Robert Lopez
Aded CMSIS5 DSP and NN folder. Needs some work
Embed:
(wiki syntax)
Show/hide line numbers
arm_correlate_q15.c
00001 /* ---------------------------------------------------------------------- 00002 * Project: CMSIS DSP Library 00003 * Title: arm_correlate_q15.c 00004 * Description: Correlation of Q15 sequences 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. 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 * @details 00050 * <b>Scaling and Overflow Behavior:</b> 00051 * 00052 * \par 00053 * The function is implemented using a 64-bit internal accumulator. 00054 * Both inputs are in 1.15 format and multiplications yield a 2.30 result. 00055 * The 2.30 intermediate results are accumulated in a 64-bit accumulator in 34.30 format. 00056 * This approach provides 33 guard bits and there is no risk of overflow. 00057 * The 34.30 result is then truncated to 34.15 format by discarding the low 15 bits and then saturated to 1.15 format. 00058 * 00059 * \par 00060 * Refer to <code>arm_correlate_fast_q15()</code> for a faster but less precise version of this function for Cortex-M3 and Cortex-M4. 00061 * 00062 * \par 00063 * Refer the function <code>arm_correlate_opt_q15()</code> for a faster implementation of this function using scratch buffers. 00064 * 00065 */ 00066 00067 void arm_correlate_q15( 00068 q15_t * pSrcA, 00069 uint32_t srcALen, 00070 q15_t * pSrcB, 00071 uint32_t srcBLen, 00072 q15_t * pDst) 00073 { 00074 00075 #if (defined(ARM_MATH_CM7) || defined(ARM_MATH_CM4) || defined(ARM_MATH_CM3)) && !defined(UNALIGNED_SUPPORT_DISABLE) 00076 00077 /* Run the below code for Cortex-M4 and Cortex-M3 */ 00078 00079 q15_t *pIn1; /* inputA pointer */ 00080 q15_t *pIn2; /* inputB pointer */ 00081 q15_t *pOut = pDst; /* output pointer */ 00082 q63_t sum, acc0, acc1, acc2, acc3; /* Accumulators */ 00083 q15_t *px; /* Intermediate inputA pointer */ 00084 q15_t *py; /* Intermediate inputB pointer */ 00085 q15_t *pSrc1; /* Intermediate pointers */ 00086 q31_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */ 00087 uint32_t j, k = 0U, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counter */ 00088 int32_t inc = 1; /* Destination address modifier */ 00089 00090 00091 /* The algorithm implementation is based on the lengths of the inputs. */ 00092 /* srcB is always made to slide across srcA. */ 00093 /* So srcBLen is always considered as shorter or equal to srcALen */ 00094 /* But CORR(x, y) is reverse of CORR(y, x) */ 00095 /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */ 00096 /* and the destination pointer modifier, inc is set to -1 */ 00097 /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */ 00098 /* But to improve the performance, 00099 * we include zeroes in the output instead of zero padding either of the the inputs*/ 00100 /* If srcALen > srcBLen, 00101 * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */ 00102 /* If srcALen < srcBLen, 00103 * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */ 00104 if (srcALen >= srcBLen) 00105 { 00106 /* Initialization of inputA pointer */ 00107 pIn1 = (pSrcA); 00108 00109 /* Initialization of inputB pointer */ 00110 pIn2 = (pSrcB); 00111 00112 /* Number of output samples is calculated */ 00113 outBlockSize = (2U * srcALen) - 1U; 00114 00115 /* When srcALen > srcBLen, zero padding is done to srcB 00116 * to make their lengths equal. 