mbed official / mbed-dsp

CMSIS DSP library

Dependents:   performance_timer Surfboard_ gps2rtty Capstone ... more

Legacy Warning

This is an mbed 2 library. To learn more about mbed OS 5, visit the docs.

cmsis_dsp/FilteringFunctions/arm_correlate_fast_q15.c@1:fdd22bb7aa52, 2012-11-28 (annotated)

Committer:
emilmont
Date:
Wed Nov 28 12:30:09 2012 +0000
Revision:
1:fdd22bb7aa52
Child:
2:da51fb522205
DSP library code

Who changed what in which revision?

UserRevisionLine numberNew contents of line
emilmont 1:fdd22bb7aa52 1 /* ----------------------------------------------------------------------
emilmont 1:fdd22bb7aa52 2 * Copyright (C) 2010 ARM Limited. All rights reserved.
emilmont 1:fdd22bb7aa52 3 *
emilmont 1:fdd22bb7aa52 4 * $Date: 15. February 2012
emilmont 1:fdd22bb7aa52 5 * $Revision: V1.1.0
emilmont 1:fdd22bb7aa52 6 *
emilmont 1:fdd22bb7aa52 7 * Project: CMSIS DSP Library
emilmont 1:fdd22bb7aa52 8 * Title: arm_correlate_fast_q15.c
emilmont 1:fdd22bb7aa52 9 *
emilmont 1:fdd22bb7aa52 10 * Description: Fast Q15 Correlation.
emilmont 1:fdd22bb7aa52 11 *
emilmont 1:fdd22bb7aa52 12 * Target Processor: Cortex-M4/Cortex-M3
emilmont 1:fdd22bb7aa52 13 *
emilmont 1:fdd22bb7aa52 14 * Version 1.1.0 2012/02/15
emilmont 1:fdd22bb7aa52 15 * Updated with more optimizations, bug fixes and minor API changes.
emilmont 1:fdd22bb7aa52 16 *
emilmont 1:fdd22bb7aa52 17 * Version 1.0.11 2011/10/18
emilmont 1:fdd22bb7aa52 18 * Bug Fix in conv, correlation, partial convolution.
emilmont 1:fdd22bb7aa52 19 *
emilmont 1:fdd22bb7aa52 20 * Version 1.0.10 2011/7/15
emilmont 1:fdd22bb7aa52 21 * Big Endian support added and Merged M0 and M3/M4 Source code.
emilmont 1:fdd22bb7aa52 22 *
emilmont 1:fdd22bb7aa52 23 * Version 1.0.3 2010/11/29
emilmont 1:fdd22bb7aa52 24 * Re-organized the CMSIS folders and updated documentation.
emilmont 1:fdd22bb7aa52 25 *
emilmont 1:fdd22bb7aa52 26 * Version 1.0.2 2010/11/11
emilmont 1:fdd22bb7aa52 27 * Documentation updated.
emilmont 1:fdd22bb7aa52 28 *
emilmont 1:fdd22bb7aa52 29 * Version 1.0.1 2010/10/05
emilmont 1:fdd22bb7aa52 30 * Production release and review comments incorporated.
emilmont 1:fdd22bb7aa52 31 *
emilmont 1:fdd22bb7aa52 32 * Version 1.0.0 2010/09/20
emilmont 1:fdd22bb7aa52 33 * Production release and review comments incorporated.
emilmont 1:fdd22bb7aa52 34 * -------------------------------------------------------------------- */
emilmont 1:fdd22bb7aa52 35
emilmont 1:fdd22bb7aa52 36 #include "arm_math.h"
emilmont 1:fdd22bb7aa52 37
emilmont 1:fdd22bb7aa52 38 /**
emilmont 1:fdd22bb7aa52 39 * @ingroup groupFilters
emilmont 1:fdd22bb7aa52 40 */
emilmont 1:fdd22bb7aa52 41
emilmont 1:fdd22bb7aa52 42 /**
emilmont 1:fdd22bb7aa52 43 * @addtogroup Corr
emilmont 1:fdd22bb7aa52 44 * @{
emilmont 1:fdd22bb7aa52 45 */
emilmont 1:fdd22bb7aa52 46
emilmont 1:fdd22bb7aa52 47 /**
emilmont 1:fdd22bb7aa52 48 * @brief Correlation of Q15 sequences (fast version) for Cortex-M3 and Cortex-M4.
emilmont 1:fdd22bb7aa52 49 * @param[in] *pSrcA points to the first input sequence.
emilmont 1:fdd22bb7aa52 50 * @param[in] srcALen length of the first input sequence.
emilmont 1:fdd22bb7aa52 51 * @param[in] *pSrcB points to the second input sequence.
emilmont 1:fdd22bb7aa52 52 * @param[in] srcBLen length of the second input sequence.
emilmont 1:fdd22bb7aa52 53 * @param[out] *pDst points to the location where the output result is written. Length 2 * max(srcALen, srcBLen) - 1.
emilmont 1:fdd22bb7aa52 54 * @return none.
emilmont 1:fdd22bb7aa52 55 *
emilmont 1:fdd22bb7aa52 56 * <b>Scaling and Overflow Behavior:</b>
emilmont 1:fdd22bb7aa52 57 *
emilmont 1:fdd22bb7aa52 58 * \par
emilmont 1:fdd22bb7aa52 59 * This fast version uses a 32-bit accumulator with 2.30 format.
emilmont 1:fdd22bb7aa52 60 * The accumulator maintains full precision of the intermediate multiplication results but provides only a single guard bit.
emilmont 1:fdd22bb7aa52 61 * There is no saturation on intermediate additions.
emilmont 1:fdd22bb7aa52 62 * Thus, if the accumulator overflows it wraps around and distorts the result.
emilmont 1:fdd22bb7aa52 63 * The input signals should be scaled down to avoid intermediate overflows.
emilmont 1:fdd22bb7aa52 64 * Scale down one of the inputs by 1/min(srcALen, srcBLen) to avoid overflow since a
emilmont 1:fdd22bb7aa52 65 * maximum of min(srcALen, srcBLen) number of additions is carried internally.
emilmont 1:fdd22bb7aa52 66 * The 2.30 accumulator is right shifted by 15 bits and then saturated to 1.15 format to yield the final result.
emilmont 1:fdd22bb7aa52 67 *
emilmont 1:fdd22bb7aa52 68 * \par
emilmont 1:fdd22bb7aa52 69 * See <code>arm_correlate_q15()</code> for a slower implementation of this function which uses a 64-bit accumulator to avoid wrap around distortion.
emilmont 1:fdd22bb7aa52 70 */
emilmont 1:fdd22bb7aa52 71
emilmont 1:fdd22bb7aa52 72 void arm_correlate_fast_q15(
emilmont 1:fdd22bb7aa52 73 q15_t * pSrcA,
emilmont 1:fdd22bb7aa52 74 uint32_t srcALen,
emilmont 1:fdd22bb7aa52 75 q15_t * pSrcB,
emilmont 1:fdd22bb7aa52 76 uint32_t srcBLen,
emilmont 1:fdd22bb7aa52 77 q15_t * pDst)
emilmont 1:fdd22bb7aa52 78 {
emilmont 1:fdd22bb7aa52 79 #ifndef UNALIGNED_SUPPORT_DISABLE
emilmont 1:fdd22bb7aa52 80
emilmont 1:fdd22bb7aa52 81 q15_t *pIn1; /* inputA pointer */
emilmont 1:fdd22bb7aa52 82 q15_t *pIn2; /* inputB pointer */
emilmont 1:fdd22bb7aa52 83 q15_t *pOut = pDst; /* output pointer */
emilmont 1:fdd22bb7aa52 84 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulators */
emilmont 1:fdd22bb7aa52 85 q15_t *px; /* Intermediate inputA pointer */
emilmont 1:fdd22bb7aa52 86 q15_t *py; /* Intermediate inputB pointer */
emilmont 1:fdd22bb7aa52 87 q15_t *pSrc1; /* Intermediate pointers */
emilmont 1:fdd22bb7aa52 88 q31_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */
emilmont 1:fdd22bb7aa52 89 uint32_t j, k = 0u, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counter */
emilmont 1:fdd22bb7aa52 90 int32_t inc = 1; /* Destination address modifier */
emilmont 1:fdd22bb7aa52 91
emilmont 1:fdd22bb7aa52 92
emilmont 1:fdd22bb7aa52 93 /* The algorithm implementation is based on the lengths of the inputs. */
emilmont 1:fdd22bb7aa52 94 /* srcB is always made to slide across srcA. */
emilmont 1:fdd22bb7aa52 95 /* So srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 96 /* But CORR(x, y) is reverse of CORR(y, x) */
emilmont 1:fdd22bb7aa52 97 /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */
emilmont 1:fdd22bb7aa52 98 /* and the destination pointer modifier, inc is set to -1 */
emilmont 1:fdd22bb7aa52 99 /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */
emilmont 1:fdd22bb7aa52 100 /* But to improve the performance,
emilmont 1:fdd22bb7aa52 101 * we include zeroes in the output instead of zero padding either of the the inputs*/
emilmont 1:fdd22bb7aa52 102 /* If srcALen > srcBLen,
emilmont 1:fdd22bb7aa52 103 * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */
emilmont 1:fdd22bb7aa52 104 /* If srcALen < srcBLen,
emilmont 1:fdd22bb7aa52 105 * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */
emilmont 1:fdd22bb7aa52 106 if(srcALen >= srcBLen)
emilmont 1:fdd22bb7aa52 107 {
emilmont 1:fdd22bb7aa52 108 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 109 pIn1 = (pSrcA);
emilmont 1:fdd22bb7aa52 110
emilmont 1:fdd22bb7aa52 111 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 112 pIn2 = (pSrcB);
emilmont 1:fdd22bb7aa52 113
emilmont 1:fdd22bb7aa52 114 /* Number of output samples is calculated */
emilmont 1:fdd22bb7aa52 115 outBlockSize = (2u * srcALen) - 1u;
emilmont 1:fdd22bb7aa52 116
emilmont 1:fdd22bb7aa52 117 /* When srcALen > srcBLen, zero padding is done to srcB
emilmont 1:fdd22bb7aa52 118 * to make their lengths equal.
