Robert Lopez
Aded CMSIS5 DSP and NN folder. Needs some work
Embed:
(wiki syntax)
Show/hide line numbers
arm_correlate_fast_q15.c
00001 /* ---------------------------------------------------------------------- 00002 * Project: CMSIS DSP Library 00003 * Title: arm_correlate_fast_q15.c 00004 * Description: Fast Q15 Correlation 00005 * 00006 * $Date: 27. January 2017 00007 * $Revision: V.1.5.1 00008 * 00009 * Target Processor: Cortex-M cores 00010 * -------------------------------------------------------------------- */ 00011 /* 00012 * Copyright (C) 2010-2017 ARM Limited or its affiliates. All rights reserved. 00013 * 00014 * SPDX-License-Identifier: Apache-2.0 00015 * 00016 * Licensed under the Apache License, Version 2.0 (the License); you may 00017 * not use this file except in compliance with the License. 00018 * You may obtain a copy of the License at 00019 * 00020 * www.apache.org/licenses/LICENSE-2.0 00021 * 00022 * Unless required by applicable law or agreed to in writing, software 00023 * distributed under the License is distributed on an AS IS BASIS, WITHOUT 00024 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 00025 * See the License for the specific language governing permissions and 00026 * limitations under the License. 00027 */ 00028 00029 #include "arm_math.h" 00030 00031 /** 00032 * @ingroup groupFilters 00033 */ 00034 00035 /** 00036 * @addtogroup Corr 00037 * @{ 00038 */ 00039 00040 /** 00041 * @brief Correlation of Q15 sequences (fast version) for Cortex-M3 and Cortex-M4. 00042 * @param[in] *pSrcA points to the first input sequence. 00043 * @param[in] srcALen length of the first input sequence. 00044 * @param[in] *pSrcB points to the second input sequence. 00045 * @param[in] srcBLen length of the second input sequence. 00046 * @param[out] *pDst points to the location where the output result is written. Length 2 * max(srcALen, srcBLen) - 1. 00047 * @return none. 00048 * 00049 * <b>Scaling and Overflow Behavior:</b> 00050 * 00051 * \par 00052 * This fast version uses a 32-bit accumulator with 2.30 format. 00053 * The accumulator maintains full precision of the intermediate multiplication results but provides only a single guard bit. 00054 * There is no saturation on intermediate additions. 00055 * Thus, if the accumulator overflows it wraps around and distorts the result. 00056 * The input signals should be scaled down to avoid intermediate overflows. 00057 * Scale down one of the inputs by 1/min(srcALen, srcBLen) to avoid overflow since a 00058 * maximum of min(srcALen, srcBLen) number of additions is carried internally. 00059 * The 2.30 accumulator is right shifted by 15 bits and then saturated to 1.15 format to yield the final result. 00060 * 00061 * \par 00062 * See <code>arm_correlate_q15()</code> for a slower implementation of this function which uses a 64-bit accumulator to avoid wrap around distortion. 00063 */ 00064 00065 void arm_correlate_fast_q15( 00066 q15_t * pSrcA, 00067 uint32_t srcALen, 00068 q15_t * pSrcB, 00069 uint32_t srcBLen, 00070 q15_t * pDst) 00071 { 00072 #ifndef UNALIGNED_SUPPORT_DISABLE 00073 00074 q15_t *pIn1; /* inputA pointer */ 00075 q15_t *pIn2; /* inputB pointer */ 00076 q15_t *pOut = pDst; /* output pointer */ 00077 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulators */ 00078 q15_t *px; /* Intermediate inputA pointer */ 00079 q15_t *py; /* Intermediate inputB pointer */ 00080 q15_t *pSrc1; /* Intermediate pointers */ 00081 q31_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */ 00082 uint32_t j, k = 0U, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counter */ 00083 int32_t inc = 1; /* Destination address modifier */ 00084 00085 00086 /* The algorithm implementation is based on the lengths of the inputs. */ 00087 /* srcB is always made to slide across srcA. */ 00088 /* So srcBLen is always considered as shorter or equal to srcALen */ 00089 /* But CORR(x, y) is reverse of CORR(y, x) */ 00090 /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */ 00091 /* and the destination pointer modifier, inc is set to -1 */ 00092 /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */ 00093 /* But to improve the performance, 00094 * we include zeroes in the output instead of zero padding either of the the inputs*/ 00095 /* If srcALen > srcBLen, 00096 * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */ 00097 /* If srcALen < srcBLen, 00098 * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */ 00099 if (srcALen >= srcBLen) 00100 { 00101 /* Initialization of inputA pointer */ 00102 pIn1 = (pSrcA); 00103 00104 /* Initialization of inputB pointer */ 00105 pIn2 = (pSrcB); 00106 00107 /* Number of output samples is calculated */ 00108 outBlockSize = (2U * srcALen) - 1U; 00109 00110 /* When srcALen > srcBLen, zero padding is done to srcB 00111 * to make their lengths equal. 