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_conv_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_q15.c
emilmont 1:fdd22bb7aa52 9 *
emilmont 1:fdd22bb7aa52 10 * Description: Convolution of Q15 sequences.
emilmont 1:fdd22bb7aa52 11 *
emilmont 1:fdd22bb7aa52 12 * Target Processor: Cortex-M4/Cortex-M3/Cortex-M0
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 * Version 0.0.7 2010/06/10
emilmont 1:fdd22bb7aa52 36 * Misra-C changes done
emilmont 1:fdd22bb7aa52 37 *
emilmont 1:fdd22bb7aa52 38 * -------------------------------------------------------------------- */
emilmont 1:fdd22bb7aa52 39
emilmont 1:fdd22bb7aa52 40 #include "arm_math.h"
emilmont 1:fdd22bb7aa52 41
emilmont 1:fdd22bb7aa52 42 /**
emilmont 1:fdd22bb7aa52 43 * @ingroup groupFilters
emilmont 1:fdd22bb7aa52 44 */
emilmont 1:fdd22bb7aa52 45
emilmont 1:fdd22bb7aa52 46 /**
emilmont 1:fdd22bb7aa52 47 * @addtogroup Conv
emilmont 1:fdd22bb7aa52 48 * @{
emilmont 1:fdd22bb7aa52 49 */
emilmont 1:fdd22bb7aa52 50
emilmont 1:fdd22bb7aa52 51 /**
emilmont 1:fdd22bb7aa52 52 * @brief Convolution of Q15 sequences.
emilmont 1:fdd22bb7aa52 53 * @param[in] *pSrcA points to the first input sequence.
emilmont 1:fdd22bb7aa52 54 * @param[in] srcALen length of the first input sequence.
emilmont 1:fdd22bb7aa52 55 * @param[in] *pSrcB points to the second input sequence.
emilmont 1:fdd22bb7aa52 56 * @param[in] srcBLen length of the second input sequence.
emilmont 1:fdd22bb7aa52 57 * @param[out] *pDst points to the location where the output result is written. Length srcALen+srcBLen-1.
emilmont 1:fdd22bb7aa52 58 * @return none.
emilmont 1:fdd22bb7aa52 59 *
emilmont 1:fdd22bb7aa52 60 * @details
emilmont 1:fdd22bb7aa52 61 * <b>Scaling and Overflow Behavior:</b>
emilmont 1:fdd22bb7aa52 62 *
emilmont 1:fdd22bb7aa52 63 * \par
emilmont 1:fdd22bb7aa52 64 * The function is implemented using a 64-bit internal accumulator.
emilmont 1:fdd22bb7aa52 65 * Both inputs are in 1.15 format and multiplications yield a 2.30 result.
emilmont 1:fdd22bb7aa52 66 * The 2.30 intermediate results are accumulated in a 64-bit accumulator in 34.30 format.
emilmont 1:fdd22bb7aa52 67 * This approach provides 33 guard bits and there is no risk of overflow.
emilmont 1:fdd22bb7aa52 68 * The 34.30 result is then truncated to 34.15 format by discarding the low 15 bits and then saturated to 1.15 format.
emilmont 1:fdd22bb7aa52 69 *
emilmont 1:fdd22bb7aa52 70 * \par
emilmont 1:fdd22bb7aa52 71 * Refer to <code>arm_conv_fast_q15()</code> for a faster but less precise version of this function for Cortex-M3 and Cortex-M4.
emilmont 1:fdd22bb7aa52 72 *
emilmont 1:fdd22bb7aa52 73 * \par
emilmont 1:fdd22bb7aa52 74 * Refer the function <code>arm_conv_opt_q15()</code> for a faster implementation of this function using scratch buffers.
