提交 | 用户 | age
|
8b51c7
|
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_mat_mult_fast_q15.c |
|
9 |
* |
|
10 |
* Description: Q15 matrix multiplication (fast variant) |
|
11 |
* |
|
12 |
* Target Processor: Cortex-M4/Cortex-M3 |
|
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 groupMatrix |
|
45 |
*/ |
|
46 |
|
|
47 |
/** |
|
48 |
* @addtogroup MatrixMult |
|
49 |
* @{ |
|
50 |
*/ |
|
51 |
|
|
52 |
|
|
53 |
/** |
|
54 |
* @brief Q15 matrix multiplication (fast variant) for Cortex-M3 and Cortex-M4 |
|
55 |
* @param[in] *pSrcA points to the first input matrix structure |
|
56 |
* @param[in] *pSrcB points to the second input matrix structure |
|
57 |
* @param[out] *pDst points to output matrix structure |
|
58 |
* @param[in] *pState points to the array for storing intermediate results |
|
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 |
* <b>Scaling and Overflow Behavior:</b> |
|
64 |
* |
|
65 |
* \par |
|
66 |
* The difference between the function arm_mat_mult_q15() and this fast variant is that |
|
67 |
* the fast variant use a 32-bit rather than a 64-bit accumulator. |
|
68 |
* The result of each 1.15 x 1.15 multiplication is truncated to |
|
69 |
* 2.30 format. These intermediate results are accumulated in a 32-bit register in 2.30 |
|
70 |
* format. Finally, the accumulator is saturated and converted to a 1.15 result. |
|
71 |
* |
|
72 |
* \par |
|
73 |
* The fast version has the same overflow behavior as the standard version but provides |
|
74 |
* less precision since it discards the low 16 bits of each multiplication result. |
|
75 |
* In order to avoid overflows completely the input signals must be scaled down. |
|
76 |
* Scale down one of the input matrices by log2(numColsA) bits to |
|
77 |
* avoid overflows, as a total of numColsA additions are computed internally for each |
|
78 |
* output element. |
|
79 |
* |
|
80 |
* \par |
|
81 |
* See <code>arm_mat_mult_q15()</code> for a slower implementation of this function |
|
82 |
* which uses 64-bit accumulation to provide higher precision. |
|
83 |
*/ |
|
84 |
|
|
85 |
arm_status arm_mat_mult_fast_q15( |
|
86 |
const arm_matrix_instance_q15 * pSrcA, |
|
87 |
const arm_matrix_instance_q15 * pSrcB, |
|
88 |
arm_matrix_instance_q15 * pDst, |
|
89 |
q15_t * pState) |
|
90 |
{ |
|
91 |
q31_t sum; /* accumulator */ |
|
92 |
q15_t *pSrcBT = pState; /* input data matrix pointer for transpose */ |
|
93 |
q15_t *pInA = pSrcA->pData; /* input data matrix pointer A of Q15 type */ |
|
94 |
q15_t *pInB = pSrcB->pData; /* input data matrix pointer B of Q15 type */ |
|
95 |
q15_t *px; /* Temporary output data matrix pointer */ |
|
96 |
uint16_t numRowsA = pSrcA->numRows; /* number of rows of input matrix A */ |
|
97 |
uint16_t numColsB = pSrcB->numCols; /* number of columns of input matrix B */ |
|
98 |
uint16_t numColsA = pSrcA->numCols; /* number of columns of input matrix A */ |
|
99 |
uint16_t numRowsB = pSrcB->numRows; /* number of rows of input matrix A */ |
|
100 |
uint16_t col, i = 0u, row = numRowsB, colCnt; /* loop counters */ |
|
101 |
arm_status status; /* status of matrix multiplication */ |
|
102 |
|
|
103 |
#ifndef UNALIGNED_SUPPORT_DISABLE |
|
104 |
|
|
105 |
q31_t in; /* Temporary variable to hold the input value */ |
|
106 |
q31_t inA1, inA2, inB1, inB2; |
|
107 |
|
|
108 |
#else |
|
109 |
|
|
110 |
q15_t in; /* Temporary variable to hold the input value */ |
|
111 |
q15_t inA1, inA2, inB1, inB2; |
|
112 |
|
|
113 |
#endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ |
|
114 |
|
|
115 |
#ifdef ARM_MATH_MATRIX_CHECK |
|
116 |
/* Check for matrix mismatch condition */ |
|
117 |
if((pSrcA->numCols != pSrcB->numRows) || |
|
118 |
(pSrcA->numRows != pDst->numRows) || (pSrcB->numCols != pDst->numCols)) |
|
119 |
{ |
|
120 |
/* Set status as ARM_MATH_SIZE_MISMATCH */ |
