1 /* ----------------------------------------------------------------------
2 * Copyright (C) 2010-2013 ARM Limited. All rights reserved.
4 * $Date: 17. January 2013
7 * Project: CMSIS DSP Library
8 * Title: arm_conv_partial_opt_q15.c
10 * Description: Partial convolution of Q15 sequences.
12 * Target Processor: Cortex-M4/Cortex-M3
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
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
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.
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 * -------------------------------------------------------------------- */
44 * @ingroup groupFilters
48 * @addtogroup PartialConv
53 * @brief Partial convolution of Q15 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.
59 * @param[in] firstIndex is the first output sample to start with.
60 * @param[in] numPoints is the number of output points to be computed.
61 * @param[in] *pScratch1 points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2.
62 * @param[in] *pScratch2 points to scratch buffer of size min(srcALen, srcBLen).
63 * @return Returns either ARM_MATH_SUCCESS if the function completed correctly or ARM_MATH_ARGUMENT_ERROR if the requested subset is not in the range [0 srcALen+srcBLen-2].
66 * If the silicon does not support unaligned memory access enable the macro UNALIGNED_SUPPORT_DISABLE
67 * In this case input, output, state buffers should be aligned by 32-bit
69 * Refer to <code>arm_conv_partial_fast_q15()</code> for a faster but less precise version of this function for Cortex-M3 and Cortex-M4.
74 #ifndef UNALIGNED_SUPPORT_DISABLE
76 arm_status arm_conv_partial_opt_q15(
88 q15_t *pOut = pDst; /* output pointer */
89 q15_t *pScr1 = pScratch1; /* Temporary pointer for scratch1 */
90 q15_t *pScr2 = pScratch2; /* Temporary pointer for scratch1 */
91 q63_t acc0, acc1, acc2, acc3; /* Accumulator */
92 q31_t x1, x2, x3; /* Temporary variables to hold state and coefficient values */
93 q31_t y1, y2; /* State variables */
94 q15_t *pIn1; /* inputA pointer */
95 q15_t *pIn2; /* inputB pointer */
96 q15_t *px; /* Intermediate inputA pointer */
97 q15_t *py; /* Intermediate inputB pointer */
98 uint32_t j, k, blkCnt; /* loop counter */
99 arm_status status; /* Status variable */
100 uint32_t tapCnt; /* loop count */
102 /* Check for range of output samples to be calculated */
103 if((firstIndex + numPoints) > ((srcALen + (srcBLen - 1u))))
105 /* Set status as ARM_MATH_ARGUMENT_ERROR */
106 status = ARM_MATH_ARGUMENT_ERROR;
111 /* The algorithm implementation is based on the lengths of the inputs. */
112 /* srcB is always made to slide across srcA. */
113 /* So srcBLen is always considered as shorter or equal to srcALen */
114 if(srcALen >= srcBLen)
116 /* Initialization of inputA pointer */
119 /* Initialization of inputB pointer */
124 /* Initialization of inputA pointer */
127 /* Initialization of inputB pointer */
130 /* srcBLen is always considered as shorter or equal to srcALen */
136 /* Temporary pointer for scratch2 */
139 /* pointer to take end of scratch2 buffer */
140 pScr2 = pScratch2 + srcBLen - 1;
142 /* points to smaller length sequence */
145 /* Apply loop unrolling and do 4 Copies simultaneously. */
148 /* First part of the processing with loop unrolling copies 4 data points at a time.
149 ** a second loop below copies for the remaining 1 to 3 samples. */
152 /* copy second buffer in reversal manner */
158 /* Decrement the loop counter */
162 /* If the count is not a multiple of 4, copy remaining samples here.