emilmont 1:fdd22bb7aa52 119 * Instead, (outBlockSize - (srcALen + srcBLen - 1))
emilmont 1:fdd22bb7aa52 120 * number of output samples are made zero */
emilmont 1:fdd22bb7aa52 121 j = outBlockSize - (srcALen + (srcBLen - 1u));
emilmont 1:fdd22bb7aa52 122
emilmont 1:fdd22bb7aa52 123 /* Updating the pointer position to non zero value */
emilmont 1:fdd22bb7aa52 124 pOut += j;
emilmont 1:fdd22bb7aa52 125
emilmont 1:fdd22bb7aa52 126 }
emilmont 1:fdd22bb7aa52 127 else
emilmont 1:fdd22bb7aa52 128 {
emilmont 1:fdd22bb7aa52 129 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 130 pIn1 = (pSrcB);
emilmont 1:fdd22bb7aa52 131
emilmont 1:fdd22bb7aa52 132 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 133 pIn2 = (pSrcA);
emilmont 1:fdd22bb7aa52 134
emilmont 1:fdd22bb7aa52 135 /* srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 136 j = srcBLen;
emilmont 1:fdd22bb7aa52 137 srcBLen = srcALen;
emilmont 1:fdd22bb7aa52 138 srcALen = j;
emilmont 1:fdd22bb7aa52 139
emilmont 1:fdd22bb7aa52 140 /* CORR(x, y) = Reverse order(CORR(y, x)) */
emilmont 1:fdd22bb7aa52 141 /* Hence set the destination pointer to point to the last output sample */
emilmont 1:fdd22bb7aa52 142 pOut = pDst + ((srcALen + srcBLen) - 2u);
emilmont 1:fdd22bb7aa52 143
emilmont 1:fdd22bb7aa52 144 /* Destination address modifier is set to -1 */
emilmont 1:fdd22bb7aa52 145 inc = -1;
emilmont 1:fdd22bb7aa52 146
emilmont 1:fdd22bb7aa52 147 }
emilmont 1:fdd22bb7aa52 148
emilmont 1:fdd22bb7aa52 149 /* The function is internally
emilmont 1:fdd22bb7aa52 150 * divided into three parts according to the number of multiplications that has to be
emilmont 1:fdd22bb7aa52 151 * taken place between inputA samples and inputB samples. In the first part of the
emilmont 1:fdd22bb7aa52 152 * algorithm, the multiplications increase by one for every iteration.
emilmont 1:fdd22bb7aa52 153 * In the second part of the algorithm, srcBLen number of multiplications are done.
emilmont 1:fdd22bb7aa52 154 * In the third part of the algorithm, the multiplications decrease by one
emilmont 1:fdd22bb7aa52 155 * for every iteration.*/
emilmont 1:fdd22bb7aa52 156 /* The algorithm is implemented in three stages.
emilmont 1:fdd22bb7aa52 157 * The loop counters of each stage is initiated here. */
emilmont 1:fdd22bb7aa52 158 blockSize1 = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 159 blockSize2 = srcALen - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 160 blockSize3 = blockSize1;
emilmont 1:fdd22bb7aa52 161
emilmont 1:fdd22bb7aa52 162 /* --------------------------
emilmont 1:fdd22bb7aa52 163 * Initializations of stage1
emilmont 1:fdd22bb7aa52 164 * -------------------------*/
emilmont 1:fdd22bb7aa52 165
emilmont 1:fdd22bb7aa52 166 /* sum = x[0] * y[srcBlen - 1]
emilmont 1:fdd22bb7aa52 167 * sum = x[0] * y[srcBlen - 2] + x[1] * y[srcBlen - 1]
emilmont 1:fdd22bb7aa52 168 * ....
emilmont 1:fdd22bb7aa52 169 * sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen - 1] * y[srcBLen - 1]
emilmont 1:fdd22bb7aa52 170 */
emilmont 1:fdd22bb7aa52 171
emilmont 1:fdd22bb7aa52 172 /* In this stage the MAC operations are increased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 173 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 174 count = 1u;
emilmont 1:fdd22bb7aa52 175
emilmont 1:fdd22bb7aa52 176 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 177 px = pIn1;
emilmont 1:fdd22bb7aa52 178
emilmont 1:fdd22bb7aa52 179 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 180 pSrc1 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 181 py = pSrc1;
emilmont 1:fdd22bb7aa52 182
emilmont 1:fdd22bb7aa52 183 /* ------------------------
emilmont 1:fdd22bb7aa52 184 * Stage1 process
emilmont 1:fdd22bb7aa52 185 * ----------------------*/
emilmont 1:fdd22bb7aa52 186
emilmont 1:fdd22bb7aa52 187 /* The first loop starts here */
emilmont 1:fdd22bb7aa52 188 while(blockSize1 > 0u)
emilmont 1:fdd22bb7aa52 189 {
emilmont 1:fdd22bb7aa52 190 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 191 sum = 0;
emilmont 1:fdd22bb7aa52 192
emilmont 1:fdd22bb7aa52 193 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 194 k = count >> 2;
emilmont 1:fdd22bb7aa52 195
emilmont 1:fdd22bb7aa52 196 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 197 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 198 while(k > 0u)
emilmont 1:fdd22bb7aa52 199 {
emilmont 1:fdd22bb7aa52 200 /* x[0] * y[srcBLen - 4] , x[1] * y[srcBLen - 3] */
emilmont 1:fdd22bb7aa52 201 sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum);
emilmont 1:fdd22bb7aa52 202 /* x[3] * y[srcBLen - 1] , x[2] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 203 sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum);
emilmont 1:fdd22bb7aa52 204
emilmont 1:fdd22bb7aa52 205 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 206 k--;
emilmont 1:fdd22bb7aa52 207 }
emilmont 1:fdd22bb7aa52 208
emilmont 1:fdd22bb7aa52 209 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 210 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 211 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 212
emilmont 1:fdd22bb7aa52 213 while(k > 0u)
emilmont 1:fdd22bb7aa52 214 {
emilmont 1:fdd22bb7aa52 215 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 216 /* x[0] * y[srcBLen - 1] */
emilmont 1:fdd22bb7aa52 217 sum = __SMLAD(*px++, *py++, sum);
emilmont 1:fdd22bb7aa52 218
emilmont 1:fdd22bb7aa52 219 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 220 k--;
emilmont 1:fdd22bb7aa52 221 }
emilmont 1:fdd22bb7aa52 222
emilmont 1:fdd22bb7aa52 223 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 224 *pOut = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 225 /* Destination pointer is updated according to the address modifier, inc */
emilmont 1:fdd22bb7aa52 226 pOut += inc;
emilmont 1:fdd22bb7aa52 227
emilmont 1:fdd22bb7aa52 228 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 229 py = pSrc1 - count;
emilmont 1:fdd22bb7aa52 230 px = pIn1;
emilmont 1:fdd22bb7aa52 231
emilmont 1:fdd22bb7aa52 232 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 233 count++;
emilmont 1:fdd22bb7aa52 234
emilmont 1:fdd22bb7aa52 235 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 236 blockSize1--;
emilmont 1:fdd22bb7aa52 237 }
emilmont 1:fdd22bb7aa52 238
emilmont 1:fdd22bb7aa52 239 /* --------------------------
emilmont 1:fdd22bb7aa52 240 * Initializations of stage2
emilmont 1:fdd22bb7aa52 241 * ------------------------*/
emilmont 1:fdd22bb7aa52 242
emilmont 1:fdd22bb7aa52 243 /* sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 244 * sum = x[1] * y[0] + x[2] * y[1] +...+ x[srcBLen] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 245 * ....