00117 * Instead, (outBlockSize - (srcALen + srcBLen - 1)) 00118 * number of output samples are made zero */ 00119 j = outBlockSize - (srcALen + (srcBLen - 1U)); 00120 00121 /* Updating the pointer position to non zero value */ 00122 pOut += j; 00123 00124 } 00125 else 00126 { 00127 /* Initialization of inputA pointer */ 00128 pIn1 = (pSrcB); 00129 00130 /* Initialization of inputB pointer */ 00131 pIn2 = (pSrcA); 00132 00133 /* srcBLen is always considered as shorter or equal to srcALen */ 00134 j = srcBLen; 00135 srcBLen = srcALen; 00136 srcALen = j; 00137 00138 /* CORR(x, y) = Reverse order(CORR(y, x)) */ 00139 /* Hence set the destination pointer to point to the last output sample */ 00140 pOut = pDst + ((srcALen + srcBLen) - 2U); 00141 00142 /* Destination address modifier is set to -1 */ 00143 inc = -1; 00144 00145 } 00146 00147 /* The function is internally 00148 * divided into three parts according to the number of multiplications that has to be 00149 * taken place between inputA samples and inputB samples. In the first part of the 00150 * algorithm, the multiplications increase by one for every iteration. 00151 * In the second part of the algorithm, srcBLen number of multiplications are done. 00152 * In the third part of the algorithm, the multiplications decrease by one 00153 * for every iteration.*/ 00154 /* The algorithm is implemented in three stages. 00155 * The loop counters of each stage is initiated here. */ 00156 blockSize1 = srcBLen - 1U; 00157 blockSize2 = srcALen - (srcBLen - 1U); 00158 blockSize3 = blockSize1; 00159 00160 /* -------------------------- 00161 * Initializations of stage1 00162 * -------------------------*/ 00163 00164 /* sum = x[0] * y[srcBlen - 1] 00165 * sum = x[0] * y[srcBlen - 2] + x[1] * y[srcBlen - 1] 00166 * .... 00167 * sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen - 1] * y[srcBLen - 1] 00168 */ 00169 00170 /* In this stage the MAC operations are increased by 1 for every iteration. 00171 The count variable holds the number of MAC operations performed */ 00172 count = 1U; 00173 00174 /* Working pointer of inputA */ 00175 px = pIn1; 00176 00177 /* Working pointer of inputB */ 00178 pSrc1 = pIn2 + (srcBLen - 1U); 00179 py = pSrc1; 00180 00181 /* ------------------------ 00182 * Stage1 process 00183 * ----------------------*/ 00184 00185 /* The first loop starts here */ 00186 while (blockSize1 > 0U) 00187 { 00188 /* Accumulator is made zero for every iteration */ 00189 sum = 0; 00190 00191 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00192 k = count >> 2; 00193 00194 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00195 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00196 while (k > 0U) 00197 { 00198 /* x[0] * y[srcBLen - 4] , x[1] * y[srcBLen - 3] */ 00199 sum = __SMLALD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00200 /* x[3] * y[srcBLen - 1] , x[2] * y[srcBLen - 2] */ 00201 sum = __SMLALD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00202 00203 /* Decrement the loop counter */ 00204 k--; 00205 } 00206 00207 /* If the count is not a multiple of 4, compute any remaining MACs here. 00208 ** No loop unrolling is used. */ 00209 k = count % 0x4U; 00210 00211 while (k > 0U) 00212 { 00213 /* Perform the multiply-accumulates */ 00214 /* x[0] * y[srcBLen - 1] */ 00215 sum = __SMLALD(*px++, *py++, sum); 00216 00217 /* Decrement the loop counter */ 00218 k--; 00219 } 00220 00221 /* Store the result in the accumulator in the destination buffer. */ 00222 *pOut = (q15_t) (__SSAT((sum >> 15), 16)); 00223 /* Destination pointer is updated according to the address modifier, inc */ 00224 pOut += inc; 00225 00226 /* Update the inputA and inputB pointers for next MAC calculation */ 00227 py = pSrc1 - count; 00228 px = pIn1; 00229 00230 /* Increment the MAC count */ 00231 count++; 00232 00233 /* Decrement the loop counter */ 00234 blockSize1--; 00235 } 00236 00237 /* -------------------------- 00238 * Initializations of stage2 00239 * ------------------------*/ 00240 00241 /* sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen-1] * y[srcBLen-1] 00242 * sum = x[1] * y[0] + x[2] * y[1] +...