mbed official / mbed-dsp

CMSIS DSP library

Dependents:   KL25Z_FFT_Demo Hat_Board_v5_1 KL25Z_FFT_Demo_tony KL25Z_FFT_Demo_tony ... more

Fork of mbed-dsp by mbed official

cmsis_dsp/FilteringFunctions/arm_conv_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_conv_fast_q15.c
emilmont 1:fdd22bb7aa52 9 *
emilmont 1:fdd22bb7aa52 10 * Description: Fast Q15 Convolution.
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 Conv
emilmont 1:fdd22bb7aa52 44 * @{
emilmont 1:fdd22bb7aa52 45 */
emilmont 1:fdd22bb7aa52 46
emilmont 1:fdd22bb7aa52 47 /**
emilmont 1:fdd22bb7aa52 48 * @brief Convolution 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 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
emilmont 1:fdd22bb7aa52 61 * but provides only a single guard bit. 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 the inputs by log2(min(srcALen, srcBLen)) (log2 is read as log to the base 2) times to avoid overflows,
emilmont 1:fdd22bb7aa52 65 * as maximum of min(srcALen, srcBLen) number of additions are 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_conv_q15()</code> for a slower implementation of this function which uses 64-bit accumulation to avoid wrap around distortion.
emilmont 1:fdd22bb7aa52 70 */
emilmont 1:fdd22bb7aa52 71
emilmont 1:fdd22bb7aa52 72 void arm_conv_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 q15_t *pIn1; /* inputA pointer */
emilmont 1:fdd22bb7aa52 81 q15_t *pIn2; /* inputB pointer */
emilmont 1:fdd22bb7aa52 82 q15_t *pOut = pDst; /* output pointer */
emilmont 1:fdd22bb7aa52 83 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
emilmont 1:fdd22bb7aa52 84 q15_t *px; /* Intermediate inputA pointer */
emilmont 1:fdd22bb7aa52 85 q15_t *py; /* Intermediate inputB pointer */
emilmont 1:fdd22bb7aa52 86 q15_t *pSrc1, *pSrc2; /* Intermediate pointers */
emilmont 1:fdd22bb7aa52 87 q31_t x0, x1, x2, x3, c0; /* Temporary variables to hold state and coefficient values */
emilmont 1:fdd22bb7aa52 88 uint32_t blockSize1, blockSize2, blockSize3, j, k, count, blkCnt; /* loop counter */
emilmont 1:fdd22bb7aa52 89
emilmont 1:fdd22bb7aa52 90 /* The algorithm implementation is based on the lengths of the inputs. */
emilmont 1:fdd22bb7aa52 91 /* srcB is always made to slide across srcA. */
emilmont 1:fdd22bb7aa52 92 /* So srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 93 if(srcALen >= srcBLen)
emilmont 1:fdd22bb7aa52 94 {
emilmont 1:fdd22bb7aa52 95 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 96 pIn1 = pSrcA;
emilmont 1:fdd22bb7aa52 97
emilmont 1:fdd22bb7aa52 98 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 99 pIn2 = pSrcB;
emilmont 1:fdd22bb7aa52 100 }
emilmont 1:fdd22bb7aa52 101 else
emilmont 1:fdd22bb7aa52 102 {
emilmont 1:fdd22bb7aa52 103 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 104 pIn1 = pSrcB;
emilmont 1:fdd22bb7aa52 105
emilmont 1:fdd22bb7aa52 106 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 107 pIn2 = pSrcA;
emilmont 1:fdd22bb7aa52 108
emilmont 1:fdd22bb7aa52 109 /* srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 110 j = srcBLen;
emilmont 1:fdd22bb7aa52 111 srcBLen = srcALen;
emilmont 1:fdd22bb7aa52 112 srcALen = j;
emilmont 1:fdd22bb7aa52 113 }
emilmont 1:fdd22bb7aa52 114
emilmont 1:fdd22bb7aa52 115 /* conv(x,y) at n = x[n] * y[0] + x[n-1] * y[1] + x[n-2] * y[2] + ...+ x[n-N+1] * y[N -1] */
emilmont 1:fdd22bb7aa52 116 /* The function is internally
emilmont 1:fdd22bb7aa52 117 * divided into three stages according to the number of multiplications that has to be
emilmont 1:fdd22bb7aa52 118 * taken place between inputA samples and inputB samples. In the first stage of the
emilmont 1:fdd22bb7aa52 119 * algorithm, the multiplications increase by one for every iteration.
emilmont 1:fdd22bb7aa52 120 * In the second stage of the algorithm, srcBLen number of multiplications are done.
emilmont 1:fdd22bb7aa52 121 * In the third stage of the algorithm, the multiplications decrease by one
emilmont 1:fdd22bb7aa52 122 * for every iteration. */
emilmont 1:fdd22bb7aa52 123
emilmont 1:fdd22bb7aa52 124 /* The algorithm is implemented in three stages.
emilmont 1:fdd22bb7aa52 125 The loop counters of each stage is initiated here. */
emilmont 1:fdd22bb7aa52 126 blockSize1 = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 127 blockSize2 = srcALen - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 128 blockSize3 = blockSize1;
emilmont 1:fdd22bb7aa52 129
emilmont 1:fdd22bb7aa52 130 /* --------------------------
emilmont 1:fdd22bb7aa52 131 * Initializations of stage1
emilmont 1:fdd22bb7aa52 132 * -------------------------*/
emilmont 1:fdd22bb7aa52 133
emilmont 1:fdd22bb7aa52 134 /* sum = x[0] * y[0]
emilmont 1:fdd22bb7aa52 135 * sum = x[0] * y[1] + x[1] * y[0]
emilmont 1:fdd22bb7aa52 136 * ....