|
121 |
status = ARM_MATH_SIZE_MISMATCH; |
|
122 |
} |
|
123 |
else |
|
124 |
#endif |
|
125 |
{ |
|
126 |
/* Matrix transpose */ |
|
127 |
do |
|
128 |
{ |
|
129 |
/* Apply loop unrolling and exchange the columns with row elements */ |
|
130 |
col = numColsB >> 2; |
|
131 |
|
|
132 |
/* The pointer px is set to starting address of the column being processed */ |
|
133 |
px = pSrcBT + i; |
|
134 |
|
|
135 |
/* First part of the processing with loop unrolling. Compute 4 outputs at a time. |
|
136 |
** a second loop below computes the remaining 1 to 3 samples. */ |
|
137 |
while(col > 0u) |
|
138 |
{ |
|
139 |
#ifndef UNALIGNED_SUPPORT_DISABLE |
|
140 |
/* Read two elements from the row */ |
|
141 |
in = *__SIMD32(pInB)++; |
|
142 |
|
|
143 |
/* Unpack and store one element in the destination */ |
|
144 |
#ifndef ARM_MATH_BIG_ENDIAN |
|
145 |
|
|
146 |
*px = (q15_t) in; |
|
147 |
|
|
148 |
#else |
|
149 |
|
|
150 |
*px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16); |
|
151 |
|
|
152 |
#endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
|
153 |
|
|
154 |
/* Update the pointer px to point to the next row of the transposed matrix */ |
|
155 |
px += numRowsB; |
|
156 |
|
|
157 |
/* Unpack and store the second element in the destination */ |
|
158 |
#ifndef ARM_MATH_BIG_ENDIAN |
|
159 |
|
|
160 |
*px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16); |
|
161 |
|
|
162 |
#else |
|
163 |
|
|
164 |
*px = (q15_t) in; |
|
165 |
|
|
166 |
#endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
|
167 |
|
|
168 |
/* Update the pointer px to point to the next row of the transposed matrix */ |
|
169 |
px += numRowsB; |
|
170 |
|
|
171 |
/* Read two elements from the row */ |
|
172 |
in = *__SIMD32(pInB)++; |
|
173 |
|
|
174 |
/* Unpack and store one element in the destination */ |
|
175 |
#ifndef ARM_MATH_BIG_ENDIAN |
|
176 |
|
|
177 |
*px = (q15_t) in; |
|
178 |
|
|
179 |
#else |
|
180 |
|
|
181 |
*px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16); |
|
182 |
|
|
183 |
#endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
|
184 |
|
|
185 |
/* Update the pointer px to point to the next row of the transposed matrix */ |
|
186 |
px += numRowsB; |
|
187 |
|
|
188 |
/* Unpack and store the second element in the destination */ |
|
189 |
|
|
190 |
#ifndef ARM_MATH_BIG_ENDIAN |
|
191 |
|
|
192 |
*px = (q15_t) ((in & (q31_t) 0xffff0000) >> 16); |
|
193 |
|
|
194 |
#else |
|
195 |
|
|
196 |
*px = (q15_t) in; |
|
197 |
|
|
198 |
#endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
|
199 |
|
|
200 |
#else |
|
201 |
|
|
202 |
/* Read one element from the row */ |
|
203 |
in = *pInB++; |
|
204 |
|
|
205 |
/* Store one element in the destination */ |
|
206 |
*px = in; |
|
207 |
|
|
208 |
/* Update the pointer px to point to the next row of the transposed matrix */ |
|
209 |
px += numRowsB; |
|
210 |
|
|
211 |
/* Read one element from the row */ |
|
212 |
in = *pInB++; |
|
213 |
|
|
214 |
/* Store one element in the destination */ |
|
215 |
*px = in; |
|
216 |
|
|
217 |
/* Update the pointer px to point to the next row of the transposed matrix */ |
|
218 |
px += numRowsB; |
|
219 |
|
|
220 |
/* Read one element from the row */ |
|
221 |
in = *pInB++; |
|
222 |
|
|
223 |
/* Store one element in the destination */ |
|
224 |
*px = in; |
|
225 |
|
|
226 |
/* Update the pointer px to point to the next row of the transposed matrix */ |
|
227 |
px += numRowsB; |
|
228 |
|
|
229 |
/* Read one element from the row */ |
|
230 |
in = *pInB++; |
|
231 |
|
|
232 |
/* Store one element in the destination */ |
|
233 |
*px = in; |
|
234 |
|
|
235 |
#endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ |
|
236 |
|
|
237 |
/* Update the pointer px to point to the next row of the transposed matrix */ |
|
238 |