163 ** No loop unrolling is used. */
168 /* copy second buffer in reversal manner for remaining samples */
171 /* Decrement the loop counter */
175 /* Initialze temporary scratch pointer */
178 /* Fill (srcBLen - 1u) zeros in scratch buffer */
179 arm_fill_q15(0, pScr1, (srcBLen - 1u));
181 /* Update temporary scratch pointer */
182 pScr1 += (srcBLen - 1u);
184 /* Copy bigger length sequence(srcALen) samples in scratch1 buffer */
186 /* Copy (srcALen) samples in scratch buffer */
187 arm_copy_q15(pIn1, pScr1, srcALen);
189 /* Update pointers */
192 /* Fill (srcBLen - 1u) zeros at end of scratch buffer */
193 arm_fill_q15(0, pScr1, (srcBLen - 1u));
196 pScr1 += (srcBLen - 1u);
198 /* Initialization of pIn2 pointer */
201 pScratch1 += firstIndex;
203 pOut = pDst + firstIndex;
205 /* Actual convolution process starts here */
206 blkCnt = (numPoints) >> 2;
210 /* Initialze temporary scratch pointer as scratch1 */
213 /* Clear Accumlators */
219 /* Read two samples from scratch1 buffer */
220 x1 = *__SIMD32(pScr1)++;
222 /* Read next two samples from scratch1 buffer */
223 x2 = *__SIMD32(pScr1)++;
225 tapCnt = (srcBLen) >> 2u;
230 /* Read four samples from smaller buffer */
231 y1 = _SIMD32_OFFSET(pIn2);
232 y2 = _SIMD32_OFFSET(pIn2 + 2u);
234 /* multiply and accumlate */
235 acc0 = __SMLALD(x1, y1, acc0);
236 acc2 = __SMLALD(x2, y1, acc2);
238 /* pack input data */
239 #ifndef ARM_MATH_BIG_ENDIAN
240 x3 = __PKHBT(x2, x1, 0);
242 x3 = __PKHBT(x1, x2, 0);
245 /* multiply and accumlate */
246 acc1 = __SMLALDX(x3, y1, acc1);
248 /* Read next two samples from scratch1 buffer */
249 x1 = _SIMD32_OFFSET(pScr1);
251 /* multiply and accumlate */
252 acc0 = __SMLALD(x2, y2, acc0);
253 acc2 = __SMLALD(x1, y2, acc2);
255 /* pack input data */
256 #ifndef ARM_MATH_BIG_ENDIAN
257 x3 = __PKHBT(x1, x2, 0);
259 x3 = __PKHBT(x2, x1, 0);
262 acc3 = __SMLALDX(x3, y1, acc3);
263 acc1 = __SMLALDX(x3, y2, acc1);
265 x2 = _SIMD32_OFFSET(pScr1 + 2u);
267 #ifndef ARM_MATH_BIG_ENDIAN
268 x3 = __PKHBT(x2, x1, 0);
270 x3 = __PKHBT(x1, x2, 0);
273 acc3 = __SMLALDX(x3, y2, acc3);
275 /* update scratch pointers */
280 /* Decrement the loop counter */
284 /* Update scratch pointer for remaining samples of smaller length sequence */
287 /* apply same above for remaining samples of smaller length sequence */
288 tapCnt = (srcBLen) & 3u;
292 /* accumlate the results */
293 acc0 += (*pScr1++ * *pIn2);
294 acc1 += (*pScr1++ * *pIn2);
295 acc2 += (*pScr1++ * *pIn2);
296 acc3 += (*pScr1++ * *pIn2++);
300 /* Decrement the loop counter */
307 /* Store the results in the accumulators in the destination buffer. */
309 #ifndef ARM_MATH_BIG_ENDIAN
312 __PKHBT(__SSAT((acc0 >> 15), 16), __SSAT((acc1 >> 15), 16), 16);
314 __PKHBT(__SSAT((acc2 >> 15), 16), __SSAT((acc3 >> 15), 16), 16);
319 __PKHBT(__SSAT((acc1 >> 15), 16), __SSAT((acc0 >> 15), 16), 16);
321 __PKHBT(__SSAT((acc3 >> 15), 16), __SSAT((acc2 >> 15), 16), 16);
323 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
325 /* Initialization of inputB pointer */
333 blkCnt = numPoints & 0x3;
335 /* Calculate convolution for remaining samples of Bigger length sequence */
338 /* Initialze temporary scratch pointer as scratch1 */
341 /* Clear Accumlators */
344 tapCnt = (srcBLen) >> 1u;
349 /* Read next two samples from scratch1 buffer */
350 x1 = *__SIMD32(pScr1)++;
352 /* Read two samples from smaller buffer */
353 y1 = *__SIMD32(pIn2)++;
355 acc0 = __SMLALD(x1, y1, acc0);
357 /* Decrement the loop counter */
361 tapCnt = (srcBLen) & 1u;
