mbed official / mbed-dsp

CMSIS DSP library

Dependents:   KL25Z_FFT_Demo Hat_Board_v5_1 KL25Z_FFT_Demo_tony KL25Z_FFT_Demo_tony ... more

Fork of mbed-dsp by mbed official

cmsis_dsp/FilteringFunctions/arm_conv_fast_q15.c@3:7a284390b0ce, 2013-11-08 (annotated)

Committer:
mbed_official
Date:
Fri Nov 08 13:45:10 2013 +0000
Revision:
3:7a284390b0ce
Parent:
2:da51fb522205 
Synchronized with git revision e69956aba2f68a2a26ac26b051f8d349deaa1ce8

Who changed what in which revision?

UserRevisionLine numberNew contents of line
emilmont 1:fdd22bb7aa52 1 /* ----------------------------------------------------------------------
mbed_official 3:7a284390b0ce 2 * Copyright (C) 2010-2013 ARM Limited. All rights reserved.
emilmont 1:fdd22bb7aa52 3 *
mbed_official 3:7a284390b0ce 4 * $Date: 17. January 2013
mbed_official 3:7a284390b0ce 5 * $Revision: V1.4.1
emilmont 1:fdd22bb7aa52 6 *
emilmont 2:da51fb522205 7 * Project: CMSIS DSP Library
emilmont 2:da51fb522205 8 * Title: arm_conv_fast_q15.c
emilmont 1:fdd22bb7aa52 9 *
emilmont 2:da51fb522205 10 * Description: Fast Q15 Convolution.
emilmont 1:fdd22bb7aa52 11 *
emilmont 1:fdd22bb7aa52 12 * Target Processor: Cortex-M4/Cortex-M3
emilmont 1:fdd22bb7aa52 13 *
mbed_official 3:7a284390b0ce 14 * Redistribution and use in source and binary forms, with or without
mbed_official 3:7a284390b0ce 15 * modification, are permitted provided that the following conditions
mbed_official 3:7a284390b0ce 16 * are met:
mbed_official 3:7a284390b0ce 17 * - Redistributions of source code must retain the above copyright
mbed_official 3:7a284390b0ce 18 * notice, this list of conditions and the following disclaimer.
mbed_official 3:7a284390b0ce 19 * - Redistributions in binary form must reproduce the above copyright
mbed_official 3:7a284390b0ce 20 * notice, this list of conditions and the following disclaimer in
mbed_official 3:7a284390b0ce 21 * the documentation and/or other materials provided with the
mbed_official 3:7a284390b0ce 22 * distribution.
mbed_official 3:7a284390b0ce 23 * - Neither the name of ARM LIMITED nor the names of its contributors
mbed_official 3:7a284390b0ce 24 * may be used to endorse or promote products derived from this
mbed_official 3:7a284390b0ce 25 * software without specific prior written permission.
emilmont 1:fdd22bb7aa52 26 *
mbed_official 3:7a284390b0ce 27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
mbed_official 3:7a284390b0ce 28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
mbed_official 3:7a284390b0ce 29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
mbed_official 3:7a284390b0ce 30 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
mbed_official 3:7a284390b0ce 31 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
mbed_official 3:7a284390b0ce 32 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
mbed_official 3:7a284390b0ce 33 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
mbed_official 3:7a284390b0ce 34 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
mbed_official 3:7a284390b0ce 35 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
mbed_official 3:7a284390b0ce 36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
mbed_official 3:7a284390b0ce 37 * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
mbed_official 3:7a284390b0ce 38 * POSSIBILITY OF SUCH DAMAGE.
emilmont 1:fdd22bb7aa52 39 * -------------------------------------------------------------------- */
emilmont 1:fdd22bb7aa52 40
emilmont 1:fdd22bb7aa52 41 #include "arm_math.h"
emilmont 1:fdd22bb7aa52 42
emilmont 1:fdd22bb7aa52 43 /**
emilmont 1:fdd22bb7aa52 44 * @ingroup groupFilters
emilmont 1:fdd22bb7aa52 45 */
emilmont 1:fdd22bb7aa52 46
emilmont 1:fdd22bb7aa52 47 /**
emilmont 1:fdd22bb7aa52 48 * @addtogroup Conv
emilmont 1:fdd22bb7aa52 49 * @{
emilmont 1:fdd22bb7aa52 50 */
emilmont 1:fdd22bb7aa52 51
emilmont 1:fdd22bb7aa52 52 /**
emilmont 1:fdd22bb7aa52 53 * @brief Convolution of Q15 sequences (fast version) for Cortex-M3 and Cortex-M4.