emilmont 1:fdd22bb7aa52 75 *
emilmont 1:fdd22bb7aa52 76 */
emilmont 1:fdd22bb7aa52 77
emilmont 1:fdd22bb7aa52 78 void arm_conv_q15(
emilmont 1:fdd22bb7aa52 79 q15_t * pSrcA,
emilmont 1:fdd22bb7aa52 80 uint32_t srcALen,
emilmont 1:fdd22bb7aa52 81 q15_t * pSrcB,
emilmont 1:fdd22bb7aa52 82 uint32_t srcBLen,
emilmont 1:fdd22bb7aa52 83 q15_t * pDst)
emilmont 1:fdd22bb7aa52 84 {
emilmont 1:fdd22bb7aa52 85
emilmont 1:fdd22bb7aa52 86 #if (defined(ARM_MATH_CM4) || defined(ARM_MATH_CM3)) && !defined(UNALIGNED_SUPPORT_DISABLE)
emilmont 1:fdd22bb7aa52 87
emilmont 1:fdd22bb7aa52 88 /* Run the below code for Cortex-M4 and Cortex-M3 */
emilmont 1:fdd22bb7aa52 89
emilmont 1:fdd22bb7aa52 90 q15_t *pIn1; /* inputA pointer */
emilmont 1:fdd22bb7aa52 91 q15_t *pIn2; /* inputB pointer */
emilmont 1:fdd22bb7aa52 92 q15_t *pOut = pDst; /* output pointer */
emilmont 1:fdd22bb7aa52 93 q63_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
emilmont 1:fdd22bb7aa52 94 q15_t *px; /* Intermediate inputA pointer */
emilmont 1:fdd22bb7aa52 95 q15_t *py; /* Intermediate inputB pointer */
emilmont 1:fdd22bb7aa52 96 q15_t *pSrc1, *pSrc2; /* Intermediate pointers */
emilmont 1:fdd22bb7aa52 97 q31_t x0, x1, x2, x3, c0; /* Temporary variables to hold state and coefficient values */
emilmont 1:fdd22bb7aa52 98 uint32_t blockSize1, blockSize2, blockSize3, j, k, count, blkCnt; /* loop counter */
emilmont 1:fdd22bb7aa52 99
emilmont 1:fdd22bb7aa52 100 /* The algorithm implementation is based on the lengths of the inputs. */
emilmont 1:fdd22bb7aa52 101 /* srcB is always made to slide across srcA. */
emilmont 1:fdd22bb7aa52 102 /* So srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 103 if(srcALen >= srcBLen)
emilmont 1:fdd22bb7aa52 104 {
emilmont 1:fdd22bb7aa52 105 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 106 pIn1 = pSrcA;
emilmont 1:fdd22bb7aa52 107
emilmont 1:fdd22bb7aa52 108 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 109 pIn2 = pSrcB;
emilmont 1:fdd22bb7aa52 110 }
emilmont 1:fdd22bb7aa52 111 else
emilmont 1:fdd22bb7aa52 112 {
emilmont 1:fdd22bb7aa52 113 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 114 pIn1 = pSrcB;
emilmont 1:fdd22bb7aa52 115
emilmont 1:fdd22bb7aa52 116 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 117 pIn2 = pSrcA;
emilmont 1:fdd22bb7aa52 118
emilmont 1:fdd22bb7aa52 119 /* srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 120 j = srcBLen;
emilmont 1:fdd22bb7aa52 121 srcBLen = srcALen;
emilmont 1:fdd22bb7aa52 122 srcALen = j;
emilmont 1:fdd22bb7aa52 123 }
emilmont 1:fdd22bb7aa52 124
emilmont 1:fdd22bb7aa52 125 /* 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 126 /* The function is internally
emilmont 1:fdd22bb7aa52 127 * divided into three stages according to the number of multiplications that has to be
emilmont 1:fdd22bb7aa52 128 * taken place between inputA samples and inputB samples. In the first stage of the
emilmont 1:fdd22bb7aa52 129 * algorithm, the multiplications increase by one for every iteration.
emilmont 1:fdd22bb7aa52 130 * In the second stage of the algorithm, srcBLen number of multiplications are done.
emilmont 1:fdd22bb7aa52 131 * In the third stage of the algorithm, the multiplications decrease by one
emilmont 1:fdd22bb7aa52 132 * for every iteration. */
emilmont 1:fdd22bb7aa52 133
emilmont 1:fdd22bb7aa52 134 /* The algorithm is implemented in three stages.
emilmont 1:fdd22bb7aa52 135 The loop counters of each stage is initiated here. */
emilmont 1:fdd22bb7aa52 136 blockSize1 = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 137 blockSize2 = srcALen - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 138
emilmont 1:fdd22bb7aa52 139 /* --------------------------
emilmont 1:fdd22bb7aa52 140 * Initializations of stage1
emilmont 1:fdd22bb7aa52 141 * -------------------------*/
emilmont 1:fdd22bb7aa52 142
emilmont 1:fdd22bb7aa52 143 /* sum = x[0] * y[0]
emilmont 1:fdd22bb7aa52 144 * sum = x[0] * y[1] + x[1] * y[0]
emilmont 1:fdd22bb7aa52 145 * ....