emilmont 1:fdd22bb7aa52 246 * sum = x[srcALen-srcBLen-2] * y[0] + x[srcALen-srcBLen-1] * y[1] +...+ x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 247 */
emilmont 1:fdd22bb7aa52 248
emilmont 1:fdd22bb7aa52 249 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 250 px = pIn1;
emilmont 1:fdd22bb7aa52 251
emilmont 1:fdd22bb7aa52 252 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 253 py = pIn2;
emilmont 1:fdd22bb7aa52 254
emilmont 1:fdd22bb7aa52 255 /* count is index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 256 count = 0u;
emilmont 1:fdd22bb7aa52 257
emilmont 1:fdd22bb7aa52 258 /* -------------------
emilmont 1:fdd22bb7aa52 259 * Stage2 process
emilmont 1:fdd22bb7aa52 260 * ------------------*/
emilmont 1:fdd22bb7aa52 261
emilmont 1:fdd22bb7aa52 262 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 263 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 264 * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */
emilmont 1:fdd22bb7aa52 265 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 266 {
emilmont 1:fdd22bb7aa52 267 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 268 blkCnt = blockSize2 >> 2u;
emilmont 1:fdd22bb7aa52 269
emilmont 1:fdd22bb7aa52 270 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 271 {
emilmont 1:fdd22bb7aa52 272 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 273 acc0 = 0;
emilmont 1:fdd22bb7aa52 274 acc1 = 0;
emilmont 1:fdd22bb7aa52 275 acc2 = 0;
emilmont 1:fdd22bb7aa52 276 acc3 = 0;
emilmont 1:fdd22bb7aa52 277
emilmont 1:fdd22bb7aa52 278 /* read x[0], x[1] samples */
emilmont 1:fdd22bb7aa52 279 x0 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 280 /* read x[1], x[2] samples */
emilmont 1:fdd22bb7aa52 281 x1 = _SIMD32_OFFSET(px + 1);
emilmont 1:fdd22bb7aa52 282 px += 2u;
emilmont 1:fdd22bb7aa52 283
emilmont 1:fdd22bb7aa52 284 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 285 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 286
emilmont 1:fdd22bb7aa52 287 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 288 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 289 do
emilmont 1:fdd22bb7aa52 290 {
emilmont 1:fdd22bb7aa52 291 /* Read the first two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 292 * y[0] and y[1] */
emilmont 1:fdd22bb7aa52 293 c0 = *__SIMD32(py)++;
emilmont 1:fdd22bb7aa52 294
emilmont 1:fdd22bb7aa52 295 /* acc0 += x[0] * y[0] + x[1] * y[1] */
emilmont 1:fdd22bb7aa52 296 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 297
emilmont 1:fdd22bb7aa52 298 /* acc1 += x[1] * y[0] + x[2] * y[1] */
emilmont 1:fdd22bb7aa52 299 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 300
emilmont 1:fdd22bb7aa52 301 /* Read x[2], x[3] */
emilmont 1:fdd22bb7aa52 302 x2 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 303
emilmont 1:fdd22bb7aa52 304 /* Read x[3], x[4] */
emilmont 1:fdd22bb7aa52 305 x3 = _SIMD32_OFFSET(px + 1);
emilmont 1:fdd22bb7aa52 306
emilmont 1:fdd22bb7aa52 307 /* acc2 += x[2] * y[0] + x[3] * y[1] */
emilmont 1:fdd22bb7aa52 308 acc2 = __SMLAD(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 309
emilmont 1:fdd22bb7aa52 310 /* acc3 += x[3] * y[0] + x[4] * y[1] */
emilmont 1:fdd22bb7aa52 311 acc3 = __SMLAD(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 312
emilmont 1:fdd22bb7aa52 313 /* Read y[2] and y[3] */
emilmont 1:fdd22bb7aa52 314 c0 = *__SIMD32(py)++;
emilmont 1:fdd22bb7aa52 315
emilmont 1:fdd22bb7aa52 316 /* acc0 += x[2] * y[2] + x[3] * y[3] */
emilmont 1:fdd22bb7aa52 317 acc0 = __SMLAD(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 318
emilmont 1:fdd22bb7aa52 319 /* acc1 += x[3] * y[2] + x[4] * y[3] */
emilmont 1:fdd22bb7aa52 320 acc1 = __SMLAD(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 321
emilmont 1:fdd22bb7aa52 322 /* Read x[4], x[5] */
emilmont 1:fdd22bb7aa52 323 x0 = _SIMD32_OFFSET(px + 2);
emilmont 1:fdd22bb7aa52 324
emilmont 1:fdd22bb7aa52 325 /* Read x[5], x[6] */
emilmont 1:fdd22bb7aa52 326 x1 = _SIMD32_OFFSET(px + 3);
emilmont 1:fdd22bb7aa52 327 px += 4u;
emilmont 1:fdd22bb7aa52 328
emilmont 1:fdd22bb7aa52 329 /* acc2 += x[4] * y[2] + x[5] * y[3] */
emilmont 1:fdd22bb7aa52 330 acc2 = __SMLAD(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 331
emilmont 1:fdd22bb7aa52 332 /* acc3 += x[5] * y[2] + x[6] * y[3] */
emilmont 1:fdd22bb7aa52 333 acc3 = __SMLAD(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 334
emilmont 1:fdd22bb7aa52 335 } while(--k);
emilmont 1:fdd22bb7aa52 336
emilmont 1:fdd22bb7aa52 337 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 338 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 339
emilmont 1:fdd22bb7aa52 340 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 341 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 342 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 343
emilmont 1:fdd22bb7aa52 344 if(k == 1u)
emilmont 1:fdd22bb7aa52 345 {
emilmont 1:fdd22bb7aa52 346 /* Read y[4] */
emilmont 1:fdd22bb7aa52 347 c0 = *py;
emilmont 1:fdd22bb7aa52 348 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 349
emilmont 1:fdd22bb7aa52 350 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 351
emilmont 1:fdd22bb7aa52 352 #else
emilmont 1:fdd22bb7aa52 353
emilmont 1:fdd22bb7aa52 354 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 355
emilmont 1:fdd22bb7aa52 356 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 357
emilmont 1:fdd22bb7aa52 358 /* Read x[7] */
emilmont 1:fdd22bb7aa52 359 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 360 px++;
emilmont 1:fdd22bb7aa52 361
emilmont 1:fdd22bb7aa52 362 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 363 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 364 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 365 acc2 = __SMLADX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 366 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 367 }
emilmont 1:fdd22bb7aa52 368
emilmont 1:fdd22bb7aa52 369 if(k == 2u)
emilmont 1:fdd22bb7aa52 370 {
emilmont 1:fdd22bb7aa52 371 /* Read y[4], y[5] */
emilmont 1:fdd22bb7aa52 372 c0 = *__SIMD32(py);
emilmont 1:fdd22bb7aa52 373
emilmont 1:fdd22bb7aa52 374 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 375 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 376
emilmont 1:fdd22bb7aa52 377 /* Read x[9] */
emilmont 1:fdd22bb7aa52 378 x2 = _SIMD32_OFFSET(px + 1);
emilmont 1:fdd22bb7aa52 379 px += 2u;
emilmont 1:fdd22bb7aa52 380
emilmont 1:fdd22bb7aa52 381 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 382 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 383 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 384 acc2 = __SMLAD(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 385 acc3 = __SMLAD(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 386 }
emilmont 1:fdd22bb7aa52 387
emilmont 1:fdd22bb7aa52 388 if(k == 3u)
emilmont 1:fdd22bb7aa52 389 {
emilmont 1:fdd22bb7aa52 390 /* Read y[4], y[5] */
emilmont 1:fdd22bb7aa52 391 c0 = *__SIMD32(py)++;
emilmont 1:fdd22bb7aa52 392
emilmont 1:fdd22bb7aa52 393 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 394 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 395
emilmont 1:fdd22bb7aa52 396 /* Read x[9] */
emilmont 1:fdd22bb7aa52 397 x2 = _SIMD32_OFFSET(px + 1);
emilmont 1:fdd22bb7aa52 398
