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_partial_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_partial_fast_q15.c
emilmont 1:fdd22bb7aa52 9 *
emilmont 1:fdd22bb7aa52 10 * Description: Fast Q15 Partial 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 PartialConv
emilmont 1:fdd22bb7aa52 44 * @{
emilmont 1:fdd22bb7aa52 45 */
emilmont 1:fdd22bb7aa52 46
emilmont 1:fdd22bb7aa52 47 /**
emilmont 1:fdd22bb7aa52 48 * @brief Partial 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.
emilmont 1:fdd22bb7aa52 54 * @param[in] firstIndex is the first output sample to start with.
emilmont 1:fdd22bb7aa52 55 * @param[in] numPoints is the number of output points to be computed.
emilmont 1:fdd22bb7aa52 56 * @return Returns either ARM_MATH_SUCCESS if the function completed correctly or ARM_MATH_ARGUMENT_ERROR if the requested subset is not in the range [0 srcALen+srcBLen-2].
emilmont 1:fdd22bb7aa52 57 *
emilmont 1:fdd22bb7aa52 58 * See <code>arm_conv_partial_q15()</code> for a slower implementation of this function which uses a 64-bit accumulator to avoid wrap around distortion.
emilmont 1:fdd22bb7aa52 59 */
emilmont 1:fdd22bb7aa52 60
emilmont 1:fdd22bb7aa52 61
emilmont 1:fdd22bb7aa52 62 arm_status arm_conv_partial_fast_q15(
emilmont 1:fdd22bb7aa52 63 q15_t * pSrcA,
emilmont 1:fdd22bb7aa52 64 uint32_t srcALen,
emilmont 1:fdd22bb7aa52 65 q15_t * pSrcB,
emilmont 1:fdd22bb7aa52 66 uint32_t srcBLen,
emilmont 1:fdd22bb7aa52 67 q15_t * pDst,
emilmont 1:fdd22bb7aa52 68 uint32_t firstIndex,
emilmont 1:fdd22bb7aa52 69 uint32_t numPoints)
emilmont 1:fdd22bb7aa52 70 {
emilmont 1:fdd22bb7aa52 71 #ifndef UNALIGNED_SUPPORT_DISABLE
emilmont 1:fdd22bb7aa52 72
emilmont 1:fdd22bb7aa52 73 q15_t *pIn1; /* inputA pointer */
emilmont 1:fdd22bb7aa52 74 q15_t *pIn2; /* inputB pointer */
emilmont 1:fdd22bb7aa52 75 q15_t *pOut = pDst; /* output pointer */
emilmont 1:fdd22bb7aa52 76 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
emilmont 1:fdd22bb7aa52 77 q15_t *px; /* Intermediate inputA pointer */
emilmont 1:fdd22bb7aa52 78 q15_t *py; /* Intermediate inputB pointer */
emilmont 1:fdd22bb7aa52 79 q15_t *pSrc1, *pSrc2; /* Intermediate pointers */
emilmont 1:fdd22bb7aa52 80 q31_t x0, x1, x2, x3, c0;
emilmont 1:fdd22bb7aa52 81 uint32_t j, k, count, check, blkCnt;
emilmont 1:fdd22bb7aa52 82 int32_t blockSize1, blockSize2, blockSize3; /* loop counters */
emilmont 1:fdd22bb7aa52 83 arm_status status; /* status of Partial convolution */
emilmont 1:fdd22bb7aa52 84
emilmont 1:fdd22bb7aa52 85 /* Check for range of output samples to be calculated */
emilmont 1:fdd22bb7aa52 86 if((firstIndex + numPoints) > ((srcALen + (srcBLen - 1u))))
emilmont 1:fdd22bb7aa52 87 {
emilmont 1:fdd22bb7aa52 88 /* Set status as ARM_MATH_ARGUMENT_ERROR */
emilmont 1:fdd22bb7aa52 89 status = ARM_MATH_ARGUMENT_ERROR;
emilmont 1:fdd22bb7aa52 90 }
emilmont 1:fdd22bb7aa52 91 else
emilmont 1:fdd22bb7aa52 92 {
emilmont 1:fdd22bb7aa52 93
emilmont 1:fdd22bb7aa52 94 /* The algorithm implementation is based on the lengths of the inputs. */
emilmont 1:fdd22bb7aa52 95 /* srcB is always made to slide across srcA. */
emilmont 1:fdd22bb7aa52 96 /* So srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 97 if(srcALen >=srcBLen)
emilmont 1:fdd22bb7aa52 98 {
emilmont 1:fdd22bb7aa52 99 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 100 pIn1 = pSrcA;
emilmont 1:fdd22bb7aa52 101
emilmont 1:fdd22bb7aa52 102 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 103 pIn2 = pSrcB;
emilmont 1:fdd22bb7aa52 104 }
emilmont 1:fdd22bb7aa52 105 else
emilmont 1:fdd22bb7aa52 106 {
emilmont 1:fdd22bb7aa52 107 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 108 pIn1 = pSrcB;
emilmont 1:fdd22bb7aa52 109
emilmont 1:fdd22bb7aa52 110 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 111 pIn2 = pSrcA;
emilmont 1:fdd22bb7aa52 112
emilmont 1:fdd22bb7aa52 113 /* srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 114 j = srcBLen;
emilmont 1:fdd22bb7aa52 115 srcBLen = srcALen;
emilmont 1:fdd22bb7aa52 116 srcALen = j;
emilmont 1:fdd22bb7aa52 117 }
emilmont 1:fdd22bb7aa52 118
emilmont 1:fdd22bb7aa52 119 /* Conditions to check which loopCounter holds
emilmont 1:fdd22bb7aa52 120 * the first and last indices of the output samples to be calculated. */
emilmont 1:fdd22bb7aa52 121 check = firstIndex + numPoints;
emilmont 1:fdd22bb7aa52 122 blockSize3 = ((int32_t) check - (int32_t) srcALen);
emilmont 1:fdd22bb7aa52 123 blockSize3 = (blockSize3 > 0) ? blockSize3 : 0;
emilmont 1:fdd22bb7aa52 124 blockSize1 = (((int32_t) srcBLen - 1) - (int32_t) firstIndex);
emilmont 1:fdd22bb7aa52 125 blockSize1 = (blockSize1 > 0) ? ((check > (srcBLen - 1u)) ? blockSize1 :
emilmont 1:fdd22bb7aa52 126 (int32_t) numPoints) : 0;
emilmont 1:fdd22bb7aa52 127 blockSize2 = (int32_t) check - ((blockSize3 + blockSize1) +
emilmont 1:fdd22bb7aa52 128 (int32_t) firstIndex);
emilmont 1:fdd22bb7aa52 129 blockSize2 = (blockSize2 > 0) ? blockSize2 : 0;
emilmont 1:fdd22bb7aa52 130
emilmont 1:fdd22bb7aa52 131 /* 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 132 /* The function is internally
emilmont 1:fdd22bb7aa52 133 * divided into three stages according to the number of multiplications that has to be
emilmont 1:fdd22bb7aa52 134 * taken place between inputA samples and inputB samples. In the first stage of the
emilmont 1:fdd22bb7aa52 135 * algorithm, the multiplications increase by one for every iteration.
emilmont 1:fdd22bb7aa52 136 * In the second stage of the algorithm, srcBLen number of multiplications are done.
emilmont 1:fdd22bb7aa52 137 * In the third stage of the algorithm, the multiplications decrease by one
emilmont 1:fdd22bb7aa52 138 * for every iteration. */
emilmont 1:fdd22bb7aa52 139
emilmont 1:fdd22bb7aa52 140 /* Set the output pointer to point to the firstIndex
emilmont 1:fdd22bb7aa52 141 * of the output sample to be calculated. */
emilmont 1:fdd22bb7aa52 142 pOut = pDst + firstIndex;
emilmont 1:fdd22bb7aa52 143
emilmont 1:fdd22bb7aa52 144 /* --------------------------
emilmont 1:fdd22bb7aa52 145 * Initializations of stage1
emilmont 1:fdd22bb7aa52 146 * -------------------------*/
emilmont 1:fdd22bb7aa52 147
emilmont 1:fdd22bb7aa52 148 /* sum = x[0] * y[0]
emilmont 1:fdd22bb7aa52 149 * sum = x[0] * y[1] + x[1] * y[0]
emilmont 1:fdd22bb7aa52 150 * ....