363 /* apply same above for remaining samples of smaller length sequence */
367 /* accumlate the results */
368 acc0 += (*pScr1++ * *pIn2++);
370 /* Decrement the loop counter */
376 /* Store the result in the accumulator in the destination buffer. */
377 *pOut++ = (q15_t) (__SSAT((acc0 >> 15), 16));
379 /* Initialization of inputB pointer */
386 /* set status as ARM_MATH_SUCCESS */
387 status = ARM_MATH_SUCCESS;
391 /* Return to application */
397 arm_status arm_conv_partial_opt_q15(
409 q15_t *pOut = pDst; /* output pointer */
410 q15_t *pScr1 = pScratch1; /* Temporary pointer for scratch1 */
411 q15_t *pScr2 = pScratch2; /* Temporary pointer for scratch1 */
412 q63_t acc0, acc1, acc2, acc3; /* Accumulator */
413 q15_t *pIn1; /* inputA pointer */
414 q15_t *pIn2; /* inputB pointer */
415 q15_t *px; /* Intermediate inputA pointer */
416 q15_t *py; /* Intermediate inputB pointer */
417 uint32_t j, k, blkCnt; /* loop counter */
418 arm_status status; /* Status variable */
419 uint32_t tapCnt; /* loop count */
420 q15_t x10, x11, x20, x21; /* Temporary variables to hold srcA buffer */
421 q15_t y10, y11; /* Temporary variables to hold srcB buffer */
424 /* Check for range of output samples to be calculated */
425 if((firstIndex + numPoints) > ((srcALen + (srcBLen - 1u))))
427 /* Set status as ARM_MATH_ARGUMENT_ERROR */
428 status = ARM_MATH_ARGUMENT_ERROR;
433 /* The algorithm implementation is based on the lengths of the inputs. */
434 /* srcB is always made to slide across srcA. */
435 /* So srcBLen is always considered as shorter or equal to srcALen */
436 if(srcALen >= srcBLen)
438 /* Initialization of inputA pointer */
441 /* Initialization of inputB pointer */
446 /* Initialization of inputA pointer */
449 /* Initialization of inputB pointer */
452 /* srcBLen is always considered as shorter or equal to srcALen */
458 /* Temporary pointer for scratch2 */
461 /* pointer to take end of scratch2 buffer */
462 pScr2 = pScratch2 + srcBLen - 1;
464 /* points to smaller length sequence */
467 /* Apply loop unrolling and do 4 Copies simultaneously. */
470 /* First part of the processing with loop unrolling copies 4 data points at a time.
471 ** a second loop below copies for the remaining 1 to 3 samples. */
474 /* copy second buffer in reversal manner */
480 /* Decrement the loop counter */
484 /* If the count is not a multiple of 4, copy remaining samples here.
485 ** No loop unrolling is used. */
490 /* copy second buffer in reversal manner for remaining samples */
493 /* Decrement the loop counter */
497 /* Initialze temporary scratch pointer */
500 /* Fill (srcBLen - 1u) zeros in scratch buffer */
501 arm_fill_q15(0, pScr1, (srcBLen - 1u));
503 /* Update temporary scratch pointer */
504 pScr1 += (srcBLen - 1u);
506 /* Copy bigger length sequence(srcALen) samples in scratch1 buffer */
509 /* Apply loop unrolling and do 4 Copies simultaneously. */
512 /* First part of the processing with loop unrolling copies 4 data points at a time.
513 ** a second loop below copies for the remaining 1 to 3 samples. */
516 /* copy second buffer in reversal manner */
522 /* Decrement the loop counter */
526 /* If the count is not a multiple of 4, copy remaining samples here.
527 ** No loop unrolling is used. */
532 /* copy second buffer in reversal manner for remaining samples */
535 /* Decrement the loop counter */
540 /* Apply loop unrolling and do 4 Copies simultaneously. */
541 k = (srcBLen - 1u) >> 2u;
543 /* First part of the processing with loop unrolling copies 4 data points at a time.