emilmont 1:fdd22bb7aa52 399 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 400 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 401 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 402 acc2 = __SMLAD(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 403 acc3 = __SMLAD(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 404
emilmont 1:fdd22bb7aa52 405 c0 = (*py);
emilmont 1:fdd22bb7aa52 406 /* Read y[6] */
emilmont 1:fdd22bb7aa52 407 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 408
emilmont 1:fdd22bb7aa52 409 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 410 #else
emilmont 1:fdd22bb7aa52 411
emilmont 1:fdd22bb7aa52 412 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 413 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 414
emilmont 1:fdd22bb7aa52 415 /* Read x[10] */
emilmont 1:fdd22bb7aa52 416 x3 = _SIMD32_OFFSET(px + 2);
emilmont 1:fdd22bb7aa52 417 px += 3u;
emilmont 1:fdd22bb7aa52 418
emilmont 1:fdd22bb7aa52 419 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 420 acc0 = __SMLADX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 421 acc1 = __SMLAD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 422 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 423 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 424 }
emilmont 1:fdd22bb7aa52 425
emilmont 1:fdd22bb7aa52 426 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 427 *pOut = (q15_t) (acc0 >> 15);
emilmont 1:fdd22bb7aa52 428 /* Destination pointer is updated according to the address modifier, inc */
emilmont 1:fdd22bb7aa52 429 pOut += inc;
emilmont 1:fdd22bb7aa52 430
emilmont 1:fdd22bb7aa52 431 *pOut = (q15_t) (acc1 >> 15);
emilmont 1:fdd22bb7aa52 432 pOut += inc;
emilmont 1:fdd22bb7aa52 433
emilmont 1:fdd22bb7aa52 434 *pOut = (q15_t) (acc2 >> 15);
emilmont 1:fdd22bb7aa52 435 pOut += inc;
emilmont 1:fdd22bb7aa52 436
emilmont 1:fdd22bb7aa52 437 *pOut = (q15_t) (acc3 >> 15);
emilmont 1:fdd22bb7aa52 438 pOut += inc;
emilmont 1:fdd22bb7aa52 439
emilmont 1:fdd22bb7aa52 440 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 441 count += 4u;
emilmont 1:fdd22bb7aa52 442
emilmont 1:fdd22bb7aa52 443 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 444 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 445 py = pIn2;
emilmont 1:fdd22bb7aa52 446
emilmont 1:fdd22bb7aa52 447
emilmont 1:fdd22bb7aa52 448 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 449 blkCnt--;
emilmont 1:fdd22bb7aa52 450 }
emilmont 1:fdd22bb7aa52 451
emilmont 1:fdd22bb7aa52 452 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 453 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 454 blkCnt = blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 455
emilmont 1:fdd22bb7aa52 456 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 457 {
emilmont 1:fdd22bb7aa52 458 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 459 sum = 0;
emilmont 1:fdd22bb7aa52 460
emilmont 1:fdd22bb7aa52 461 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 462 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 463
emilmont 1:fdd22bb7aa52 464 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 465 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 466 while(k > 0u)
emilmont 1:fdd22bb7aa52 467 {
emilmont 1:fdd22bb7aa52 468 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 469 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 470 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 471 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 472 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 473
emilmont 1:fdd22bb7aa52 474 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 475 k--;
emilmont 1:fdd22bb7aa52 476 }
emilmont 1:fdd22bb7aa52 477
emilmont 1:fdd22bb7aa52 478 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 479 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 480 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 481
emilmont 1:fdd22bb7aa52 482 while(k > 0u)
emilmont 1:fdd22bb7aa52 483 {
emilmont 1:fdd22bb7aa52 484 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 485 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 486
emilmont 1:fdd22bb7aa52 487 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 488 k--;
emilmont 1:fdd22bb7aa52 489 }
emilmont 1:fdd22bb7aa52 490
emilmont 1:fdd22bb7aa52 491 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 492 *pOut = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 493 /* Destination pointer is updated according to the address modifier, inc */
emilmont 1:fdd22bb7aa52 494 pOut += inc;
emilmont 1:fdd22bb7aa52 495
emilmont 1:fdd22bb7aa52 496 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 497 count++;
emilmont 1:fdd22bb7aa52 498
emilmont 1:fdd22bb7aa52 499 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 500 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 501 py = pIn2;
emilmont 1:fdd22bb7aa52 502
emilmont 1:fdd22bb7aa52 503 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 504 blkCnt--;
emilmont 1:fdd22bb7aa52 505 }
emilmont 1:fdd22bb7aa52 506 }
emilmont 1:fdd22bb7aa52 507 else
emilmont 1:fdd22bb7aa52 508 {
emilmont 1:fdd22bb7aa52 509 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 510 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 511 blkCnt = blockSize2;
emilmont 1:fdd22bb7aa52 512
emilmont 1:fdd22bb7aa52 513 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 514 {
emilmont 1:fdd22bb7aa52 515 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 516 sum = 0;
emilmont 1:fdd22bb7aa52 517
emilmont 1:fdd22bb7aa52 518 /* Loop over srcBLen */
emilmont 1:fdd22bb7aa52 519 k = srcBLen;
emilmont 1:fdd22bb7aa52 520
emilmont 1:fdd22bb7aa52 521 while(k > 0u)
emilmont 1:fdd22bb7aa52 522 {
emilmont 1:fdd22bb7aa52 523 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 524 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 525
emilmont 1:fdd22bb7aa52 526 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 527 k--;
emilmont 1:fdd22bb7aa52 528 }
emilmont 1:fdd22bb7aa52 529
emilmont 1:fdd22bb7aa52 530 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 531 *pOut = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 532 /* Destination pointer is updated according to the address modifier, inc */
emilmont 1:fdd22bb7aa52 533 pOut += inc;
emilmont 1:fdd22bb7aa52 534
emilmont 1:fdd22bb7aa52 535 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 536 count++;
emilmont 1:fdd22bb7aa52 537
emilmont 1:fdd22bb7aa52 538 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 539 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 540 py = pIn2;
emilmont 1:fdd22bb7aa52 541
emilmont 1:fdd22bb7aa52 542 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 543 blkCnt--;
emilmont 1:fdd22bb7aa52 544 }
emilmont 1:fdd22bb7aa52 545 }
emilmont 1:fdd22bb7aa52 546
emilmont 1:fdd22bb7aa52 547 /* --------------------------
emilmont 1:fdd22bb7aa52 548 * Initializations of stage3
emilmont 1:fdd22bb7aa52 549 * -------------------------*/
emilmont 1:fdd22bb7aa52 550
emilmont 1:fdd22bb7aa52 551 /* sum += x[srcALen-srcBLen+1] * y[0] + x[srcALen-srcBLen+2] * y[1] +...+ x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 552 * sum += x[srcALen-srcBLen+2] * y[0] + x[srcALen-srcBLen+3] * y[1] +...+ x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 553 * ....