emilmont 1:fdd22bb7aa52 151 * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
emilmont 1:fdd22bb7aa52 152 */
emilmont 1:fdd22bb7aa52 153
emilmont 1:fdd22bb7aa52 154 /* In this stage the MAC operations are increased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 155 The count variable holds the number of MAC operations performed.
emilmont 1:fdd22bb7aa52 156 Since the partial convolution starts from firstIndex
emilmont 1:fdd22bb7aa52 157 Number of Macs to be performed is firstIndex + 1 */
emilmont 1:fdd22bb7aa52 158 count = 1u + firstIndex;
emilmont 1:fdd22bb7aa52 159
emilmont 1:fdd22bb7aa52 160 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 161 px = pIn1;
emilmont 1:fdd22bb7aa52 162
emilmont 1:fdd22bb7aa52 163 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 164 pSrc2 = pIn2 + firstIndex;
emilmont 1:fdd22bb7aa52 165 py = pSrc2;
emilmont 1:fdd22bb7aa52 166
emilmont 1:fdd22bb7aa52 167 /* ------------------------
emilmont 1:fdd22bb7aa52 168 * Stage1 process
emilmont 1:fdd22bb7aa52 169 * ----------------------*/
emilmont 1:fdd22bb7aa52 170
emilmont 1:fdd22bb7aa52 171 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 172 /* First part of this stage computes the MAC operations less than 4 */
emilmont 1:fdd22bb7aa52 173 /* Second part of this stage computes the MAC operations greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 174
emilmont 1:fdd22bb7aa52 175 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 176 while((count < 4u) && (blockSize1 > 0))
emilmont 1:fdd22bb7aa52 177 {
emilmont 1:fdd22bb7aa52 178 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 179 sum = 0;
emilmont 1:fdd22bb7aa52 180
emilmont 1:fdd22bb7aa52 181 /* Loop over number of MAC operations between
emilmont 1:fdd22bb7aa52 182 * inputA samples and inputB samples */
emilmont 1:fdd22bb7aa52 183 k = count;
emilmont 1:fdd22bb7aa52 184
emilmont 1:fdd22bb7aa52 185 while(k > 0u)
emilmont 1:fdd22bb7aa52 186 {
emilmont 1:fdd22bb7aa52 187 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 188 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 189
emilmont 1:fdd22bb7aa52 190 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 191 k--;
emilmont 1:fdd22bb7aa52 192 }
emilmont 1:fdd22bb7aa52 193
emilmont 1:fdd22bb7aa52 194 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 195 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 196
emilmont 1:fdd22bb7aa52 197 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 198 py = ++pSrc2;
emilmont 1:fdd22bb7aa52 199 px = pIn1;
emilmont 1:fdd22bb7aa52 200
emilmont 1:fdd22bb7aa52 201 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 202 count++;
emilmont 1:fdd22bb7aa52 203
emilmont 1:fdd22bb7aa52 204 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 205 blockSize1--;
emilmont 1:fdd22bb7aa52 206 }
emilmont 1:fdd22bb7aa52 207
emilmont 1:fdd22bb7aa52 208 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 209 /* The internal loop, over count, is unrolled by 4 */
emilmont 1:fdd22bb7aa52 210 /* To, read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 211 * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */
emilmont 1:fdd22bb7aa52 212 py = py - 1;
emilmont 1:fdd22bb7aa52 213
emilmont 1:fdd22bb7aa52 214 while(blockSize1 > 0)
emilmont 1:fdd22bb7aa52 215 {
emilmont 1:fdd22bb7aa52 216 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 217 sum = 0;
emilmont 1:fdd22bb7aa52 218
emilmont 1:fdd22bb7aa52 219 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 220 k = count >> 2u;
emilmont 1:fdd22bb7aa52 221
emilmont 1:fdd22bb7aa52 222 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 223 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 224 while(k > 0u)
emilmont 1:fdd22bb7aa52 225 {
emilmont 1:fdd22bb7aa52 226 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 227 /* x[0], x[1] are multiplied with y[srcBLen - 1], y[srcBLen - 2] respectively */
emilmont 1:fdd22bb7aa52 228 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 229 /* x[2], x[3] are multiplied with y[srcBLen - 3], y[srcBLen - 4] respectively */
emilmont 1:fdd22bb7aa52 230 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 231
emilmont 1:fdd22bb7aa52 232 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 233 k--;
emilmont 1:fdd22bb7aa52 234 }
emilmont 1:fdd22bb7aa52 235
emilmont 1:fdd22bb7aa52 236 /* For the next MAC operations, the pointer py is used without SIMD
emilmont 1:fdd22bb7aa52 237 * So, py is incremented by 1 */
emilmont 1:fdd22bb7aa52 238 py = py + 1u;
emilmont 1:fdd22bb7aa52 239
emilmont 1:fdd22bb7aa52 240 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 241 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 242 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 243
emilmont 1:fdd22bb7aa52 244 while(k > 0u)
emilmont 1:fdd22bb7aa52 245 {
emilmont 1:fdd22bb7aa52 246 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 247 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 248
emilmont 1:fdd22bb7aa52 249 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 250 k--;
emilmont 1:fdd22bb7aa52 251 }
emilmont 1:fdd22bb7aa52 252
emilmont 1:fdd22bb7aa52 253 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 254 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 255
emilmont 1:fdd22bb7aa52 256 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 257 py = ++pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 258 px = pIn1;
emilmont 1:fdd22bb7aa52 259
emilmont 1:fdd22bb7aa52 260 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 261 count++;
emilmont 1:fdd22bb7aa52 262
emilmont 1:fdd22bb7aa52 263 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 264 blockSize1--;
emilmont 1:fdd22bb7aa52 265 }
emilmont 1:fdd22bb7aa52 266
emilmont 1:fdd22bb7aa52 267 /* --------------------------
emilmont 1:fdd22bb7aa52 268 * Initializations of stage2
emilmont 1:fdd22bb7aa52 269 * ------------------------*/
emilmont 1:fdd22bb7aa52 270
emilmont 1:fdd22bb7aa52 271 /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
emilmont 1:fdd22bb7aa52 272 * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
emilmont 1:fdd22bb7aa52 273 * ....
