CMSIS DSP Library from CMSIS 2.0. See http://www.onarm.com/cmsis/ for full details
Dependents: K22F_DSP_Matrix_least_square BNO055-ELEC3810 1BNO055 ECE4180Project--Slave2 ... more
arm_correlate_fast_q15.c
00001 /* ---------------------------------------------------------------------- 00002 * Copyright (C) 2010 ARM Limited. All rights reserved. 00003 * 00004 * $Date: 29. November 2010 00005 * $Revision: V1.0.3 00006 * 00007 * Project: CMSIS DSP Library 00008 * Title: arm_correlate_fast_q15.c 00009 * 00010 * Description: Fast Q15 Correlation. 00011 * 00012 * Target Processor: Cortex-M4/Cortex-M3 00013 * 00014 * Version 1.0.3 2010/11/29 00015 * Re-organized the CMSIS folders and updated documentation. 00016 * 00017 * Version 1.0.2 2010/11/11 00018 * Documentation updated. 00019 * 00020 * Version 1.0.1 2010/10/05 00021 * Production release and review comments incorporated. 00022 * 00023 * Version 1.0.0 2010/09/20 00024 * Production release and review comments incorporated. 00025 * -------------------------------------------------------------------- */ 00026 00027 #include "arm_math.h" 00028 00029 /** 00030 * @ingroup groupFilters 00031 */ 00032 00033 /** 00034 * @addtogroup Corr 00035 * @{ 00036 */ 00037 00038 /** 00039 * @brief Correlation of Q15 sequences (fast version). 00040 * @param[in] *pSrcA points to the first input sequence. 00041 * @param[in] srcALen length of the first input sequence. 00042 * @param[in] *pSrcB points to the second input sequence. 00043 * @param[in] srcBLen length of the second input sequence. 00044 * @param[out] *pDst points to the location where the output result is written. Length 2 * max(srcALen, srcBLen) - 1. 00045 * @return none. 00046 * 00047 * <b>Scaling and Overflow Behavior:</b> 00048 * 00049 * \par 00050 * This fast version uses a 32-bit accumulator with 2.30 format. 00051 * The accumulator maintains full precision of the intermediate multiplication results but provides only a single guard bit. 00052 * There is no saturation on intermediate additions. 00053 * Thus, if the accumulator overflows it wraps around and distorts the result. 00054 * The input signals should be scaled down to avoid intermediate overflows. 00055 * Scale down one of the inputs by 1/min(srcALen, srcBLen) to avoid overflow since a 00056 * maximum of min(srcALen, srcBLen) number of additions is carried internally. 00057 * The 2.30 accumulator is right shifted by 15 bits and then saturated to 1.15 format to yield the final result. 00058 * 00059 * \par 00060 * See <code>arm_correlate_q15()</code> for a slower implementation of this function which uses a 64-bit accumulator to avoid wrap around distortion. 00061 */ 00062 00063 void arm_correlate_fast_q15( 00064 q15_t * pSrcA, 00065 uint32_t srcALen, 00066 q15_t * pSrcB, 00067 uint32_t srcBLen, 00068 q15_t * pDst) 00069 { 00070 q15_t *pIn1; /* inputA pointer */ 00071 q15_t *pIn2; /* inputB pointer */ 00072 q15_t *pOut = pDst; /* output pointer */ 00073 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulators */ 00074 q15_t *px; /* Intermediate inputA pointer */ 00075 q15_t *py; /* Intermediate inputB pointer */ 00076 q15_t *pSrc1; /* Intermediate pointers */ 00077 q31_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */ 00078 uint32_t j, k = 0u, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counter */ 00079 int32_t inc = 1; /* Destination address modifier */ 00080 q31_t *pb; /* 32 bit pointer for inputB buffer */ 00081 00082 00083 /* The algorithm implementation is based on the lengths of the inputs. */ 00084 /* srcB is always made to slide across srcA. */ 00085 /* So srcBLen is always considered as shorter or equal to srcALen */ 00086 /* But CORR(x, y) is reverse of CORR(y, x) */ 00087 /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */ 00088 /* and the destination pointer modifier, inc is set to -1 */ 00089 /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */ 00090 /* But to improve the performance, 00091 * we include zeroes in the output instead of zero padding either of the the inputs*/ 00092 /* If srcALen > srcBLen, 00093 * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */ 00094 /* If srcALen < srcBLen, 00095 * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */ 00096 if(srcALen >= srcBLen) 00097 { 00098 /* Initialization of inputA pointer */ 00099 pIn1 = (pSrcA); 00100 00101 /* Initialization of inputB pointer */ 00102 pIn2 = (pSrcB); 00103 00104 /* Number of output samples is calculated */ 00105 outBlockSize = (2u * srcALen) - 1u; 00106 00107 /* When srcALen > srcBLen, zero padding is done to srcB 00108 * to make their lengths equal. 00109 * Instead, (outBlockSize - (srcALen + srcBLen - 1)) 00110 * number of output samples are made zero */ 00111 j = outBlockSize - (srcALen + (srcBLen - 1u)); 00112 00113 while(j > 0u) 00114 { 00115 /* Zero is stored in the destination buffer */ 00116 *pOut++ = 0; 00117 00118 /* Decrement the loop counter */ 00119 j--; 00120 } 00121 00122 } 00123 else 00124 { 00125 /* Initialization of inputA pointer */ 00126 pIn1 = (pSrcB); 00127 00128 /* Initialization of inputB pointer */ 00129 pIn2 = (pSrcA); 00130 00131 /* srcBLen is always considered as shorter or equal to srcALen */ 00132 j = srcBLen; 00133 srcBLen = srcALen; 00134 srcALen = j; 00135 00136 /* CORR(x, y) = Reverse order(CORR(y, x)) */ 00137 /* Hence set the destination pointer to point to the last output sample */ 00138 pOut = pDst + ((srcALen + srcBLen) - 2u); 00139 00140 /* Destination address modifier is set to -1 */ 00141 inc = -1; 00142 00143 } 00144 00145 /* The function is internally 00146 * divided into three parts according to the number of multiplications that has to be 00147 * taken place between inputA samples and inputB samples. In the first part of the 00148 * algorithm, the multiplications increase by one for every iteration. 00149 * In the second part of the algorithm, srcBLen number of multiplications are done. 00150 * In the third part of the algorithm, the multiplications decrease by one 00151 * for every iteration.*/ 00152 /* The algorithm is implemented in three stages. 00153 * The loop counters of each stage is initiated here. */ 00154 blockSize1 = srcBLen - 1u; 00155 blockSize2 = srcALen - (srcBLen - 1u); 00156 blockSize3 = blockSize1; 00157 00158 /* -------------------------- 00159 * Initializations of stage1 00160 * -------------------------*/ 00161 00162 /* sum = x[0] * y[srcBlen - 1] 00163 * sum = x[0] * y[srcBlen - 2] + x[1] * y[srcBlen - 1] 00164 * .... 00165 * sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen - 1] * y[srcBLen - 1] 00166 */ 00167 00168 /* In this stage the MAC operations are increased by 1 for every iteration. 00169 The count variable holds the number of MAC operations performed */ 00170 count = 1u; 00171 00172 /* Working pointer of inputA */ 00173 px = pIn1; 00174 00175 /* Working pointer of inputB */ 00176 pSrc1 = pIn2 + (srcBLen - 1u); 00177 py = pSrc1; 00178 00179 /* ------------------------ 00180 * Stage1 process 00181 * ----------------------*/ 00182 00183 /* The first loop starts here */ 00184 while(blockSize1 > 0u) 00185 { 00186 /* Accumulator is made zero for every iteration */ 00187 sum = 0; 00188 00189 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00190 k = count >> 2; 00191 00192 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00193 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00194 while(k > 0u) 00195 { 00196 /* x[0] * y[srcBLen - 4] , x[1] * y[srcBLen - 3] */ 