emilmont 1:fdd22bb7aa52 554 * sum += x[srcALen-2] * y[0] + x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 555 * sum += x[srcALen-1] * y[0]
emilmont 1:fdd22bb7aa52 556 */
emilmont 1:fdd22bb7aa52 557
emilmont 1:fdd22bb7aa52 558 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 559 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 560 count = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 561
emilmont 1:fdd22bb7aa52 562 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 563 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 564 px = pSrc1;
emilmont 1:fdd22bb7aa52 565
emilmont 1:fdd22bb7aa52 566 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 567 py = pIn2;
emilmont 1:fdd22bb7aa52 568
emilmont 1:fdd22bb7aa52 569 /* -------------------
emilmont 1:fdd22bb7aa52 570 * Stage3 process
emilmont 1:fdd22bb7aa52 571 * ------------------*/
emilmont 1:fdd22bb7aa52 572
emilmont 1:fdd22bb7aa52 573 while(blockSize3 > 0u)
emilmont 1:fdd22bb7aa52 574 {
emilmont 1:fdd22bb7aa52 575 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 576 sum = 0;
emilmont 1:fdd22bb7aa52 577
emilmont 1:fdd22bb7aa52 578 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 579 k = count >> 2u;
emilmont 1:fdd22bb7aa52 580
emilmont 1:fdd22bb7aa52 581 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 582 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 583 while(k > 0u)
emilmont 1:fdd22bb7aa52 584 {
emilmont 1:fdd22bb7aa52 585 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 586 /* sum += x[srcALen - srcBLen + 4] * y[3] , sum += x[srcALen - srcBLen + 3] * y[2] */
emilmont 1:fdd22bb7aa52 587 sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum);
emilmont 1:fdd22bb7aa52 588 /* sum += x[srcALen - srcBLen + 2] * y[1] , sum += x[srcALen - srcBLen + 1] * y[0] */
emilmont 1:fdd22bb7aa52 589 sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum);
emilmont 1:fdd22bb7aa52 590
emilmont 1:fdd22bb7aa52 591 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 592 k--;
emilmont 1:fdd22bb7aa52 593 }
emilmont 1:fdd22bb7aa52 594
emilmont 1:fdd22bb7aa52 595 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 596 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 597 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 598
emilmont 1:fdd22bb7aa52 599 while(k > 0u)
emilmont 1:fdd22bb7aa52 600 {
emilmont 1:fdd22bb7aa52 601 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 602 sum = __SMLAD(*px++, *py++, sum);
emilmont 1:fdd22bb7aa52 603
emilmont 1:fdd22bb7aa52 604 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 605 k--;
emilmont 1:fdd22bb7aa52 606 }
emilmont 1:fdd22bb7aa52 607
emilmont 1:fdd22bb7aa52 608 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 609 *pOut = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 610 /* Destination pointer is updated according to the address modifier, inc */
emilmont 1:fdd22bb7aa52 611 pOut += inc;
emilmont 1:fdd22bb7aa52 612
emilmont 1:fdd22bb7aa52 613 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 614 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 615 py = pIn2;
emilmont 1:fdd22bb7aa52 616
emilmont 1:fdd22bb7aa52 617 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 618 count--;
emilmont 1:fdd22bb7aa52 619
emilmont 1:fdd22bb7aa52 620 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 621 blockSize3--;
emilmont 1:fdd22bb7aa52 622 }
emilmont 1:fdd22bb7aa52 623
emilmont 1:fdd22bb7aa52 624 #else
emilmont 1:fdd22bb7aa52 625
emilmont 1:fdd22bb7aa52 626 q15_t *pIn1; /* inputA pointer */
emilmont 1:fdd22bb7aa52 627 q15_t *pIn2; /* inputB pointer */
emilmont 1:fdd22bb7aa52 628 q15_t *pOut = pDst; /* output pointer */
emilmont 1:fdd22bb7aa52 629 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulators */
emilmont 1:fdd22bb7aa52 630 q15_t *px; /* Intermediate inputA pointer */
emilmont 1:fdd22bb7aa52 631 q15_t *py; /* Intermediate inputB pointer */
emilmont 1:fdd22bb7aa52 632 q15_t *pSrc1; /* Intermediate pointers */
emilmont 1:fdd22bb7aa52 633 q31_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */
emilmont 1:fdd22bb7aa52 634 uint32_t j, k = 0u, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counter */
emilmont 1:fdd22bb7aa52 635 int32_t inc = 1; /* Destination address modifier */
emilmont 1:fdd22bb7aa52 636 q15_t a, b;
emilmont 1:fdd22bb7aa52 637
emilmont 1:fdd22bb7aa52 638
emilmont 1:fdd22bb7aa52 639 /* The algorithm implementation is based on the lengths of the inputs. */
emilmont 1:fdd22bb7aa52 640 /* srcB is always made to slide across srcA. */
emilmont 1:fdd22bb7aa52 641 /* So srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 642 /* But CORR(x, y) is reverse of CORR(y, x) */
emilmont 1:fdd22bb7aa52 643 /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */
emilmont 1:fdd22bb7aa52 644 /* and the destination pointer modifier, inc is set to -1 */
emilmont 1:fdd22bb7aa52 645 /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */
emilmont 1:fdd22bb7aa52 646 /* But to improve the performance,
emilmont 1:fdd22bb7aa52 647 * we include zeroes in the output instead of zero padding either of the the inputs*/
emilmont 1:fdd22bb7aa52 648 /* If srcALen > srcBLen,
emilmont 1:fdd22bb7aa52 649 * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */
emilmont 1:fdd22bb7aa52 650 /* If srcALen < srcBLen,
emilmont 1:fdd22bb7aa52 651 * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */
emilmont 1:fdd22bb7aa52 652 if(srcALen >= srcBLen)
emilmont 1:fdd22bb7aa52 653 {
emilmont 1:fdd22bb7aa52 654 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 655 pIn1 = (pSrcA);
emilmont 1:fdd22bb7aa52 656
emilmont 1:fdd22bb7aa52 657 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 658 pIn2 = (pSrcB);
emilmont 1:fdd22bb7aa52 659
emilmont 1:fdd22bb7aa52 660 /* Number of output samples is calculated */
emilmont 1:fdd22bb7aa52 661 outBlockSize = (2u * srcALen) - 1u;
emilmont 1:fdd22bb7aa52 662
emilmont 1:fdd22bb7aa52 663 /* When srcALen > srcBLen, zero padding is done to srcB
emilmont 1:fdd22bb7aa52 664 * to make their lengths equal.
emilmont 1:fdd22bb7aa52 665 * Instead, (outBlockSize - (srcALen + srcBLen - 1))
emilmont 1:fdd22bb7aa52 666 * number of output samples are made zero */
emilmont 1:fdd22bb7aa52 667 j = outBlockSize - (srcALen + (srcBLen - 1u));
emilmont 1:fdd22bb7aa52 668
emilmont 1:fdd22bb7aa52 669 /* Updating the pointer position to non zero value */
emilmont 1:fdd22bb7aa52 670 pOut += j;
emilmont 1:fdd22bb7aa52 671
emilmont 1:fdd22bb7aa52 672 }
emilmont 1:fdd22bb7aa52 673 else
emilmont 1:fdd22bb7aa52 674 {
emilmont 1:fdd22bb7aa52 675 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 676 pIn1 = (pSrcB);
emilmont 1:fdd22bb7aa52 677
emilmont 1:fdd22bb7aa52 678 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 679 pIn2 = (pSrcA);
emilmont 1:fdd22bb7aa52 680
emilmont 1:fdd22bb7aa52 681 /* srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 682 j = srcBLen;
emilmont 1:fdd22bb7aa52 683 srcBLen = srcALen;
emilmont 1:fdd22bb7aa52 684 srcALen = j;
emilmont 1:fdd22bb7aa52 685
emilmont 1:fdd22bb7aa52 686 /* CORR(x, y) = Reverse order(CORR(y, x)) */
emilmont 1:fdd22bb7aa52 687 /* Hence set the destination pointer to point to the last output sample */
emilmont 1:fdd22bb7aa52 688 pOut = pDst + ((srcALen + srcBLen) - 2u);
emilmont 1:fdd22bb7aa52 689
emilmont 1:fdd22bb7aa52 690 /* Destination address modifier is set to -1 */
emilmont 1:fdd22bb7aa52 691 inc = -1;
emilmont 1:fdd22bb7aa52 692
emilmont 1:fdd22bb7aa52 693 }
emilmont 1:fdd22bb7aa52 694
emilmont 1:fdd22bb7aa52 695 /* The function is internally
emilmont 1:fdd22bb7aa52 696 * divided into three parts according to the number of multiplications that has to be
emilmont 1:fdd22bb7aa52 697 * taken place between inputA samples and inputB samples. In the first part of the
emilmont 1:fdd22bb7aa52 698 * algorithm, the multiplications increase by one for every iteration.
emilmont 1:fdd22bb7aa52 699 * In the second part of the algorithm, srcBLen number of multiplications are done.
emilmont 1:fdd22bb7aa52 700 * In the third part of the algorithm, the multiplications decrease by one
emilmont 1:fdd22bb7aa52 701 * for every iteration.*/
emilmont 1:fdd22bb7aa52 702 /* The algorithm is implemented in three stages.
emilmont 1:fdd22bb7aa52 703 * The loop counters of each stage is initiated here. */
emilmont 1:fdd22bb7aa52 704 blockSize1 = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 705 blockSize2 = srcALen - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 706 blockSize3 = blockSize1;
emilmont 1:fdd22bb7aa52 707
emilmont 1:fdd22bb7aa52 708 /* --------------------------
emilmont 1:fdd22bb7aa52 709 * Initializations of stage1
emilmont 1:fdd22bb7aa52 710 * -------------------------*/
emilmont 1:fdd22bb7aa52 711
emilmont 1:fdd22bb7aa52 712 /* sum = x[0] * y[srcBlen - 1]
emilmont 1:fdd22bb7aa52 713 * sum = x[0] * y[srcBlen - 2] + x[1] * y[srcBlen - 1]
emilmont 1:fdd22bb7aa52 714 * ....