emilmont 1:fdd22bb7aa52 274 * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
emilmont 1:fdd22bb7aa52 275 */
emilmont 1:fdd22bb7aa52 276
emilmont 1:fdd22bb7aa52 277 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 278 px = pIn1;
emilmont 1:fdd22bb7aa52 279
emilmont 1:fdd22bb7aa52 280 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 281 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 282 py = pSrc2;
emilmont 1:fdd22bb7aa52 283
emilmont 1:fdd22bb7aa52 284 /* count is the index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 285 count = 0u;
emilmont 1:fdd22bb7aa52 286
emilmont 1:fdd22bb7aa52 287
emilmont 1:fdd22bb7aa52 288 /* --------------------
emilmont 1:fdd22bb7aa52 289 * Stage2 process
emilmont 1:fdd22bb7aa52 290 * -------------------*/
emilmont 1:fdd22bb7aa52 291
emilmont 1:fdd22bb7aa52 292 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 293 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 294 * srcBLen should be greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 295 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 296 {
emilmont 1:fdd22bb7aa52 297 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 298 blkCnt = ((uint32_t) blockSize2 >> 2u);
emilmont 1:fdd22bb7aa52 299
emilmont 1:fdd22bb7aa52 300 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 301 {
emilmont 1:fdd22bb7aa52 302 py = py - 1u;
emilmont 1:fdd22bb7aa52 303
emilmont 1:fdd22bb7aa52 304 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 305 acc0 = 0;
emilmont 1:fdd22bb7aa52 306 acc1 = 0;
emilmont 1:fdd22bb7aa52 307 acc2 = 0;
emilmont 1:fdd22bb7aa52 308 acc3 = 0;
emilmont 1:fdd22bb7aa52 309
emilmont 1:fdd22bb7aa52 310
emilmont 1:fdd22bb7aa52 311 /* read x[0], x[1] samples */
emilmont 1:fdd22bb7aa52 312 x0 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 313 /* read x[1], x[2] samples */
emilmont 1:fdd22bb7aa52 314 x1 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 315 px+= 2u;
emilmont 1:fdd22bb7aa52 316
emilmont 1:fdd22bb7aa52 317
emilmont 1:fdd22bb7aa52 318 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 319 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 320
emilmont 1:fdd22bb7aa52 321 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 322 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 323 do
emilmont 1:fdd22bb7aa52 324 {
emilmont 1:fdd22bb7aa52 325 /* Read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 326 * y[srcBLen - 1] and y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 327 c0 = *__SIMD32(py)--;
emilmont 1:fdd22bb7aa52 328
emilmont 1:fdd22bb7aa52 329 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 330 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 331
emilmont 1:fdd22bb7aa52 332 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 333 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 334
emilmont 1:fdd22bb7aa52 335 /* Read x[2], x[3] */
emilmont 1:fdd22bb7aa52 336 x2 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 337
emilmont 1:fdd22bb7aa52 338 /* Read x[3], x[4] */
emilmont 1:fdd22bb7aa52 339 x3 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 340
emilmont 1:fdd22bb7aa52 341 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 342 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 343
emilmont 1:fdd22bb7aa52 344 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 345 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 346
emilmont 1:fdd22bb7aa52 347 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 348 c0 = *__SIMD32(py)--;
emilmont 1:fdd22bb7aa52 349
emilmont 1:fdd22bb7aa52 350 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 351 acc0 = __SMLADX(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 352
emilmont 1:fdd22bb7aa52 353 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 354 acc1 = __SMLADX(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 355
emilmont 1:fdd22bb7aa52 356 /* Read x[4], x[5] */
emilmont 1:fdd22bb7aa52 357 x0 = _SIMD32_OFFSET(px+2);
emilmont 1:fdd22bb7aa52 358
emilmont 1:fdd22bb7aa52 359 /* Read x[5], x[6] */
emilmont 1:fdd22bb7aa52 360 x1 = _SIMD32_OFFSET(px+3);
emilmont 1:fdd22bb7aa52 361 px += 4u;
emilmont 1:fdd22bb7aa52 362
emilmont 1:fdd22bb7aa52 363 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 364 acc2 = __SMLADX(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 365
emilmont 1:fdd22bb7aa52 366 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 367 acc3 = __SMLADX(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 368
emilmont 1:fdd22bb7aa52 369 } while(--k);
emilmont 1:fdd22bb7aa52 370
emilmont 1:fdd22bb7aa52 371 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 372 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 373
emilmont 1:fdd22bb7aa52 374 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 375 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 376 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 377
emilmont 1:fdd22bb7aa52 378 if(k == 1u)
emilmont 1:fdd22bb7aa52 379 {
emilmont 1:fdd22bb7aa52 380 /* Read y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 381 c0 = *(py+1);
emilmont 1:fdd22bb7aa52 382 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 383
emilmont 1:fdd22bb7aa52 384 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 385
emilmont 1:fdd22bb7aa52 386 #else
emilmont 1:fdd22bb7aa52 387
emilmont 1:fdd22bb7aa52 388 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 389
emilmont 1:fdd22bb7aa52 390 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 391
emilmont 1:fdd22bb7aa52 392 /* Read x[7] */
emilmont 1:fdd22bb7aa52 393 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 394 px++;
emilmont 1:fdd22bb7aa52 395
emilmont 1:fdd22bb7aa52 396 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 397 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 398 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 399 acc2 = __SMLADX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 400 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 401 }
emilmont 1:fdd22bb7aa52 402
emilmont 1:fdd22bb7aa52 403 if(k == 2u)
emilmont 1:fdd22bb7aa52 404 {
emilmont 1:fdd22bb7aa52 405 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 406 c0 = _SIMD32_OFFSET(py);
emilmont 1:fdd22bb7aa52 407
emilmont 1:fdd22bb7aa52 408 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 409 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 410
emilmont 1:fdd22bb7aa52 411 /* Read x[9] */
emilmont 1:fdd22bb7aa52 412 x2 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 413 px += 2u;
emilmont 1:fdd22bb7aa52 414
emilmont 1:fdd22bb7aa52 415 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 416 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 417 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 418 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 419 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 420 }
emilmont 1:fdd22bb7aa52 421
emilmont 1:fdd22bb7aa52 422 if(k == 3u)
emilmont 1:fdd22bb7aa52 423 {
emilmont 1:fdd22bb7aa52 424 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 425 c0 = _SIMD32_OFFSET(py);
emilmont 1:fdd22bb7aa52 426
emilmont 1:fdd22bb7aa52 427 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 428 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 429
emilmont 1:fdd22bb7aa52 430 /* Read x[9] */
emilmont 1:fdd22bb7aa52 431 x2 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 432
emilmont 1:fdd22bb7aa52 433 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 434 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 435 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 436 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 437 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 438
emilmont 1:fdd22bb7aa52 439 c0 = *(py-1);
emilmont 1:fdd22bb7aa52 440 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 441
emilmont 1:fdd22bb7aa52 442 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 443 #else
emilmont 1:fdd22bb7aa52 444
emilmont 1:fdd22bb7aa52 445 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 446 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 447
emilmont 1:fdd22bb7aa52 448 /* Read x[10] */
emilmont 1:fdd22bb7aa52 449 x3 = _SIMD32_OFFSET(px+2);
emilmont 1:fdd22bb7aa52 450 px += 3u;
emilmont 1:fdd22bb7aa52 451
emilmont 1:fdd22bb7aa52 452 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 453 acc0 = __SMLADX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 454 acc1 = __SMLAD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 455 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 456 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 457 }
emilmont 1:fdd22bb7aa52 458
emilmont 1:fdd22bb7aa52 459 /* Store the results in the accumulators in the destination buffer. */
emilmont 1:fdd22bb7aa52 460 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 461
emilmont 1:fdd22bb7aa52 462 *__SIMD32(pOut)++ = __PKHBT(acc0 >> 15, acc1 >> 15, 16);
emilmont 1:fdd22bb7aa52 463 *__SIMD32(pOut)++ = __PKHBT(acc2 >> 15, acc3 >> 15, 16);
emilmont 1:fdd22bb7aa52 464
emilmont 1:fdd22bb7aa52 465 #else
emilmont 1:fdd22bb7aa52 466
emilmont 1:fdd22bb7aa52 467 *__SIMD32(pOut)++ = __PKHBT(acc1 >> 15, acc0 >> 15, 16);
emilmont 1:fdd22bb7aa52 468 *__SIMD32(pOut)++ = __PKHBT(acc3 >> 15, acc2 >> 15, 16);
emilmont 1:fdd22bb7aa52 469
emilmont 1:fdd22bb7aa52 470 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 471