00112 * Instead, (outBlockSize - (srcALen + srcBLen - 1)) 00113 * number of output samples are made zero */ 00114 j = outBlockSize - (srcALen + (srcBLen - 1U)); 00115 00116 /* Updating the pointer position to non zero value */ 00117 pOut += j; 00118 00119 } 00120 else 00121 { 00122 /* Initialization of inputA pointer */ 00123 pIn1 = (pSrcB); 00124 00125 /* Initialization of inputB pointer */ 00126 pIn2 = (pSrcA); 00127 00128 /* srcBLen is always considered as shorter or equal to srcALen */ 00129 j = srcBLen; 00130 srcBLen = srcALen; 00131 srcALen = j; 00132 00133 /* CORR(x, y) = Reverse order(CORR(y, x)) */ 00134 /* Hence set the destination pointer to point to the last output sample */ 00135 pOut = pDst + ((srcALen + srcBLen) - 2U); 00136 00137 /* Destination address modifier is set to -1 */ 00138 inc = -1; 00139 00140 } 00141 00142 /* The function is internally 00143 * divided into three parts according to the number of multiplications that has to be 00144 * taken place between inputA samples and inputB samples. In the first part of the 00145 * algorithm, the multiplications increase by one for every iteration. 00146 * In the second part of the algorithm, srcBLen number of multiplications are done. 00147 * In the third part of the algorithm, the multiplications decrease by one 00148 * for every iteration.*/ 00149 /* The algorithm is implemented in three stages. 00150 * The loop counters of each stage is initiated here. */ 00151 blockSize1 = srcBLen - 1U; 00152 blockSize2 = srcALen - (srcBLen - 1U); 00153 blockSize3 = blockSize1; 00154 00155 /* -------------------------- 00156 * Initializations of stage1 00157 * -------------------------*/ 00158 00159 /* sum = x[0] * y[srcBlen - 1] 00160 * sum = x[0] * y[srcBlen - 2] + x[1] * y[srcBlen - 1] 00161 * .... 00162 * sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen - 1] * y[srcBLen - 1] 00163 */ 00164 00165 /* In this stage the MAC operations are increased by 1 for every iteration. 00166 The count variable holds the number of MAC operations performed */ 00167 count = 1U; 00168 00169 /* Working pointer of inputA */ 00170 px = pIn1; 00171 00172 /* Working pointer of inputB */ 00173 pSrc1 = pIn2 + (srcBLen - 1U); 00174 py = pSrc1; 00175 00176 /* ------------------------ 00177 * Stage1 process 00178 * ----------------------*/ 00179 00180 /* The first loop starts here */ 00181 while (blockSize1 > 0U) 00182 { 00183 /* Accumulator is made zero for every iteration */ 00184 sum = 0; 00185 00186 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00187 k = count >> 2; 00188 00189 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00190 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00191 while (k > 0U) 00192 { 00193 /* x[0] * y[srcBLen - 4] , x[1] * y[srcBLen - 3] */ 00194 sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00195 /* x[3] * y[srcBLen - 1] , x[2] * y[srcBLen - 2] */ 00196 sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00197 00198 /* Decrement the loop counter */ 00199 k--; 00200 } 00201 00202 /* If the count is not a multiple of 4, compute any remaining MACs here. 00203 ** No loop unrolling is used. */ 00204 k = count % 0x4U; 00205 00206 while (k > 0U) 00207 { 00208 /* Perform the multiply-accumulates */ 00209 /* x[0] * y[srcBLen - 1] */ 00210 sum = __SMLAD(*px++, *py++, sum); 00211 00212 /* Decrement the loop counter */ 00213 k--; 00214 } 00215 00216 /* Store the result in the accumulator in the destination buffer. */ 00217 *pOut = (q15_t) (sum >> 15); 00218 /* Destination pointer is updated according to the address modifier, inc */ 00219 pOut += inc; 00220 00221 /* Update the inputA and inputB pointers for next MAC calculation */ 00222 py = pSrc1 - count; 00223 px = pIn1; 00224 00225 /* Increment the MAC count */ 00226 count++; 00227 00228 /* Decrement the loop counter */ 00229 blockSize1--; 00230 } 00231 00232 /* -------------------------- 00233 * Initializations of stage2 00234 * ------------------------*/ 00235 00236 /* sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen-1] * y[srcBLen-1] 00237 * sum = x[1] * y[0] + x[2] * y[1] +...+ x[srcBLen] * y[srcBLen-1] 00238 * .... 