emilmont 1:fdd22bb7aa52 54 * @param[in] *pSrcA points to the first input sequence.
emilmont 1:fdd22bb7aa52 55 * @param[in] srcALen length of the first input sequence.
emilmont 1:fdd22bb7aa52 56 * @param[in] *pSrcB points to the second input sequence.
emilmont 1:fdd22bb7aa52 57 * @param[in] srcBLen length of the second input sequence.
emilmont 1:fdd22bb7aa52 58 * @param[out] *pDst points to the location where the output result is written. Length srcALen+srcBLen-1.
emilmont 1:fdd22bb7aa52 59 * @return none.
emilmont 1:fdd22bb7aa52 60 *
emilmont 1:fdd22bb7aa52 61 * <b>Scaling and Overflow Behavior:</b>
emilmont 1:fdd22bb7aa52 62 *
emilmont 1:fdd22bb7aa52 63 * \par
emilmont 1:fdd22bb7aa52 64 * This fast version uses a 32-bit accumulator with 2.30 format.
emilmont 1:fdd22bb7aa52 65 * The accumulator maintains full precision of the intermediate multiplication results
emilmont 1:fdd22bb7aa52 66 * but provides only a single guard bit. There is no saturation on intermediate additions.
emilmont 1:fdd22bb7aa52 67 * Thus, if the accumulator overflows it wraps around and distorts the result.
emilmont 1:fdd22bb7aa52 68 * The input signals should be scaled down to avoid intermediate overflows.
emilmont 1:fdd22bb7aa52 69 * Scale down the inputs by log2(min(srcALen, srcBLen)) (log2 is read as log to the base 2) times to avoid overflows,
emilmont 1:fdd22bb7aa52 70 * as maximum of min(srcALen, srcBLen) number of additions are carried internally.
emilmont 1:fdd22bb7aa52 71 * The 2.30 accumulator is right shifted by 15 bits and then saturated to 1.15 format to yield the final result.
emilmont 1:fdd22bb7aa52 72 *
emilmont 1:fdd22bb7aa52 73 * \par
emilmont 1:fdd22bb7aa52 74 * See <code>arm_conv_q15()</code> for a slower implementation of this function which uses 64-bit accumulation to avoid wrap around distortion.
emilmont 1:fdd22bb7aa52 75 */
emilmont 1:fdd22bb7aa52 76
emilmont 1:fdd22bb7aa52 77 void arm_conv_fast_q15(
emilmont 1:fdd22bb7aa52 78 q15_t * pSrcA,
emilmont 1:fdd22bb7aa52 79 uint32_t srcALen,
emilmont 1:fdd22bb7aa52 80 q15_t * pSrcB,
emilmont 1:fdd22bb7aa52 81 uint32_t srcBLen,
emilmont 1:fdd22bb7aa52 82 q15_t * pDst)
emilmont 1:fdd22bb7aa52 83 {
emilmont 1:fdd22bb7aa52 84 #ifndef UNALIGNED_SUPPORT_DISABLE
emilmont 1:fdd22bb7aa52 85 q15_t *pIn1; /* inputA pointer */
emilmont 1:fdd22bb7aa52 86 q15_t *pIn2; /* inputB pointer */
emilmont 1:fdd22bb7aa52 87 q15_t *pOut = pDst; /* output pointer */
emilmont 1:fdd22bb7aa52 88 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
emilmont 1:fdd22bb7aa52 89 q15_t *px; /* Intermediate inputA pointer */
emilmont 1:fdd22bb7aa52 90 q15_t *py; /* Intermediate inputB pointer */
emilmont 1:fdd22bb7aa52 91 q15_t *pSrc1, *pSrc2; /* Intermediate pointers */
emilmont 1:fdd22bb7aa52 92 q31_t x0, x1, x2, x3, c0; /* Temporary variables to hold state and coefficient values */