emilmont 1:fdd22bb7aa52 472 /* Increment the pointer pIn1 index, count by 4 */
emilmont 1:fdd22bb7aa52 473 count += 4u;
emilmont 1:fdd22bb7aa52 474
emilmont 1:fdd22bb7aa52 475 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 476 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 477 py = pSrc2;
emilmont 1:fdd22bb7aa52 478
emilmont 1:fdd22bb7aa52 479 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 480 blkCnt--;
emilmont 1:fdd22bb7aa52 481 }
emilmont 1:fdd22bb7aa52 482
emilmont 1:fdd22bb7aa52 483 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 484 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 485 blkCnt = (uint32_t) blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 486
emilmont 1:fdd22bb7aa52 487 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 488 {
emilmont 1:fdd22bb7aa52 489 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 490 sum = 0;
emilmont 1:fdd22bb7aa52 491
emilmont 1:fdd22bb7aa52 492 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 493 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 494
emilmont 1:fdd22bb7aa52 495 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 496 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 497 while(k > 0u)
emilmont 1:fdd22bb7aa52 498 {
emilmont 1:fdd22bb7aa52 499 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 500 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 501 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 502 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 503 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 504
emilmont 1:fdd22bb7aa52 505 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 506 k--;
emilmont 1:fdd22bb7aa52 507 }
emilmont 1:fdd22bb7aa52 508
emilmont 1:fdd22bb7aa52 509 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 510 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 511 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 512
emilmont 1:fdd22bb7aa52 513 while(k > 0u)
emilmont 1:fdd22bb7aa52 514 {
emilmont 1:fdd22bb7aa52 515 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 516 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 517
emilmont 1:fdd22bb7aa52 518 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 519 k--;
emilmont 1:fdd22bb7aa52 520 }
emilmont 1:fdd22bb7aa52 521
emilmont 1:fdd22bb7aa52 522 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 523 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 524
emilmont 1:fdd22bb7aa52 525 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 526 count++;
emilmont 1:fdd22bb7aa52 527
emilmont 1:fdd22bb7aa52 528 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 529 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 530 py = pSrc2;
emilmont 1:fdd22bb7aa52 531
emilmont 1:fdd22bb7aa52 532 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 533 blkCnt--;
emilmont 1:fdd22bb7aa52 534 }
emilmont 1:fdd22bb7aa52 535 }
emilmont 1:fdd22bb7aa52 536 else
emilmont 1:fdd22bb7aa52 537 {
emilmont 1:fdd22bb7aa52 538 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 539 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 540 blkCnt = (uint32_t) blockSize2;
emilmont 1:fdd22bb7aa52 541
emilmont 1:fdd22bb7aa52 542 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 543 {
emilmont 1:fdd22bb7aa52 544 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 545 sum = 0;
emilmont 1:fdd22bb7aa52 546
emilmont 1:fdd22bb7aa52 547 /* srcBLen number of MACS should be performed */
emilmont 1:fdd22bb7aa52 548 k = srcBLen;
emilmont 1:fdd22bb7aa52 549
emilmont 1:fdd22bb7aa52 550 while(k > 0u)
emilmont 1:fdd22bb7aa52 551 {
emilmont 1:fdd22bb7aa52 552 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 553 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 554
emilmont 1:fdd22bb7aa52 555 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 556 k--;
emilmont 1:fdd22bb7aa52 557 }
emilmont 1:fdd22bb7aa52 558
emilmont 1:fdd22bb7aa52 559 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 560 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 561
emilmont 1:fdd22bb7aa52 562 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 563 count++;
emilmont 1:fdd22bb7aa52 564
emilmont 1:fdd22bb7aa52 565 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 566 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 567 py = pSrc2;
emilmont 1:fdd22bb7aa52 568
emilmont 1:fdd22bb7aa52 569 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 570 blkCnt--;
emilmont 1:fdd22bb7aa52 571 }
emilmont 1:fdd22bb7aa52 572 }
emilmont 1:fdd22bb7aa52 573
emilmont 1:fdd22bb7aa52 574
emilmont 1:fdd22bb7aa52 575 /* --------------------------
emilmont 1:fdd22bb7aa52 576 * Initializations of stage3
emilmont 1:fdd22bb7aa52 577 * -------------------------*/
emilmont 1:fdd22bb7aa52 578
emilmont 1:fdd22bb7aa52 579 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 580 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emilmont 1:fdd22bb7aa52 581 * ....
emilmont 1:fdd22bb7aa52 582 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emilmont 1:fdd22bb7aa52 583 * sum += x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 584 */
emilmont 1:fdd22bb7aa52 585
emilmont 1:fdd22bb7aa52 586 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 587 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 588 count = 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 = count >> 2u;
emilmont 1:fdd22bb7aa52 609
emilmont 1:fdd22bb7aa52 610 while((j > 0u) && (blockSize3 > 0))
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 = count >> 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 = __SMLADX(*__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 = __SMLADX(*__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 count 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 = count % 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 = __SMLAD(*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) (sum >> 15);
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 MAC count */
emilmont 1:fdd22bb7aa52 658 count--;
emilmont 1:fdd22bb7aa52 659
emilmont 1:fdd22bb7aa52 660 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 661 blockSize3--;
emilmont 1:fdd22bb7aa52 662
emilmont 1:fdd22bb7aa52 663 j--;
emilmont 1:fdd22bb7aa52 664 }
emilmont 1:fdd22bb7aa52 665
emilmont 1:fdd22bb7aa52 666 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 667 /* SIMD is not used for the next MAC operations,
emilmont 1:fdd22bb7aa52 668 * so pointer py is updated to read only one sample at a time */
emilmont 1:fdd22bb7aa52 669 py = py + 1u;
emilmont 1:fdd22bb7aa52 670
emilmont 1:fdd22bb7aa52 671 while(blockSize3 > 0)
emilmont 1:fdd22bb7aa52 672 {
emilmont 1:fdd22bb7aa52 673 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 674 sum = 0;
emilmont 1:fdd22bb7aa52 675
emilmont 1:fdd22bb7aa52 676 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 677 k = count;
emilmont 1:fdd22bb7aa52 678
emilmont 1:fdd22bb7aa52 679 while(k > 0u)
emilmont 1:fdd22bb7aa52 680 {
emilmont 1:fdd22bb7aa52 681 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 682 /* sum += x[srcALen-1] * y[srcBLen-1] */
emilmont 1:fdd22bb7aa52 683 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 684
emilmont 1:fdd22bb7aa52 685 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 686 k--;
emilmont 1:fdd22bb7aa52 687 }
emilmont 1:fdd22bb7aa52 688
emilmont 1:fdd22bb7aa52 689 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 690 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 691
emilmont 1:fdd22bb7aa52 692 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 693 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 694 py = pSrc2;
emilmont 1:fdd22bb7aa52 695
emilmont 1:fdd22bb7aa52 696 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 697 count--;
emilmont 1:fdd22bb7aa52 698
emilmont 1:fdd22bb7aa52 699 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 700 blockSize3--;
emilmont 1:fdd22bb7aa52 701 }
emilmont 1:fdd22bb7aa52 702
emilmont 1:fdd22bb7aa52 703 /* set status as ARM_MATH_SUCCESS */
emilmont 1:fdd22bb7aa52 704 status = ARM_MATH_SUCCESS;
emilmont 1:fdd22bb7aa52 705 }
emilmont 1:fdd22bb7aa52 706
emilmont 1:fdd22bb7aa52 707 /* Return to application */
emilmont 1:fdd22bb7aa52 708 return (status);
emilmont 1:fdd22bb7aa52 709
emilmont 1:fdd22bb7aa52 710 #else
emilmont 1:fdd22bb7aa52 711
emilmont 1:fdd22bb7aa52 712 q15_t *pIn1; /* inputA pointer */
emilmont 1:fdd22bb7aa52 713 q15_t *pIn2; /* inputB pointer */
emilmont 1:fdd22bb7aa52 714 q15_t *pOut = pDst; /* output pointer */
emilmont 1:fdd22bb7aa52 715 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
emilmont 1:fdd22bb7aa52 716 q15_t *px; /* Intermediate inputA pointer */
emilmont 1:fdd22bb7aa52 717 q15_t *py; /* Intermediate inputB pointer */
emilmont 1:fdd22bb7aa52 718 q15_t *pSrc1, *pSrc2; /* Intermediate pointers */
emilmont 1:fdd22bb7aa52 719 q31_t x0, x1, x2, x3, c0;
emilmont 1:fdd22bb7aa52 720 uint32_t j, k, count, check, blkCnt;
emilmont 1:fdd22bb7aa52 721 int32_t blockSize1, blockSize2, blockSize3; /* loop counters */
emilmont 1:fdd22bb7aa52 722 arm_status status; /* status of Partial convolution */
emilmont 1:fdd22bb7aa52 723 q15_t a, b;
emilmont 1:fdd22bb7aa52 724
emilmont 1:fdd22bb7aa52 725 /* Check for range of output samples to be calculated */
emilmont 1:fdd22bb7aa52 726 if((firstIndex + numPoints) > ((srcALen + (srcBLen - 1u))))
emilmont 1:fdd22bb7aa52 727 {
emilmont 1:fdd22bb7aa52 728 /* Set status as ARM_MATH_ARGUMENT_ERROR */