emilmont 1:fdd22bb7aa52 715 * sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen - 1] * y[srcBLen - 1]
emilmont 1:fdd22bb7aa52 716 */
emilmont 1:fdd22bb7aa52 717
emilmont 1:fdd22bb7aa52 718 /* In this stage the MAC operations are increased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 719 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 720 count = 1u;
emilmont 1:fdd22bb7aa52 721
emilmont 1:fdd22bb7aa52 722 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 723 px = pIn1;
emilmont 1:fdd22bb7aa52 724
emilmont 1:fdd22bb7aa52 725 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 726 pSrc1 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 727 py = pSrc1;
emilmont 1:fdd22bb7aa52 728
emilmont 1:fdd22bb7aa52 729 /* ------------------------
emilmont 1:fdd22bb7aa52 730 * Stage1 process
emilmont 1:fdd22bb7aa52 731 * ----------------------*/
emilmont 1:fdd22bb7aa52 732
emilmont 1:fdd22bb7aa52 733 /* The first loop starts here */
emilmont 1:fdd22bb7aa52 734 while(blockSize1 > 0u)
emilmont 1:fdd22bb7aa52 735 {
emilmont 1:fdd22bb7aa52 736 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 737 sum = 0;
emilmont 1:fdd22bb7aa52 738
emilmont 1:fdd22bb7aa52 739 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 740 k = count >> 2;
emilmont 1:fdd22bb7aa52 741
emilmont 1:fdd22bb7aa52 742 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 743 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 744 while(k > 0u)
emilmont 1:fdd22bb7aa52 745 {
emilmont 1:fdd22bb7aa52 746 /* x[0] * y[srcBLen - 4] , x[1] * y[srcBLen - 3] */
emilmont 1:fdd22bb7aa52 747 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 748 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 749 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 750 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 751
emilmont 1:fdd22bb7aa52 752 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 753 k--;
emilmont 1:fdd22bb7aa52 754 }
emilmont 1:fdd22bb7aa52 755
emilmont 1:fdd22bb7aa52 756 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 757 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 758 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 759
emilmont 1:fdd22bb7aa52 760 while(k > 0u)
emilmont 1:fdd22bb7aa52 761 {
emilmont 1:fdd22bb7aa52 762 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 763 /* x[0] * y[srcBLen - 1] */
emilmont 1:fdd22bb7aa52 764 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 765
emilmont 1:fdd22bb7aa52 766 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 767 k--;
emilmont 1:fdd22bb7aa52 768 }
emilmont 1:fdd22bb7aa52 769
emilmont 1:fdd22bb7aa52 770 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 771 *pOut = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 772 /* Destination pointer is updated according to the address modifier, inc */
emilmont 1:fdd22bb7aa52 773 pOut += inc;
emilmont 1:fdd22bb7aa52 774
emilmont 1:fdd22bb7aa52 775 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 776 py = pSrc1 - count;
emilmont 1:fdd22bb7aa52 777 px = pIn1;
emilmont 1:fdd22bb7aa52 778
emilmont 1:fdd22bb7aa52 779 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 780 count++;
emilmont 1:fdd22bb7aa52 781
emilmont 1:fdd22bb7aa52 782 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 783 blockSize1--;
emilmont 1:fdd22bb7aa52 784 }
emilmont 1:fdd22bb7aa52 785
emilmont 1:fdd22bb7aa52 786 /* --------------------------
emilmont 1:fdd22bb7aa52 787 * Initializations of stage2
emilmont 1:fdd22bb7aa52 788 * ------------------------*/
emilmont 1:fdd22bb7aa52 789
emilmont 1:fdd22bb7aa52 790 /* sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 791 * sum = x[1] * y[0] + x[2] * y[1] +...+ x[srcBLen] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 792 * ....
emilmont 1:fdd22bb7aa52 793 * sum = x[srcALen-srcBLen-2] * y[0] + x[srcALen-srcBLen-1] * y[1] +...+ x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 794 */
emilmont 1:fdd22bb7aa52 795
emilmont 1:fdd22bb7aa52 796 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 797 px = pIn1;
emilmont 1:fdd22bb7aa52 798
emilmont 1:fdd22bb7aa52 799 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 800 py = pIn2;
emilmont 1:fdd22bb7aa52 801
emilmont 1:fdd22bb7aa52 802 /* count is index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 803 count = 0u;
emilmont 1:fdd22bb7aa52 804
emilmont 1:fdd22bb7aa52 805 /* -------------------
emilmont 1:fdd22bb7aa52 806 * Stage2 process
emilmont 1:fdd22bb7aa52 807 * ------------------*/
emilmont 1:fdd22bb7aa52 808
emilmont 1:fdd22bb7aa52 809 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 810 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 811 * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */
emilmont 1:fdd22bb7aa52 812 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 813 {
emilmont 1:fdd22bb7aa52 814 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 815 blkCnt = blockSize2 >> 2u;
emilmont 1:fdd22bb7aa52 816
emilmont 1:fdd22bb7aa52 817 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 818 {
emilmont 1:fdd22bb7aa52 819 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 820 acc0 = 0;
emilmont 1:fdd22bb7aa52 821 acc1 = 0;
emilmont 1:fdd22bb7aa52 822 acc2 = 0;
emilmont 1:fdd22bb7aa52 823 acc3 = 0;
emilmont 1:fdd22bb7aa52 824
emilmont 1:fdd22bb7aa52 825 /* read x[0], x[1], x[2] samples */
emilmont 1:fdd22bb7aa52 826 a = *px;
emilmont 1:fdd22bb7aa52 827 b = *(px + 1);
emilmont 1:fdd22bb7aa52 828
emilmont 1:fdd22bb7aa52 829 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 830
emilmont 1:fdd22bb7aa52 831 x0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 832 a = *(px + 2);
emilmont 1:fdd22bb7aa52 833 x1 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 834
emilmont 1:fdd22bb7aa52 835 #else
emilmont 1:fdd22bb7aa52 836
emilmont 1:fdd22bb7aa52 837 x0 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 838 a = *(px + 2);
emilmont 1:fdd22bb7aa52 839 x1 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 840
emilmont 1:fdd22bb7aa52 841 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 842
emilmont 1:fdd22bb7aa52 843 px += 2u;
emilmont 1:fdd22bb7aa52 844
emilmont 1:fdd22bb7aa52 845 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 846 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 847
emilmont 1:fdd22bb7aa52 848 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 849 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 850 do
emilmont 1:fdd22bb7aa52 851 {
emilmont 1:fdd22bb7aa52 852 /* Read the first two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 853 * y[0] and y[1] */
emilmont 1:fdd22bb7aa52 854 a = *py;
emilmont 1:fdd22bb7aa52 855 b = *(py + 1);
emilmont 1:fdd22bb7aa52 856
emilmont 1:fdd22bb7aa52 857 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 858
emilmont 1:fdd22bb7aa52 859 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 860
emilmont 1:fdd22bb7aa52 861 #else
emilmont 1:fdd22bb7aa52 862
emilmont 1:fdd22bb7aa52 863 c0 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 864
emilmont 1:fdd22bb7aa52 865 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 866
emilmont 1:fdd22bb7aa52 867 /* acc0 += x[0] * y[0] + x[1] * y[1] */
emilmont 1:fdd22bb7aa52 868 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 869
emilmont 1:fdd22bb7aa52 870 /* acc1 += x[1] * y[0] + x[2] * y[1] */
emilmont 1:fdd22bb7aa52 871 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 872
emilmont 1:fdd22bb7aa52 873 /* Read x[2], x[3], x[4] */
emilmont 1:fdd22bb7aa52 874 a = *px;
emilmont 1:fdd22bb7aa52 875 b = *(px + 1);
emilmont 1:fdd22bb7aa52 876
emilmont 1:fdd22bb7aa52 877 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 878
emilmont 1:fdd22bb7aa52 879 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 880 a = *(px + 2);
emilmont 1:fdd22bb7aa52 881 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 882
emilmont 1:fdd22bb7aa52 883 #else
emilmont 1:fdd22bb7aa52 884
emilmont 1:fdd22bb7aa52 885 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 886 a = *(px + 2);
emilmont 1:fdd22bb7aa52 887 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 888
emilmont 1:fdd22bb7aa52 889 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 890
emilmont 1:fdd22bb7aa52 891 /* acc2 += x[2] * y[0] + x[3] * y[1] */
emilmont 1:fdd22bb7aa52 892 acc2 = __SMLAD(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 893
emilmont 1:fdd22bb7aa52 894 /* acc3 += x[3] * y[0] + x[4] * y[1] */
emilmont 1:fdd22bb7aa52 895 acc3 = __SMLAD(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 896
emilmont 1:fdd22bb7aa52 897 /* Read y[2] and y[3] */
emilmont 1:fdd22bb7aa52 898 a = *(py + 2);
emilmont 1:fdd22bb7aa52 899 b = *(py + 3);
emilmont 1:fdd22bb7aa52 900
emilmont 1:fdd22bb7aa52 901 py += 4u;
emilmont 1:fdd22bb7aa52 902
emilmont 1:fdd22bb7aa52 903 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 904
emilmont 1:fdd22bb7aa52 905 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 906
emilmont 1:fdd22bb7aa52 907 #else
emilmont 1:fdd22bb7aa52 908
emilmont 1:fdd22bb7aa52 909 c0 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 910
emilmont 1:fdd22bb7aa52 911 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 912
emilmont 1:fdd22bb7aa52 913 /* acc0 += x[2] * y[2] + x[3] * y[3] */