emilmont 1:fdd22bb7aa52 146 * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
emilmont 1:fdd22bb7aa52 147 */
emilmont 1:fdd22bb7aa52 148
emilmont 1:fdd22bb7aa52 149 /* In this stage the MAC operations are increased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 150 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 151 count = 1u;
emilmont 1:fdd22bb7aa52 152
emilmont 1:fdd22bb7aa52 153 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 154 px = pIn1;
emilmont 1:fdd22bb7aa52 155
emilmont 1:fdd22bb7aa52 156 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 157 py = pIn2;
emilmont 1:fdd22bb7aa52 158
emilmont 1:fdd22bb7aa52 159
emilmont 1:fdd22bb7aa52 160 /* ------------------------
emilmont 1:fdd22bb7aa52 161 * Stage1 process
emilmont 1:fdd22bb7aa52 162 * ----------------------*/
emilmont 1:fdd22bb7aa52 163
emilmont 1:fdd22bb7aa52 164 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 165 /* First part of this stage computes the MAC operations less than 4 */
emilmont 1:fdd22bb7aa52 166 /* Second part of this stage computes the MAC operations greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 167
emilmont 1:fdd22bb7aa52 168 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 169 while((count < 4u) && (blockSize1 > 0u))
emilmont 1:fdd22bb7aa52 170 {
emilmont 1:fdd22bb7aa52 171 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 172 sum = 0;
emilmont 1:fdd22bb7aa52 173
emilmont 1:fdd22bb7aa52 174 /* Loop over number of MAC operations between
emilmont 1:fdd22bb7aa52 175 * inputA samples and inputB samples */
emilmont 1:fdd22bb7aa52 176 k = count;
emilmont 1:fdd22bb7aa52 177
emilmont 1:fdd22bb7aa52 178 while(k > 0u)
emilmont 1:fdd22bb7aa52 179 {
emilmont 1:fdd22bb7aa52 180 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 181 sum = __SMLALD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 182
emilmont 1:fdd22bb7aa52 183 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 184 k--;
emilmont 1:fdd22bb7aa52 185 }
emilmont 1:fdd22bb7aa52 186
emilmont 1:fdd22bb7aa52 187 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 188 *pOut++ = (q15_t) (__SSAT((sum >> 15), 16));
emilmont 1:fdd22bb7aa52 189
emilmont 1:fdd22bb7aa52 190 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 191 py = pIn2 + count;
emilmont 1:fdd22bb7aa52 192 px = pIn1;
emilmont 1:fdd22bb7aa52 193
emilmont 1:fdd22bb7aa52 194 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 195 count++;
emilmont 1:fdd22bb7aa52 196
emilmont 1:fdd22bb7aa52 197 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 198 blockSize1--;
emilmont 1:fdd22bb7aa52 199 }
emilmont 1:fdd22bb7aa52 200
emilmont 1:fdd22bb7aa52 201 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 202 /* The internal loop, over count, is unrolled by 4 */
emilmont 1:fdd22bb7aa52 203 /* To, read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 204 * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */
emilmont 1:fdd22bb7aa52 205 py = py - 1;
emilmont 1:fdd22bb7aa52 206
emilmont 1:fdd22bb7aa52 207 while(blockSize1 > 0u)
emilmont 1:fdd22bb7aa52 208 {
emilmont 1:fdd22bb7aa52 209 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 210 sum = 0;
emilmont 1:fdd22bb7aa52 211
emilmont 1:fdd22bb7aa52 212 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 213 k = count >> 2u;
emilmont 1:fdd22bb7aa52 214
emilmont 1:fdd22bb7aa52 215 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 216 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 217 while(k > 0u)
emilmont 1:fdd22bb7aa52 218 {
emilmont 1:fdd22bb7aa52 219 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 220 /* x[0], x[1] are multiplied with y[srcBLen - 1], y[srcBLen - 2] respectively */
emilmont 1:fdd22bb7aa52 221 sum = __SMLALDX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 222 /* x[2], x[3] are multiplied with y[srcBLen - 3], y[srcBLen - 4] respectively */
emilmont 1:fdd22bb7aa52 223 sum = __SMLALDX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 224
emilmont 1:fdd22bb7aa52 225 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 226 k--;
emilmont 1:fdd22bb7aa52 227 }
emilmont 1:fdd22bb7aa52 228
emilmont 1:fdd22bb7aa52 229 /* For the next MAC operations, the pointer py is used without SIMD
emilmont 1:fdd22bb7aa52 230 * So, py is incremented by 1 */
emilmont 1:fdd22bb7aa52 231 py = py + 1u;
emilmont 1:fdd22bb7aa52 232
emilmont 1:fdd22bb7aa52 233 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 234 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 235 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 236
emilmont 1:fdd22bb7aa52 237 while(k > 0u)
emilmont 1:fdd22bb7aa52 238 {
emilmont 1:fdd22bb7aa52 239 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 240 sum = __SMLALD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 241
emilmont 1:fdd22bb7aa52 242 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 243 k--;
emilmont 1:fdd22bb7aa52 244 }
emilmont 1:fdd22bb7aa52 245
emilmont 1:fdd22bb7aa52 246 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 247 *pOut++ = (q15_t) (__SSAT((sum >> 15), 16));
emilmont 1:fdd22bb7aa52 248
emilmont 1:fdd22bb7aa52 249 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 250 py = pIn2 + (count - 1u);
emilmont 1:fdd22bb7aa52 251 px = pIn1;
emilmont 1:fdd22bb7aa52 252
emilmont 1:fdd22bb7aa52 253 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 254 count++;
emilmont 1:fdd22bb7aa52 255
emilmont 1:fdd22bb7aa52 256 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 257 blockSize1--;
emilmont 1:fdd22bb7aa52 258 }
emilmont 1:fdd22bb7aa52 259
emilmont 1:fdd22bb7aa52 260 /* --------------------------
emilmont 1:fdd22bb7aa52 261 * Initializations of stage2
emilmont 1:fdd22bb7aa52 262 * ------------------------*/
emilmont 1:fdd22bb7aa52 263
emilmont 1:fdd22bb7aa52 264 /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
emilmont 1:fdd22bb7aa52 265 * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
emilmont 1:fdd22bb7aa52 266 * ....