00239 * sum = x[srcALen-srcBLen-2] * y[0] + x[srcALen-srcBLen-1] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] 00240 */ 00241 00242 /* Working pointer of inputA */ 00243 px = pIn1; 00244 00245 /* Working pointer of inputB */ 00246 py = pIn2; 00247 00248 /* count is index by which the pointer pIn1 to be incremented */ 00249 count = 0U; 00250 00251 /* ------------------- 00252 * Stage2 process 00253 * ------------------*/ 00254 00255 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed. 00256 * So, to loop unroll over blockSize2, 00257 * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */ 00258 if (srcBLen >= 4U) 00259 { 00260 /* Loop unroll over blockSize2, by 4 */ 00261 blkCnt = blockSize2 >> 2U; 00262 00263 while (blkCnt > 0U) 00264 { 00265 /* Set all accumulators to zero */ 00266 acc0 = 0; 00267 acc1 = 0; 00268 acc2 = 0; 00269 acc3 = 0; 00270 00271 /* read x[0], x[1] samples */ 00272 x0 = *__SIMD32(px); 00273 /* read x[1], x[2] samples */ 00274 x1 = _SIMD32_OFFSET(px + 1); 00275 px += 2U; 00276 00277 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00278 k = srcBLen >> 2U; 00279 00280 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00281 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00282 do 00283 { 00284 /* Read the first two inputB samples using SIMD: 00285 * y[0] and y[1] */ 00286 c0 = *__SIMD32(py)++; 00287 00288 /* acc0 += x[0] * y[0] + x[1] * y[1] */ 00289 acc0 = __SMLAD(x0, c0, acc0); 00290 00291 /* acc1 += x[1] * y[0] + x[2] * y[1] */ 00292 acc1 = __SMLAD(x1, c0, acc1); 00293 00294 /* Read x[2], x[3] */ 00295 x2 = *__SIMD32(px); 00296 00297 /* Read x[3], x[4] */ 00298 x3 = _SIMD32_OFFSET(px + 1); 00299 00300 /* acc2 += x[2] * y[0] + x[3] * y[1] */ 00301 acc2 = __SMLAD(x2, c0, acc2); 00302 00303 /* acc3 += x[3] * y[0] + x[4] * y[1] */ 00304 acc3 = __SMLAD(x3, c0, acc3); 00305 00306 /* Read y[2] and y[3] */ 00307 c0 = *__SIMD32(py)++; 00308 00309 /* acc0 += x[2] * y[2] + x[3] * y[3] */ 00310 acc0 = __SMLAD(x2, c0, acc0); 00311 00312 /* acc1 += x[3] * y[2] + x[4] * y[3] */ 00313 acc1 = __SMLAD(x3, c0, acc1); 00314 00315 /* Read x[4], x[5] */ 00316 x0 = _SIMD32_OFFSET(px + 2); 00317 00318 /* Read x[5], x[6] */ 00319 x1 = _SIMD32_OFFSET(px + 3); 00320 px += 4U; 00321 00322 /* acc2 += x[4] * y[2] + x[5] * y[3] */ 00323 acc2 = __SMLAD(x0, c0, acc2); 00324 00325 /* acc3 += x[5] * y[2] + x[6] * y[3] */ 00326 acc3 = __SMLAD(x1, c0, acc3); 00327 00328 } while (--k); 00329 00330 /* For the next MAC operations, SIMD is not used 00331 * So, the 16 bit pointer if inputB, py is updated */ 00332 00333 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. 00334 ** No loop unrolling is used. */ 00335 k = srcBLen % 0x4U; 00336 00337 if (k == 1U) 00338 { 00339 /* Read y[4] */ 00340 c0 = *py; 00341 #ifdef ARM_MATH_BIG_ENDIAN 00342 00343 c0 = c0 << 16U; 00344 00345 #else 00346 00347 c0 = c0 & 0x0000FFFF; 00348 00349 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ 00350 00351 /* Read x[7] */ 00352 x3 = *__SIMD32(px); 00353 px++; 00354 00355 /* Perform the multiply-accumulates */ 00356 acc0 = __SMLAD(x0, c0, acc0); 00357 acc1 = __SMLAD(x1, c0, acc1); 00358 acc2 = __SMLADX(x1, c0, acc2); 00359 acc3 = __SMLADX(x3, c0, acc3); 00360 } 00361 00362 if (k == 2U) 00363 { 00364 /* Read y[4], y[5] */ 00365 c0 = *__SIMD32(py); 00366 00367 /* Read x[7], x[8] */ 00368 x3 = *__SIMD32(px); 00369 00370 /* Read x[9] */ 00371 x2 = _SIMD32_OFFSET(px + 1); 00372 px += 2U; 00373 00374 /* Perform the multiply-accumulates */ 00375 acc0 = __SMLAD(x0, c0, acc0); 00376 acc1 = __SMLAD(x1, c0, acc1); 00377 acc2 = __SMLAD(x3, c0, acc2); 00378 acc3 = __SMLAD(x2, c0, acc3); 00379 } 00380 00381 if (k == 3U) 00382 { 00383 /* Read y[4], y[5] */ 00384 c0 = *__SIMD32(py)++; 00385 00386 /* Read x[7], x[8] */ 00387 x3 = *__SIMD32(px); 00388 00389 /* Read x[9] */ 00390 x2 = _SIMD32_OFFSET(px + 1); 00391 00392 /* Perform the multiply-accumulates */ 00393 acc0 = __SMLAD(x0, c0, acc0); 00394 acc1 = __SMLAD(x1, c0, acc1); 00395 acc2 = __SMLAD(x3, c0, acc2); 00396 acc3 = __SMLAD(x2, c0, acc3); 00397 00398 c0 = (*py); 00399 /* Read y[6] */ 00400 #ifdef ARM_MATH_BIG_ENDIAN 00401 00402 c0 = c0 << 16U; 00403 #else 00404 00405 c0 = c0 & 0x0000FFFF; 00406 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ 00407 00408 /* Read x[10] */ 00409 x3 = _SIMD32_OFFSET(px + 2); 00410 px += 3U; 00411 00412 /* Perform the multiply-accumulates */ 00413 acc0 = __SMLADX(x1, c0, acc0); 00414 acc1 = __SMLAD(x2, c0, acc1); 00415 acc2 = __SMLADX(x2, c0, acc2); 00416 acc3 = __SMLADX(x3, c0, acc3); 00417 } 00418 00419 /* Store the result in the accumulator in the destination buffer. */ 00420 *pOut = (q15_t) (acc0 >> 15); 00421 /* Destination pointer is updated according to the address modifier, inc */ 00422 pOut += inc; 00423 00424 *pOut = (q15_t) (acc1 >> 15); 00425 pOut += inc; 00426 00427 *pOut = (q15_t) (acc2 >> 15); 00428 pOut += inc; 00429 00430 *pOut = (q15_t) (acc3 >> 15); 00431 pOut += inc; 00432 00433 /* Increment the pointer pIn1 index, count by 1 */ 00434 count += 4U; 00435 00436 /* Update the inputA and inputB pointers for next MAC calculation */ 00437 px = pIn1 + count; 00438 py = pIn2; 00439 00440 00441 /* Decrement the loop counter */ 00442 blkCnt--; 00443 } 00444 00445 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here. 