emilmont 1:fdd22bb7aa52 729 status = ARM_MATH_ARGUMENT_ERROR;
emilmont 1:fdd22bb7aa52 730 }
emilmont 1:fdd22bb7aa52 731 else
emilmont 1:fdd22bb7aa52 732 {
emilmont 1:fdd22bb7aa52 733
emilmont 1:fdd22bb7aa52 734 /* The algorithm implementation is based on the lengths of the inputs. */
emilmont 1:fdd22bb7aa52 735 /* srcB is always made to slide across srcA. */
emilmont 1:fdd22bb7aa52 736 /* So srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 737 if(srcALen >=srcBLen)
emilmont 1:fdd22bb7aa52 738 {
emilmont 1:fdd22bb7aa52 739 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 740 pIn1 = pSrcA;
emilmont 1:fdd22bb7aa52 741
emilmont 1:fdd22bb7aa52 742 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 743 pIn2 = pSrcB;
emilmont 1:fdd22bb7aa52 744 }
emilmont 1:fdd22bb7aa52 745 else
emilmont 1:fdd22bb7aa52 746 {
emilmont 1:fdd22bb7aa52 747 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 748 pIn1 = pSrcB;
emilmont 1:fdd22bb7aa52 749
emilmont 1:fdd22bb7aa52 750 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 751 pIn2 = pSrcA;
emilmont 1:fdd22bb7aa52 752
emilmont 1:fdd22bb7aa52 753 /* srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 754 j = srcBLen;
emilmont 1:fdd22bb7aa52 755 srcBLen = srcALen;
emilmont 1:fdd22bb7aa52 756 srcALen = j;
emilmont 1:fdd22bb7aa52 757 }
emilmont 1:fdd22bb7aa52 758
emilmont 1:fdd22bb7aa52 759 /* Conditions to check which loopCounter holds
emilmont 1:fdd22bb7aa52 760 * the first and last indices of the output samples to be calculated. */
emilmont 1:fdd22bb7aa52 761 check = firstIndex + numPoints;
emilmont 1:fdd22bb7aa52 762 blockSize3 = ((int32_t) check - (int32_t) srcALen);
emilmont 1:fdd22bb7aa52 763 blockSize3 = (blockSize3 > 0) ? blockSize3 : 0;
emilmont 1:fdd22bb7aa52 764 blockSize1 = (((int32_t) srcBLen - 1) - (int32_t) firstIndex);
emilmont 1:fdd22bb7aa52 765 blockSize1 = (blockSize1 > 0) ? ((check > (srcBLen - 1u)) ? blockSize1 :
emilmont 1:fdd22bb7aa52 766 (int32_t) numPoints) : 0;
emilmont 1:fdd22bb7aa52 767 blockSize2 = (int32_t) check - ((blockSize3 + blockSize1) +
emilmont 1:fdd22bb7aa52 768 (int32_t) firstIndex);
emilmont 1:fdd22bb7aa52 769 blockSize2 = (blockSize2 > 0) ? blockSize2 : 0;
emilmont 1:fdd22bb7aa52 770
emilmont 1:fdd22bb7aa52 771 /* 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 772 /* The function is internally
emilmont 1:fdd22bb7aa52 773 * divided into three stages according to the number of multiplications that has to be
emilmont 1:fdd22bb7aa52 774 * taken place between inputA samples and inputB samples. In the first stage of the
emilmont 1:fdd22bb7aa52 775 * algorithm, the multiplications increase by one for every iteration.
emilmont 1:fdd22bb7aa52 776 * In the second stage of the algorithm, srcBLen number of multiplications are done.
emilmont 1:fdd22bb7aa52 777 * In the third stage of the algorithm, the multiplications decrease by one
emilmont 1:fdd22bb7aa52 778 * for every iteration. */
emilmont 1:fdd22bb7aa52 779
emilmont 1:fdd22bb7aa52 780 /* Set the output pointer to point to the firstIndex
emilmont 1:fdd22bb7aa52 781 * of the output sample to be calculated. */
emilmont 1:fdd22bb7aa52 782 pOut = pDst + firstIndex;
emilmont 1:fdd22bb7aa52 783
emilmont 1:fdd22bb7aa52 784 /* --------------------------
emilmont 1:fdd22bb7aa52 785 * Initializations of stage1
emilmont 1:fdd22bb7aa52 786 * -------------------------*/
emilmont 1:fdd22bb7aa52 787
emilmont 1:fdd22bb7aa52 788 /* sum = x[0] * y[0]
emilmont 1:fdd22bb7aa52 789 * sum = x[0] * y[1] + x[1] * y[0]
emilmont 1:fdd22bb7aa52 790 * ....
emilmont 1:fdd22bb7aa52 791 * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
emilmont 1:fdd22bb7aa52 792 */
emilmont 1:fdd22bb7aa52 793
emilmont 1:fdd22bb7aa52 794 /* In this stage the MAC operations are increased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 795 The count variable holds the number of MAC operations performed.
emilmont 1:fdd22bb7aa52 796 Since the partial convolution starts from firstIndex
emilmont 1:fdd22bb7aa52 797 Number of Macs to be performed is firstIndex + 1 */
emilmont 1:fdd22bb7aa52 798 count = 1u + firstIndex;
emilmont 1:fdd22bb7aa52 799
emilmont 1:fdd22bb7aa52 800 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 801 px = pIn1;
emilmont 1:fdd22bb7aa52 802
emilmont 1:fdd22bb7aa52 803 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 804 pSrc2 = pIn2 + firstIndex;
emilmont 1:fdd22bb7aa52 805 py = pSrc2;
emilmont 1:fdd22bb7aa52 806
emilmont 1:fdd22bb7aa52 807 /* ------------------------
emilmont 1:fdd22bb7aa52 808 * Stage1 process
emilmont 1:fdd22bb7aa52 809 * ----------------------*/
emilmont 1:fdd22bb7aa52 810
emilmont 1:fdd22bb7aa52 811 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 812 /* First part of this stage computes the MAC operations less than 4 */
emilmont 1:fdd22bb7aa52 813 /* Second part of this stage computes the MAC operations greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 814
emilmont 1:fdd22bb7aa52 815 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 816 while((count < 4u) && (blockSize1 > 0u))
emilmont 1:fdd22bb7aa52 817 {
emilmont 1:fdd22bb7aa52 818 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 819 sum = 0;
emilmont 1:fdd22bb7aa52 820
emilmont 1:fdd22bb7aa52 821 /* Loop over number of MAC operations between
emilmont 1:fdd22bb7aa52 822 * inputA samples and inputB samples */
emilmont 1:fdd22bb7aa52 823 k = count;
emilmont 1:fdd22bb7aa52 824
emilmont 1:fdd22bb7aa52 825 while(k > 0u)
emilmont 1:fdd22bb7aa52 826 {
emilmont 1:fdd22bb7aa52 827 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 828 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 829
emilmont 1:fdd22bb7aa52 830 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 831 k--;
emilmont 1:fdd22bb7aa52 832 }
emilmont 1:fdd22bb7aa52 833
emilmont 1:fdd22bb7aa52 834 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 835 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 836
emilmont 1:fdd22bb7aa52 837 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 838 py = ++pSrc2;
emilmont 1:fdd22bb7aa52 839 px = pIn1;
emilmont 1:fdd22bb7aa52 840
emilmont 1:fdd22bb7aa52 841 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 842 count++;
emilmont 1:fdd22bb7aa52 843
emilmont 1:fdd22bb7aa52 844 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 845 blockSize1--;
emilmont 1:fdd22bb7aa52 846 }
emilmont 1:fdd22bb7aa52 847
emilmont 1:fdd22bb7aa52 848 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 849 /* The internal loop, over count, is unrolled by 4 */
emilmont 1:fdd22bb7aa52 850 /* To, read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 851 * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */
emilmont 1:fdd22bb7aa52 852 py = py - 1;
emilmont 1:fdd22bb7aa52 853
emilmont 1:fdd22bb7aa52 854 while(blockSize1 > 0u)
emilmont 1:fdd22bb7aa52 855 {
emilmont 1:fdd22bb7aa52 856 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 857 sum = 0;
emilmont 1:fdd22bb7aa52 858
emilmont 1:fdd22bb7aa52 859 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 860 k = count >> 2u;
emilmont 1:fdd22bb7aa52 861
emilmont 1:fdd22bb7aa52 862 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 863 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 864 py++;
emilmont 1:fdd22bb7aa52 865
emilmont 1:fdd22bb7aa52 866 while(k > 0u)
emilmont 1:fdd22bb7aa52 867 {
emilmont 1:fdd22bb7aa52 868 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 869 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 870 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 871 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 872 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 873
emilmont 1:fdd22bb7aa52 874 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 875 k--;
emilmont 1:fdd22bb7aa52 876 }
emilmont 1:fdd22bb7aa52 877
emilmont 1:fdd22bb7aa52 878 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 879 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 880 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 881
emilmont 1:fdd22bb7aa52 882 while(k > 0u)
emilmont 1:fdd22bb7aa52 883 {
emilmont 1:fdd22bb7aa52 884 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 885 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 886
emilmont 1:fdd22bb7aa52 887 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 888 k--;
emilmont 1:fdd22bb7aa52 889 }
emilmont 1:fdd22bb7aa52 890
emilmont 1:fdd22bb7aa52 891 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 892 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 893
emilmont 1:fdd22bb7aa52 894 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 895 py = ++pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 896 px = pIn1;
emilmont 1:fdd22bb7aa52 897
emilmont 1:fdd22bb7aa52 898 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 899 count++;
emilmont 1:fdd22bb7aa52 900
emilmont 1:fdd22bb7aa52 901 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 902 blockSize1--;
emilmont 1:fdd22bb7aa52 903 }
emilmont 1:fdd22bb7aa52 904
emilmont 1:fdd22bb7aa52 905 /* --------------------------
emilmont 1:fdd22bb7aa52 906 * Initializations of stage2
emilmont 1:fdd22bb7aa52 907 * ------------------------*/
emilmont 1:fdd22bb7aa52 908
emilmont 1:fdd22bb7aa52 909 /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
emilmont 1:fdd22bb7aa52 910 * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
emilmont 1:fdd22bb7aa52 911 * ....