+ x[srcBLen] * y[srcBLen-1] 00243 * .... 00244 * sum = x[srcALen-srcBLen-2] * y[0] + x[srcALen-srcBLen-1] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] 00245 */ 00246 00247 /* Working pointer of inputA */ 00248 px = pIn1; 00249 00250 /* Working pointer of inputB */ 00251 py = pIn2; 00252 00253 /* count is index by which the pointer pIn1 to be incremented */ 00254 count = 0U; 00255 00256 /* ------------------- 00257 * Stage2 process 00258 * ------------------*/ 00259 00260 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed. 00261 * So, to loop unroll over blockSize2, 00262 * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */ 00263 if (srcBLen >= 4U) 00264 { 00265 /* Loop unroll over blockSize2, by 4 */ 00266 blkCnt = blockSize2 >> 2U; 00267 00268 while (blkCnt > 0U) 00269 { 00270 /* Set all accumulators to zero */ 00271 acc0 = 0; 00272 acc1 = 0; 00273 acc2 = 0; 00274 acc3 = 0; 00275 00276 /* read x[0], x[1] samples */ 00277 x0 = *__SIMD32(px); 00278 /* read x[1], x[2] samples */ 00279 x1 = _SIMD32_OFFSET(px + 1); 00280 px += 2U; 00281 00282 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00283 k = srcBLen >> 2U; 00284 00285 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00286 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00287 do 00288 { 00289 /* Read the first two inputB samples using SIMD: 00290 * y[0] and y[1] */ 00291 c0 = *__SIMD32(py)++; 00292 00293 /* acc0 += x[0] * y[0] + x[1] * y[1] */ 00294 acc0 = __SMLALD(x0, c0, acc0); 00295 00296 /* acc1 += x[1] * y[0] + x[2] * y[1] */ 00297 acc1 = __SMLALD(x1, c0, acc1); 00298 00299 /* Read x[2], x[3] */ 00300 x2 = *__SIMD32(px); 00301 00302 /* Read x[3], x[4] */ 00303 x3 = _SIMD32_OFFSET(px + 1); 00304 00305 /* acc2 += x[2] * y[0] + x[3] * y[1] */ 00306 acc2 = __SMLALD(x2, c0, acc2); 00307 00308 /* acc3 += x[3] * y[0] + x[4] * y[1] */ 00309 acc3 = __SMLALD(x3, c0, acc3); 00310 00311 /* Read y[2] and y[3] */ 00312 c0 = *__SIMD32(py)++; 00313 00314 /* acc0 += x[2] * y[2] + x[3] * y[3] */ 00315 acc0 = __SMLALD(x2, c0, acc0); 00316 00317 /* acc1 += x[3] * y[2] + x[4] * y[3] */ 00318 acc1 = __SMLALD(x3, c0, acc1); 00319 00320 /* Read x[4], x[5] */ 00321 x0 = _SIMD32_OFFSET(px + 2); 00322 00323 /* Read x[5], x[6] */ 00324 x1 = _SIMD32_OFFSET(px + 3); 00325 00326 px += 4U; 00327 00328 /* acc2 += x[4] * y[2] + x[5] * y[3] */ 00329 acc2 = __SMLALD(x0, c0, acc2); 00330 00331 /* acc3 += x[5] * y[2] + x[6] * y[3] */ 00332 acc3 = __SMLALD(x1, c0, acc3); 00333 00334 } while (--k); 00335 00336 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. 