emilmont 1:fdd22bb7aa52 137 * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
emilmont 1:fdd22bb7aa52 138 */
emilmont 1:fdd22bb7aa52 139
emilmont 1:fdd22bb7aa52 140 /* In this stage the MAC operations are increased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 141 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 142 count = 1u;
emilmont 1:fdd22bb7aa52 143
emilmont 1:fdd22bb7aa52 144 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 145 px = pIn1;
emilmont 1:fdd22bb7aa52 146
emilmont 1:fdd22bb7aa52 147 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 148 py = pIn2;
emilmont 1:fdd22bb7aa52 149
emilmont 1:fdd22bb7aa52 150
emilmont 1:fdd22bb7aa52 151 /* ------------------------
emilmont 1:fdd22bb7aa52 152 * Stage1 process
emilmont 1:fdd22bb7aa52 153 * ----------------------*/
emilmont 1:fdd22bb7aa52 154
emilmont 1:fdd22bb7aa52 155 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 156 /* First part of this stage computes the MAC operations less than 4 */
emilmont 1:fdd22bb7aa52 157 /* Second part of this stage computes the MAC operations greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 158
emilmont 1:fdd22bb7aa52 159 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 160 while((count < 4u) && (blockSize1 > 0u))
emilmont 1:fdd22bb7aa52 161 {
emilmont 1:fdd22bb7aa52 162 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 163 sum = 0;
emilmont 1:fdd22bb7aa52 164
emilmont 1:fdd22bb7aa52 165 /* Loop over number of MAC operations between
emilmont 1:fdd22bb7aa52 166 * inputA samples and inputB samples */
emilmont 1:fdd22bb7aa52 167 k = count;
emilmont 1:fdd22bb7aa52 168
emilmont 1:fdd22bb7aa52 169 while(k > 0u)
emilmont 1:fdd22bb7aa52 170 {
emilmont 1:fdd22bb7aa52 171 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 172 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 173
emilmont 1:fdd22bb7aa52 174 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 175 k--;
emilmont 1:fdd22bb7aa52 176 }
emilmont 1:fdd22bb7aa52 177
emilmont 1:fdd22bb7aa52 178 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 179 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 180
emilmont 1:fdd22bb7aa52 181 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 182 py = pIn2 + count;
emilmont 1:fdd22bb7aa52 183 px = pIn1;
emilmont 1:fdd22bb7aa52 184
emilmont 1:fdd22bb7aa52 185 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 186 count++;
emilmont 1:fdd22bb7aa52 187
emilmont 1:fdd22bb7aa52 188 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 189 blockSize1--;
emilmont 1:fdd22bb7aa52 190 }
emilmont 1:fdd22bb7aa52 191
emilmont 1:fdd22bb7aa52 192 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 193 /* The internal loop, over count, is unrolled by 4 */
emilmont 1:fdd22bb7aa52 194 /* To, read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 195 * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */
emilmont 1:fdd22bb7aa52 196 py = py - 1;
emilmont 1:fdd22bb7aa52 197
emilmont 1:fdd22bb7aa52 198 while(blockSize1 > 0u)
emilmont 1:fdd22bb7aa52 199 {
emilmont 1:fdd22bb7aa52 200 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 201 sum = 0;
emilmont 1:fdd22bb7aa52 202
emilmont 1:fdd22bb7aa52 203 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 204 k = count >> 2u;
emilmont 1:fdd22bb7aa52 205
emilmont 1:fdd22bb7aa52 206 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 207 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 208 while(k > 0u)
emilmont 1:fdd22bb7aa52 209 {
emilmont 1:fdd22bb7aa52 210 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 211 /* x[0], x[1] are multiplied with y[srcBLen - 1], y[srcBLen - 2] respectively */
emilmont 1:fdd22bb7aa52 212 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 213 /* x[2], x[3] are multiplied with y[srcBLen - 3], y[srcBLen - 4] respectively */
emilmont 1:fdd22bb7aa52 214 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 215
emilmont 1:fdd22bb7aa52 216 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 217 k--;
emilmont 1:fdd22bb7aa52 218 }
emilmont 1:fdd22bb7aa52 219
emilmont 1:fdd22bb7aa52 220 /* For the next MAC operations, the pointer py is used without SIMD
emilmont 1:fdd22bb7aa52 221 * So, py is incremented by 1 */
emilmont 1:fdd22bb7aa52 222 py = py + 1u;
emilmont 1:fdd22bb7aa52 223
emilmont 1:fdd22bb7aa52 224 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 225 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 226 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 227
emilmont 1:fdd22bb7aa52 228 while(k > 0u)
emilmont 1:fdd22bb7aa52 229 {
emilmont 1:fdd22bb7aa52 230 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 231 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 232
emilmont 1:fdd22bb7aa52 233 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 234 k--;
emilmont 1:fdd22bb7aa52 235 }
emilmont 1:fdd22bb7aa52 236
emilmont 1:fdd22bb7aa52 237 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 238 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 239
emilmont 1:fdd22bb7aa52 240 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 241 py = pIn2 + (count - 1u);
emilmont 1:fdd22bb7aa52 242 px = pIn1;
emilmont 1:fdd22bb7aa52 243
emilmont 1:fdd22bb7aa52 244 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 245 count++;
emilmont 1:fdd22bb7aa52 246
emilmont 1:fdd22bb7aa52 247 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 248 blockSize1--;
emilmont 1:fdd22bb7aa52 249 }
emilmont 1:fdd22bb7aa52 250
emilmont 1:fdd22bb7aa52 251 /* --------------------------
emilmont 1:fdd22bb7aa52 252 * Initializations of stage2
emilmont 1:fdd22bb7aa52 253 * ------------------------*/
emilmont 1:fdd22bb7aa52 254
emilmont 1:fdd22bb7aa52 255 /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
emilmont 1:fdd22bb7aa52 256 * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
emilmont 1:fdd22bb7aa52 257 * ....
emilmont 1:fdd22bb7aa52 258 * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
emilmont 1:fdd22bb7aa52 259 */
emilmont 1:fdd22bb7aa52 260
emilmont 1:fdd22bb7aa52 261 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 262 px = pIn1;
emilmont 1:fdd22bb7aa52 263
emilmont 1:fdd22bb7aa52 264 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 265 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 266 py = pSrc2;
emilmont 1:fdd22bb7aa52 267
emilmont 1:fdd22bb7aa52 268 /* count is the index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 269 count = 0u;
emilmont 1:fdd22bb7aa52 270
emilmont 1:fdd22bb7aa52 271
emilmont 1:fdd22bb7aa52 272 /* --------------------
emilmont 1:fdd22bb7aa52 273 * Stage2 process
emilmont 1:fdd22bb7aa52 274 * -------------------*/
emilmont 1:fdd22bb7aa52 275
emilmont 1:fdd22bb7aa52 276 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 277 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 278 * srcBLen should be greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 279 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 280 {
emilmont 1:fdd22bb7aa52 281 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 282 blkCnt = blockSize2 >> 2u;
emilmont 1:fdd22bb7aa52 283
emilmont 1:fdd22bb7aa52 284 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 285 {
emilmont 1:fdd22bb7aa52 286 py = py - 1u;
emilmont 1:fdd22bb7aa52 287
emilmont 1:fdd22bb7aa52 288 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 289 acc0 = 0;
emilmont 1:fdd22bb7aa52 290 acc1 = 0;
emilmont 1:fdd22bb7aa52 291 acc2 = 0;
emilmont 1:fdd22bb7aa52 292 acc3 = 0;
emilmont 1:fdd22bb7aa52 293
emilmont 1:fdd22bb7aa52 294
emilmont 1:fdd22bb7aa52 295 /* read x[0], x[1] samples */
emilmont 1:fdd22bb7aa52 296 x0 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 297 /* read x[1], x[2] samples */
emilmont 1:fdd22bb7aa52 298 x1 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 299 px+= 2u;
emilmont 1:fdd22bb7aa52 300
emilmont 1:fdd22bb7aa52 301
emilmont 1:fdd22bb7aa52 302 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 303 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 304
emilmont 1:fdd22bb7aa52 305 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 306 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 307 do
emilmont 1:fdd22bb7aa52 308 {
emilmont 1:fdd22bb7aa52 309 /* Read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 310 * y[srcBLen - 1] and y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 311 c0 = *__SIMD32(py)--;
emilmont 1:fdd22bb7aa52 312
emilmont 1:fdd22bb7aa52 313 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 314 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 315
emilmont 1:fdd22bb7aa52 316 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 317 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 318
emilmont 1:fdd22bb7aa52 319 /* Read x[2], x[3] */
emilmont 1:fdd22bb7aa52 320 x2 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 321
emilmont 1:fdd22bb7aa52 322 /* Read x[3], x[4] */
emilmont 1:fdd22bb7aa52 323 x3 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 324
emilmont 1:fdd22bb7aa52 325 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 326 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 327
emilmont 1:fdd22bb7aa52 328 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 329 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 330
emilmont 1:fdd22bb7aa52 331 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 332 c0 = *__SIMD32(py)--;
emilmont 1:fdd22bb7aa52 333
emilmont 1:fdd22bb7aa52 334 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 335 acc0 = __SMLADX(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 336
emilmont 1:fdd22bb7aa52 337 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 338 acc1 = __SMLADX(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 339
emilmont 1:fdd22bb7aa52 340 /* Read x[4], x[5] */
emilmont 1:fdd22bb7aa52 341 x0 = _SIMD32_OFFSET(px+2);
emilmont 1:fdd22bb7aa52 342
emilmont 1:fdd22bb7aa52 343 /* Read x[5], x[6] */