00197 sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00198 /* x[3] * y[srcBLen - 1] , x[2] * y[srcBLen - 2] */ 00199 sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00200 00201 /* Decrement the loop counter */ 00202 k--; 00203 } 00204 00205 /* If the count is not a multiple of 4, compute any remaining MACs here. 00206 ** No loop unrolling is used. */ 00207 k = count % 0x4u; 00208 00209 while(k > 0u) 00210 { 00211 /* Perform the multiply-accumulates */ 00212 /* x[0] * y[srcBLen - 1] */ 00213 sum = __SMLAD(*px++, *py++, sum); 00214 00215 /* Decrement the loop counter */ 00216 k--; 00217 } 00218 00219 /* Store the result in the accumulator in the destination buffer. */ 00220 *pOut = (q15_t) (sum >> 15); 00221 /* Destination pointer is updated according to the address modifier, inc */ 00222 pOut += inc; 00223 00224 /* Update the inputA and inputB pointers for next MAC calculation */ 00225 py = pSrc1 - count; 00226 px = pIn1; 00227 00228 /* Increment the MAC count */ 00229 count++; 00230 00231 /* Decrement the loop counter */ 00232 blockSize1--; 00233 } 00234 00235 /* -------------------------- 00236 * Initializations of stage2 00237 * ------------------------*/ 00238 00239 /* sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen-1] * y[srcBLen-1] 00240 * sum = x[1] * y[0] + x[2] * y[1] +...+ x[srcBLen] * y[srcBLen-1] 00241 * .... 00242 * sum = x[srcALen-srcBLen-2] * y[0] + x[srcALen-srcBLen-1] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] 00243 */ 00244 00245 /* Working pointer of inputA */ 00246 px = pIn1; 00247 00248 /* Working pointer of inputB */ 00249 py = pIn2; 00250 00251 /* Initialize inputB pointer of type q31 */ 00252 pb = (q31_t *) (py); 00253 00254 /* count is index by which the pointer pIn1 to be incremented */ 00255 count = 0u; 00256 00257 /* ------------------- 00258 * Stage2 process 00259 * ------------------*/ 00260 00261 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed. 00262 * So, to loop unroll over blockSize2, 00263 * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */ 00264 if(srcBLen >= 4u) 00265 { 00266 /* Loop unroll over blockSize2, by 4 */ 00267 blkCnt = blockSize2 >> 2u; 00268 00269 while(blkCnt > 0u) 00270 { 00271 /* Set all accumulators to zero */ 00272 acc0 = 0; 00273 acc1 = 0; 00274 acc2 = 0; 00275 acc3 = 0; 00276 00277 /* read x[0], x[1] samples */ 00278 x0 = *(q31_t *) (px++); 00279 /* read x[1], x[2] samples */ 00280 x1 = *(q31_t *) (px++); 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 = *(pb++); 00292 00293 /* acc0 += x[0] * y[0] + x[1] * y[1] */ 00294 acc0 = __SMLAD(x0, c0, acc0); 00295 00296 /* acc1 += x[1] * y[0] + x[2] * y[1] */ 00297 acc1 = __SMLAD(x1, c0, acc1); 00298 00299 /* Read x[2], x[3] */ 00300 x2 = *(q31_t *) (px++); 00301 00302 /* Read x[3], x[4] */ 00303 x3 = *(q31_t *) (px++); 00304 00305 /* acc2 += x[2] * y[0] + x[3] * y[1] */ 00306 acc2 = __SMLAD(x2, c0, acc2); 00307 00308 /* acc3 += x[3] * y[0] + x[4] * y[1] */ 00309 acc3 = __SMLAD(x3, c0, acc3); 00310 00311 /* Read y[2] and y[3] */ 00312 c0 = *(pb++); 00313 00314 /* acc0 += x[2] * y[2] + x[3] * y[3] */ 00315 acc0 = __SMLAD(x2, c0, acc0); 00316 00317 /* acc1 += x[3] * y[2] + x[4] * y[3] */ 00318 acc1 = __SMLAD(x3, c0, acc1); 00319 00320 /* Read x[4], x[5] */ 00321 x0 = *(q31_t *) (px++); 00322 00323 /* Read x[5], x[6] */ 00324 x1 = *(q31_t *) (px++); 00325 00326 /* acc2 += x[4] * y[2] + x[5] * y[3] */ 00327 acc2 = __SMLAD(x0, c0, acc2); 00328 00329 /* acc3 += x[5] * y[2] + x[6] * y[3] */ 00330 acc3 = __SMLAD(x1, c0, acc3); 00331 00332 } while(--k); 00333 00334 /* For the next MAC operations, SIMD is not used 00335 * So, the 16 bit pointer if inputB, py is updated */ 00336 py = (q15_t *) (pb); 00337 00338 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. 