emilmont 1:fdd22bb7aa52 912 * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
emilmont 1:fdd22bb7aa52 913 */
emilmont 1:fdd22bb7aa52 914
emilmont 1:fdd22bb7aa52 915 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 916 px = pIn1;
emilmont 1:fdd22bb7aa52 917
emilmont 1:fdd22bb7aa52 918 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 919 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 920 py = pSrc2;
emilmont 1:fdd22bb7aa52 921
emilmont 1:fdd22bb7aa52 922 /* count is the index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 923 count = 0u;
emilmont 1:fdd22bb7aa52 924
emilmont 1:fdd22bb7aa52 925
emilmont 1:fdd22bb7aa52 926 /* --------------------
emilmont 1:fdd22bb7aa52 927 * Stage2 process
emilmont 1:fdd22bb7aa52 928 * -------------------*/
emilmont 1:fdd22bb7aa52 929
emilmont 1:fdd22bb7aa52 930 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 931 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 932 * srcBLen should be greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 933 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 934 {
emilmont 1:fdd22bb7aa52 935 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 936 blkCnt = ((uint32_t) blockSize2 >> 2u);
emilmont 1:fdd22bb7aa52 937
emilmont 1:fdd22bb7aa52 938 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 939 {
emilmont 1:fdd22bb7aa52 940 py = py - 1u;
emilmont 1:fdd22bb7aa52 941
emilmont 1:fdd22bb7aa52 942 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 943 acc0 = 0;
emilmont 1:fdd22bb7aa52 944 acc1 = 0;
emilmont 1:fdd22bb7aa52 945 acc2 = 0;
emilmont 1:fdd22bb7aa52 946 acc3 = 0;
emilmont 1:fdd22bb7aa52 947
emilmont 1:fdd22bb7aa52 948 /* read x[0], x[1] samples */
emilmont 1:fdd22bb7aa52 949 a = *px++;
emilmont 1:fdd22bb7aa52 950 b = *px++;
emilmont 1:fdd22bb7aa52 951
emilmont 1:fdd22bb7aa52 952 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 953
emilmont 1:fdd22bb7aa52 954 x0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 955 a = *px;
emilmont 1:fdd22bb7aa52 956 x1 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 957
emilmont 1:fdd22bb7aa52 958 #else
emilmont 1:fdd22bb7aa52 959
emilmont 1:fdd22bb7aa52 960 x0 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 961 a = *px;
emilmont 1:fdd22bb7aa52 962 x1 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 963
emilmont 1:fdd22bb7aa52 964 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 965
emilmont 1:fdd22bb7aa52 966 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 967 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 968
emilmont 1:fdd22bb7aa52 969 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 970 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 971 do
emilmont 1:fdd22bb7aa52 972 {
emilmont 1:fdd22bb7aa52 973 /* Read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 974 * y[srcBLen - 1] and y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 975 a = *py;
emilmont 1:fdd22bb7aa52 976 b = *(py+1);
emilmont 1:fdd22bb7aa52 977 py -= 2;
emilmont 1:fdd22bb7aa52 978
emilmont 1:fdd22bb7aa52 979 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 980
emilmont 1:fdd22bb7aa52 981 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 982
emilmont 1:fdd22bb7aa52 983 #else
emilmont 1:fdd22bb7aa52 984
emilmont 1:fdd22bb7aa52 985 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 986
emilmont 1:fdd22bb7aa52 987 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 988
emilmont 1:fdd22bb7aa52 989 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 990 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 991
emilmont 1:fdd22bb7aa52 992 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 993 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 994
emilmont 1:fdd22bb7aa52 995 a = *px;
emilmont 1:fdd22bb7aa52 996 b = *(px + 1);
emilmont 1:fdd22bb7aa52 997
emilmont 1:fdd22bb7aa52 998 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 999
emilmont 1:fdd22bb7aa52 1000 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1001 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1002 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1003
emilmont 1:fdd22bb7aa52 1004 #else
emilmont 1:fdd22bb7aa52 1005
emilmont 1:fdd22bb7aa52 1006 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1007 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1008 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1009
emilmont 1:fdd22bb7aa52 1010 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1011
emilmont 1:fdd22bb7aa52 1012 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 1013 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 1014
emilmont 1:fdd22bb7aa52 1015 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 1016 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1017
emilmont 1:fdd22bb7aa52 1018 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1019 a = *py;
emilmont 1:fdd22bb7aa52 1020 b = *(py+1);
emilmont 1:fdd22bb7aa52 1021 py -= 2;
emilmont 1:fdd22bb7aa52 1022
emilmont 1:fdd22bb7aa52 1023 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1024
emilmont 1:fdd22bb7aa52 1025 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1026
emilmont 1:fdd22bb7aa52 1027 #else
emilmont 1:fdd22bb7aa52 1028
emilmont 1:fdd22bb7aa52 1029 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1030
emilmont 1:fdd22bb7aa52 1031 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1032
emilmont 1:fdd22bb7aa52 1033 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1034 acc0 = __SMLADX(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 1035
emilmont 1:fdd22bb7aa52 1036 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1037 acc1 = __SMLADX(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 1038
emilmont 1:fdd22bb7aa52 1039 /* Read x[4], x[5], x[6] */
emilmont 1:fdd22bb7aa52 1040 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1041 b = *(px + 3);
emilmont 1:fdd22bb7aa52 1042
emilmont 1:fdd22bb7aa52 1043 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1044
emilmont 1:fdd22bb7aa52 1045 x0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1046 a = *(px + 4);
emilmont 1:fdd22bb7aa52 1047 x1 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1048
emilmont 1:fdd22bb7aa52 1049 #else
emilmont 1:fdd22bb7aa52 1050
emilmont 1:fdd22bb7aa52 1051 x0 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1052 a = *(px + 4);
emilmont 1:fdd22bb7aa52 1053 x1 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1054
emilmont 1:fdd22bb7aa52 1055 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1056
emilmont 1:fdd22bb7aa52 1057 px += 4u;
emilmont 1:fdd22bb7aa52 1058
emilmont 1:fdd22bb7aa52 1059 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1060 acc2 = __SMLADX(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 1061
emilmont 1:fdd22bb7aa52 1062 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1063 acc3 = __SMLADX(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 1064
emilmont 1:fdd22bb7aa52 1065 } while(--k);
emilmont 1:fdd22bb7aa52 1066
emilmont 1:fdd22bb7aa52 1067 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 1068 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 1069
emilmont 1:fdd22bb7aa52 1070 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1071 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1072 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 1073
emilmont 1:fdd22bb7aa52 1074 if(k == 1u)
emilmont 1:fdd22bb7aa52 1075 {
emilmont 1:fdd22bb7aa52 1076 /* Read y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 1077 c0 = *(py+1);
emilmont 1:fdd22bb7aa52 1078
emilmont 1:fdd22bb7aa52 1079 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1080
emilmont 1:fdd22bb7aa52 1081 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 1082
emilmont 1:fdd22bb7aa52 1083 #else
emilmont 1:fdd22bb7aa52 1084
emilmont 1:fdd22bb7aa52 1085 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 1086
emilmont 1:fdd22bb7aa52 1087 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1088
emilmont 1:fdd22bb7aa52 1089 /* Read x[7] */
emilmont 1:fdd22bb7aa52 1090 a = *px;
emilmont 1:fdd22bb7aa52 1091 b = *(px+1);
emilmont 1:fdd22bb7aa52 1092 px++;
emilmont 1:fdd22bb7aa52 1093
emilmont 1:fdd22bb7aa52 1094 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1095
emilmont 1:fdd22bb7aa52 1096 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1097
emilmont 1:fdd22bb7aa52 1098 #else
emilmont 1:fdd22bb7aa52 1099
emilmont 1:fdd22bb7aa52 1100 x3 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1101
emilmont 1:fdd22bb7aa52 1102 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1103
emilmont 1:fdd22bb7aa52 1104
emilmont 1:fdd22bb7aa52 1105 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1106 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1107 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1108 acc2 = __SMLADX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 1109 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1110 }
emilmont 1:fdd22bb7aa52 1111
emilmont 1:fdd22bb7aa52 1112 if(k == 2u)
emilmont 1:fdd22bb7aa52 1113 {
emilmont 1:fdd22bb7aa52 1114 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 1115 a = *py;
emilmont 1:fdd22bb7aa52 1116 b = *(py+1);
emilmont 1:fdd22bb7aa52 1117
emilmont 1:fdd22bb7aa52 1118 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1119
emilmont 1:fdd22bb7aa52 1120 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1121
emilmont 1:fdd22bb7aa52 1122 #else
emilmont 1:fdd22bb7aa52 1123
emilmont 1:fdd22bb7aa52 1124 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1125
emilmont 1:fdd22bb7aa52 1126 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1127
emilmont 1:fdd22bb7aa52 1128 /* Read x[7], x[8], x[9] */
emilmont 1:fdd22bb7aa52 1129 a = *px;