emilmont 1:fdd22bb7aa52 93 uint32_t blockSize1, blockSize2, blockSize3, j, k, count, blkCnt; /* loop counter */
emilmont 1:fdd22bb7aa52 94
emilmont 1:fdd22bb7aa52 95 /* The algorithm implementation is based on the lengths of the inputs. */
emilmont 1:fdd22bb7aa52 96 /* srcB is always made to slide across srcA. */
emilmont 1:fdd22bb7aa52 97 /* So srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 98 if(srcALen >= srcBLen)
emilmont 1:fdd22bb7aa52 99 {
emilmont 1:fdd22bb7aa52 100 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 101 pIn1 = pSrcA;
emilmont 1:fdd22bb7aa52 102
emilmont 1:fdd22bb7aa52 103 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 104 pIn2 = pSrcB;
emilmont 1:fdd22bb7aa52 105 }
emilmont 1:fdd22bb7aa52 106 else
emilmont 1:fdd22bb7aa52 107 {
emilmont 1:fdd22bb7aa52 108 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 109 pIn1 = pSrcB;
emilmont 1:fdd22bb7aa52 110
emilmont 1:fdd22bb7aa52 111 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 112 pIn2 = pSrcA;
emilmont 1:fdd22bb7aa52 113
emilmont 1:fdd22bb7aa52 114 /* srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 115 j = srcBLen;
emilmont 1:fdd22bb7aa52 116 srcBLen = srcALen;
emilmont 1:fdd22bb7aa52 117 srcALen = j;
emilmont 1:fdd22bb7aa52 118 }
emilmont 1:fdd22bb7aa52 119
emilmont 1:fdd22bb7aa52 120 /* 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 121 /* The function is internally
emilmont 1:fdd22bb7aa52 122 * divided into three stages according to the number of multiplications that has to be
emilmont 1:fdd22bb7aa52 123 * taken place between inputA samples and inputB samples. In the first stage of the
emilmont 1:fdd22bb7aa52 124 * algorithm, the multiplications increase by one for every iteration.
emilmont 1:fdd22bb7aa52 125 * In the second stage of the algorithm, srcBLen number of multiplications are done.
emilmont 1:fdd22bb7aa52 126 * In the third stage of the algorithm, the multiplications decrease by one
emilmont 1:fdd22bb7aa52 127 * for every iteration. */
emilmont 1:fdd22bb7aa52 128
emilmont 1:fdd22bb7aa52 129 /* The algorithm is implemented in three stages.
emilmont 1:fdd22bb7aa52 130 The loop counters of each stage is initiated here. */
emilmont 1:fdd22bb7aa52 131 blockSize1 = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 132 blockSize2 = srcALen - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 133 blockSize3 = blockSize1;
emilmont 1:fdd22bb7aa52 134
emilmont 1:fdd22bb7aa52 135 /* --------------------------
emilmont 1:fdd22bb7aa52 136 * Initializations of stage1
emilmont 1:fdd22bb7aa52 137 * -------------------------*/
emilmont 1:fdd22bb7aa52 138
emilmont 1:fdd22bb7aa52 139 /* sum = x[0] * y[0]
emilmont 1:fdd22bb7aa52 140 * sum = x[0] * y[1] + x[1] * y[0]
emilmont 1:fdd22bb7aa52 141 * ....