emilmont 1:fdd22bb7aa52 914 acc0 = __SMLAD(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 915
emilmont 1:fdd22bb7aa52 916 /* acc1 += x[3] * y[2] + x[4] * y[3] */
emilmont 1:fdd22bb7aa52 917 acc1 = __SMLAD(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 918
emilmont 1:fdd22bb7aa52 919 /* Read x[4], x[5], x[6] */
emilmont 1:fdd22bb7aa52 920 a = *(px + 2);
emilmont 1:fdd22bb7aa52 921 b = *(px + 3);
emilmont 1:fdd22bb7aa52 922
emilmont 1:fdd22bb7aa52 923 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 924
emilmont 1:fdd22bb7aa52 925 x0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 926 a = *(px + 4);
emilmont 1:fdd22bb7aa52 927 x1 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 928
emilmont 1:fdd22bb7aa52 929 #else
emilmont 1:fdd22bb7aa52 930
emilmont 1:fdd22bb7aa52 931 x0 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 932 a = *(px + 4);
emilmont 1:fdd22bb7aa52 933 x1 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 934
emilmont 1:fdd22bb7aa52 935 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 936
emilmont 1:fdd22bb7aa52 937 px += 4u;
emilmont 1:fdd22bb7aa52 938
emilmont 1:fdd22bb7aa52 939 /* acc2 += x[4] * y[2] + x[5] * y[3] */
emilmont 1:fdd22bb7aa52 940 acc2 = __SMLAD(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 941
emilmont 1:fdd22bb7aa52 942 /* acc3 += x[5] * y[2] + x[6] * y[3] */
emilmont 1:fdd22bb7aa52 943 acc3 = __SMLAD(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 944
emilmont 1:fdd22bb7aa52 945 } while(--k);
emilmont 1:fdd22bb7aa52 946
emilmont 1:fdd22bb7aa52 947 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 948 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 949
emilmont 1:fdd22bb7aa52 950 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 951 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 952 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 953
emilmont 1:fdd22bb7aa52 954 if(k == 1u)
emilmont 1:fdd22bb7aa52 955 {
emilmont 1:fdd22bb7aa52 956 /* Read y[4] */
emilmont 1:fdd22bb7aa52 957 c0 = *py;
emilmont 1:fdd22bb7aa52 958 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 959
emilmont 1:fdd22bb7aa52 960 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 961
emilmont 1:fdd22bb7aa52 962 #else
emilmont 1:fdd22bb7aa52 963
emilmont 1:fdd22bb7aa52 964 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 965
emilmont 1:fdd22bb7aa52 966 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 967
emilmont 1:fdd22bb7aa52 968 /* Read x[7] */
emilmont 1:fdd22bb7aa52 969 a = *px;
emilmont 1:fdd22bb7aa52 970 b = *(px + 1);
emilmont 1:fdd22bb7aa52 971
emilmont 1:fdd22bb7aa52 972 px++;;
emilmont 1:fdd22bb7aa52 973
emilmont 1:fdd22bb7aa52 974 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 975
emilmont 1:fdd22bb7aa52 976 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 977
emilmont 1:fdd22bb7aa52 978 #else
emilmont 1:fdd22bb7aa52 979
emilmont 1:fdd22bb7aa52 980 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 981
emilmont 1:fdd22bb7aa52 982 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 983
emilmont 1:fdd22bb7aa52 984 px++;
emilmont 1:fdd22bb7aa52 985
emilmont 1:fdd22bb7aa52 986 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 987 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 988 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 989 acc2 = __SMLADX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 990 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 991 }
emilmont 1:fdd22bb7aa52 992
emilmont 1:fdd22bb7aa52 993 if(k == 2u)
emilmont 1:fdd22bb7aa52 994 {
emilmont 1:fdd22bb7aa52 995 /* Read y[4], y[5] */
emilmont 1:fdd22bb7aa52 996 a = *py;
emilmont 1:fdd22bb7aa52 997 b = *(py + 1);
emilmont 1:fdd22bb7aa52 998
emilmont 1:fdd22bb7aa52 999 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1000
emilmont 1:fdd22bb7aa52 1001 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1002
emilmont 1:fdd22bb7aa52 1003 #else
emilmont 1:fdd22bb7aa52 1004
emilmont 1:fdd22bb7aa52 1005 c0 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1006
emilmont 1:fdd22bb7aa52 1007 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1008
emilmont 1:fdd22bb7aa52 1009 /* Read x[7], x[8], x[9] */
emilmont 1:fdd22bb7aa52 1010 a = *px;
emilmont 1:fdd22bb7aa52 1011 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1012
emilmont 1:fdd22bb7aa52 1013 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1014
emilmont 1:fdd22bb7aa52 1015 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1016 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1017 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1018
emilmont 1:fdd22bb7aa52 1019 #else
emilmont 1:fdd22bb7aa52 1020
emilmont 1:fdd22bb7aa52 1021 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1022 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1023 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1024
emilmont 1:fdd22bb7aa52 1025 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1026
emilmont 1:fdd22bb7aa52 1027 px += 2u;
emilmont 1:fdd22bb7aa52 1028
emilmont 1:fdd22bb7aa52 1029 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1030 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1031 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1032 acc2 = __SMLAD(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 1033 acc3 = __SMLAD(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 1034 }
emilmont 1:fdd22bb7aa52 1035
emilmont 1:fdd22bb7aa52 1036 if(k == 3u)
emilmont 1:fdd22bb7aa52 1037 {
emilmont 1:fdd22bb7aa52 1038 /* Read y[4], y[5] */
emilmont 1:fdd22bb7aa52 1039 a = *py;
emilmont 1:fdd22bb7aa52 1040 b = *(py + 1);
emilmont 1:fdd22bb7aa52 1041
emilmont 1:fdd22bb7aa52 1042 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1043
emilmont 1:fdd22bb7aa52 1044 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1045
emilmont 1:fdd22bb7aa52 1046 #else
emilmont 1:fdd22bb7aa52 1047
emilmont 1:fdd22bb7aa52 1048 c0 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1049
emilmont 1:fdd22bb7aa52 1050 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1051
emilmont 1:fdd22bb7aa52 1052 py += 2u;
emilmont 1:fdd22bb7aa52 1053
emilmont 1:fdd22bb7aa52 1054 /* Read x[7], x[8], x[9] */
emilmont 1:fdd22bb7aa52 1055 a = *px;
emilmont 1:fdd22bb7aa52 1056 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1057
emilmont 1:fdd22bb7aa52 1058 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1059
emilmont 1:fdd22bb7aa52 1060 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1061 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1062 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1063
emilmont 1:fdd22bb7aa52 1064 #else
emilmont 1:fdd22bb7aa52 1065
emilmont 1:fdd22bb7aa52 1066 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1067 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1068 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1069
emilmont 1:fdd22bb7aa52 1070 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1071
emilmont 1:fdd22bb7aa52 1072 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1073 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1074 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1075 acc2 = __SMLAD(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 1076 acc3 = __SMLAD(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 1077
emilmont 1:fdd22bb7aa52 1078 c0 = (*py);
emilmont 1:fdd22bb7aa52 1079 /* Read y[6] */
emilmont 1:fdd22bb7aa52 1080 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1081
emilmont 1:fdd22bb7aa52 1082 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 1083 #else
emilmont 1:fdd22bb7aa52 1084
emilmont 1:fdd22bb7aa52 1085 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 1086 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1087
emilmont 1:fdd22bb7aa52 1088 /* Read x[10] */
emilmont 1:fdd22bb7aa52 1089 b = *(px + 3);
emilmont 1:fdd22bb7aa52 1090
emilmont 1:fdd22bb7aa52 1091 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1092
emilmont 1:fdd22bb7aa52 1093 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1094
emilmont 1:fdd22bb7aa52 1095 #else
emilmont 1:fdd22bb7aa52 1096
emilmont 1:fdd22bb7aa52 1097 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1098
emilmont 1:fdd22bb7aa52 1099 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1100
emilmont 1:fdd22bb7aa52 1101 px += 3u;
emilmont 1:fdd22bb7aa52 1102
emilmont 1:fdd22bb7aa52 1103 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1104 acc0 = __SMLADX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 1105 acc1 = __SMLAD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 1106 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 1107 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1108 }
emilmont 1:fdd22bb7aa52 1109
emilmont 1:fdd22bb7aa52 1110 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1111 *pOut = (q15_t) (acc0 >> 15);
emilmont 1:fdd22bb7aa52 1112 /* Destination pointer is updated according to the address modifier, inc */
emilmont 1:fdd22bb7aa52 1113 pOut += inc;
emilmont 1:fdd22bb7aa52 1114
emilmont 1:fdd22bb7aa52 1115 *pOut = (q15_t) (acc1 >> 15);
emilmont 1:fdd22bb7aa52 1116 pOut += inc;
emilmont 1:fdd22bb7aa52 1117