emilmont 1:fdd22bb7aa52 1130 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1131
emilmont 1:fdd22bb7aa52 1132 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1133
emilmont 1:fdd22bb7aa52 1134 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1135 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1136 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1137
emilmont 1:fdd22bb7aa52 1138 #else
emilmont 1:fdd22bb7aa52 1139
emilmont 1:fdd22bb7aa52 1140 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1141 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1142 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1143
emilmont 1:fdd22bb7aa52 1144 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1145 px += 2u;
emilmont 1:fdd22bb7aa52 1146
emilmont 1:fdd22bb7aa52 1147 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1148 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1149 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1150 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 1151 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 1152 }
emilmont 1:fdd22bb7aa52 1153
emilmont 1:fdd22bb7aa52 1154 if(k == 3u)
emilmont 1:fdd22bb7aa52 1155 {
emilmont 1:fdd22bb7aa52 1156 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 1157 a = *py;
emilmont 1:fdd22bb7aa52 1158 b = *(py+1);
emilmont 1:fdd22bb7aa52 1159
emilmont 1:fdd22bb7aa52 1160 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1161
emilmont 1:fdd22bb7aa52 1162 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1163
emilmont 1:fdd22bb7aa52 1164 #else
emilmont 1:fdd22bb7aa52 1165
emilmont 1:fdd22bb7aa52 1166 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1167
emilmont 1:fdd22bb7aa52 1168 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1169
emilmont 1:fdd22bb7aa52 1170 /* Read x[7], x[8], x[9] */
emilmont 1:fdd22bb7aa52 1171 a = *px;
emilmont 1:fdd22bb7aa52 1172 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1173
emilmont 1:fdd22bb7aa52 1174 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1175
emilmont 1:fdd22bb7aa52 1176 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1177 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1178 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1179
emilmont 1:fdd22bb7aa52 1180 #else
emilmont 1:fdd22bb7aa52 1181
emilmont 1:fdd22bb7aa52 1182 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1183 a = *(px + 2);
emilmont 1:fdd22bb7aa52 1184 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1185
emilmont 1:fdd22bb7aa52 1186 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1187
emilmont 1:fdd22bb7aa52 1188 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1189 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1190 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1191 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 1192 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 1193
emilmont 1:fdd22bb7aa52 1194 /* Read y[srcBLen - 7] */
emilmont 1:fdd22bb7aa52 1195 c0 = *(py-1);
emilmont 1:fdd22bb7aa52 1196 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1197
emilmont 1:fdd22bb7aa52 1198 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 1199 #else
emilmont 1:fdd22bb7aa52 1200
emilmont 1:fdd22bb7aa52 1201 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 1202 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1203
emilmont 1:fdd22bb7aa52 1204 /* Read x[10] */
emilmont 1:fdd22bb7aa52 1205 a = *(px+2);
emilmont 1:fdd22bb7aa52 1206 b = *(px+3);
emilmont 1:fdd22bb7aa52 1207
emilmont 1:fdd22bb7aa52 1208 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1209
emilmont 1:fdd22bb7aa52 1210 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1211
emilmont 1:fdd22bb7aa52 1212 #else
emilmont 1:fdd22bb7aa52 1213
emilmont 1:fdd22bb7aa52 1214 x3 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1215
emilmont 1:fdd22bb7aa52 1216 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1217
emilmont 1:fdd22bb7aa52 1218 px += 3u;
emilmont 1:fdd22bb7aa52 1219
emilmont 1:fdd22bb7aa52 1220 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1221 acc0 = __SMLADX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 1222 acc1 = __SMLAD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 1223 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 1224 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1225 }
emilmont 1:fdd22bb7aa52 1226
emilmont 1:fdd22bb7aa52 1227 /* Store the results in the accumulators in the destination buffer. */
emilmont 1:fdd22bb7aa52 1228 *pOut++ = (q15_t)(acc0 >> 15);
emilmont 1:fdd22bb7aa52 1229 *pOut++ = (q15_t)(acc1 >> 15);
emilmont 1:fdd22bb7aa52 1230 *pOut++ = (q15_t)(acc2 >> 15);
emilmont 1:fdd22bb7aa52 1231 *pOut++ = (q15_t)(acc3 >> 15);
emilmont 1:fdd22bb7aa52 1232
emilmont 1:fdd22bb7aa52 1233 /* Increment the pointer pIn1 index, count by 4 */
emilmont 1:fdd22bb7aa52 1234 count += 4u;
emilmont 1:fdd22bb7aa52 1235
emilmont 1:fdd22bb7aa52 1236 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1237 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1238 py = pSrc2;
emilmont 1:fdd22bb7aa52 1239
emilmont 1:fdd22bb7aa52 1240 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1241 blkCnt--;
emilmont 1:fdd22bb7aa52 1242 }
emilmont 1:fdd22bb7aa52 1243
emilmont 1:fdd22bb7aa52 1244 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 1245 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1246 blkCnt = (uint32_t) blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 1247
emilmont 1:fdd22bb7aa52 1248 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 1249 {
emilmont 1:fdd22bb7aa52 1250 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1251 sum = 0;
emilmont 1:fdd22bb7aa52 1252
emilmont 1:fdd22bb7aa52 1253 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1254 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 1255
emilmont 1:fdd22bb7aa52 1256 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 1257 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 1258 while(k > 0u)
emilmont 1:fdd22bb7aa52 1259 {
emilmont 1:fdd22bb7aa52 1260 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1261 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1262 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1263 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1264 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1265
emilmont 1:fdd22bb7aa52 1266 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1267 k--;
emilmont 1:fdd22bb7aa52 1268 }
emilmont 1:fdd22bb7aa52 1269
emilmont 1:fdd22bb7aa52 1270 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1271 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1272 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 1273
emilmont 1:fdd22bb7aa52 1274 while(k > 0u)
emilmont 1:fdd22bb7aa52 1275 {
emilmont 1:fdd22bb7aa52 1276 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1277 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1278
emilmont 1:fdd22bb7aa52 1279 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1280 k--;
emilmont 1:fdd22bb7aa52 1281 }
emilmont 1:fdd22bb7aa52 1282
emilmont 1:fdd22bb7aa52 1283 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1284 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1285
emilmont 1:fdd22bb7aa52 1286 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 1287 count++;
emilmont 1:fdd22bb7aa52 1288
emilmont 1:fdd22bb7aa52 1289 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1290 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1291 py = pSrc2;
emilmont 1:fdd22bb7aa52 1292
emilmont 1:fdd22bb7aa52 1293 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1294 blkCnt--;
emilmont 1:fdd22bb7aa52 1295 }
emilmont 1:fdd22bb7aa52 1296 }
emilmont 1:fdd22bb7aa52 1297 else
emilmont 1:fdd22bb7aa52 1298 {
emilmont 1:fdd22bb7aa52 1299 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 1300 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 1301 blkCnt = (uint32_t) blockSize2;
emilmont 1:fdd22bb7aa52 1302
emilmont 1:fdd22bb7aa52 1303 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 1304 {
emilmont 1:fdd22bb7aa52 1305 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1306 sum = 0;
emilmont 1:fdd22bb7aa52 1307
emilmont 1:fdd22bb7aa52 1308 /* srcBLen number of MACS should be performed */
emilmont 1:fdd22bb7aa52 1309 k = srcBLen;
emilmont 1:fdd22bb7aa52 1310
emilmont 1:fdd22bb7aa52 1311 while(k > 0u)
emilmont 1:fdd22bb7aa52 1312 {
emilmont 1:fdd22bb7aa52 1313 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 1314 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1315
emilmont 1:fdd22bb7aa52 1316 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1317 k--;
emilmont 1:fdd22bb7aa52 1318 }
emilmont 1:fdd22bb7aa52 1319
emilmont 1:fdd22bb7aa52 1320 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1321 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1322
emilmont 1:fdd22bb7aa52 1323 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 1324 count++;
emilmont 1:fdd22bb7aa52 1325
emilmont 1:fdd22bb7aa52 1326 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1327 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1328 py = pSrc2;
emilmont 1:fdd22bb7aa52 1329
emilmont 1:fdd22bb7aa52 1330 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1331 blkCnt--;
emilmont 1:fdd22bb7aa52 1332 }
emilmont 1:fdd22bb7aa52 1333 }
emilmont 1:fdd22bb7aa52 1334
emilmont 1:fdd22bb7aa52 1335
emilmont 1:fdd22bb7aa52 1336 /* --------------------------
emilmont 1:fdd22bb7aa52 1337 * Initializations of stage3
emilmont 1:fdd22bb7aa52 1338 * -------------------------*/
emilmont 1:fdd22bb7aa52 1339
emilmont 1:fdd22bb7aa52 1340 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 1341 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emilmont 1:fdd22bb7aa52 1342 * ....
