointer of inputA */
|
|
575 |
pSrc1 = pIn1 + (srcALen - (srcBLen - 1u));
|
|
576 |
px = pSrc1;
|
|
577 |
|
|
578 |
/* Working pointer of inputB */
|
|
579 |
pSrc2 = pIn2 + (srcBLen - 1u);
|
|
580 |
py = pSrc2;
|
|
581 |
|
|
582 |
/* -------------------
|
|
583 |
* Stage3 process
|
|
584 |
* ------------------*/
|
|
585 |
|
|
586 |
while(blockSize3 > 0u)
|
|
587 |
{
|
|
588 |
/* Accumulator is made zero for every iteration */
|
|
589 |
sum = 0;
|
|
590 |
|
|
591 |
/* Apply loop unrolling and compute 4 MACs simultaneously. */
|
|
592 |
k = blockSize3 >> 2u;
|
|
593 |
|
|
594 |
/* First part of the processing with loop unrolling. Compute 4 MACs at a time.
|
|
595 |
** a second loop below computes MACs for the remaining 1 to 3 samples. */
|
|
596 |
while(k > 0u)
|
|
597 |
{
|
|
598 |
/* Reading two inputs, x[srcALen - srcBLen + 1] and x[srcALen - srcBLen + 2] of SrcA buffer and packing */
|
|
599 |
in1 = (q15_t) * px++;
|
|
600 |
in2 = (q15_t) * px++;
|
|
601 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
602 |
|
|
603 |
/* Reading two inputs, y[srcBLen - 1] and y[srcBLen - 2] of SrcB buffer and packing */
|
|
604 |
in1 = (q15_t) * py--;
|
|
605 |
in2 = (q15_t) * py--;
|
|
606 |
input2 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
607 |
|
|
608 |
/* sum += x[srcALen - srcBLen + 1] * y[srcBLen - 1] */
|
|
609 |
/* sum += x[srcALen - srcBLen + 2] * y[srcBLen - 2] */
|
|
610 |
sum = __SMLAD(input1, input2, sum);
|
|
611 |
|
|
612 |
/* Reading two inputs, x[srcALen - srcBLen + 3] and x[srcALen - srcBLen + 4] of SrcA buffer and packing */
|
|
613 |
in1 = (q15_t) * px++;
|
|
614 |
in2 = (q15_t) * px++;
|
|
615 |
input1 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
616 |
|
|
617 |
/* Reading two inputs, y[srcBLen - 3] and y[srcBLen - 4] of SrcB buffer and packing */
|
|
618 |
in1 = (q15_t) * py--;
|
|
619 |
in2 = (q15_t) * py--;
|
|
620 |
input2 = ((q31_t) in1 & 0x0000FFFF) | ((q31_t) in2 << 16u);
|
|
621 |
|
|
622 |
/* sum += x[srcALen - srcBLen + 3] * y[srcBLen - 3] */
|
|
623 |
/* sum += x[srcALen - srcBLen + 4] * y[srcBLen - 4] */
|
|
624 |
sum = __SMLAD(input1, input2, sum);
|
|
625 |
|
|
626 |
/* Decrement the loop counter */
|
|
627 |
k--;
|
|
628 |
}
|
|
629 |
|
|
630 |
/* If the blockSize3 is not a multiple of 4, compute any remaining MACs here.
|
|
631 |
** No loop unrolling is used. */
|
|
632 |
k = blockSize3 % 0x4u;
|
|
633 |
|
|
634 |
while(k > 0u)
|
|
635 |
{
|
|
636 |
/* Perform the multiply-accumulates */
|
|
637 |
sum += ((q15_t) * px++ * *py--);
|
|
638 |
|
|
639 |
/* Decrement the loop counter */
|
|
640 |
k--;
|
|
641 |
}
|
|
642 |
|
|
643 |
/* Store the result in the accumulator in the destination buffer. */
|
|
644 |
*pOut++ = (q7_t) (__SSAT(sum >> 7u, 8));
|
|
645 |
|
|
646 |
/* Update the inputA and inputB pointers for next MAC calculation */
|
|
647 |
px = ++pSrc1;
|
|
648 |
py = pSrc2;
|
|
649 |
|
|
650 |
/* Decrement the loop counter */
|
|
651 |
blockSize3--;
|
|
652 |
}
|
|
653 |
|
|
654 |
#else
|
|
655 |
|
|
656 |
/* Run the below code for Cortex-M0 */
|
|
657 |
|
|
658 |
q7_t *pIn1 = pSrcA; /* input pointer */
|
|
659 |
q7_t *pIn2 = pSrcB; /* coefficient pointer */
|
|
660 |
q31_t sum; /* Accumulator */
|
|
661 |
uint32_t i, j; /* loop counter */
|
|
662 |
|
|
663 |
/* Loop to calculate output of convolution for output length number of times */
|
|
664 |
for (i = 0; i < (srcALen + srcBLen - 1); i++)
|
|
665 |
{
|
|
666 |
/* Initialize sum with zero to carry on MAC operations */
|
|
667 |
sum = 0;
|
|
668 |
|
|
669 |
/* Loop to perform MAC operations according to convolution equation */
|
|
670 |
for (j = 0; j <= i; j++)
|
|
671 |
{
|
|
672 |
/* Check the array limitations */
|
|
673 |
if(((i - j) < srcBLen) && (j < srcALen))
|
|
674 |
{
|
|
675 |
/* z[i] += x[i-j] * y[j] */
|
|
676 |
sum += (q15_t) pIn1[j] * (pIn2[i - j]);
|
|
677 |
}
|
|
678 |
}
|
|
679 |
|
|
680 |
/* Store the output in the destination buffer */
|
|
681 |
pDst[i] = (q7_t) __SSAT((sum >> 7u), 8u);
|
|
682 |
}
|
|
683 |
|
|
684 |
#endif /* #ifndef ARM_MATH_CM0_FAMILY */
|
|
685 |
|
|
686 |
}
|
|
687 |
|
|
688 |
/**
|
|
689 |
* @} end of Conv group
|
|
690 |
*/
|