emilmont 1:fdd22bb7aa52 344 x1 = _SIMD32_OFFSET(px+3);
emilmont 1:fdd22bb7aa52 345 px += 4u;
emilmont 1:fdd22bb7aa52 346
emilmont 1:fdd22bb7aa52 347 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 348 acc2 = __SMLADX(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 349
emilmont 1:fdd22bb7aa52 350 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 351 acc3 = __SMLADX(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 352
emilmont 1:fdd22bb7aa52 353 } while(--k);
emilmont 1:fdd22bb7aa52 354
emilmont 1:fdd22bb7aa52 355 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 356 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 357
emilmont 1:fdd22bb7aa52 358 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 359 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 360 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 361
emilmont 1:fdd22bb7aa52 362 if(k == 1u)
emilmont 1:fdd22bb7aa52 363 {
emilmont 1:fdd22bb7aa52 364 /* Read y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 365 c0 = *(py+1);
emilmont 1:fdd22bb7aa52 366
emilmont 1:fdd22bb7aa52 367 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 368
emilmont 1:fdd22bb7aa52 369 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 370
emilmont 1:fdd22bb7aa52 371 #else
emilmont 1:fdd22bb7aa52 372
emilmont 1:fdd22bb7aa52 373 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 374
emilmont 1:fdd22bb7aa52 375 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 376
emilmont 1:fdd22bb7aa52 377 /* Read x[7] */
emilmont 1:fdd22bb7aa52 378 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 379 px++;
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 = __SMLADX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 385 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 386 }
emilmont 1:fdd22bb7aa52 387
emilmont 1:fdd22bb7aa52 388 if(k == 2u)
emilmont 1:fdd22bb7aa52 389 {
emilmont 1:fdd22bb7aa52 390 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 391 c0 = _SIMD32_OFFSET(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 px += 2u;
emilmont 1:fdd22bb7aa52 399
emilmont 1:fdd22bb7aa52 400 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 401 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 402 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 403 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 404 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 405 }
emilmont 1:fdd22bb7aa52 406
emilmont 1:fdd22bb7aa52 407 if(k == 3u)
emilmont 1:fdd22bb7aa52 408 {
emilmont 1:fdd22bb7aa52 409 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 410 c0 = _SIMD32_OFFSET(py);
emilmont 1:fdd22bb7aa52 411
emilmont 1:fdd22bb7aa52 412 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 413 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 414
emilmont 1:fdd22bb7aa52 415 /* Read x[9] */
emilmont 1:fdd22bb7aa52 416 x2 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 417
emilmont 1:fdd22bb7aa52 418 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 419 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 420 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 421 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 422 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 423
emilmont 1:fdd22bb7aa52 424 /* Read y[srcBLen - 7] */
emilmont 1:fdd22bb7aa52 425 c0 = *(py-1);
emilmont 1:fdd22bb7aa52 426 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 427
emilmont 1:fdd22bb7aa52 428 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 429 #else
emilmont 1:fdd22bb7aa52 430
emilmont 1:fdd22bb7aa52 431 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 432 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 433
emilmont 1:fdd22bb7aa52 434 /* Read x[10] */
emilmont 1:fdd22bb7aa52 435 x3 = _SIMD32_OFFSET(px+2);
emilmont 1:fdd22bb7aa52 436 px += 3u;
emilmont 1:fdd22bb7aa52 437
emilmont 1:fdd22bb7aa52 438 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 439 acc0 = __SMLADX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 440 acc1 = __SMLAD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 441 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 442 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 443 }
emilmont 1:fdd22bb7aa52 444
emilmont 1:fdd22bb7aa52 445 /* Store the results in the accumulators in the destination buffer. */
emilmont 1:fdd22bb7aa52 446 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 447
emilmont 1:fdd22bb7aa52 448 *__SIMD32(pOut)++ = __PKHBT((acc0 >> 15), (acc1 >> 15), 16);
emilmont 1:fdd22bb7aa52 449 *__SIMD32(pOut)++ = __PKHBT((acc2 >> 15), (acc3 >> 15), 16);
emilmont 1:fdd22bb7aa52 450
emilmont 1:fdd22bb7aa52 451 #else
emilmont 1:fdd22bb7aa52 452
emilmont 1:fdd22bb7aa52 453 *__SIMD32(pOut)++ = __PKHBT((acc1 >> 15), (acc0 >> 15), 16);
emilmont 1:fdd22bb7aa52 454 *__SIMD32(pOut)++ = __PKHBT((acc3 >> 15), (acc2 >> 15), 16);
emilmont 1:fdd22bb7aa52 455
emilmont 1:fdd22bb7aa52 456 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 457
emilmont 1:fdd22bb7aa52 458 /* Increment the pointer pIn1 index, count by 4 */
emilmont 1:fdd22bb7aa52 459 count += 4u;
emilmont 1:fdd22bb7aa52 460
emilmont 1:fdd22bb7aa52 461 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 462 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 463 py = pSrc2;
emilmont 1:fdd22bb7aa52 464
emilmont 1:fdd22bb7aa52 465 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 466 blkCnt--;
emilmont 1:fdd22bb7aa52 467 }
emilmont 1:fdd22bb7aa52 468
emilmont 1:fdd22bb7aa52 469 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 470 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 471 blkCnt = blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 472
emilmont 1:fdd22bb7aa52 473 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 474 {
emilmont 1:fdd22bb7aa52 475 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 476 sum = 0;
emilmont 1:fdd22bb7aa52 477
emilmont 1:fdd22bb7aa52 478 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 479 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 480
emilmont 1:fdd22bb7aa52 481 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 482 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 483 while(k > 0u)
emilmont 1:fdd22bb7aa52 484 {
emilmont 1:fdd22bb7aa52 485 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 486 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 487 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 488 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 489 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 490
emilmont 1:fdd22bb7aa52 491 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 492 k--;
emilmont 1:fdd22bb7aa52 493 }
emilmont 1:fdd22bb7aa52 494
emilmont 1:fdd22bb7aa52 495 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 496 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 497 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 498
emilmont 1:fdd22bb7aa52 499 while(k > 0u)
emilmont 1:fdd22bb7aa52 500 {
emilmont 1:fdd22bb7aa52 501 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 502 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 503
emilmont 1:fdd22bb7aa52 504 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 505 k--;
emilmont 1:fdd22bb7aa52 506 }
emilmont 1:fdd22bb7aa52 507
emilmont 1:fdd22bb7aa52 508 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 509 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 510
emilmont 1:fdd22bb7aa52 511 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 512 count++;
emilmont 1:fdd22bb7aa52 513
emilmont 1:fdd22bb7aa52 514 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 515 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 516 py = pSrc2;
emilmont 1:fdd22bb7aa52 517
emilmont 1:fdd22bb7aa52 518 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 519 blkCnt--;
emilmont 1:fdd22bb7aa52 520 }
emilmont 1:fdd22bb7aa52 521 }
emilmont 1:fdd22bb7aa52 522 else
emilmont 1:fdd22bb7aa52 523 {
emilmont 1:fdd22bb7aa52 524 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 525 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 526 blkCnt = blockSize2;
emilmont 1:fdd22bb7aa52 527
emilmont 1:fdd22bb7aa52 528 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 529 {
emilmont 1:fdd22bb7aa52 530 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 531 sum = 0;
emilmont 1:fdd22bb7aa52 532
emilmont 1:fdd22bb7aa52 533 /* srcBLen number of MACS should be performed */
emilmont 1:fdd22bb7aa52 534 k = srcBLen;
emilmont 1:fdd22bb7aa52 535
emilmont 1:fdd22bb7aa52 536 while(k > 0u)
emilmont 1:fdd22bb7aa52 537 {
emilmont 1:fdd22bb7aa52 538 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 539 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 540
emilmont 1:fdd22bb7aa52 541 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 542 k--;
emilmont 1:fdd22bb7aa52 543 }
emilmont 1:fdd22bb7aa52 544
emilmont 1:fdd22bb7aa52 545 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 546 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 547
emilmont 1:fdd22bb7aa52 548 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 549 count++;
emilmont 1:fdd22bb7aa52 550
emilmont 1:fdd22bb7aa52 551 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 552 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 553 py = pSrc2;
emilmont 1:fdd22bb7aa52 554
emilmont 1:fdd22bb7aa52 555 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 556 blkCnt--;
emilmont 1:fdd22bb7aa52 557 }
emilmont 1:fdd22bb7aa52 558 }
emilmont 1:fdd22bb7aa52 559
emilmont 1:fdd22bb7aa52 560
emilmont 1:fdd22bb7aa52 561 /* --------------------------
emilmont 1:fdd22bb7aa52 562 * Initializations of stage3
emilmont 1:fdd22bb7aa52 563 * -------------------------*/
emilmont 1:fdd22bb7aa52 564
emilmont 1:fdd22bb7aa52 565 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 566 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emilmont 1:fdd22bb7aa52 567 * ....