00446 ** No loop unrolling is used. */ 00447 blkCnt = blockSize2 % 0x4U; 00448 00449 while (blkCnt > 0U) 00450 { 00451 /* Accumulator is made zero for every iteration */ 00452 sum = 0; 00453 00454 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00455 k = srcBLen >> 2U; 00456 00457 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00458 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00459 while (k > 0U) 00460 { 00461 /* Perform the multiply-accumulates */ 00462 sum += ((q31_t) * px++ * *py++); 00463 sum += ((q31_t) * px++ * *py++); 00464 sum += ((q31_t) * px++ * *py++); 00465 sum += ((q31_t) * px++ * *py++); 00466 00467 /* Decrement the loop counter */ 00468 k--; 00469 } 00470 00471 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. 00472 ** No loop unrolling is used. */ 00473 k = srcBLen % 0x4U; 00474 00475 while (k > 0U) 00476 { 00477 /* Perform the multiply-accumulates */ 00478 sum += ((q31_t) * px++ * *py++); 00479 00480 /* Decrement the loop counter */ 00481 k--; 00482 } 00483 00484 /* Store the result in the accumulator in the destination buffer. */ 00485 *pOut = (q15_t) (sum >> 15); 00486 /* Destination pointer is updated according to the address modifier, inc */ 00487 pOut += inc; 00488 00489 /* Increment the pointer pIn1 index, count by 1 */ 00490 count++; 00491 00492 /* Update the inputA and inputB pointers for next MAC calculation */ 00493 px = pIn1 + count; 00494 py = pIn2; 00495 00496 /* Decrement the loop counter */ 00497 blkCnt--; 00498 } 00499 } 00500 else 00501 { 00502 /* If the srcBLen is not a multiple of 4, 00503 * the blockSize2 loop cannot be unrolled by 4 */ 00504 blkCnt = blockSize2; 00505 00506 while (blkCnt > 0U) 00507 { 00508 /* Accumulator is made zero for every iteration */ 00509 sum = 0; 00510 00511 /* Loop over srcBLen */ 00512 k = srcBLen; 00513 00514 while (k > 0U) 00515 { 00516 /* Perform the multiply-accumulate */ 00517 sum += ((q31_t) * px++ * *py++); 00518 00519 /* Decrement the loop counter */ 00520 k--; 00521 } 00522 00523 /* Store the result in the accumulator in the destination buffer. */ 00524 *pOut = (q15_t) (sum >> 15); 00525 /* Destination pointer is updated according to the address modifier, inc */ 00526 pOut += inc; 00527 00528 /* Increment the MAC count */ 00529 count++; 00530 00531 /* Update the inputA and inputB pointers for next MAC calculation */ 00532 px = pIn1 + count; 00533 py = pIn2; 00534 00535 /* Decrement the loop counter */ 00536 blkCnt--; 00537 } 00538 } 00539 00540 /* -------------------------- 00541 * Initializations of stage3 00542 * -------------------------*/ 00543 00544 /* sum += x[srcALen-srcBLen+1] * y[0] + x[srcALen-srcBLen+2] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] 00545 * sum += x[srcALen-srcBLen+2] * y[0] + x[srcALen-srcBLen+3] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] 00546 * .... 00547 * sum += x[srcALen-2] * y[0] + x[srcALen-1] * y[1] 00548 * sum += x[srcALen-1] * y[0] 00549 */ 00550 00551 /* In this stage the MAC operations are decreased by 1 for every iteration. 00552 The count variable holds the number of MAC operations performed */ 00553 count = srcBLen - 1U; 00554 00555 /* Working pointer of inputA */ 00556 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1U); 00557 px = pSrc1; 00558 00559 /* Working pointer of inputB */ 00560 py = pIn2; 00561 00562 /* ------------------- 00563 * Stage3 process 00564 * ------------------*/ 00565 00566 while (blockSize3 > 0U) 00567 { 00568 /* Accumulator is made zero for every iteration */ 00569 sum = 0; 00570 00571 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00572 k = count >> 2U; 00573 00574 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00575 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00576 while (k > 0U) 00577 { 00578 /* Perform the multiply-accumulates */ 00579 /* sum += x[srcALen - srcBLen + 4] * y[3] , sum += x[srcALen - srcBLen + 3] * y[2] */ 00580 sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00581 /* sum += x[srcALen - srcBLen + 2] * y[1] , sum += x[srcALen - srcBLen + 1] * y[0] */ 00582 sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00583 00584 /* Decrement the loop counter */ 00585 k--; 00586 } 00587 00588 /* If the count is not a multiple of 4, compute any remaining MACs here. 00589 ** No loop unrolling is used. */ 00590 k = count % 0x4U; 00591 00592 while (k > 0U) 00593 { 00594 /* Perform the multiply-accumulates */ 00595 sum = __SMLAD(*px++, *py++, sum); 00596 00597 /* Decrement the loop counter */ 00598 k--; 00599 } 00600 00601 /* Store the result in the accumulator in the destination buffer. */ 00602 *pOut = (q15_t) (sum >> 15); 00603 /* Destination pointer is updated according to the address modifier, inc */ 00604 pOut += inc; 