emilmont 1:fdd22bb7aa52 267 * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
emilmont 1:fdd22bb7aa52 268 */
emilmont 1:fdd22bb7aa52 269
emilmont 1:fdd22bb7aa52 270 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 271 px = pIn1;
emilmont 1:fdd22bb7aa52 272
emilmont 1:fdd22bb7aa52 273 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 274 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 275 py = pSrc2;
emilmont 1:fdd22bb7aa52 276
emilmont 1:fdd22bb7aa52 277 /* count is the index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 278 count = 0u;
emilmont 1:fdd22bb7aa52 279
emilmont 1:fdd22bb7aa52 280
emilmont 1:fdd22bb7aa52 281 /* --------------------
emilmont 1:fdd22bb7aa52 282 * Stage2 process
emilmont 1:fdd22bb7aa52 283 * -------------------*/
emilmont 1:fdd22bb7aa52 284
emilmont 1:fdd22bb7aa52 285 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 286 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 287 * srcBLen should be greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 288 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 289 {
emilmont 1:fdd22bb7aa52 290 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 291 blkCnt = blockSize2 >> 2u;
emilmont 1:fdd22bb7aa52 292
emilmont 1:fdd22bb7aa52 293 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 294 {
emilmont 1:fdd22bb7aa52 295 py = py - 1u;
emilmont 1:fdd22bb7aa52 296
emilmont 1:fdd22bb7aa52 297 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 298 acc0 = 0;
emilmont 1:fdd22bb7aa52 299 acc1 = 0;
emilmont 1:fdd22bb7aa52 300 acc2 = 0;
emilmont 1:fdd22bb7aa52 301 acc3 = 0;
emilmont 1:fdd22bb7aa52 302
emilmont 1:fdd22bb7aa52 303
emilmont 1:fdd22bb7aa52 304 /* read x[0], x[1] samples */
emilmont 1:fdd22bb7aa52 305 x0 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 306 /* read x[1], x[2] samples */
emilmont 1:fdd22bb7aa52 307 x1 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 308 px+= 2u;
emilmont 1:fdd22bb7aa52 309
emilmont 1:fdd22bb7aa52 310
emilmont 1:fdd22bb7aa52 311 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 312 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 313
emilmont 1:fdd22bb7aa52 314 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 315 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 316 do
emilmont 1:fdd22bb7aa52 317 {
emilmont 1:fdd22bb7aa52 318 /* Read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 319 * y[srcBLen - 1] and y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 320 c0 = *__SIMD32(py)--;
emilmont 1:fdd22bb7aa52 321
emilmont 1:fdd22bb7aa52 322 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 323 acc0 = __SMLALDX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 324
emilmont 1:fdd22bb7aa52 325 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 326 acc1 = __SMLALDX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 327
emilmont 1:fdd22bb7aa52 328 /* Read x[2], x[3] */
emilmont 1:fdd22bb7aa52 329 x2 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 330
emilmont 1:fdd22bb7aa52 331 /* Read x[3], x[4] */
emilmont 1:fdd22bb7aa52 332 x3 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 333
emilmont 1:fdd22bb7aa52 334 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 335 acc2 = __SMLALDX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 336
emilmont 1:fdd22bb7aa52 337 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 338 acc3 = __SMLALDX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 339
emilmont 1:fdd22bb7aa52 340 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 341 c0 = *__SIMD32(py)--;
emilmont 1:fdd22bb7aa52 342
emilmont 1:fdd22bb7aa52 343 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 344 acc0 = __SMLALDX(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 345
emilmont 1:fdd22bb7aa52 346 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 347 acc1 = __SMLALDX(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 348
emilmont 1:fdd22bb7aa52 349 /* Read x[4], x[5] */
emilmont 1:fdd22bb7aa52 350 x0 = _SIMD32_OFFSET(px+2);
emilmont 1:fdd22bb7aa52 351
emilmont 1:fdd22bb7aa52 352 /* Read x[5], x[6] */
emilmont 1:fdd22bb7aa52 353 x1 = _SIMD32_OFFSET(px+3);
emilmont 1:fdd22bb7aa52 354 px += 4u;
emilmont 1:fdd22bb7aa52 355
emilmont 1:fdd22bb7aa52 356 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 357 acc2 = __SMLALDX(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 358
emilmont 1:fdd22bb7aa52 359 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 360 acc3 = __SMLALDX(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 361
emilmont 1:fdd22bb7aa52 362 } while(--k);
emilmont 1:fdd22bb7aa52 363
emilmont 1:fdd22bb7aa52 364 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 365 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 366
emilmont 1:fdd22bb7aa52 367 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 368 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 369 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 370