emilmont 1:fdd22bb7aa52 568 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emilmont 1:fdd22bb7aa52 569 * sum += x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 570 */
emilmont 1:fdd22bb7aa52 571
emilmont 1:fdd22bb7aa52 572 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 573 The blockSize3 variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 574
emilmont 1:fdd22bb7aa52 575 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 576 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 577 px = pSrc1;
emilmont 1:fdd22bb7aa52 578
emilmont 1:fdd22bb7aa52 579 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 580 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 581 pIn2 = pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 582 py = pIn2;
emilmont 1:fdd22bb7aa52 583
emilmont 1:fdd22bb7aa52 584 /* -------------------
emilmont 1:fdd22bb7aa52 585 * Stage3 process
emilmont 1:fdd22bb7aa52 586 * ------------------*/
emilmont 1:fdd22bb7aa52 587
emilmont 1:fdd22bb7aa52 588 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 589 /* First part of this stage computes the MAC operations greater than 4 */
emilmont 1:fdd22bb7aa52 590 /* Second part of this stage computes the MAC operations less than or equal to 4 */
emilmont 1:fdd22bb7aa52 591
emilmont 1:fdd22bb7aa52 592 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 593 j = blockSize3 >> 2u;
emilmont 1:fdd22bb7aa52 594
emilmont 1:fdd22bb7aa52 595 while((j > 0u) && (blockSize3 > 0u))
emilmont 1:fdd22bb7aa52 596 {
emilmont 1:fdd22bb7aa52 597 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 598 sum = 0;
emilmont 1:fdd22bb7aa52 599
emilmont 1:fdd22bb7aa52 600 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 601 k = blockSize3 >> 2u;
emilmont 1:fdd22bb7aa52 602
emilmont 1:fdd22bb7aa52 603 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 604 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 605 while(k > 0u)
emilmont 1:fdd22bb7aa52 606 {
emilmont 1:fdd22bb7aa52 607 /* x[srcALen - srcBLen + 1], x[srcALen - srcBLen + 2] are multiplied
emilmont 1:fdd22bb7aa52 608 * with y[srcBLen - 1], y[srcBLen - 2] respectively */
emilmont 1:fdd22bb7aa52 609 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 610 /* x[srcALen - srcBLen + 3], x[srcALen - srcBLen + 4] are multiplied
emilmont 1:fdd22bb7aa52 611 * with y[srcBLen - 3], y[srcBLen - 4] respectively */
emilmont 1:fdd22bb7aa52 612 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 613
emilmont 1:fdd22bb7aa52 614 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 615 k--;
emilmont 1:fdd22bb7aa52 616 }
emilmont 1:fdd22bb7aa52 617
emilmont 1:fdd22bb7aa52 618 /* For the next MAC operations, the pointer py is used without SIMD
emilmont 1:fdd22bb7aa52 619 * So, py is incremented by 1 */
emilmont 1:fdd22bb7aa52 620 py = py + 1u;
emilmont 1:fdd22bb7aa52 621
emilmont 1:fdd22bb7aa52 622 /* If the blockSize3 is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 623 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 624 k = blockSize3 % 0x4u;
emilmont 1:fdd22bb7aa52 625
emilmont 1:fdd22bb7aa52 626 while(k > 0u)
emilmont 1:fdd22bb7aa52 627 {
emilmont 1:fdd22bb7aa52 628 /* sum += x[srcALen - srcBLen + 5] * y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 629 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 630
emilmont 1:fdd22bb7aa52 631 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 632 k--;
emilmont 1:fdd22bb7aa52 633 }
emilmont 1:fdd22bb7aa52 634
emilmont 1:fdd22bb7aa52 635 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 636 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 637
emilmont 1:fdd22bb7aa52 638 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 639 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 640 py = pIn2;
emilmont 1:fdd22bb7aa52 641
emilmont 1:fdd22bb7aa52 642 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 643 blockSize3--;
emilmont 1:fdd22bb7aa52 644
emilmont 1:fdd22bb7aa52 645 j--;
emilmont 1:fdd22bb7aa52 646 }
emilmont 1:fdd22bb7aa52 647
emilmont 1:fdd22bb7aa52 648 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 649 /* SIMD is not used for the next MAC operations,
emilmont 1:fdd22bb7aa52 650 * so pointer py is updated to read only one sample at a time */
emilmont 1:fdd22bb7aa52 651 py = py + 1u;
emilmont 1:fdd22bb7aa52 652
emilmont 1:fdd22bb7aa52 653 while(blockSize3 > 0u)
emilmont 1:fdd22bb7aa52 654 {
emilmont 1:fdd22bb7aa52 655 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 656 sum = 0;
emilmont 1:fdd22bb7aa52 657
emilmont 1:fdd22bb7aa52 658 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 659 k = blockSize3;
emilmont 1:fdd22bb7aa52 660
emilmont 1:fdd22bb7aa52 661 while(k > 0u)
emilmont 1:fdd22bb7aa52 662 {
emilmont 1:fdd22bb7aa52 663 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 664 /* sum += x[srcALen-1] * y[srcBLen-1] */
emilmont 1:fdd22bb7aa52 665 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 666
emilmont 1:fdd22bb7aa52 667 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 668 k--;
emilmont 1:fdd22bb7aa52 669 }
emilmont 1:fdd22bb7aa52 670
emilmont 1:fdd22bb7aa52 671 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 672 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 673
emilmont 1:fdd22bb7aa52 674 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 675 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 676 py = pSrc2;
emilmont 1:fdd22bb7aa52 677
emilmont 1:fdd22bb7aa52 678 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 679 blockSize3--;
emilmont 1:fdd22bb7aa52 680 }
emilmont 1:fdd22bb7aa52 681
emilmont 1:fdd22bb7aa52 682 #else
emilmont 1:fdd22bb7aa52 683 q15_t *pIn1; /* inputA pointer */
emilmont 1:fdd22bb7aa52 684 q15_t *pIn2; /* inputB pointer */
emilmont 1:fdd22bb7aa52 685 q15_t *pOut = pDst; /* output pointer */
emilmont 1:fdd22bb7aa52 686 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
emilmont 1:fdd22bb7aa52 687 q15_t *px; /* Intermediate inputA pointer */
emilmont 1:fdd22bb7aa52 688 q15_t *py; /* Intermediate inputB pointer */
emilmont 1:fdd22bb7aa52 689 q15_t *pSrc1, *pSrc2; /* Intermediate pointers */
emilmont 1:fdd22bb7aa52 690 q31_t x0, x1, x2, x3, c0; /* Temporary variables to hold state and coefficient values */
emilmont 1:fdd22bb7aa52 691 uint32_t blockSize1, blockSize2, blockSize3, j, k, count, blkCnt; /* loop counter */
emilmont 1:fdd22bb7aa52 692 q15_t a, b;
emilmont 1:fdd22bb7aa52 693
emilmont 1:fdd22bb7aa52 694 /* The algorithm implementation is based on the lengths of the inputs. */
emilmont 1:fdd22bb7aa52 695 /* srcB is always made to slide across srcA. */
emilmont 1:fdd22bb7aa52 696 /* So srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 697 if(srcALen >= srcBLen)
emilmont 1:fdd22bb7aa52 698 {
emilmont 1:fdd22bb7aa52 699 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 700 pIn1 = pSrcA;
emilmont 1:fdd22bb7aa52 701
emilmont 1:fdd22bb7aa52 702 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 703 pIn2 = pSrcB;
emilmont 1:fdd22bb7aa52 704 }
emilmont 1:fdd22bb7aa52 705 else
emilmont 1:fdd22bb7aa52 706 {
emilmont 1:fdd22bb7aa52 707 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 708 pIn1 = pSrcB;
emilmont 1:fdd22bb7aa52 709
emilmont 1:fdd22bb7aa52 710 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 711 pIn2 = pSrcA;
emilmont 1:fdd22bb7aa52 712
emilmont 1:fdd22bb7aa52 713 /* srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 714 j = srcBLen;
emilmont 1:fdd22bb7aa52 715 srcBLen = srcALen;
emilmont 1:fdd22bb7aa52 716 srcALen = j;
emilmont 1:fdd22bb7aa52 717 }
emilmont 1:fdd22bb7aa52 718
emilmont 1:fdd22bb7aa52 719 /* conv(x,y) at n = x[n] * y[0] + x[n-1] * y[1] + x[n-2] * y[2] + ...+ x[n-N+1] * y[N -1] */
emilmont 1:fdd22bb7aa52 720 /* The function is internally
emilmont 1:fdd22bb7aa52 721 * divided into three stages according to the number of multiplications that has to be
emilmont 1:fdd22bb7aa52 722 * taken place between inputA samples and inputB samples. In the first stage of the
emilmont 1:fdd22bb7aa52 723 * algorithm, the multiplications increase by one for every iteration.