emilmont 1:fdd22bb7aa52 1343 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emilmont 1:fdd22bb7aa52 1344 * sum += x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 1345 */
emilmont 1:fdd22bb7aa52 1346
emilmont 1:fdd22bb7aa52 1347 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 1348 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 1349 count = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 1350
emilmont 1:fdd22bb7aa52 1351 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 1352 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 1353 px = pSrc1;
emilmont 1:fdd22bb7aa52 1354
emilmont 1:fdd22bb7aa52 1355 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 1356 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 1357 pIn2 = pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 1358 py = pIn2;
emilmont 1:fdd22bb7aa52 1359
emilmont 1:fdd22bb7aa52 1360 /* -------------------
emilmont 1:fdd22bb7aa52 1361 * Stage3 process
emilmont 1:fdd22bb7aa52 1362 * ------------------*/
emilmont 1:fdd22bb7aa52 1363
emilmont 1:fdd22bb7aa52 1364 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 1365 /* First part of this stage computes the MAC operations greater than 4 */
emilmont 1:fdd22bb7aa52 1366 /* Second part of this stage computes the MAC operations less than or equal to 4 */
emilmont 1:fdd22bb7aa52 1367
emilmont 1:fdd22bb7aa52 1368 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 1369 j = count >> 2u;
emilmont 1:fdd22bb7aa52 1370
emilmont 1:fdd22bb7aa52 1371 while((j > 0u) && (blockSize3 > 0))
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 = count >> 2u;
emilmont 1:fdd22bb7aa52 1378
emilmont 1:fdd22bb7aa52 1379 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 1380 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 1381 py++;
emilmont 1:fdd22bb7aa52 1382
emilmont 1:fdd22bb7aa52 1383 while(k > 0u)
emilmont 1:fdd22bb7aa52 1384 {
emilmont 1:fdd22bb7aa52 1385 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1386 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1387 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1388 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1389 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1390 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1391 k--;
emilmont 1:fdd22bb7aa52 1392 }
emilmont 1:fdd22bb7aa52 1393
emilmont 1:fdd22bb7aa52 1394
emilmont 1:fdd22bb7aa52 1395 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1396 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1397 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 1398
emilmont 1:fdd22bb7aa52 1399 while(k > 0u)
emilmont 1:fdd22bb7aa52 1400 {
emilmont 1:fdd22bb7aa52 1401 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1402 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1403
emilmont 1:fdd22bb7aa52 1404 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1405 k--;
emilmont 1:fdd22bb7aa52 1406 }
emilmont 1:fdd22bb7aa52 1407
emilmont 1:fdd22bb7aa52 1408 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1409 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1410
emilmont 1:fdd22bb7aa52 1411 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1412 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 1413 py = pIn2;
emilmont 1:fdd22bb7aa52 1414
emilmont 1:fdd22bb7aa52 1415 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 1416 count--;
emilmont 1:fdd22bb7aa52 1417
emilmont 1:fdd22bb7aa52 1418 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1419 blockSize3--;
emilmont 1:fdd22bb7aa52 1420
emilmont 1:fdd22bb7aa52 1421 j--;
emilmont 1:fdd22bb7aa52 1422 }
emilmont 1:fdd22bb7aa52 1423
emilmont 1:fdd22bb7aa52 1424 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 1425 /* SIMD is not used for the next MAC operations,
emilmont 1:fdd22bb7aa52 1426 * so pointer py is updated to read only one sample at a time */
emilmont 1:fdd22bb7aa52 1427 py = py + 1u;
emilmont 1:fdd22bb7aa52 1428
emilmont 1:fdd22bb7aa52 1429 while(blockSize3 > 0u)
emilmont 1:fdd22bb7aa52 1430 {
emilmont 1:fdd22bb7aa52 1431 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1432 sum = 0;
emilmont 1:fdd22bb7aa52 1433
emilmont 1:fdd22bb7aa52 1434 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1435 k = count;
emilmont 1:fdd22bb7aa52 1436
emilmont 1:fdd22bb7aa52 1437 while(k > 0u)
emilmont 1:fdd22bb7aa52 1438 {
emilmont 1:fdd22bb7aa52 1439 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1440 /* sum += x[srcALen-1] * y[srcBLen-1] */
emilmont 1:fdd22bb7aa52 1441 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1442
emilmont 1:fdd22bb7aa52 1443 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1444 k--;
emilmont 1:fdd22bb7aa52 1445 }
emilmont 1:fdd22bb7aa52 1446
emilmont 1:fdd22bb7aa52 1447 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1448 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1449
emilmont 1:fdd22bb7aa52 1450 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1451 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 1452 py = pSrc2;
emilmont 1:fdd22bb7aa52 1453
emilmont 1:fdd22bb7aa52 1454 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 1455 count--;
emilmont 1:fdd22bb7aa52 1456
emilmont 1:fdd22bb7aa52 1457 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1458 blockSize3--;
emilmont 1:fdd22bb7aa52 1459 }
emilmont 1:fdd22bb7aa52 1460
emilmont 1:fdd22bb7aa52 1461 /* set status as ARM_MATH_SUCCESS */
emilmont 1:fdd22bb7aa52 1462 status = ARM_MATH_SUCCESS;
emilmont 1:fdd22bb7aa52 1463 }
emilmont 1:fdd22bb7aa52 1464
emilmont 1:fdd22bb7aa52 1465 /* Return to application */
emilmont 1:fdd22bb7aa52 1466 return (status);
emilmont 1:fdd22bb7aa52 1467
emilmont 1:fdd22bb7aa52 1468 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
emilmont 1:fdd22bb7aa52 1469 }
emilmont 1:fdd22bb7aa52 1470
emilmont 1:fdd22bb7aa52 1471 /**
emilmont 1:fdd22bb7aa52 1472 * @} end of PartialConv group
emilmont 1:fdd22bb7aa52 1473 */