emilmont 1:fdd22bb7aa52 371 if(k == 1u)
emilmont 1:fdd22bb7aa52 372 {
emilmont 1:fdd22bb7aa52 373 /* Read y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 374 c0 = *(py+1);
emilmont 1:fdd22bb7aa52 375
emilmont 1:fdd22bb7aa52 376 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 377
emilmont 1:fdd22bb7aa52 378 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 379
emilmont 1:fdd22bb7aa52 380 #else
emilmont 1:fdd22bb7aa52 381
emilmont 1:fdd22bb7aa52 382 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 383
emilmont 1:fdd22bb7aa52 384 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 385 /* Read x[7] */
emilmont 1:fdd22bb7aa52 386 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 387 px++;
emilmont 1:fdd22bb7aa52 388
emilmont 1:fdd22bb7aa52 389 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 390 acc0 = __SMLALD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 391 acc1 = __SMLALD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 392 acc2 = __SMLALDX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 393 acc3 = __SMLALDX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 394 }
emilmont 1:fdd22bb7aa52 395
emilmont 1:fdd22bb7aa52 396 if(k == 2u)
emilmont 1:fdd22bb7aa52 397 {
emilmont 1:fdd22bb7aa52 398 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 399 c0 = _SIMD32_OFFSET(py);
emilmont 1:fdd22bb7aa52 400
emilmont 1:fdd22bb7aa52 401 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 402 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 403
emilmont 1:fdd22bb7aa52 404 /* Read x[9] */
emilmont 1:fdd22bb7aa52 405 x2 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 406 px += 2u;
emilmont 1:fdd22bb7aa52 407
emilmont 1:fdd22bb7aa52 408 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 409 acc0 = __SMLALDX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 410 acc1 = __SMLALDX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 411 acc2 = __SMLALDX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 412 acc3 = __SMLALDX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 413 }
emilmont 1:fdd22bb7aa52 414
emilmont 1:fdd22bb7aa52 415 if(k == 3u)
emilmont 1:fdd22bb7aa52 416 {
emilmont 1:fdd22bb7aa52 417 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 418 c0 = _SIMD32_OFFSET(py);
emilmont 1:fdd22bb7aa52 419
emilmont 1:fdd22bb7aa52 420 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 421 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 422
emilmont 1:fdd22bb7aa52 423 /* Read x[9] */
emilmont 1:fdd22bb7aa52 424 x2 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 425
emilmont 1:fdd22bb7aa52 426 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 427 acc0 = __SMLALDX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 428 acc1 = __SMLALDX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 429 acc2 = __SMLALDX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 430 acc3 = __SMLALDX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 431
emilmont 1:fdd22bb7aa52 432 c0 = *(py-1);
emilmont 1:fdd22bb7aa52 433
emilmont 1:fdd22bb7aa52 434 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 435
emilmont 1:fdd22bb7aa52 436 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 437 #else
emilmont 1:fdd22bb7aa52 438
emilmont 1:fdd22bb7aa52 439 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 440 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 441 /* Read x[10] */
emilmont 1:fdd22bb7aa52 442 x3 = _SIMD32_OFFSET(px+2);
emilmont 1:fdd22bb7aa52 443 px += 3u;
emilmont 1:fdd22bb7aa52 444
emilmont 1:fdd22bb7aa52 445 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 446 acc0 = __SMLALDX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 447 acc1 = __SMLALD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 448 acc2 = __SMLALDX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 449 acc3 = __SMLALDX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 450 }
emilmont 1:fdd22bb7aa52 451
emilmont 1:fdd22bb7aa52 452
emilmont 1:fdd22bb7aa52 453 /* Store the results in the accumulators in the destination buffer. */
emilmont 1:fdd22bb7aa52 454
emilmont 1:fdd22bb7aa52 455 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 456
emilmont 1:fdd22bb7aa52 457 *__SIMD32(pOut)++ =
emilmont 1:fdd22bb7aa52 458 __PKHBT(__SSAT((acc0 >> 15), 16), __SSAT((acc1 >> 15), 16), 16);
emilmont 1:fdd22bb7aa52 459 *__SIMD32(pOut)++ =
emilmont 1:fdd22bb7aa52 460 __PKHBT(__SSAT((acc2 >> 15), 16), __SSAT((acc3 >> 15), 16), 16);
emilmont 1:fdd22bb7aa52 461
emilmont 1:fdd22bb7aa52 462 #else
emilmont 1:fdd22bb7aa52 463
emilmont 1:fdd22bb7aa52 464 *__SIMD32(pOut)++ =
emilmont 1:fdd22bb7aa52 465 __PKHBT(__SSAT((acc1 >> 15), 16), __SSAT((acc0 >> 15), 16), 16);
emilmont 1:fdd22bb7aa52 466 *__SIMD32(pOut)++ =
emilmont 1:fdd22bb7aa52 467 __PKHBT(__SSAT((acc3 >> 15), 16), __SSAT((acc2 >> 15), 16), 16);
emilmont 1:fdd22bb7aa52 468