emilmont 1:fdd22bb7aa52 724 * In the second stage of the algorithm, srcBLen number of multiplications are done.
emilmont 1:fdd22bb7aa52 725 * In the third stage of the algorithm, the multiplications decrease by one
emilmont 1:fdd22bb7aa52 726 * for every iteration. */
emilmont 1:fdd22bb7aa52 727
emilmont 1:fdd22bb7aa52 728 /* The algorithm is implemented in three stages.
emilmont 1:fdd22bb7aa52 729 The loop counters of each stage is initiated here. */
emilmont 1:fdd22bb7aa52 730 blockSize1 = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 731 blockSize2 = srcALen - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 732 blockSize3 = blockSize1;
emilmont 1:fdd22bb7aa52 733
emilmont 1:fdd22bb7aa52 734 /* --------------------------
emilmont 1:fdd22bb7aa52 735 * Initializations of stage1
emilmont 1:fdd22bb7aa52 736 * -------------------------*/
emilmont 1:fdd22bb7aa52 737
emilmont 1:fdd22bb7aa52 738 /* sum = x[0] * y[0]
emilmont 1:fdd22bb7aa52 739 * sum = x[0] * y[1] + x[1] * y[0]
emilmont 1:fdd22bb7aa52 740 * ....
emilmont 1:fdd22bb7aa52 741 * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
emilmont 1:fdd22bb7aa52 742 */
emilmont 1:fdd22bb7aa52 743
emilmont 1:fdd22bb7aa52 744 /* In this stage the MAC operations are increased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 745 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 746 count = 1u;
emilmont 1:fdd22bb7aa52 747
emilmont 1:fdd22bb7aa52 748 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 749 px = pIn1;
emilmont 1:fdd22bb7aa52 750
emilmont 1:fdd22bb7aa52 751 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 752 py = pIn2;
emilmont 1:fdd22bb7aa52 753
emilmont 1:fdd22bb7aa52 754
emilmont 1:fdd22bb7aa52 755 /* ------------------------
emilmont 1:fdd22bb7aa52 756 * Stage1 process
emilmont 1:fdd22bb7aa52 757 * ----------------------*/
emilmont 1:fdd22bb7aa52 758
emilmont 1:fdd22bb7aa52 759 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 760 /* First part of this stage computes the MAC operations less than 4 */
emilmont 1:fdd22bb7aa52 761 /* Second part of this stage computes the MAC operations greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 762
emilmont 1:fdd22bb7aa52 763 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 764 while((count < 4u) && (blockSize1 > 0u))
emilmont 1:fdd22bb7aa52 765 {
emilmont 1:fdd22bb7aa52 766 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 767 sum = 0;
emilmont 1:fdd22bb7aa52 768
emilmont 1:fdd22bb7aa52 769 /* Loop over number of MAC operations between
emilmont 1:fdd22bb7aa52 770 * inputA samples and inputB samples */
emilmont 1:fdd22bb7aa52 771 k = count;
emilmont 1:fdd22bb7aa52 772
emilmont 1:fdd22bb7aa52 773 while(k > 0u)
emilmont 1:fdd22bb7aa52 774 {
emilmont 1:fdd22bb7aa52 775 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 776 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 777
emilmont 1:fdd22bb7aa52 778 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 779 k--;
emilmont 1:fdd22bb7aa52 780 }
emilmont 1:fdd22bb7aa52 781
emilmont 1:fdd22bb7aa52 782 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 783 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 784
emilmont 1:fdd22bb7aa52 785 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 786 py = pIn2 + count;
emilmont 1:fdd22bb7aa52 787 px = pIn1;
emilmont 1:fdd22bb7aa52 788
emilmont 1:fdd22bb7aa52 789 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 790 count++;
emilmont 1:fdd22bb7aa52 791
emilmont 1:fdd22bb7aa52 792 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 793 blockSize1--;
emilmont 1:fdd22bb7aa52 794 }
emilmont 1:fdd22bb7aa52 795
emilmont 1:fdd22bb7aa52 796 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 797 /* The internal loop, over count, is unrolled by 4 */
emilmont 1:fdd22bb7aa52 798 /* To, read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 799 * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */
emilmont 1:fdd22bb7aa52 800 py = py - 1;
emilmont 1:fdd22bb7aa52 801
emilmont 1:fdd22bb7aa52 802 while(blockSize1 > 0u)
emilmont 1:fdd22bb7aa52 803 {
emilmont 1:fdd22bb7aa52 804 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 805 sum = 0;
emilmont 1:fdd22bb7aa52 806
emilmont 1:fdd22bb7aa52 807 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 808 k = count >> 2u;
emilmont 1:fdd22bb7aa52 809
emilmont 1:fdd22bb7aa52 810 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 811 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 812 py++;
emilmont 1:fdd22bb7aa52 813
emilmont 1:fdd22bb7aa52 814 while(k > 0u)
emilmont 1:fdd22bb7aa52 815 {
emilmont 1:fdd22bb7aa52 816 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 817 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 818 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 819 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 820 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 821
emilmont 1:fdd22bb7aa52 822 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 823 k--;
emilmont 1:fdd22bb7aa52 824 }
emilmont 1:fdd22bb7aa52 825
emilmont 1:fdd22bb7aa52 826 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 827 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 828 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 829
emilmont 1:fdd22bb7aa52 830 while(k > 0u)
emilmont 1:fdd22bb7aa52 831 {
emilmont 1:fdd22bb7aa52 832 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 833 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 834
emilmont 1:fdd22bb7aa52 835 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 836 k--;
emilmont 1:fdd22bb7aa52 837 }
emilmont 1:fdd22bb7aa52 838
emilmont 1:fdd22bb7aa52 839 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 840 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 841
emilmont 1:fdd22bb7aa52 842 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 843 py = pIn2 + (count - 1u);
emilmont 1:fdd22bb7aa52 844 px = pIn1;
emilmont 1:fdd22bb7aa52 845
emilmont 1:fdd22bb7aa52 846 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 847 count++;
emilmont 1:fdd22bb7aa52 848
emilmont 1:fdd22bb7aa52 849 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 850 blockSize1--;
emilmont 1:fdd22bb7aa52 851 }
emilmont 1:fdd22bb7aa52 852
emilmont 1:fdd22bb7aa52 853 /* --------------------------
emilmont 1:fdd22bb7aa52 854 * Initializations of stage2
emilmont 1:fdd22bb7aa52 855 * ------------------------*/
emilmont 1:fdd22bb7aa52 856
emilmont 1:fdd22bb7aa52 857 /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
emilmont 1:fdd22bb7aa52 858 * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
emilmont 1:fdd22bb7aa52 859 * ....