emilmont 1:fdd22bb7aa52 1118 *pOut = (q15_t) (acc2 >> 15);
emilmont 1:fdd22bb7aa52 1119 pOut += inc;
emilmont 1:fdd22bb7aa52 1120
emilmont 1:fdd22bb7aa52 1121 *pOut = (q15_t) (acc3 >> 15);
emilmont 1:fdd22bb7aa52 1122 pOut += inc;
emilmont 1:fdd22bb7aa52 1123
emilmont 1:fdd22bb7aa52 1124 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 1125 count += 4u;
emilmont 1:fdd22bb7aa52 1126
emilmont 1:fdd22bb7aa52 1127 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1128 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1129 py = pIn2;
emilmont 1:fdd22bb7aa52 1130
emilmont 1:fdd22bb7aa52 1131
emilmont 1:fdd22bb7aa52 1132 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1133 blkCnt--;
emilmont 1:fdd22bb7aa52 1134 }
emilmont 1:fdd22bb7aa52 1135
emilmont 1:fdd22bb7aa52 1136 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 1137 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1138 blkCnt = blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 1139
emilmont 1:fdd22bb7aa52 1140 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 1141 {
emilmont 1:fdd22bb7aa52 1142 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1143 sum = 0;
emilmont 1:fdd22bb7aa52 1144
emilmont 1:fdd22bb7aa52 1145 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1146 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 1147
emilmont 1:fdd22bb7aa52 1148 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 1149 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 1150 while(k > 0u)
emilmont 1:fdd22bb7aa52 1151 {
emilmont 1:fdd22bb7aa52 1152 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1153 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 1154 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 1155 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 1156 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 1157
emilmont 1:fdd22bb7aa52 1158 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1159 k--;
emilmont 1:fdd22bb7aa52 1160 }
emilmont 1:fdd22bb7aa52 1161
emilmont 1:fdd22bb7aa52 1162 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1163 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1164 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 1165
emilmont 1:fdd22bb7aa52 1166 while(k > 0u)
emilmont 1:fdd22bb7aa52 1167 {
emilmont 1:fdd22bb7aa52 1168 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1169 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 1170
emilmont 1:fdd22bb7aa52 1171 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1172 k--;
emilmont 1:fdd22bb7aa52 1173 }
emilmont 1:fdd22bb7aa52 1174
emilmont 1:fdd22bb7aa52 1175 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1176 *pOut = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1177 /* Destination pointer is updated according to the address modifier, inc */
emilmont 1:fdd22bb7aa52 1178 pOut += inc;
emilmont 1:fdd22bb7aa52 1179
emilmont 1:fdd22bb7aa52 1180 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 1181 count++;
emilmont 1:fdd22bb7aa52 1182
emilmont 1:fdd22bb7aa52 1183 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1184 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1185 py = pIn2;
emilmont 1:fdd22bb7aa52 1186
emilmont 1:fdd22bb7aa52 1187 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1188 blkCnt--;
emilmont 1:fdd22bb7aa52 1189 }
emilmont 1:fdd22bb7aa52 1190 }
emilmont 1:fdd22bb7aa52 1191 else
emilmont 1:fdd22bb7aa52 1192 {
emilmont 1:fdd22bb7aa52 1193 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 1194 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 1195 blkCnt = blockSize2;
emilmont 1:fdd22bb7aa52 1196
emilmont 1:fdd22bb7aa52 1197 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 1198 {
emilmont 1:fdd22bb7aa52 1199 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1200 sum = 0;
emilmont 1:fdd22bb7aa52 1201
emilmont 1:fdd22bb7aa52 1202 /* Loop over srcBLen */
emilmont 1:fdd22bb7aa52 1203 k = srcBLen;
emilmont 1:fdd22bb7aa52 1204
emilmont 1:fdd22bb7aa52 1205 while(k > 0u)
emilmont 1:fdd22bb7aa52 1206 {
emilmont 1:fdd22bb7aa52 1207 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 1208 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 1209
emilmont 1:fdd22bb7aa52 1210 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1211 k--;
emilmont 1:fdd22bb7aa52 1212 }
emilmont 1:fdd22bb7aa52 1213
emilmont 1:fdd22bb7aa52 1214 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1215 *pOut = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1216 /* Destination pointer is updated according to the address modifier, inc */
emilmont 1:fdd22bb7aa52 1217 pOut += inc;
emilmont 1:fdd22bb7aa52 1218
emilmont 1:fdd22bb7aa52 1219 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 1220 count++;
emilmont 1:fdd22bb7aa52 1221
emilmont 1:fdd22bb7aa52 1222 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1223 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1224 py = pIn2;
emilmont 1:fdd22bb7aa52 1225
emilmont 1:fdd22bb7aa52 1226 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1227 blkCnt--;
emilmont 1:fdd22bb7aa52 1228 }
emilmont 1:fdd22bb7aa52 1229 }
emilmont 1:fdd22bb7aa52 1230
emilmont 1:fdd22bb7aa52 1231 /* --------------------------
emilmont 1:fdd22bb7aa52 1232 * Initializations of stage3
emilmont 1:fdd22bb7aa52 1233 * -------------------------*/
emilmont 1:fdd22bb7aa52 1234
emilmont 1:fdd22bb7aa52 1235 /* sum += x[srcALen-srcBLen+1] * y[0] + x[srcALen-srcBLen+2] * y[1] +...+ x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 1236 * sum += x[srcALen-srcBLen+2] * y[0] + x[srcALen-srcBLen+3] * y[1] +...+ x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 1237 * ....
emilmont 1:fdd22bb7aa52 1238 * sum += x[srcALen-2] * y[0] + x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 1239 * sum += x[srcALen-1] * y[0]
emilmont 1:fdd22bb7aa52 1240 */
emilmont 1:fdd22bb7aa52 1241
emilmont 1:fdd22bb7aa52 1242 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 1243 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 1244 count = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 1245
emilmont 1:fdd22bb7aa52 1246 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 1247 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 1248 px = pSrc1;
emilmont 1:fdd22bb7aa52 1249
emilmont 1:fdd22bb7aa52 1250 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 1251 py = pIn2;
emilmont 1:fdd22bb7aa52 1252
emilmont 1:fdd22bb7aa52 1253 /* -------------------
emilmont 1:fdd22bb7aa52 1254 * Stage3 process
emilmont 1:fdd22bb7aa52 1255 * ------------------*/
emilmont 1:fdd22bb7aa52 1256
emilmont 1:fdd22bb7aa52 1257 while(blockSize3 > 0u)
emilmont 1:fdd22bb7aa52 1258 {
emilmont 1:fdd22bb7aa52 1259 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1260 sum = 0;
emilmont 1:fdd22bb7aa52 1261
emilmont 1:fdd22bb7aa52 1262 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1263 k = count >> 2u;
emilmont 1:fdd22bb7aa52 1264
emilmont 1:fdd22bb7aa52 1265 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 1266 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 1267 while(k > 0u)
emilmont 1:fdd22bb7aa52 1268 {
emilmont 1:fdd22bb7aa52 1269 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1270 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 1271 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 1272 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 1273 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 1274
emilmont 1:fdd22bb7aa52 1275 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1276 k--;
emilmont 1:fdd22bb7aa52 1277 }
emilmont 1:fdd22bb7aa52 1278
emilmont 1:fdd22bb7aa52 1279 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1280 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1281 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 1282
emilmont 1:fdd22bb7aa52 1283 while(k > 0u)
emilmont 1:fdd22bb7aa52 1284 {
emilmont 1:fdd22bb7aa52 1285 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1286 sum += ((q31_t) * px++ * *py++);
emilmont 1:fdd22bb7aa52 1287
emilmont 1:fdd22bb7aa52 1288 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1289 k--;
emilmont 1:fdd22bb7aa52 1290 }
emilmont 1:fdd22bb7aa52 1291
emilmont 1:fdd22bb7aa52 1292 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1293 *pOut = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1294 /* Destination pointer is updated according to the address modifier, inc */
emilmont 1:fdd22bb7aa52 1295 pOut += inc;
emilmont 1:fdd22bb7aa52 1296
emilmont 1:fdd22bb7aa52 1297 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1298 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 1299 py = pIn2;
emilmont 1:fdd22bb7aa52 1300
emilmont 1:fdd22bb7aa52 1301 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 1302 count--;
emilmont 1:fdd22bb7aa52 1303
emilmont 1:fdd22bb7aa52 1304 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1305 blockSize3--;
emilmont 1:fdd22bb7aa52 1306 }
emilmont 1:fdd22bb7aa52 1307
emilmont 1:fdd22bb7aa52 1308 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
emilmont 1:fdd22bb7aa52 1309
emilmont 1:fdd22bb7aa52 1310 }
emilmont 1:fdd22bb7aa52 1311
emilmont 1:fdd22bb7aa52 1312 /**
emilmont 1:fdd22bb7aa52 1313 * @} end of Corr group
emilmont 1:fdd22bb7aa52 1314 */