emilmont 1:fdd22bb7aa52 469 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 470
emilmont 1:fdd22bb7aa52 471 /* Increment the pointer pIn1 index, count by 4 */
emilmont 1:fdd22bb7aa52 472 count += 4u;
emilmont 1:fdd22bb7aa52 473
emilmont 1:fdd22bb7aa52 474 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 475 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 476 py = pSrc2;
emilmont 1:fdd22bb7aa52 477
emilmont 1:fdd22bb7aa52 478 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 479 blkCnt--;
emilmont 1:fdd22bb7aa52 480 }
emilmont 1:fdd22bb7aa52 481
emilmont 1:fdd22bb7aa52 482 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 483 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 484 blkCnt = blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 485
emilmont 1:fdd22bb7aa52 486 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 487 {
emilmont 1:fdd22bb7aa52 488 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 489 sum = 0;
emilmont 1:fdd22bb7aa52 490
emilmont 1:fdd22bb7aa52 491 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 492 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 493
emilmont 1:fdd22bb7aa52 494 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 495 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 496 while(k > 0u)
emilmont 1:fdd22bb7aa52 497 {
emilmont 1:fdd22bb7aa52 498 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 499 sum += (q63_t) ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 500 sum += (q63_t) ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 501 sum += (q63_t) ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 502 sum += (q63_t) ((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 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 509 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 510 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 511
emilmont 1:fdd22bb7aa52 512 while(k > 0u)
emilmont 1:fdd22bb7aa52 513 {
emilmont 1:fdd22bb7aa52 514 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 515 sum += (q63_t) ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 516
emilmont 1:fdd22bb7aa52 517 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 518 k--;
emilmont 1:fdd22bb7aa52 519 }
emilmont 1:fdd22bb7aa52 520
emilmont 1:fdd22bb7aa52 521 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 522 *pOut++ = (q15_t) (__SSAT(sum >> 15, 16));
emilmont 1:fdd22bb7aa52 523
emilmont 1:fdd22bb7aa52 524 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 525 count++;
emilmont 1:fdd22bb7aa52 526
emilmont 1:fdd22bb7aa52 527 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 528 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 529 py = pSrc2;
emilmont 1:fdd22bb7aa52 530
emilmont 1:fdd22bb7aa52 531 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 532 blkCnt--;
emilmont 1:fdd22bb7aa52 533 }
emilmont 1:fdd22bb7aa52 534 }
emilmont 1:fdd22bb7aa52 535 else
emilmont 1:fdd22bb7aa52 536 {
emilmont 1:fdd22bb7aa52 537 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 538 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 539 blkCnt = blockSize2;
emilmont 1:fdd22bb7aa52 540
emilmont 1:fdd22bb7aa52 541 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 542 {
emilmont 1:fdd22bb7aa52 543 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 544 sum = 0;
emilmont 1:fdd22bb7aa52 545
emilmont 1:fdd22bb7aa52 546 /* srcBLen number of MACS should be performed */
emilmont 1:fdd22bb7aa52 547 k = srcBLen;
emilmont 1:fdd22bb7aa52 548
emilmont 1:fdd22bb7aa52 549 while(k > 0u)
emilmont 1:fdd22bb7aa52 550 {
emilmont 1:fdd22bb7aa52 551 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 552 sum += (q63_t) ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 553
emilmont 1:fdd22bb7aa52 554 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 555 k--;
emilmont 1:fdd22bb7aa52 556 }
emilmont 1:fdd22bb7aa52 557
emilmont 1:fdd22bb7aa52 558 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 559 *pOut++ = (q15_t) (__SSAT(sum >> 15, 16));
emilmont 1:fdd22bb7aa52 560
emilmont 1:fdd22bb7aa52 561 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 562 count++;
emilmont 1:fdd22bb7aa52 563
emilmont 1:fdd22bb7aa52 564 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 565 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 566 py = pSrc2;
emilmont 1:fdd22bb7aa52 567
emilmont 1:fdd22bb7aa52 568 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 569 blkCnt--;
emilmont 1:fdd22bb7aa52 570 }
emilmont 1:fdd22bb7aa52 571 }
emilmont 1:fdd22bb7aa52 572
emilmont 1:fdd22bb7aa52 573
emilmont 1:fdd22bb7aa52 574 /* --------------------------
emilmont 1:fdd22bb7aa52 575 * Initializations of stage3
emilmont 1:fdd22bb7aa52 576 * -------------------------*/
emilmont 1:fdd22bb7aa52 577
emilmont 1:fdd22bb7aa52 578 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 579 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emilmont 1:fdd22bb7aa52 580 * ....