00337 ** No loop unrolling is used. */ 00338 k = srcBLen % 0x4U; 00339 00340 if (k == 1U) 00341 { 00342 /* Read y[4] */ 00343 c0 = *py; 00344 #ifdef ARM_MATH_BIG_ENDIAN 00345 00346 c0 = c0 << 16U; 00347 00348 #else 00349 00350 c0 = c0 & 0x0000FFFF; 00351 00352 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ 00353 /* Read x[7] */ 00354 x3 = *__SIMD32(px); 00355 px++; 00356 00357 /* Perform the multiply-accumulates */ 00358 acc0 = __SMLALD(x0, c0, acc0); 00359 acc1 = __SMLALD(x1, c0, acc1); 00360 acc2 = __SMLALDX(x1, c0, acc2); 00361 acc3 = __SMLALDX(x3, c0, acc3); 00362 } 00363 00364 if (k == 2U) 00365 { 00366 /* Read y[4], y[5] */ 00367 c0 = *__SIMD32(py); 00368 00369 /* Read x[7], x[8] */ 00370 x3 = *__SIMD32(px); 00371 00372 /* Read x[9] */ 00373 x2 = _SIMD32_OFFSET(px + 1); 00374 px += 2U; 00375 00376 /* Perform the multiply-accumulates */ 00377 acc0 = __SMLALD(x0, c0, acc0); 00378 acc1 = __SMLALD(x1, c0, acc1); 00379 acc2 = __SMLALD(x3, c0, acc2); 00380 acc3 = __SMLALD(x2, c0, acc3); 00381 } 00382 00383 if (k == 3U) 00384 { 00385 /* Read y[4], y[5] */ 00386 c0 = *__SIMD32(py)++; 00387 00388 /* Read x[7], x[8] */ 00389 x3 = *__SIMD32(px); 00390 00391 /* Read x[9] */ 00392 x2 = _SIMD32_OFFSET(px + 1); 00393 00394 /* Perform the multiply-accumulates */ 00395 acc0 = __SMLALD(x0, c0, acc0); 00396 acc1 = __SMLALD(x1, c0, acc1); 00397 acc2 = __SMLALD(x3, c0, acc2); 00398 acc3 = __SMLALD(x2, c0, acc3); 00399 00400 c0 = (*py); 00401 00402 /* Read y[6] */ 00403 #ifdef ARM_MATH_BIG_ENDIAN 00404 00405 c0 = c0 << 16U; 00406 #else 00407 00408 c0 = c0 & 0x0000FFFF; 00409 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ 00410 /* Read x[10] */ 00411 x3 = _SIMD32_OFFSET(px + 2); 00412 px += 3U; 00413 00414 /* Perform the multiply-accumulates */ 00415 acc0 = __SMLALDX(x1, c0, acc0); 00416 acc1 = __SMLALD(x2, c0, acc1); 00417 acc2 = __SMLALDX(x2, c0, acc2); 00418 acc3 = __SMLALDX(x3, c0, acc3); 00419 } 00420 00421 /* Store the result in the accumulator in the destination buffer. */ 00422 *pOut = (q15_t) (__SSAT(acc0 >> 15, 16)); 00423 /* Destination pointer is updated according to the address modifier, inc */ 00424 pOut += inc; 00425 00426 *pOut = (q15_t) (__SSAT(acc1 >> 15, 16)); 00427 pOut += inc; 00428 00429 *pOut = (q15_t) (__SSAT(acc2 >> 15, 16)); 00430 pOut += inc; 00431 00432 *pOut = (q15_t) (__SSAT(acc3 >> 15, 16)); 00433 pOut += inc; 00434 00435 /* Increment the count by 4 as 4 output values are computed */ 00436 count += 4U; 00437 00438 /* Update the inputA and inputB pointers for next MAC calculation */ 00439 px = pIn1 + count; 00440 py = pIn2; 00441 00442 /* Decrement the loop counter */ 00443 blkCnt--; 00444 } 00445 00446 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here. 00447 ** No loop unrolling is used. */ 00448 blkCnt = blockSize2 % 0x4U; 00449 00450 while (blkCnt > 0U) 00451 { 00452 /* Accumulator is made zero for every iteration */ 00453 sum = 0; 00454 00455 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00456 k = srcBLen >> 2U; 00457 00458 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00459 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00460 while (k > 0U) 00461 { 00462 /* Perform the multiply-accumulates */ 00463 sum += ((q63_t) * px++ * *py++); 00464 sum += ((q63_t) * px++ * *py++); 00465 sum += ((q63_t) * px++ * *py++); 00466 sum += ((q63_t) * px++ * *py++); 00467 00468 /* Decrement the loop counter */ 00469 k--; 00470 } 00471 00472 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. 