00605 00606 /* Update the inputA and inputB pointers for next MAC calculation */ 00607 px = ++pSrc1; 00608 py = pIn2; 00609 00610 /* Decrement the MAC count */ 00611 count--; 00612 00613 /* Decrement the loop counter */ 00614 blockSize3--; 00615 } 00616 00617 #else 00618 00619 q15_t *pIn1; /* inputA pointer */ 00620 q15_t *pIn2; /* inputB pointer */ 00621 q15_t *pOut = pDst; /* output pointer */ 00622 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulators */ 00623 q15_t *px; /* Intermediate inputA pointer */ 00624 q15_t *py; /* Intermediate inputB pointer */ 00625 q15_t *pSrc1; /* Intermediate pointers */ 00626 q31_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */ 00627 uint32_t j, k = 0U, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counter */ 00628 int32_t inc = 1; /* Destination address modifier */ 00629 q15_t a, b; 00630 00631 00632 /* The algorithm implementation is based on the lengths of the inputs. */ 00633 /* srcB is always made to slide across srcA. */ 00634 /* So srcBLen is always considered as shorter or equal to srcALen */ 00635 /* But CORR(x, y) is reverse of CORR(y, x) */ 00636 /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */ 00637 /* and the destination pointer modifier, inc is set to -1 */ 00638 /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */ 00639 /* But to improve the performance, 00640 * we include zeroes in the output instead of zero padding either of the the inputs*/ 00641 /* If srcALen > srcBLen, 00642 * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */ 00643 /* If srcALen < srcBLen, 00644 * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */ 00645 if (srcALen >= srcBLen) 00646 { 00647 /* Initialization of inputA pointer */ 00648 pIn1 = (pSrcA); 00649 00650 /* Initialization of inputB pointer */ 00651 pIn2 = (pSrcB); 00652 00653 /* Number of output samples is calculated */ 00654 outBlockSize = (2U * srcALen) - 1U; 00655 00656 /* When srcALen > srcBLen, zero padding is done to srcB 00657 * to make their lengths equal. 00658 * Instead, (outBlockSize - (srcALen + srcBLen - 1)) 00659 * number of output samples are made zero */ 00660 j = outBlockSize - (srcALen + (srcBLen - 1U)); 00661 00662 /* Updating the pointer position to non zero value */ 00663 pOut += j; 00664 00665 } 00666 else 00667 { 00668 /* Initialization of inputA pointer */ 00669 pIn1 = (pSrcB); 00670 00671 /* Initialization of inputB pointer */ 00672 pIn2 = (pSrcA); 00673 00674 /* srcBLen is always considered as shorter or equal to srcALen */ 00675 j = srcBLen; 00676 srcBLen = srcALen; 00677 srcALen = j; 00678 00679 /* CORR(x, y) = Reverse order(CORR(y, x)) */ 00680 /* Hence set the destination pointer to point to the last output sample */ 00681 pOut = pDst + ((srcALen + srcBLen) - 2U); 00682 00683 /* Destination address modifier is set to -1 */ 00684 inc = -1; 00685 00686 } 00687 00688 /* The function is internally 00689 * divided into three parts according to the number of multiplications that has to be 00690 * taken place between inputA samples and inputB samples. In the first part of the 00691 * algorithm, the multiplications increase by one for every iteration. 00692 * In the second part of the algorithm, srcBLen number of multiplications are done. 00693 * In the third part of the algorithm, the multiplications decrease by one 00694 * for every iteration.*/ 00695 /* The algorithm is implemented in three stages. 00696 * The loop counters of each stage is initiated here. */ 00697 blockSize1 = srcBLen - 1U; 00698 blockSize2 = srcALen - (srcBLen - 1U); 00699 blockSize3 = blockSize1; 00700 00701 /* -------------------------- 00702 * Initializations of stage1 00703 * -------------------------*/ 00704 00705 /* sum = x[0] * y[srcBlen - 1] 00706 * sum = x[0] * y[srcBlen - 2] + x[1] * y[srcBlen - 1] 00707 * .... 00708 * sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen - 1] * y[srcBLen - 1] 00709 */ 00710 00711 /* In this stage the MAC operations are increased by 1 for every iteration. 00712 The count variable holds the number of MAC operations performed */ 00713 count = 1U; 00714 00715 /* Working pointer of inputA */ 00716 px = pIn1; 00717 00718 /* Working pointer of inputB */ 00719 pSrc1 = pIn2 + (srcBLen - 1U); 00720 py = pSrc1; 00721 00722 /* ------------------------ 00723 * Stage1 process 00724 * ----------------------*/ 00725 00726 /* The first loop starts here */ 00727 while (blockSize1 > 0U) 00728 { 00729 /* Accumulator is made zero for every iteration */ 00730 sum = 0; 00731 00732 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00733 k = count >> 2; 00734 00735 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00736 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00737 while (k > 0U) 00738 { 00739 /* x[0] * y[srcBLen - 4] , x[1] * y[srcBLen - 3] */ 00740 sum += ((q31_t) * px++ * *py++); 00741 sum += ((q31_t) * px++ * *py++); 00742 sum += ((q31_t) * px++ * *py++); 00743 sum += ((q31_t) * px++ * *py++); 00744 00745 /* Decrement the loop counter */ 00746 k--; 00747 } 00748 00749 /* If the count is not a multiple of 4, compute any remaining MACs here. 00750 ** No loop unrolling is used. */ 00751 k = count % 0x4U; 00752 00753 while (k > 0U) 00754 { 00755 /* Perform the multiply-accumulates */ 00756 /* x[0] * y[srcBLen - 1] */ 00757 sum += ((q31_t) * px++ * *py++); 00758 00759 /* Decrement the loop counter */ 00760 k--; 00761 } 00762 00763 /* Store the result in the accumulator in the destination buffer. */ 00764 *pOut = (q15_t) (sum >> 15); 00765 /* Destination pointer is updated according to the address modifier, inc */ 00766 pOut += inc; 00767 00768 /* Update the inputA and inputB pointers for next MAC calculation */ 00769 py = pSrc1 - count; 00770 px = pIn1; 00771 00772 /* Increment the MAC count */ 00773 count++; 00774 00775 /* Decrement the loop counter */ 00776 blockSize1--; 00777 } 00778 00779 /* -------------------------- 00780 * Initializations of stage2 00781 * ------------------------*/ 00782 00783 /* sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen-1] * y[srcBLen-1] 00784 * sum = x[1] * y[0] + x[2] * y[1] +...+ x[srcBLen] * y[srcBLen-1] 00785 * .... 00786 * sum = x[srcALen-srcBLen-2] * y[0] + x[srcALen-srcBLen-1] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] 00787 */ 00788 00789 /* Working pointer of inputA */ 00790 px = pIn1; 00791 00792 /* Working pointer of inputB */ 00793 py = pIn2; 00794 00795 /* count is index by which the pointer pIn1 to be incremented */ 00796 count = 0U; 00797 00798 /* ------------------- 00799 * Stage2 process 00800 * ------------------*/ 00801 00802 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed. 00803 * So, to loop unroll over blockSize2, 00804 * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */ 00805 if (srcBLen >= 4U) 00806 { 00807 /* Loop unroll over blockSize2, by 4 */ 00808 blkCnt = blockSize2 >> 2U; 00809 00810 while (blkCnt > 0U) 00811 { 00812 /* Set all accumulators to zero */ 00813 acc0 = 0; 00814 acc1 = 0; 00815 acc2 = 0; 00816 acc3 = 0; 00817 00818 /* read x[0], x[1], x[2] samples */ 00819 a = *px; 00820 b = *(px + 1); 00821 00822 #ifndef ARM_MATH_BIG_ENDIAN 00823 00824 x0 = __PKHBT(a, b, 16); 00825 a = *(px + 2); 00826 x1 = __PKHBT(b, a, 16); 00827 00828 #else 00829 00830 x0 = __PKHBT(b, a, 16); 00831 a = *(px + 2); 00832 x1 = __PKHBT(a, b, 16); 00833 00834 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ 00835 00836 px += 2U; 00837 00838 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00839 k = srcBLen >> 2U; 00840 00841 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00842 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00843 do 00844 { 00845 /* Read the first two inputB samples using SIMD: 00846 * y[0] and y[1] */ 00847 a = *py; 00848 b = *(py + 1); 00849 00850 #ifndef ARM_MATH_BIG_ENDIAN 00851 00852 c0 = __PKHBT(a, b, 16); 00853 00854 #else 00855 00856 c0 = __PKHBT(b, a, 16); 00857 00858 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ 00859 00860 /* acc0 += x[0] * y[0] + x[1] * y[1] */ 00861 acc0 = __SMLAD(x0, c0, acc0); 00862 00863 /* acc1 += x[1] * y[0] + x[2] * y[1] */ 00864 acc1 = __SMLAD(x1, c0, acc1); 00865 00866 /* Read x[2], x[3], x[4] */ 00867 a = *px; 00868 b = *(px + 1); 00869 00870 #ifndef ARM_MATH_BIG_ENDIAN 00871 00872 x2 = __PKHBT(a, b, 16); 00873 a = *(px + 2); 00874 x3 = __PKHBT(b, a, 16); 00875 00876 #else 00877 00878 x2 = __PKHBT(b, a, 16); 00879 a = *(px + 2); 00880 x3 = __PKHBT(a, b, 16); 00881 00882 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ 00883 00884 /* acc2 += x[2] * y[0] + x[3] * y[1] */ 00885 acc2 = __SMLAD(x2, c0, acc2); 00886 00887 /* acc3 += x[3] * y[0] + x[4] * y[1] */ 00888 acc3 = __SMLAD(x3, c0, acc3); 00889 00890 /* Read y[2] and y[3] */ 00891 a = *(py + 2); 00892 b = *(py + 3); 00893 00894 py += 4U; 00895 00896 #ifndef ARM_MATH_BIG_ENDIAN 00897 00898 c0 = __PKHBT(a, b, 16); 00899 00900 #else 00901 00902 c0 = __PKHBT(b, a, 16); 00903 00904 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ 00905 00906 /* acc0 += x[2] * y[2] + x[3] * y[3] */ 00907 acc0 = __SMLAD(x2, c0, acc0); 00908 00909 /* acc1 += x[3] * y[2] + x[4] * y[3] */ 00910 acc1 = __SMLAD(x3, c0, acc1); 00911 00912 /* Read x[4], x[5], x[6] */ 00913 a = *(px + 2); 00914 b = *(px + 3); 00915 00916 #ifndef ARM_MATH_BIG_ENDIAN 00917 00918 x0 = __PKHBT(a, b, 16); 00919 a = *(px + 4); 00920 x1 = __PKHBT(b, a, 16); 00921 00922 #else 00923 00924 x0 = __PKHBT(b, a, 16); 00925 a = *(px + 4); 00926 x1 = __PKHBT(a, b, 16); 00927 00928 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ 00929 00930 px += 4U; 00931 00932 /* acc2 += x[4] * y[2] + x[5] * y[3] */ 00933 acc2 = __SMLAD(x0, c0, acc2); 00934 00935 /* acc3 += x[5] * y[2] + x[6] * y[3] */ 00936 acc3 = __SMLAD(x1, c0, acc3); 00937 00938 } while (--k); 00939 00940 /* For the next MAC operations, SIMD is not used 00941 * So, the 16 bit pointer if inputB, py is updated */ 00942 00943 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. 