emilmont 1:fdd22bb7aa52 581 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emilmont 1:fdd22bb7aa52 582 * sum += x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 583 */
emilmont 1:fdd22bb7aa52 584
emilmont 1:fdd22bb7aa52 585 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 586 The blockSize3 variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 587
emilmont 1:fdd22bb7aa52 588 blockSize3 = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 589
emilmont 1:fdd22bb7aa52 590 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 591 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 592 px = pSrc1;
emilmont 1:fdd22bb7aa52 593
emilmont 1:fdd22bb7aa52 594 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 595 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 596 pIn2 = pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 597 py = pIn2;
emilmont 1:fdd22bb7aa52 598
emilmont 1:fdd22bb7aa52 599 /* -------------------
emilmont 1:fdd22bb7aa52 600 * Stage3 process
emilmont 1:fdd22bb7aa52 601 * ------------------*/
emilmont 1:fdd22bb7aa52 602
emilmont 1:fdd22bb7aa52 603 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 604 /* First part of this stage computes the MAC operations greater than 4 */
emilmont 1:fdd22bb7aa52 605 /* Second part of this stage computes the MAC operations less than or equal to 4 */
emilmont 1:fdd22bb7aa52 606
emilmont 1:fdd22bb7aa52 607 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 608 j = blockSize3 >> 2u;
emilmont 1:fdd22bb7aa52 609
emilmont 1:fdd22bb7aa52 610 while((j > 0u) && (blockSize3 > 0u))
emilmont 1:fdd22bb7aa52 611 {
emilmont 1:fdd22bb7aa52 612 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 613 sum = 0;
emilmont 1:fdd22bb7aa52 614
emilmont 1:fdd22bb7aa52 615 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 616 k = blockSize3 >> 2u;
emilmont 1:fdd22bb7aa52 617
emilmont 1:fdd22bb7aa52 618 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 619 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 620 while(k > 0u)
emilmont 1:fdd22bb7aa52 621 {
emilmont 1:fdd22bb7aa52 622 /* x[srcALen - srcBLen + 1], x[srcALen - srcBLen + 2] are multiplied
emilmont 1:fdd22bb7aa52 623 * with y[srcBLen - 1], y[srcBLen - 2] respectively */
emilmont 1:fdd22bb7aa52 624 sum = __SMLALDX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 625 /* x[srcALen - srcBLen + 3], x[srcALen - srcBLen + 4] are multiplied
emilmont 1:fdd22bb7aa52 626 * with y[srcBLen - 3], y[srcBLen - 4] respectively */
emilmont 1:fdd22bb7aa52 627 sum = __SMLALDX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 628
emilmont 1:fdd22bb7aa52 629 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 630 k--;
emilmont 1:fdd22bb7aa52 631 }
emilmont 1:fdd22bb7aa52 632
emilmont 1:fdd22bb7aa52 633 /* For the next MAC operations, the pointer py is used without SIMD
emilmont 1:fdd22bb7aa52 634 * So, py is incremented by 1 */
emilmont 1:fdd22bb7aa52 635 py = py + 1u;
emilmont 1:fdd22bb7aa52 636
emilmont 1:fdd22bb7aa52 637 /* If the blockSize3 is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 638 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 639 k = blockSize3 % 0x4u;
emilmont 1:fdd22bb7aa52 640
emilmont 1:fdd22bb7aa52 641 while(k > 0u)
emilmont 1:fdd22bb7aa52 642 {
emilmont 1:fdd22bb7aa52 643 /* sum += x[srcALen - srcBLen + 5] * y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 644 sum = __SMLALD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 645
emilmont 1:fdd22bb7aa52 646 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 647 k--;
emilmont 1:fdd22bb7aa52 648 }
emilmont 1:fdd22bb7aa52 649
emilmont 1:fdd22bb7aa52 650 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 651 *pOut++ = (q15_t) (__SSAT((sum >> 15), 16));
emilmont 1:fdd22bb7aa52 652