emilmont 1:fdd22bb7aa52 860 * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
emilmont 1:fdd22bb7aa52 861 */
emilmont 1:fdd22bb7aa52 862
emilmont 1:fdd22bb7aa52 863 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 864 px = pIn1;
emilmont 1:fdd22bb7aa52 865
emilmont 1:fdd22bb7aa52 866 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 867 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 868 py = pSrc2;
emilmont 1:fdd22bb7aa52 869
emilmont 1:fdd22bb7aa52 870 /* count is the index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 871 count = 0u;
emilmont 1:fdd22bb7aa52 872
emilmont 1:fdd22bb7aa52 873
emilmont 1:fdd22bb7aa52 874 /* --------------------
emilmont 1:fdd22bb7aa52 875 * Stage2 process
emilmont 1:fdd22bb7aa52 876 * -------------------*/
emilmont 1:fdd22bb7aa52 877
emilmont 1:fdd22bb7aa52 878 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 879 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 880 * srcBLen should be greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 881 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 882 {
emilmont 1:fdd22bb7aa52 883 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 884 blkCnt = blockSize2 >> 2u;
emilmont 1:fdd22bb7aa52 885
emilmont 1:fdd22bb7aa52 886 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 887 {
emilmont 1:fdd22bb7aa52 888 py = py - 1u;
emilmont 1:fdd22bb7aa52 889
emilmont 1:fdd22bb7aa52 890 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 891 acc0 = 0;
emilmont 1:fdd22bb7aa52 892 acc1 = 0;
emilmont 1:fdd22bb7aa52 893 acc2 = 0;
emilmont 1:fdd22bb7aa52 894 acc3 = 0;
emilmont 1:fdd22bb7aa52 895
emilmont 1:fdd22bb7aa52 896 /* read x[0], x[1] samples */
emilmont 1:fdd22bb7aa52 897 a = *px++;
emilmont 1:fdd22bb7aa52 898 b = *px++;
emilmont 1:fdd22bb7aa52 899
emilmont 1:fdd22bb7aa52 900 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 901
emilmont 1:fdd22bb7aa52 902 x0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 903 a = *px;
emilmont 1:fdd22bb7aa52 904 x1 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 905
emilmont 1:fdd22bb7aa52 906 #else
emilmont 1:fdd22bb7aa52 907
emilmont 1:fdd22bb7aa52 908 x0 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 909 a = *px;
emilmont 1:fdd22bb7aa52 910 x1 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 911
emilmont 1:fdd22bb7aa52 912 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 913
emilmont 1:fdd22bb7aa52 914 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 915 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 916
emilmont 1:fdd22bb7aa52 917 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 918 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 919 do
emilmont 1:fdd22bb7aa52 920 {
emilmont 1:fdd22bb7aa52 921 /* Read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 922 * y[srcBLen - 1] and y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 923 a = *py;
emilmont 1:fdd22bb7aa52 924 b = *(py+1);
emilmont 1:fdd22bb7aa52 925 py -= 2;
emilmont 1:fdd22bb7aa52 926
emilmont 1:fdd22bb7aa52 927 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 928
emilmont 1:fdd22bb7aa52 929 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 930
emilmont 1:fdd22bb7aa52 931 #else
emilmont 1:fdd22bb7aa52 932
emilmont 1:fdd22bb7aa52 933 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 934
emilmont 1:fdd22bb7aa52 935 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 936
emilmont 1:fdd22bb7aa52 937 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 938 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 939
emilmont 1:fdd22bb7aa52 940 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 941 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 942
emilmont 1:fdd22bb7aa52 943 a = *px;
emilmont 1:fdd22bb7aa52 944 b = *(px + 1);
emilmont 1:fdd22bb7aa52 945
emilmont 1:fdd22bb7aa52 946 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 947
emilmont 1:fdd22bb7aa52 948 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 949 a = *(px + 2);
emilmont 1:fdd22bb7aa52 950 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 951
emilmont 1:fdd22bb7aa52 952 #else
emilmont 1:fdd22bb7aa52 953
emilmont 1:fdd22bb7aa52 954 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 955 a = *(px + 2);
emilmont 1:fdd22bb7aa52 956 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 957
emilmont 1:fdd22bb7aa52 958 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 959
emilmont 1:fdd22bb7aa52 960 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 961 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 962
emilmont 1:fdd22bb7aa52 963 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 964 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 965
emilmont 1:fdd22bb7aa52 966 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 967 a = *py;
emilmont 1:fdd22bb7aa52 968 b = *(py+1);
emilmont 1:fdd22bb7aa52 969 py -= 2;
emilmont 1:fdd22bb7aa52 970
emilmont 1:fdd22bb7aa52 971 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 972
emilmont 1:fdd22bb7aa52 973 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 974
emilmont 1:fdd22bb7aa52 975 #else
emilmont 1:fdd22bb7aa52 976
emilmont 1:fdd22bb7aa52 977 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 978
emilmont 1:fdd22bb7aa52 979 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 980
emilmont 1:fdd22bb7aa52 981 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 982 acc0 = __SMLADX(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 983
emilmont 1:fdd22bb7aa52 984 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 985 acc1 = __SMLADX(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 986
emilmont 1:fdd22bb7aa52 987 /* Read x[4], x[5], x[6] */
emilmont 1:fdd22bb7aa52 988 a = *(px + 2);
emilmont 1:fdd22bb7aa52 989 b = *(px + 3);
emilmont 1:fdd22bb7aa52 990
emilmont 1:fdd22bb7aa52 991 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 992
emilmont 1:fdd22bb7aa52 993 x0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 994 a = *(px + 4);
emilmont 1:fdd22bb7aa52 995 x1 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 996
emilmont 1:fdd22bb7aa52 997 #else
emilmont 1:fdd22bb7aa52 998
emilmont 1:fdd22bb7aa52 999 x0 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1000 a = *(px + 4);
emilmont 1:fdd22bb7aa52 1001 x1 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1002
emilmont 1:fdd22bb7aa52 1003 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1004
emilmont 1:fdd22bb7aa52 1005 px += 4u;
emilmont 1:fdd22bb7aa52 1006
emilmont 1:fdd22bb7aa52 1007 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1008 acc2 = __SMLADX(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 1009
emilmont 1:fdd22bb7aa52 1010 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1011 acc3 = __SMLADX(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 1012
emilmont 1:fdd22bb7aa52 1013 } while(--k);
emilmont 1:fdd22bb7aa52 1014
emilmont 1:fdd22bb7aa52 1015 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 1016 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 1017
emilmont 1:fdd22bb7aa52 1018 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1019 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1020 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 1021
emilmont 1:fdd22bb7aa52 1022 if(k == 1u)
emilmont 1:fdd22bb7aa52 1023 {
emilmont 1:fdd22bb7aa52 1024 /* Read y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 1025 c0 = *(py+1);
emilmont 1:fdd22bb7aa52 1026
emilmont 1:fdd22bb7aa52 1027 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1028
emilmont 1:fdd22bb7aa52 1029 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 1030
emilmont 1:fdd22bb7aa52 1031 #else
emilmont 1:fdd22bb7aa52 1032
emilmont 1:fdd22bb7aa52 1033 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 1034
emilmont 1:fdd22bb7aa52 1035 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1036
emilmont 1:fdd22bb7aa52 1037 /* Read x[7] */
emilmont 1:fdd22bb7aa52 1038 a = *px;
emilmont 1:fdd22bb7aa52 1039 b = *(px+1);
emilmont 1:fdd22bb7aa52 1040 px++;
emilmont 1:fdd22bb7aa52 1041
emilmont 1:fdd22bb7aa52 1042 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1043
emilmont 1:fdd22bb7aa52 1044 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1045