00944 ** No loop unrolling is used. */ 00945 k = srcBLen % 0x4U; 00946 00947 if (k == 1U) 00948 { 00949 /* Read y[4] */ 00950 c0 = *py; 00951 #ifdef ARM_MATH_BIG_ENDIAN 00952 00953 c0 = c0 << 16U; 00954 00955 #else 00956 00957 c0 = c0 & 0x0000FFFF; 00958 00959 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ 00960 00961 /* Read x[7] */ 00962 a = *px; 00963 b = *(px + 1); 00964 00965 px++;; 00966 00967 #ifndef ARM_MATH_BIG_ENDIAN 00968 00969 x3 = __PKHBT(a, b, 16); 00970 00971 #else 00972 00973 x3 = __PKHBT(b, a, 16); 00974 00975 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ 00976 00977 px++; 00978 00979 /* Perform the multiply-accumulates */ 00980 acc0 = __SMLAD(x0, c0, acc0); 00981 acc1 = __SMLAD(x1, c0, acc1); 00982 acc2 = __SMLADX(x1, c0, acc2); 00983 acc3 = __SMLADX(x3, c0, acc3); 00984 } 00985 00986 if (k == 2U) 00987 { 00988 /* Read y[4], y[5] */ 00989 a = *py; 00990 b = *(py + 1); 00991 00992 #ifndef ARM_MATH_BIG_ENDIAN 00993 00994 c0 = __PKHBT(a, b, 16); 00995 00996 #else 00997 00998 c0 = __PKHBT(b, a, 16); 00999 01000 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ 01001 01002 /* Read x[7], x[8], x[9] */ 01003 a = *px; 01004 b = *(px + 1); 01005 01006 #ifndef ARM_MATH_BIG_ENDIAN 01007 01008 x3 = __PKHBT(a, b, 16); 01009 a = *(px + 2); 01010 x2 = __PKHBT(b, a, 16); 01011 01012 #else 01013 01014 x3 = __PKHBT(b, a, 16); 01015 a = *(px + 2); 01016 x2 = __PKHBT(a, b, 16); 01017 01018 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ 01019 01020 px += 2U; 01021 01022 /* Perform the multiply-accumulates */ 01023 acc0 = __SMLAD(x0, c0, acc0); 01024 acc1 = __SMLAD(x1, c0, acc1); 01025 acc2 = __SMLAD(x3, c0, acc2); 01026 acc3 = __SMLAD(x2, c0, acc3); 01027 } 01028 01029 if (k == 3U) 01030 { 01031 /* Read y[4], y[5] */ 01032 a = *py; 01033 b = *(py + 1); 01034 01035 #ifndef ARM_MATH_BIG_ENDIAN 01036 01037 c0 = __PKHBT(a, b, 16); 01038 01039 #else 01040 01041 c0 = __PKHBT(b, a, 16); 01042 01043 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ 01044 01045 py += 2U; 01046 01047 /* Read x[7], x[8], x[9] */ 01048 a = *px; 01049 b = *(px + 1); 01050 01051 #ifndef ARM_MATH_BIG_ENDIAN 01052 01053 x3 = __PKHBT(a, b, 16); 01054 a = *(px + 2); 01055 x2 = __PKHBT(b, a, 16); 01056 01057 #else 01058 01059 x3 = __PKHBT(b, a, 16); 01060 a = *(px + 2); 01061 x2 = __PKHBT(a, b, 16); 01062 01063 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ 01064 01065 /* Perform the multiply-accumulates */ 01066 acc0 = __SMLAD(x0, c0, acc0); 01067 acc1 = __SMLAD(x1, c0, acc1); 01068 acc2 = __SMLAD(x3, c0, acc2); 01069 acc3 = __SMLAD(x2, c0, acc3); 01070 01071 c0 = (*py); 01072 /* Read y[6] */ 01073 #ifdef ARM_MATH_BIG_ENDIAN 01074 01075 c0 = c0 << 16U; 01076 #else 01077 01078 c0 = c0 & 0x0000FFFF; 01079 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ 01080 01081 /* Read x[10] */ 01082 b = *(px + 3); 01083 01084 #ifndef ARM_MATH_BIG_ENDIAN 01085 01086 x3 = __PKHBT(a, b, 16); 01087 01088 #else 01089 01090 x3 = __PKHBT(b, a, 16); 01091 01092 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ 01093 01094 px += 3U; 01095 01096 /* Perform the multiply-accumulates */ 01097 acc0 = __SMLADX(x1, c0, acc0); 01098 acc1 = __SMLAD(x2, c0, acc1); 01099 acc2 = __SMLADX(x2, c0, acc2); 01100 acc3 = __SMLADX(x3, c0, acc3); 01101 } 01102 01103 /* Store the result in the accumulator in the destination buffer. */ 01104 *pOut = (q15_t) (acc0 >> 15); 01105 /* Destination pointer is updated according to the address modifier, inc */ 01106 pOut += inc; 01107 01108 *pOut = (q15_t) (acc1 >> 15); 01109 pOut += inc; 01110 01111 *pOut = (q15_t) (acc2 >> 15); 01112 pOut += inc; 01113 01114 *pOut = (q15_t) (acc3 >> 15); 01115 pOut += inc; 01116 01117 /* Increment the pointer pIn1 index, count by 1 */ 01118 count += 4U; 01119 01120 /* Update the inputA and inputB pointers for next MAC calculation */ 01121 px = pIn1 + count; 01122 py = pIn2; 01123 01124 01125 /* Decrement the loop counter */ 01126 blkCnt--; 01127 } 01128 01129 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here. 01130 ** No loop unrolling is used. */ 01131 blkCnt = blockSize2 % 0x4U; 01132 01133 while (blkCnt > 0U) 01134 { 01135 /* Accumulator is made zero for every iteration */ 01136 sum = 0; 01137 01138 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 01139 k = srcBLen >> 2U; 01140 01141 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 01142 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 01143 while (k > 0U) 01144 { 01145 /* Perform the multiply-accumulates */ 01146 sum += ((q31_t) * px++ * *py++); 01147 sum += ((q31_t) * px++ * *py++); 01148 sum += ((q31_t) * px++ * *py++); 01149 sum += ((q31_t) * px++ * *py++); 01150 01151 /* Decrement the loop counter */ 01152 k--; 01153 } 01154 01155 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. 