emilmont 1:fdd22bb7aa52 142 * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
emilmont 1:fdd22bb7aa52 143 */
emilmont 1:fdd22bb7aa52 144
emilmont 1:fdd22bb7aa52 145 /* In this stage the MAC operations are increased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 146 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 147 count = 1u;
emilmont 1:fdd22bb7aa52 148
emilmont 1:fdd22bb7aa52 149 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 150 px = pIn1;
emilmont 1:fdd22bb7aa52 151
emilmont 1:fdd22bb7aa52 152 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 153 py = pIn2;
emilmont 1:fdd22bb7aa52 154
emilmont 1:fdd22bb7aa52 155
emilmont 1:fdd22bb7aa52 156 /* ------------------------
emilmont 1:fdd22bb7aa52 157 * Stage1 process
emilmont 1:fdd22bb7aa52 158 * ----------------------*/
emilmont 1:fdd22bb7aa52 159
emilmont 1:fdd22bb7aa52 160 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 161 /* First part of this stage computes the MAC operations less than 4 */
emilmont 1:fdd22bb7aa52 162 /* Second part of this stage computes the MAC operations greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 163
emilmont 1:fdd22bb7aa52 164 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 165 while((count < 4u) && (blockSize1 > 0u))
emilmont 1:fdd22bb7aa52 166 {
emilmont 1:fdd22bb7aa52 167 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 168 sum = 0;
emilmont 1:fdd22bb7aa52 169
emilmont 1:fdd22bb7aa52 170 /* Loop over number of MAC operations between
emilmont 1:fdd22bb7aa52 171 * inputA samples and inputB samples */
emilmont 1:fdd22bb7aa52 172 k = count;
emilmont 1:fdd22bb7aa52 173
emilmont 1:fdd22bb7aa52 174 while(k > 0u)
emilmont 1:fdd22bb7aa52 175 {
emilmont 1:fdd22bb7aa52 176 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 177 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 178
emilmont 1:fdd22bb7aa52 179 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 180 k--;
emilmont 1:fdd22bb7aa52 181 }
emilmont 1:fdd22bb7aa52 182
emilmont 1:fdd22bb7aa52 183 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 184 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 185
emilmont 1:fdd22bb7aa52 186 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 187 py = pIn2 + count;
emilmont 1:fdd22bb7aa52 188 px = pIn1;
emilmont 1:fdd22bb7aa52 189
emilmont 1:fdd22bb7aa52 190 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 191 count++;
emilmont 1:fdd22bb7aa52 192
emilmont 1:fdd22bb7aa52 193 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 194 blockSize1--;
emilmont 1:fdd22bb7aa52 195 }
emilmont 1:fdd22bb7aa52 196
emilmont 1:fdd22bb7aa52 197 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 198 /* The internal loop, over count, is unrolled by 4 */
emilmont 1:fdd22bb7aa52 199 /* To, read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 200 * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */
emilmont 1:fdd22bb7aa52 201 py = py - 1;
emilmont 1:fdd22bb7aa52 202
emilmont 1:fdd22bb7aa52 203 while(blockSize1 > 0u)
emilmont 1:fdd22bb7aa52 204 {
emilmont 1:fdd22bb7aa52 205 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 206 sum = 0;
emilmont 1:fdd22bb7aa52 207
emilmont 1:fdd22bb7aa52 208 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 209 k = count >> 2u;
emilmont 1:fdd22bb7aa52 210
emilmont 1:fdd22bb7aa52 211 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 212 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 213 while(k > 0u)
emilmont 1:fdd22bb7aa52 214 {
emilmont 1:fdd22bb7aa52 215 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 216 /* x[0], x[1] are multiplied with y[srcBLen - 1], y[srcBLen - 2] respectively */
emilmont 1:fdd22bb7aa52 217 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 218 /* x[2], x[3] are multiplied with y[srcBLen - 3], y[srcBLen - 4] respectively */
emilmont 1:fdd22bb7aa52 219 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 220
emilmont 1:fdd22bb7aa52 221 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 222 k--;
emilmont 1:fdd22bb7aa52 223 }
emilmont 1:fdd22bb7aa52 224
emilmont 1:fdd22bb7aa52 225 /* For the next MAC operations, the pointer py is used without SIMD
emilmont 1:fdd22bb7aa52 226 * So, py is incremented by 1 */
emilmont 1:fdd22bb7aa52 227 py = py + 1u;
emilmont 1:fdd22bb7aa52 228
emilmont 1:fdd22bb7aa52 229 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 230 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 231 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 232
emilmont 1:fdd22bb7aa52 233 while(k > 0u)
emilmont 1:fdd22bb7aa52 234 {
emilmont 1:fdd22bb7aa52 235 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 236 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 237
emilmont 1:fdd22bb7aa52 238 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 239 k--;
emilmont 1:fdd22bb7aa52 240 }
emilmont 1:fdd22bb7aa52 241
emilmont 1:fdd22bb7aa52 242 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 243 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 244
emilmont 1:fdd22bb7aa52 245 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 246 py = pIn2 + (count - 1u);
emilmont 1:fdd22bb7aa52 247 px = pIn1;
emilmont 1:fdd22bb7aa52 248
emilmont 1:fdd22bb7aa52 249 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 250 count++;
emilmont 1:fdd22bb7aa52 251
emilmont 1:fdd22bb7aa52 252 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 253 blockSize1--;
emilmont 1:fdd22bb7aa52 254 }
emilmont 1:fdd22bb7aa52 255
emilmont 1:fdd22bb7aa52 256 /* --------------------------
emilmont 1:fdd22bb7aa52 257 * Initializations of stage2
emilmont 1:fdd22bb7aa52 258 * ------------------------*/
emilmont 1:fdd22bb7aa52 259
emilmont 1:fdd22bb7aa52 260 /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
emilmont 1:fdd22bb7aa52 261 * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
emilmont 1:fdd22bb7aa52 262 * ....