emilmont 1:fdd22bb7aa52 1046 #else
emilmont 1:fdd22bb7aa52 1047
emilmont 1:fdd22bb7aa52 1048 x3 = __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
emilmont 1:fdd22bb7aa52 1053 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1054 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1055 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1056 acc2 = __SMLADX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 1057 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1058 }
emilmont 1:fdd22bb7aa52 1059
emilmont 1:fdd22bb7aa52 1060 if(k == 2u)
emilmont 1:fdd22bb7aa52 1061 {
emilmont 1:fdd22bb7aa52 1062 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 1063 a = *py;
emilmont 1:fdd22bb7aa52 1064 b = *(py+1);
emilmont 1:fdd22bb7aa52 1065
emilmont 1:fdd22bb7aa52 1066 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1067
emilmont 1:fdd22bb7aa52 1068 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1069
emilmont 1:fdd22bb7aa52 1070 #else
emilmont 1:fdd22bb7aa52 1071
emilmont 1:fdd22bb7aa52 1072 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1073
emilmont 1:fdd22bb7aa52 1074 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1075
emilmont 1:fdd22bb7aa52 1076 /* Read x[7], x[8], x[9] */
emilmont 1:fdd22bb7aa52 1077 a = *px;
emilmont 1:fdd22bb7aa52 1078 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1079
emilmont 1:fdd22bb7aa52 1080 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1081
emilmont 1:fdd22bb7aa52 1082 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1083 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1084 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1085
emilmont 1:fdd22bb7aa52 1086 #else
emilmont 1:fdd22bb7aa52 1087
emilmont 1:fdd22bb7aa52 1088 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1089 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1090 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1091
emilmont 1:fdd22bb7aa52 1092 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1093 px += 2u;
emilmont 1:fdd22bb7aa52 1094
emilmont 1:fdd22bb7aa52 1095 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1096 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1097 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1098 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 1099 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 1100 }
emilmont 1:fdd22bb7aa52 1101
emilmont 1:fdd22bb7aa52 1102 if(k == 3u)
emilmont 1:fdd22bb7aa52 1103 {
emilmont 1:fdd22bb7aa52 1104 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 1105 a = *py;
emilmont 1:fdd22bb7aa52 1106 b = *(py+1);
emilmont 1:fdd22bb7aa52 1107
emilmont 1:fdd22bb7aa52 1108 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1109
emilmont 1:fdd22bb7aa52 1110 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1111
emilmont 1:fdd22bb7aa52 1112 #else
emilmont 1:fdd22bb7aa52 1113
emilmont 1:fdd22bb7aa52 1114 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1115
emilmont 1:fdd22bb7aa52 1116 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1117
emilmont 1:fdd22bb7aa52 1118 /* Read x[7], x[8], x[9] */
emilmont 1:fdd22bb7aa52 1119 a = *px;
emilmont 1:fdd22bb7aa52 1120 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1121
emilmont 1:fdd22bb7aa52 1122 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1123
emilmont 1:fdd22bb7aa52 1124 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1125 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1126 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1127
emilmont 1:fdd22bb7aa52 1128 #else
emilmont 1:fdd22bb7aa52 1129
emilmont 1:fdd22bb7aa52 1130 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1131 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1132 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1133
emilmont 1:fdd22bb7aa52 1134 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1135
emilmont 1:fdd22bb7aa52 1136 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1137 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1138 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1139 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 1140 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 1141
emilmont 1:fdd22bb7aa52 1142 /* Read y[srcBLen - 7] */
emilmont 1:fdd22bb7aa52 1143 c0 = *(py-1);
emilmont 1:fdd22bb7aa52 1144 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1145
emilmont 1:fdd22bb7aa52 1146 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 1147 #else
emilmont 1:fdd22bb7aa52 1148
emilmont 1:fdd22bb7aa52 1149 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 1150 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1151
emilmont 1:fdd22bb7aa52 1152 /* Read x[10] */
emilmont 1:fdd22bb7aa52 1153 a = *(px+2);
emilmont 1:fdd22bb7aa52 1154 b = *(px+3);
emilmont 1:fdd22bb7aa52 1155
emilmont 1:fdd22bb7aa52 1156 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1157
emilmont 1:fdd22bb7aa52 1158 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1159
emilmont 1:fdd22bb7aa52 1160 #else
emilmont 1:fdd22bb7aa52 1161
emilmont 1:fdd22bb7aa52 1162 x3 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1163
emilmont 1:fdd22bb7aa52 1164 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1165
emilmont 1:fdd22bb7aa52 1166 px += 3u;
emilmont 1:fdd22bb7aa52 1167
emilmont 1:fdd22bb7aa52 1168 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1169 acc0 = __SMLADX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 1170 acc1 = __SMLAD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 1171 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 1172 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1173 }
emilmont 1:fdd22bb7aa52 1174
emilmont 1:fdd22bb7aa52 1175 /* Store the results in the accumulators in the destination buffer. */
emilmont 1:fdd22bb7aa52 1176 *pOut++ = (q15_t)(acc0 >> 15);
emilmont 1:fdd22bb7aa52 1177 *pOut++ = (q15_t)(acc1 >> 15);
emilmont 1:fdd22bb7aa52 1178 *pOut++ = (q15_t)(acc2 >> 15);
emilmont 1:fdd22bb7aa52 1179 *pOut++ = (q15_t)(acc3 >> 15);
emilmont 1:fdd22bb7aa52 1180
emilmont 1:fdd22bb7aa52 1181 /* Increment the pointer pIn1 index, count by 4 */
emilmont 1:fdd22bb7aa52 1182 count += 4u;
emilmont 1:fdd22bb7aa52 1183
emilmont 1:fdd22bb7aa52 1184 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1185 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1186 py = pSrc2;
emilmont 1:fdd22bb7aa52 1187
emilmont 1:fdd22bb7aa52 1188 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1189 blkCnt--;
emilmont 1:fdd22bb7aa52 1190 }
emilmont 1:fdd22bb7aa52 1191
emilmont 1:fdd22bb7aa52 1192 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 1193 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1194 blkCnt = blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 1195
emilmont 1:fdd22bb7aa52 1196 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 1197 {
emilmont 1:fdd22bb7aa52 1198 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1199 sum = 0;
emilmont 1:fdd22bb7aa52 1200
emilmont 1:fdd22bb7aa52 1201 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1202 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 1203
emilmont 1:fdd22bb7aa52 1204 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 1205 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 1206 while(k > 0u)
emilmont 1:fdd22bb7aa52 1207 {
emilmont 1:fdd22bb7aa52 1208 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1209 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1210 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1211 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1212 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1213
emilmont 1:fdd22bb7aa52 1214 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1215 k--;
emilmont 1:fdd22bb7aa52 1216 }
emilmont 1:fdd22bb7aa52 1217
emilmont 1:fdd22bb7aa52 1218 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1219 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1220 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 1221
emilmont 1:fdd22bb7aa52 1222 while(k > 0u)
emilmont 1:fdd22bb7aa52 1223 {
emilmont 1:fdd22bb7aa52 1224 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1225 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1226
emilmont 1:fdd22bb7aa52 1227 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1228 k--;