px += numRowsB; |
|
239 |
|
|
240 |
/* Decrement the column loop counter */ |
|
241 |
col--; |
|
242 |
} |
|
243 |
|
|
244 |
/* If the columns of pSrcB is not a multiple of 4, compute any remaining output samples here. |
|
245 |
** No loop unrolling is used. */ |
|
246 |
col = numColsB % 0x4u; |
|
247 |
|
|
248 |
while(col > 0u) |
|
249 |
{ |
|
250 |
/* Read and store the input element in the destination */ |
|
251 |
*px = *pInB++; |
|
252 |
|
|
253 |
/* Update the pointer px to point to the next row of the transposed matrix */ |
|
254 |
px += numRowsB; |
|
255 |
|
|
256 |
/* Decrement the column loop counter */ |
|
257 |
col--; |
|
258 |
} |
|
259 |
|
|
260 |
i++; |
|
261 |
|
|
262 |
/* Decrement the row loop counter */ |
|
263 |
row--; |
|
264 |
|
|
265 |
} while(row > 0u); |
|
266 |
|
|
267 |
/* Reset the variables for the usage in the following multiplication process */ |
|
268 |
row = numRowsA; |
|
269 |
i = 0u; |
|
270 |
px = pDst->pData; |
|
271 |
|
|
272 |
/* The following loop performs the dot-product of each row in pSrcA with each column in pSrcB */ |
|
273 |
/* row loop */ |
|
274 |
do |
|
275 |
{ |
|
276 |
/* For every row wise process, the column loop counter is to be initiated */ |
|
277 |
col = numColsB; |
|
278 |
|
|
279 |
/* For every row wise process, the pIn2 pointer is set |
|
280 |
** to the starting address of the transposed pSrcB data */ |
|
281 |
pInB = pSrcBT; |
|
282 |
|
|
283 |
/* column loop */ |
|
284 |
do |
|
285 |
{ |
|
286 |
/* Set the variable sum, that acts as accumulator, to zero */ |
|
287 |
sum = 0; |
|
288 |
|
|
289 |
/* Apply loop unrolling and compute 2 MACs simultaneously. */ |
|
290 |
colCnt = numColsA >> 2; |
|
291 |
|
|
292 |
/* Initiate the pointer pIn1 to point to the starting address of the column being processed */ |
|
293 |
pInA = pSrcA->pData + i; |
|
294 |
|
|
295 |
/* matrix multiplication */ |
|
296 |
while(colCnt > 0u) |
|
297 |
{ |
|
298 |
/* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */ |
|
299 |
#ifndef UNALIGNED_SUPPORT_DISABLE |
|
300 |
|
|
301 |
inA1 = *__SIMD32(pInA)++; |
|
302 |
inB1 = *__SIMD32(pInB)++; |
|
303 |
inA2 = *__SIMD32(pInA)++; |
|
304 |
inB2 = *__SIMD32(pInB)++; |
|
305 |
|
|
306 |
sum = __SMLAD(inA1, inB1, sum); |
|
307 |
sum = __SMLAD(inA2, inB2, sum); |
|
308 |
|
|
309 |
#else |
|
310 |
|
|
311 |
inA1 = *pInA++; |
|
312 |
inB1 = *pInB++; |
|
313 |
inA2 = *pInA++; |
|
314 |
sum += inA1 * inB1; |
|
315 |
inB2 = *pInB++; |
|
316 |
|
|
317 |
inA1 = *pInA++; |
|
318 |
inB1 = *pInB++; |
|
319 |
sum += inA2 * inB2; |
|
320 |
inA2 = *pInA++; |
|
321 |
inB2 = *pInB++; |
|
322 |
|
|
323 |
sum += inA1 * inB1; |
|
324 |
sum += inA2 * inB2; |
|
325 |
|
|
326 |
#endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ |
|
327 |
|
|
328 |
/* Decrement the loop counter */ |
|
329 |
colCnt--; |
|
330 |
} |
|
331 |
|
|
332 |
/* process odd column samples */ |
|
333 |
colCnt = numColsA % 0x4u; |
|
334 |
|
|
335 |
while(colCnt > 0u) |
|
336 |
{ |
|
337 |
/* c(m,n) = a(1,1)*b(1,1) + a(1,2) * b(2,1) + .... + a(m,p)*b(p,n) */ |
|
338 |
sum += (q31_t) (*pInA++) * (*pInB++); |
|
339 |
|
|
340 |
colCnt--; |
|
341 |
} |
|
342 |
|
|
343 |
/* Saturate and store the result in the destination buffer */ |
|
344 |
*px = (q15_t) (sum >> 15); |
|
345 |
px++; |
|
346 |
|
|
347 |
/* Decrement the column loop counter */ |
|
348 |
col--; |
|
349 |
|
|
350 |
} while(col > 0u); |
|
351 |
|
|
352 |
i = i + numColsA; |
|
353 |
|
|
354 |
/* Decrement the row loop counter */ |
|
355 |
row--; |
|
356 |
|
|
357 |
} while(row > 0u); |
|
358 |
|
|
359 |
/* set status as ARM_MATH_SUCCESS */ |
|
360 |
status = ARM_MATH_SUCCESS; |
|
361 |
} |
|
362 |
|
|
363 |
/* Return to application */ |
|
364 |
return (status); |
|
365 |
} |
|
366 |
|
|
367 |
/** |
|
368 |
* @} end of MatrixMult group |
|
369 |
*/ |