00339 ** No loop unrolling is used. */ 00340 k = srcBLen % 0x4u; 00341 00342 if(k == 1u) 00343 { 00344 /* Read y[4] */ 00345 c0 = *py; 00346 c0 = c0 & 0x0000FFFF; 00347 00348 /* Read x[7] */ 00349 x3 = *(q31_t *) px++; 00350 00351 /* Perform the multiply-accumulates */ 00352 acc0 = __SMLAD(x0, c0, acc0); 00353 acc1 = __SMLAD(x1, c0, acc1); 00354 acc2 = __SMLADX(x1, c0, acc2); 00355 acc3 = __SMLADX(x3, c0, acc3); 00356 } 00357 00358 if(k == 2u) 00359 { 00360 /* Read y[4], y[5] */ 00361 c0 = *(pb); 00362 00363 /* Read x[7], x[8] */ 00364 x3 = *(q31_t *) px++; 00365 00366 /* Read x[9] */ 00367 x2 = *(q31_t *) px++; 00368 00369 /* Perform the multiply-accumulates */ 00370 acc0 = __SMLAD(x0, c0, acc0); 00371 acc1 = __SMLAD(x1, c0, acc1); 00372 acc2 = __SMLAD(x3, c0, acc2); 00373 acc3 = __SMLAD(x2, c0, acc3); 00374 } 00375 00376 if(k == 3u) 00377 { 00378 /* Read y[4], y[5] */ 00379 c0 = *pb++; 00380 00381 /* Read x[7], x[8] */ 00382 x3 = *(q31_t *) px++; 00383 00384 /* Read x[9] */ 00385 x2 = *(q31_t *) px++; 00386 00387 /* Perform the multiply-accumulates */ 00388 acc0 = __SMLAD(x0, c0, acc0); 00389 acc1 = __SMLAD(x1, c0, acc1); 00390 acc2 = __SMLAD(x3, c0, acc2); 00391 acc3 = __SMLAD(x2, c0, acc3); 00392 00393 /* Read y[6] */ 00394 c0 = (q15_t) (*pb); 00395 c0 = c0 & 0x0000FFFF; 00396 00397 /* Read x[10] */ 00398 x3 = *(q31_t *) px++; 00399 00400 /* Perform the multiply-accumulates */ 00401 acc0 = __SMLADX(x1, c0, acc0); 00402 acc1 = __SMLAD(x2, c0, acc1); 00403 acc2 = __SMLADX(x2, c0, acc2); 00404 acc3 = __SMLADX(x3, c0, acc3); 00405 } 00406 00407 /* Store the result in the accumulator in the destination buffer. */ 00408 *pOut = (q15_t) (acc0 >> 15); 00409 /* Destination pointer is updated according to the address modifier, inc */ 00410 pOut += inc; 00411 00412 *pOut = (q15_t) (acc1 >> 15); 00413 pOut += inc; 00414 00415 *pOut = (q15_t) (acc2 >> 15); 00416 pOut += inc; 00417 00418 *pOut = (q15_t) (acc3 >> 15); 00419 pOut += inc; 00420 00421 /* Increment the pointer pIn1 index, count by 1 */ 00422 count += 4u; 00423 00424 /* Update the inputA and inputB pointers for next MAC calculation */ 00425 px = pIn1 + count; 00426 py = pIn2; 00427 pb = (q31_t *) (py); 00428 00429 00430 /* Decrement the loop counter */ 00431 blkCnt--; 00432 } 00433 00434 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here. 00435 ** No loop unrolling is used. */ 00436 blkCnt = blockSize2 % 0x4u; 00437 00438 while(blkCnt > 0u) 00439 { 00440 /* Accumulator is made zero for every iteration */ 00441 sum = 0; 00442 00443 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00444 k = srcBLen >> 2u; 00445 00446 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00447 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00448 while(k > 0u) 00449 { 00450 /* Perform the multiply-accumulates */ 00451 sum += ((q31_t) * px++ * *py++); 00452 sum += ((q31_t) * px++ * *py++); 00453 sum += ((q31_t) * px++ * *py++); 00454 sum += ((q31_t) * px++ * *py++); 00455 00456 /* Decrement the loop counter */ 00457 k--; 00458 } 00459 00460 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. 00461 ** No loop unrolling is used. */ 00462 k = srcBLen % 0x4u; 00463 00464 while(k > 0u) 00465 { 00466 /* Perform the multiply-accumulates */ 00467 sum += ((q31_t) * px++ * *py++); 00468 00469 /* Decrement the loop counter */ 00470 k--; 00471 } 00472 00473 /* Store the result in the accumulator in the destination buffer. */ 00474 *pOut = (q15_t) (sum >> 15); 00475 /* Destination pointer is updated according to the address modifier, inc */ 00476 pOut += inc; 00477 00478 /* Increment the pointer pIn1 index, count by 1 */ 00479 count++; 00480 00481 /* Update the inputA and inputB pointers