emilmont 1:fdd22bb7aa52 1229 }
emilmont 1:fdd22bb7aa52 1230
emilmont 1:fdd22bb7aa52 1231 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1232 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1233
emilmont 1:fdd22bb7aa52 1234 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 1235 count++;
emilmont 1:fdd22bb7aa52 1236
emilmont 1:fdd22bb7aa52 1237 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1238 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1239 py = pSrc2;
emilmont 1:fdd22bb7aa52 1240
emilmont 1:fdd22bb7aa52 1241 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1242 blkCnt--;
emilmont 1:fdd22bb7aa52 1243 }
emilmont 1:fdd22bb7aa52 1244 }
emilmont 1:fdd22bb7aa52 1245 else
emilmont 1:fdd22bb7aa52 1246 {
emilmont 1:fdd22bb7aa52 1247 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 1248 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 1249 blkCnt = blockSize2;
emilmont 1:fdd22bb7aa52 1250
emilmont 1:fdd22bb7aa52 1251 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 1252 {
emilmont 1:fdd22bb7aa52 1253 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1254 sum = 0;
emilmont 1:fdd22bb7aa52 1255
emilmont 1:fdd22bb7aa52 1256 /* srcBLen number of MACS should be performed */
emilmont 1:fdd22bb7aa52 1257 k = srcBLen;
emilmont 1:fdd22bb7aa52 1258
emilmont 1:fdd22bb7aa52 1259 while(k > 0u)
emilmont 1:fdd22bb7aa52 1260 {
emilmont 1:fdd22bb7aa52 1261 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 1262 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1263
emilmont 1:fdd22bb7aa52 1264 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1265 k--;
emilmont 1:fdd22bb7aa52 1266 }
emilmont 1:fdd22bb7aa52 1267
emilmont 1:fdd22bb7aa52 1268 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1269 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1270
emilmont 1:fdd22bb7aa52 1271 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 1272 count++;
emilmont 1:fdd22bb7aa52 1273
emilmont 1:fdd22bb7aa52 1274 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1275 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1276 py = pSrc2;
emilmont 1:fdd22bb7aa52 1277
emilmont 1:fdd22bb7aa52 1278 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1279 blkCnt--;
emilmont 1:fdd22bb7aa52 1280 }
emilmont 1:fdd22bb7aa52 1281 }
emilmont 1:fdd22bb7aa52 1282
emilmont 1:fdd22bb7aa52 1283
emilmont 1:fdd22bb7aa52 1284 /* --------------------------
emilmont 1:fdd22bb7aa52 1285 * Initializations of stage3
emilmont 1:fdd22bb7aa52 1286 * -------------------------*/
emilmont 1:fdd22bb7aa52 1287
emilmont 1:fdd22bb7aa52 1288 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 1289 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emilmont 1:fdd22bb7aa52 1290 * ....
emilmont 1:fdd22bb7aa52 1291 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emilmont 1:fdd22bb7aa52 1292 * sum += x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 1293 */
emilmont 1:fdd22bb7aa52 1294
emilmont 1:fdd22bb7aa52 1295 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 1296 The blockSize3 variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 1297
emilmont 1:fdd22bb7aa52 1298 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 1299 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 1300 px = pSrc1;
emilmont 1:fdd22bb7aa52 1301
emilmont 1:fdd22bb7aa52 1302 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 1303 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 1304 pIn2 = pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 1305 py = pIn2;
emilmont 1:fdd22bb7aa52 1306
emilmont 1:fdd22bb7aa52 1307 /* -------------------
emilmont 1:fdd22bb7aa52 1308 * Stage3 process
emilmont 1:fdd22bb7aa52 1309 * ------------------*/
emilmont 1:fdd22bb7aa52 1310
emilmont 1:fdd22bb7aa52 1311 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 1312 /* First part of this stage computes the MAC operations greater than 4 */
emilmont 1:fdd22bb7aa52 1313 /* Second part of this stage computes the MAC operations less than or equal to 4 */
emilmont 1:fdd22bb7aa52 1314
emilmont 1:fdd22bb7aa52 1315 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 1316 j = blockSize3 >> 2u;
emilmont 1:fdd22bb7aa52 1317
emilmont 1:fdd22bb7aa52 1318 while((j > 0u) && (blockSize3 > 0u))
emilmont 1:fdd22bb7aa52 1319 {
emilmont 1:fdd22bb7aa52 1320 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1321 sum = 0;
emilmont 1:fdd22bb7aa52 1322
emilmont 1:fdd22bb7aa52 1323 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1324 k = blockSize3 >> 2u;
emilmont 1:fdd22bb7aa52 1325
emilmont 1:fdd22bb7aa52 1326 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 1327 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 1328 py++;
emilmont 1:fdd22bb7aa52 1329
emilmont 1:fdd22bb7aa52 1330 while(k > 0u)
emilmont 1:fdd22bb7aa52 1331 {
emilmont 1:fdd22bb7aa52 1332 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1333 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1334 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1335 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1336 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1337 k--;
emilmont 1:fdd22bb7aa52 1338 }
emilmont 1:fdd22bb7aa52 1339
emilmont 1:fdd22bb7aa52 1340 /* If the blockSize3 is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1341 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1342 k = blockSize3 % 0x4u;
emilmont 1:fdd22bb7aa52 1343
emilmont 1:fdd22bb7aa52 1344 while(k > 0u)
emilmont 1:fdd22bb7aa52 1345 {
emilmont 1:fdd22bb7aa52 1346 /* sum += x[srcALen - srcBLen + 5] * y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 1347 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1348
emilmont 1:fdd22bb7aa52 1349 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1350 k--;
emilmont 1:fdd22bb7aa52 1351 }
emilmont 1:fdd22bb7aa52 1352
emilmont 1:fdd22bb7aa52 1353 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1354 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1355
emilmont 1:fdd22bb7aa52 1356 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1357 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 1358 py = pIn2;
emilmont 1:fdd22bb7aa52 1359
emilmont 1:fdd22bb7aa52 1360 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1361 blockSize3--;
emilmont 1:fdd22bb7aa52 1362
emilmont 1:fdd22bb7aa52 1363 j--;
emilmont 1:fdd22bb7aa52 1364 }
emilmont 1:fdd22bb7aa52 1365
emilmont 1:fdd22bb7aa52 1366 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 1367 /* SIMD is not used for the next MAC operations,
emilmont 1:fdd22bb7aa52 1368 * so pointer py is updated to read only one sample at a time */
emilmont 1:fdd22bb7aa52 1369 py = py + 1u;
emilmont 1:fdd22bb7aa52 1370
emilmont 1:fdd22bb7aa52 1371 while(blockSize3 > 0u)
emilmont 1:fdd22bb7aa52 1372 {
emilmont 1:fdd22bb7aa52 1373 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1374 sum = 0;
emilmont 1:fdd22bb7aa52 1375
emilmont 1:fdd22bb7aa52 1376 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1377 k = blockSize3;
emilmont 1:fdd22bb7aa52 1378
emilmont 1:fdd22bb7aa52 1379 while(k > 0u)
emilmont 1:fdd22bb7aa52 1380 {
emilmont 1:fdd22bb7aa52 1381 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1382 /* sum += x[srcALen-1] * y[srcBLen-1] */
emilmont 1:fdd22bb7aa52 1383 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1384
emilmont 1:fdd22bb7aa52 1385 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1386 k--;
emilmont 1:fdd22bb7aa52 1387 }
emilmont 1:fdd22bb7aa52 1388
emilmont 1:fdd22bb7aa52 1389 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1390 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1391
emilmont 1:fdd22bb7aa52 1392 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1393 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 1394 py = pSrc2;
emilmont 1:fdd22bb7aa52 1395
emilmont 1:fdd22bb7aa52 1396 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1397 blockSize3--;
emilmont 1:fdd22bb7aa52 1398 }
emilmont 1:fdd22bb7aa52 1399
emilmont 1:fdd22bb7aa52 1400 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
emilmont 1:fdd22bb7aa52 1401 }
emilmont 1:fdd22bb7aa52 1402
emilmont 1:fdd22bb7aa52 1403 /**
emilmont 1:fdd22bb7aa52 1404 * @} end of Conv group
emilmont 1:fdd22bb7aa52 1405 */