emilmont 1:fdd22bb7aa52 653 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 654 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 655 py = pIn2;
emilmont 1:fdd22bb7aa52 656
emilmont 1:fdd22bb7aa52 657 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 658 blockSize3--;
emilmont 1:fdd22bb7aa52 659
emilmont 1:fdd22bb7aa52 660 j--;
emilmont 1:fdd22bb7aa52 661 }
emilmont 1:fdd22bb7aa52 662
emilmont 1:fdd22bb7aa52 663 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 664 /* SIMD is not used for the next MAC operations,
emilmont 1:fdd22bb7aa52 665 * so pointer py is updated to read only one sample at a time */
emilmont 1:fdd22bb7aa52 666 py = py + 1u;
emilmont 1:fdd22bb7aa52 667
emilmont 1:fdd22bb7aa52 668 while(blockSize3 > 0u)
emilmont 1:fdd22bb7aa52 669 {
emilmont 1:fdd22bb7aa52 670 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 671 sum = 0;
emilmont 1:fdd22bb7aa52 672
emilmont 1:fdd22bb7aa52 673 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 674 k = blockSize3;
emilmont 1:fdd22bb7aa52 675
emilmont 1:fdd22bb7aa52 676 while(k > 0u)
emilmont 1:fdd22bb7aa52 677 {
emilmont 1:fdd22bb7aa52 678 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 679 /* sum += x[srcALen-1] * y[srcBLen-1] */
emilmont 1:fdd22bb7aa52 680 sum = __SMLALD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 681
emilmont 1:fdd22bb7aa52 682 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 683 k--;
emilmont 1:fdd22bb7aa52 684 }
emilmont 1:fdd22bb7aa52 685
emilmont 1:fdd22bb7aa52 686 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 687 *pOut++ = (q15_t) (__SSAT((sum >> 15), 16));
emilmont 1:fdd22bb7aa52 688
emilmont 1:fdd22bb7aa52 689 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 690 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 691 py = pSrc2;
emilmont 1:fdd22bb7aa52 692
emilmont 1:fdd22bb7aa52 693 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 694 blockSize3--;
emilmont 1:fdd22bb7aa52 695 }
emilmont 1:fdd22bb7aa52 696
emilmont 1:fdd22bb7aa52 697 #else
emilmont 1:fdd22bb7aa52 698
emilmont 1:fdd22bb7aa52 699 /* Run the below code for Cortex-M0 */
emilmont 1:fdd22bb7aa52 700
emilmont 1:fdd22bb7aa52 701 q15_t *pIn1 = pSrcA; /* input pointer */
emilmont 1:fdd22bb7aa52 702 q15_t *pIn2 = pSrcB; /* coefficient pointer */
emilmont 1:fdd22bb7aa52 703 q63_t sum; /* Accumulator */
emilmont 1:fdd22bb7aa52 704 uint32_t i, j; /* loop counter */
emilmont 1:fdd22bb7aa52 705
emilmont 1:fdd22bb7aa52 706 /* Loop to calculate output of convolution for output length number of times */
emilmont 1:fdd22bb7aa52 707 for (i = 0; i < (srcALen + srcBLen - 1); i++)
emilmont 1:fdd22bb7aa52 708 {
emilmont 1:fdd22bb7aa52 709 /* Initialize sum with zero to carry on MAC operations */
emilmont 1:fdd22bb7aa52 710 sum = 0;
emilmont 1:fdd22bb7aa52 711
emilmont 1:fdd22bb7aa52 712 /* Loop to perform MAC operations according to convolution equation */
emilmont 1:fdd22bb7aa52 713 for (j = 0; j <= i; j++)
emilmont 1:fdd22bb7aa52 714 {
emilmont 1:fdd22bb7aa52 715 /* Check the array limitations */
emilmont 1:fdd22bb7aa52 716 if(((i - j) < srcBLen) && (j < srcALen))
emilmont 1:fdd22bb7aa52 717 {
emilmont 1:fdd22bb7aa52 718 /* z[i] += x[i-j] * y[j] */
emilmont 1:fdd22bb7aa52 719 sum += (q31_t) pIn1[j] * (pIn2[i - j]);
emilmont 1:fdd22bb7aa52 720 }
emilmont 1:fdd22bb7aa52 721 }
emilmont 1:fdd22bb7aa52 722
emilmont 1:fdd22bb7aa52 723 /* Store the output in the destination buffer */
emilmont 1:fdd22bb7aa52 724 pDst[i] = (q15_t) __SSAT((sum >> 15u), 16u);
emilmont 1:fdd22bb7aa52 725 }
emilmont 1:fdd22bb7aa52 726
emilmont 1:fdd22bb7aa52 727 #endif /* #if (defined(ARM_MATH_CM4) || defined(ARM_MATH_CM3)) && !defined(UNALIGNED_SUPPORT_DISABLE)*/
emilmont 1:fdd22bb7aa52 728
emilmont 1:fdd22bb7aa52 729 }
emilmont 1:fdd22bb7aa52 730
emilmont 1:fdd22bb7aa52 731 /**
emilmont 1:fdd22bb7aa52 732 * @} end of Conv group
emilmont 1:fdd22bb7aa52 733 */