00001 /* ---------------------------------------------------------------------- 00002 * Copyright (C) 2010 ARM Limited. All rights reserved. 00003 * 00004 * $Date: 15. July 2011 00005 * $Revision: V1.0.10 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 1.0.10 2011/7/15 00015 * Big Endian support added and Merged M0 and M3/M4 Source code. 00016 * 00017 * Version 1.0.3 2010/11/29 00018 * Re-organized the CMSIS folders and updated documentation. 00019 * 00020 * Version 1.0.2 2010/11/11 00021 * Documentation updated. 00022 * 00023 * Version 1.0.1 2010/10/05 00024 * Production release and review comments incorporated. 00025 * 00026 * Version 1.0.0 2010/09/20 00027 * Production release and review comments incorporated. 00028 * -------------------------------------------------------------------- */ 00029 00030 #include "arm_math.h" 00031 00066 void arm_correlate_fast_q15( 00067 q15_t * pSrcA, 00068 uint32_t srcALen, 00069 q15_t * pSrcB, 00070 uint32_t srcBLen, 00071 q15_t * pDst) 00072 { 00073 q15_t *pIn1; /* inputA pointer */ 00074 q15_t *pIn2; /* inputB pointer */ 00075 q15_t *pOut = pDst; /* output pointer */ 00076 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulators */ 00077 q15_t *px; /* Intermediate inputA pointer */ 00078 q15_t *py; /* Intermediate inputB pointer */ 00079 q15_t *pSrc1; /* Intermediate pointers */ 00080 q31_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */ 00081 uint32_t j, k = 0u, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counter */ 00082 int32_t inc = 1; /* Destination address modifier */ 00083 q31_t *pb; /* 32 bit pointer for inputB buffer */ 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 /* Initialize inputB pointer of type q31 */ 00249 pb = (q31_t *) (py); 00250 00251 /* count is index by which the pointer pIn1 to be incremented */ 00252 count = 0u; 00253 00254 /* ------------------- 00255 * Stage2 process 00256 * ------------------*/ 00257 00258 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed. 00259 * So, to loop unroll over blockSize2, 00260 * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */ 00261 if(srcBLen >= 4u) 00262 { 00263 /* Loop unroll over blockSize2, by 4 */ 00264 blkCnt = blockSize2 >> 2u; 00265 00266 while(blkCnt > 0u) 00267 { 00268 /* Set all accumulators to zero */ 00269 acc0 = 0; 00270 acc1 = 0; 00271 acc2 = 0; 00272 acc3 = 0; 00273 00274 /* read x[0], x[1] samples */ 00275 x0 = *(q31_t *) (px++); 00276 /* read x[1], x[2] samples */ 00277 x1 = *(q31_t *) (px++); 00278 00279 /* Apply loop unrolling and compute 4 MACs simultaneously. */ 00280 k = srcBLen >> 2u; 00281 00282 /* First part of the processing with loop unrolling. Compute 4 MACs at a time. 00283 ** a second loop below computes MACs for the remaining 1 to 3 samples. */ 00284 do 00285 { 00286 /* Read the first two inputB samples using SIMD: 00287 * y[0] and y[1] */ 00288 c0 = *(pb++); 00289 00290 /* acc0 += x[0] * y[0] + x[1] * y[1] */ 00291 acc0 = __SMLAD(x0, c0, acc0); 00292 00293 /* acc1 += x[1] * y[0] + x[2] * y[1] */ 00294 acc1 = __SMLAD(x1, c0, acc1); 00295 00296 /* Read x[2], x[3] */ 00297 x2 = *(q31_t *) (px++); 00298 00299 /* Read x[3], x[4] */ 00300 x3 = *(q31_t *) (px++); 00301 00302 /* acc2 += x[2] * y[0] + x[3] * y[1] */ 00303 acc2 = __SMLAD(x2, c0, acc2); 00304 00305 /* acc3 += x[3] * y[0] + x[4] * y[1] */ 00306 acc3 = __SMLAD(x3, c0, acc3); 00307 00308 /* Read y[2] and y[3] */ 00309 c0 = *(pb++); 00310 00311 /* acc0 += x[2] * y[2] + x[3] * y[3] */ 00312 acc0 = __SMLAD(x2, c0, acc0); 00313 00314 /* acc1 += x[3] * y[2] + x[4] * y[3] */ 00315 acc1 = __SMLAD(x3, c0, acc1); 00316 00317 /* Read x[4], x[5] */ 00318 x0 = *(q31_t *) (px++); 00319 00320 /* Read x[5], x[6] */ 00321 x1 = *(q31_t *) (px++); 00322 00323 /* acc2 += x[4] * y[2] + x[5] * y[3] */ 00324 acc2 = __SMLAD(x0, c0, acc2); 00325 00326 /* acc3 += x[5] * y[2] + x[6] * y[3] */ 00327 acc3 = __SMLAD(x1, c0, acc3); 00328 00329 } while(--k); 00330 00331 /* For the next MAC operations, SIMD is not used 00332 * So, the 16 bit pointer if inputB, py is updated */ 00333 py = (q15_t *) (pb); 00334 00335 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. 00336 ** No loop unrolling is used. */ 00337 k = srcBLen % 0x4u; 00338 00339 if(k == 1u) 00340 { 00341 /* Read y[4] */ 00342 c0 = *py; 00343 #ifdef ARM_MATH_BIG_ENDIAN 00344 00345 c0 = c0 << 16u; 00346 00347 #else 00348 00349 c0 = c0 & 0x0000FFFF; 00350 00351 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ 00352 00353 /* Read x[7] */ 00354 x3 = *(q31_t *) px++; 00355 00356 /* Perform the multiply-accumulates */ 00357 acc0 = __SMLAD(x0, c0, acc0); 00358 acc1 = __SMLAD(x1, c0, acc1); 00359 acc2 = __SMLADX(x1, c0, acc2); 00360 acc3 = __SMLADX(x3, c0, acc3); 00361 } 00362 00363 if(k == 2u) 00364 { 00365 /* Read y[4], y[5] */ 00366 c0 = *(pb); 00367 00368 /* Read x[7], x[8] */ 00369 x3 = *(q31_t *) px++; 00370 00371 /* Read x[9] */ 00372 x2 = *(q31_t *) px++; 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 = *pb++; 00385 00386 /* Read x[7], x[8] */ 00387 x3 = *(q31_t *) px++; 00388 00389 /* Read x[9] */ 00390 x2 = *(q31_t *) px++; 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 /* Read y[6] */ 00399 #ifdef ARM_MATH_BIG_ENDIAN 00400 c0 = (*pb); 00401 c0 = c0 & 0xFFFF0000; 00402 00403 #else 00404 c0 = (q15_t) (*pb); 00405 c0 = c0 & 0x0000FFFF; 00406 00407 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ 00408 00409 /* Read x[10] */ 00410 x3 = *(q31_t *) px++; 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 pb = (q31_t *) (py); 00440 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 += ((q31_t) * px++ * *py++); 00464 sum += ((q31_t) * px++ * *py++); 00465 sum += ((q31_t) * px++ * *py++); 00466 sum += ((q31_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 += ((q31_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) (sum >> 15); 00487 /* Destination pointer is updated according to the address modifier, inc */ 00488 pOut += inc; 00489 00490 /* Increment the pointer pIn1 index, count by 1 */ 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 += ((q31_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) (sum >> 15); 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 = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); 00582 /* sum += x[srcALen - srcBLen + 2] * y[1] , sum += x[srcALen - srcBLen + 1] * y[0] */ 00583 sum = __SMLAD(*__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 = __SMLAD(*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) (sum >> 15); 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 } 00619