544 ** a second loop below copies for the remaining 1 to 3 samples. */
547 /* copy second buffer in reversal manner */
553 /* Decrement the loop counter */
557 /* If the count is not a multiple of 4, copy remaining samples here.
558 ** No loop unrolling is used. */
559 k = (srcBLen - 1u) % 0x4u;
563 /* copy second buffer in reversal manner for remaining samples */
566 /* Decrement the loop counter */
571 /* Initialization of pIn2 pointer */
574 pScratch1 += firstIndex;
576 pOut = pDst + firstIndex;
578 /* Actual convolution process starts here */
579 blkCnt = (numPoints) >> 2;
583 /* Initialze temporary scratch pointer as scratch1 */
586 /* Clear Accumlators */
592 /* Read two samples from scratch1 buffer */
596 /* Read next two samples from scratch1 buffer */
600 tapCnt = (srcBLen) >> 2u;
605 /* Read two samples from smaller buffer */
609 /* multiply and accumlate */
610 acc0 += (q63_t) x10 *y10;
611 acc0 += (q63_t) x11 *y11;
612 acc2 += (q63_t) x20 *y10;
613 acc2 += (q63_t) x21 *y11;
615 /* multiply and accumlate */
616 acc1 += (q63_t) x11 *y10;
617 acc1 += (q63_t) x20 *y11;
619 /* Read next two samples from scratch1 buffer */
623 /* multiply and accumlate */
624 acc3 += (q63_t) x21 *y10;
625 acc3 += (q63_t) x10 *y11;
627 /* Read next two samples from scratch2 buffer */
631 /* multiply and accumlate */
632 acc0 += (q63_t) x20 *y10;
633 acc0 += (q63_t) x21 *y11;
634 acc2 += (q63_t) x10 *y10;
635 acc2 += (q63_t) x11 *y11;
636 acc1 += (q63_t) x21 *y10;
637 acc1 += (q63_t) x10 *y11;
639 /* Read next two samples from scratch1 buffer */
643 /* multiply and accumlate */
644 acc3 += (q63_t) x11 *y10;
645 acc3 += (q63_t) x20 *y11;
647 /* update scratch pointers */
651 /* Decrement the loop counter */
655 /* Update scratch pointer for remaining samples of smaller length sequence */
658 /* apply same above for remaining samples of smaller length sequence */
659 tapCnt = (srcBLen) & 3u;
663 /* accumlate the results */
664 acc0 += (*pScr1++ * *pIn2);
665 acc1 += (*pScr1++ * *pIn2);
666 acc2 += (*pScr1++ * *pIn2);
667 acc3 += (*pScr1++ * *pIn2++);
671 /* Decrement the loop counter */
678 /* Store the results in the accumulators in the destination buffer. */
679 *pOut++ = __SSAT((acc0 >> 15), 16);
680 *pOut++ = __SSAT((acc1 >> 15), 16);
681 *pOut++ = __SSAT((acc2 >> 15), 16);
682 *pOut++ = __SSAT((acc3 >> 15), 16);
685 /* Initialization of inputB pointer */
693 blkCnt = numPoints & 0x3;
695 /* Calculate convolution for remaining samples of Bigger length sequence */
698 /* Initialze temporary scratch pointer as scratch1 */
701 /* Clear Accumlators */
704 tapCnt = (srcBLen) >> 1u;
709 /* Read next two samples from scratch1 buffer */
713 /* Read two samples from smaller buffer */
717 /* multiply and accumlate */
718 acc0 += (q63_t) x10 *y10;
719 acc0 += (q63_t) x11 *y11;
721 /* Decrement the loop counter */
725 tapCnt = (srcBLen) & 1u;
727 /* apply same above for remaining samples of smaller length sequence */
731 /* accumlate the results */
732 acc0 += (*pScr1++ * *pIn2++);
734 /* Decrement the loop counter */
740 /* Store the result in the accumulator in the destination buffer. */
741 *pOut++ = (q15_t) (__SSAT((acc0 >> 15), 16));
744 /* Initialization of inputB pointer */
751 /* set status as ARM_MATH_SUCCESS */
752 status = ARM_MATH_SUCCESS;
756 /* Return to application */
760 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
764 * @} end of PartialConv group