for next MAC calculation */ 00482 px = pIn1 + count; 00483 py = pIn2; 00484 00485 /* Decrement the loop counter */ 00486 blkCnt--; 00487 } 00488 } 00489 else 00490 { 00491 /* If the srcBLen is not a multiple of 4, 00492 * the blockSize2 loop cannot be unrolled by 4 */ 00493 blkCnt = blockSize2; 00494 00495 while(blkCnt > 0u) 00496 { 00497 /* Accumulator is made zero for every iteration */ 00498 sum = 0; 00499 00500 /* Loop over srcBLen */ 00501 k = srcBLen; 00502 00503 while(k > 0u) 00504 { 00505 /* Perform the multiply-accumulate */ 00506 sum += ((q31_t) * px++ * *py++); 00507 00508 /* Decrement the loop counter */ 00509 k--; 00510 } 00511 00512 /* Store the result in the accumulator in the destination buffer. */ 00513 *pOut = (q15_t) (sum >> 15); 00514 /* Destination pointer is updated according to the address modifier, inc */ 00515 pOut += inc; 00516 00517 /* Increment the MAC count */ 00518 count++; 00519 00520 /* Update the inputA and inputB pointers for next MAC calculation */ 00521 px = pIn1 + count; 00522 py = pIn2; 00523 00524 /* Decrement the loop counter */ 00525 blkCnt--; 00526 } 00527 } 00528 00529 /* -------------------------- 00530 * Initializations of stage3 00531 * -------------------------*/ 00532 00533 /* sum += x[srcALen-srcBLen+1] * y[0] + x[srcALen-srcBLen+2] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] 00534 * sum += x[srcALen-srcBLen+2] * y[0] + x[srcALen-srcBLen+3] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] 00535 * .... 00536 * sum += x[srcALen-2] * y[0] + x[srcALen-1] * y[1] 00537 * sum += x[srcALen-1] * y[0] 00538 */ 00539 00540 /* In this stage the MAC operations are decreased by 1 for every iteration. 00541 The count variable holds the number of MAC operations performed */ 00542 count = srcBLen - 1u; 00543 00544 /* Working pointer of inputA */ 00545 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u); 00546 px = pSrc1; 00547 00548 /* Working pointer of inputB */ 00549 py = pIn2; 00550 00551 /* ------------------- 00552 * Stage3 process 00553 * ------------------*/ 00554 00555 while(blockSize3 > 0u) 00556 { 00557 /* Accumulator is made zero for every iteration */ 00558 sum = 0; 00559 00560 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00561 k = count >> 2u; 00562 00563 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00564 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00565 while(k > 0u) 00566 { 00567 /* Perform the multiply-accumulates */ 00568 /* sum += x[srcALen - srcBLen + 4] * y[3] , sum += x[srcALen - srcBLen + 3] * y[2] */ 00569 sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00570 /* sum += x[srcALen - srcBLen + 2] * y[1] , sum += x[srcALen - srcBLen + 1] * y[0] */ 00571 sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00572 00573 /* Decrement the loop counter */ 00574 k--; 00575 } 00576 00577 /* If the count is not a multiple of 4, compute any remaining MACs here. 00578 ** No loop unrolling is used. */ 00579 k = count % 0x4u; 00580 00581 while(k > 0u) 00582 { 00583 /* Perform the multiply-accumulates */ 00584 sum = __SMLAD(*px++, *py++, sum); 00585 00586 /* Decrement the loop counter */ 00587 k--; 00588 } 00589 00590 /* Store the result in the accumulator in the destination buffer. */ 00591 *pOut = (q15_t) (sum >> 15); 00592 /* Destination pointer is updated according to the address modifier, inc */ 00593 pOut += inc; 00594 00595 /* Update the inputA and inputB pointers for next MAC calculation */ 00596 px = ++pSrc1; 00597 py = pIn2; 00598 00599 /* Decrement the MAC count */ 00600 count--; 00601 00602 /* Decrement the loop counter */ 00603 blockSize3--; 00604 } 00605 00606 } 00607 00608 /** 00609 * @} end of Corr group 00610 */
Generated on Tue Jul 12 2022 14:13:52 by 1.7.2