00473 ** No loop unrolling is used. */ 00474 k = srcBLen % 0x4U; 00475 00476 while (k > 0U) 00477 { 00478 /* Perform the multiply-accumulates */ 00479 sum += ((q63_t) * px++ * *py++); 00480 00481 /* Decrement the loop counter */ 00482 k--; 00483 } 00484 00485 /* Store the result in the accumulator in the destination buffer. */ 00486 *pOut = (q15_t) (__SSAT(sum >> 15, 16)); 00487 /* Destination pointer is updated according to the address modifier, inc */ 00488 pOut += inc; 00489 00490 /* Increment count by 1, as one output value is computed */ 00491 count++; 00492 00493 /* Update the inputA and inputB pointers for next MAC calculation */ 00494 px = pIn1 + count; 00495 py = pIn2; 00496 00497 /* Decrement the loop counter */ 00498 blkCnt--; 00499 } 00500 } 00501 else 00502 { 00503 /* If the srcBLen is not a multiple of 4, 00504 * the blockSize2 loop cannot be unrolled by 4 */ 00505 blkCnt = blockSize2; 00506 00507 while (blkCnt > 0U) 00508 { 00509 /* Accumulator is made zero for every iteration */ 00510 sum = 0; 00511 00512 /* Loop over srcBLen */ 00513 k = srcBLen; 00514 00515 while (k > 0U) 00516 { 00517 /* Perform the multiply-accumulate */ 00518 sum += ((q63_t) * px++ * *py++); 00519 00520 /* Decrement the loop counter */ 00521 k--; 00522 } 00523 00524 /* Store the result in the accumulator in the destination buffer. */ 00525 *pOut = (q15_t) (__SSAT(sum >> 15, 16)); 00526 /* Destination pointer is updated according to the address modifier, inc */ 00527 pOut += inc; 00528 00529 /* Increment the MAC count */ 00530 count++; 00531 00532 /* Update the inputA and inputB pointers for next MAC calculation */ 00533 px = pIn1 + count; 00534 py = pIn2; 00535 00536 /* Decrement the loop counter */ 00537 blkCnt--; 00538 } 00539 } 00540 00541 /* -------------------------- 00542 * Initializations of stage3 00543 * -------------------------*/ 00544 00545 /* sum += x[srcALen-srcBLen+1] * y[0] + x[srcALen-srcBLen+2] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] 00546 * sum += x[srcALen-srcBLen+2] * y[0] + x[srcALen-srcBLen+3] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] 00547 * .... 00548 * sum += x[srcALen-2] * y[0] + x[srcALen-1] * y[1] 00549 * sum += x[srcALen-1] * y[0] 00550 */ 00551 00552 /* In this stage the MAC operations are decreased by 1 for every iteration. 00553 The count variable holds the number of MAC operations performed */ 00554 count = srcBLen - 1U; 00555 00556 /* Working pointer of inputA */ 00557 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1U); 00558 px = pSrc1; 00559 00560 /* Working pointer of inputB */ 00561 py = pIn2; 00562 00563 /* ------------------- 00564 * Stage3 process 00565 * ------------------*/ 00566 00567 while (blockSize3 > 0U) 00568 { 00569 /* Accumulator is made zero for every iteration */ 00570 sum = 0; 00571 00572 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00573 k = count >> 2U; 00574 00575 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00576 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00577 while (k > 0U) 00578 { 00579 /* Perform the multiply-accumulates */ 00580 /* sum += x[srcALen - srcBLen + 4] * y[3] , sum += x[srcALen - srcBLen + 3] * y[2] */ 00581 sum = __SMLALD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00582 /* sum += x[srcALen - srcBLen + 2] * y[1] , sum += x[srcALen - srcBLen + 1] * y[0] */ 00583 sum = __SMLALD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00584 00585 /* Decrement the loop counter */ 00586 k--; 00587 } 00588 00589 /* If the count is not a multiple of 4, compute any remaining MACs here. 00590 ** No loop unrolling is used. */ 00591 k = count % 0x4U; 00592 00593 while (k > 0U) 00594 { 00595 /* Perform the multiply-accumulates */ 00596 sum = __SMLALD(*px++, *py++, sum); 00597 00598 /* Decrement the loop counter */ 00599 k--; 00600 } 00601 00602 /* Store the result in the accumulator in the destination buffer. */ 00603 *pOut = (q15_t) (__SSAT((sum >> 15), 16)); 00604 /* Destination pointer is updated according to the address modifier, inc */ 00605 pOut += inc; 00606 00607 /* Update the inputA and inputB pointers for next MAC calculation */ 00608 px = ++pSrc1; 00609 py = pIn2; 00610 00611 /* Decrement the MAC count */ 00612 count--; 00613 00614 /* Decrement the loop counter */ 00615 blockSize3--; 00616 } 00617 00618 #else 00619 00620 /* Run the below code for Cortex-M0 */ 00621 00622 q15_t *pIn1 = pSrcA; /* inputA pointer */ 00623 q15_t *pIn2 = pSrcB + (srcBLen - 1U); /* inputB pointer */ 00624 q63_t sum; /* Accumulators */ 00625 uint32_t i = 0U, j; /* loop counters */ 00626 uint32_t inv = 0U; /* Reverse order flag */ 00627 uint32_t tot = 0U; /* Length */ 00628 00629 /* The algorithm implementation is based on the lengths of the inputs. */ 00630 /* srcB is always made to slide across srcA. */ 00631 /* So srcBLen is always considered as shorter or equal to srcALen */ 00632 /* But CORR(x, y) is reverse of CORR(y, x) */ 00633 /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */ 00634 /* and a varaible, inv is set to 1 */ 00635 /* If lengths are not equal then zero pad has to be done to make the two 00636 * inputs of same length. But to improve the performance, we include zeroes 00637 * in the output instead of zero padding either of the the inputs*/ 00638 /* If srcALen > srcBLen, (srcALen - srcBLen) zeroes has to included in the 00639 * starting of the output buffer */ 00640 /* If srcALen < srcBLen, (srcALen - srcBLen) zeroes has to included in the 00641 * ending of the output buffer */ 00642 /* Once the zero padding is done the remaining of the output is calcualted 00643 * using convolution but with the shorter signal time shifted. */ 00644 00645 /* Calculate the length of the remaining sequence */ 00646 tot = ((srcALen + srcBLen) - 2U); 00647 00648 if (srcALen > srcBLen) 00649 { 00650 /* Calculating the number of zeros to be padded to the output */ 00651 j = srcALen - srcBLen; 00652 00653 /* Initialise the pointer after zero padding */ 00654 pDst += j; 00655 } 00656 00657 else if (srcALen < srcBLen) 00658 { 00659 /* Initialization to inputB pointer */ 00660 pIn1 = pSrcB; 00661 00662 /* Initialization to the end of inputA pointer */ 00663 pIn2 = pSrcA + (srcALen - 1U); 00664 00665 /* Initialisation of the pointer after zero padding */ 00666 pDst = pDst + tot; 00667 00668 /* Swapping the lengths */ 00669 j = srcALen; 00670 srcALen = srcBLen; 00671 srcBLen = j; 00672 00673 /* Setting the reverse flag */ 00674 inv = 1; 00675 00676 } 00677 00678 /* Loop to calculate convolution for output length number of times */ 00679 for (i = 0U; i <= tot; i++) 00680 { 00681 /* Initialize sum with zero to carry on MAC operations */ 00682 sum = 0; 00683 00684 /* Loop to perform MAC operations according to convolution equation */ 00685 for (j = 0U; j <= i; j++) 00686 { 00687 /* Check the array limitations */ 00688 if ((((i - j) < srcBLen) && (j < srcALen))) 00689 { 00690 /* z[i] += x[i-j] * y[j] */ 00691 sum += ((q31_t) pIn1[j] * pIn2[-((int32_t) i - j)]); 00692 } 00693 } 00694 /* Store the output in the destination buffer */ 00695 if (inv == 1) 00696 *pDst-- = (q15_t) __SSAT((sum >> 15U), 16U); 00697 else 00698 *pDst++ = (q15_t) __SSAT((sum >> 15U), 16U); 00699 } 00700 00701 #endif /* #if (defined(ARM_MATH_CM7) || defined(ARM_MATH_CM4) || defined(ARM_MATH_CM3)) && !defined(UNALIGNED_SUPPORT_DISABLE) */ 00702 00703 } 00704 00705 /** 00706 * @} end of Corr group 00707 */ 00708
Generated on Tue Jul 12 2022 16:46:23 by
1.7.2