01156 ** No loop unrolling is used. */ 01157 k = srcBLen % 0x4U; 01158 01159 while (k > 0U) 01160 { 01161 /* Perform the multiply-accumulates */ 01162 sum += ((q31_t) * px++ * *py++); 01163 01164 /* Decrement the loop counter */ 01165 k--; 01166 } 01167 01168 /* Store the result in the accumulator in the destination buffer. */ 01169 *pOut = (q15_t) (sum >> 15); 01170 /* Destination pointer is updated according to the address modifier, inc */ 01171 pOut += inc; 01172 01173 /* Increment the pointer pIn1 index, count by 1 */ 01174 count++; 01175 01176 /* Update the inputA and inputB pointers for next MAC calculation */ 01177 px = pIn1 + count; 01178 py = pIn2; 01179 01180 /* Decrement the loop counter */ 01181 blkCnt--; 01182 } 01183 } 01184 else 01185 { 01186 /* If the srcBLen is not a multiple of 4, 01187 * the blockSize2 loop cannot be unrolled by 4 */ 01188 blkCnt = blockSize2; 01189 01190 while (blkCnt > 0U) 01191 { 01192 /* Accumulator is made zero for every iteration */ 01193 sum = 0; 01194 01195 /* Loop over srcBLen */ 01196 k = srcBLen; 01197 01198 while (k > 0U) 01199 { 01200 /* Perform the multiply-accumulate */ 01201 sum += ((q31_t) * px++ * *py++); 01202 01203 /* Decrement the loop counter */ 01204 k--; 01205 } 01206 01207 /* Store the result in the accumulator in the destination buffer. */ 01208 *pOut = (q15_t) (sum >> 15); 01209 /* Destination pointer is updated according to the address modifier, inc */ 01210 pOut += inc; 01211 01212 /* Increment the MAC count */ 01213 count++; 01214 01215 /* Update the inputA and inputB pointers for next MAC calculation */ 01216 px = pIn1 + count; 01217 py = pIn2; 01218 01219 /* Decrement the loop counter */ 01220 blkCnt--; 01221 } 01222 } 01223 01224 /* -------------------------- 01225 * Initializations of stage3 01226 * -------------------------*/ 01227 01228 /* sum += x[srcALen-srcBLen+1] * y[0] + x[srcALen-srcBLen+2] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] 01229 * sum += x[srcALen-srcBLen+2] * y[0] + x[srcALen-srcBLen+3] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] 01230 * .... 01231 * sum += x[srcALen-2] * y[0] + x[srcALen-1] * y[1] 01232 * sum += x[srcALen-1] * y[0] 01233 */ 01234 01235 /* In this stage the MAC operations are decreased by 1 for every iteration. 01236 The count variable holds the number of MAC operations performed */ 01237 count = srcBLen - 1U; 01238 01239 /* Working pointer of inputA */ 01240 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1U); 01241 px = pSrc1; 01242 01243 /* Working pointer of inputB */ 01244 py = pIn2; 01245 01246 /* ------------------- 01247 * Stage3 process 01248 * ------------------*/ 01249 01250 while (blockSize3 > 0U) 01251 { 01252 /* Accumulator is made zero for every iteration */ 01253 sum = 0; 01254 01255 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 01256 k = count >> 2U; 01257 01258 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 01259 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 01260 while (k > 0U) 01261 { 01262 /* Perform the multiply-accumulates */ 01263 sum += ((q31_t) * px++ * *py++); 01264 sum += ((q31_t) * px++ * *py++); 01265 sum += ((q31_t) * px++ * *py++); 01266 sum += ((q31_t) * px++ * *py++); 01267 01268 /* Decrement the loop counter */ 01269 k--; 01270 } 01271 01272 /* If the count is not a multiple of 4, compute any remaining MACs here. 01273 ** No loop unrolling is used. */ 01274 k = count % 0x4U; 01275 01276 while (k > 0U) 01277 { 01278 /* Perform the multiply-accumulates */ 01279 sum += ((q31_t) * px++ * *py++); 01280 01281 /* Decrement the loop counter */ 01282 k--; 01283 } 01284 01285 /* Store the result in the accumulator in the destination buffer. */ 01286 *pOut = (q15_t) (sum >> 15); 01287 /* Destination pointer is updated according to the address modifier, inc */ 01288 pOut += inc; 01289 01290 /* Update the inputA and inputB pointers for next MAC calculation */ 01291 px = ++pSrc1; 01292 py = pIn2; 01293 01294 /* Decrement the MAC count */ 01295 count--; 01296 01297 /* Decrement the loop counter */ 01298 blockSize3--; 01299 } 01300 01301 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ 01302 01303 } 01304 01305 /** 01306 * @} end of Corr group 01307 */ 01308
Generated on Tue Jul 12 2022 16:46:23 by
1.7.2