emilmont 1:fdd22bb7aa52 263 * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
emilmont 1:fdd22bb7aa52 264 */
emilmont 1:fdd22bb7aa52 265
emilmont 1:fdd22bb7aa52 266 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 267 px = pIn1;
emilmont 1:fdd22bb7aa52 268
emilmont 1:fdd22bb7aa52 269 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 270 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 271 py = pSrc2;
emilmont 1:fdd22bb7aa52 272
emilmont 1:fdd22bb7aa52 273 /* count is the index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 274 count = 0u;
emilmont 1:fdd22bb7aa52 275
emilmont 1:fdd22bb7aa52 276
emilmont 1:fdd22bb7aa52 277 /* --------------------
emilmont 1:fdd22bb7aa52 278 * Stage2 process
emilmont 1:fdd22bb7aa52 279 * -------------------*/
emilmont 1:fdd22bb7aa52 280
emilmont 1:fdd22bb7aa52 281 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 282 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 283 * srcBLen should be greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 284 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 285 {
emilmont 1:fdd22bb7aa52 286 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 287 blkCnt = blockSize2 >> 2u;
emilmont 1:fdd22bb7aa52 288
emilmont 1:fdd22bb7aa52 289 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 290 {
emilmont 1:fdd22bb7aa52 291 py = py - 1u;
emilmont 1:fdd22bb7aa52 292
emilmont 1:fdd22bb7aa52 293 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 294 acc0 = 0;
emilmont 1:fdd22bb7aa52 295 acc1 = 0;
emilmont 1:fdd22bb7aa52 296 acc2 = 0;
emilmont 1:fdd22bb7aa52 297 acc3 = 0;
emilmont 1:fdd22bb7aa52 298
emilmont 1:fdd22bb7aa52 299
emilmont 1:fdd22bb7aa52 300 /* read x[0], x[1] samples */
emilmont 1:fdd22bb7aa52 301 x0 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 302 /* read x[1], x[2] samples */
emilmont 1:fdd22bb7aa52 303 x1 = _SIMD32_OFFSET(px+1);
emilmont 2:da51fb522205 304 px+= 2u;
emilmont 1:fdd22bb7aa52 305
emilmont 1:fdd22bb7aa52 306
emilmont 1:fdd22bb7aa52 307 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 308 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 309
emilmont 1:fdd22bb7aa52 310 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 311 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 312 do
emilmont 1:fdd22bb7aa52 313 {
emilmont 1:fdd22bb7aa52 314 /* Read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 315 * y[srcBLen - 1] and y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 316 c0 = *__SIMD32(py)--;
emilmont 1:fdd22bb7aa52 317
emilmont 1:fdd22bb7aa52 318 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 319 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 320
emilmont 1:fdd22bb7aa52 321 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 322 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 323
emilmont 1:fdd22bb7aa52 324 /* Read x[2], x[3] */
emilmont 1:fdd22bb7aa52 325 x2 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 326
emilmont 1:fdd22bb7aa52 327 /* Read x[3], x[4] */
emilmont 1:fdd22bb7aa52 328 x3 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 329
emilmont 1:fdd22bb7aa52 330 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 331 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 332
emilmont 1:fdd22bb7aa52 333 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 334 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 335
emilmont 1:fdd22bb7aa52 336 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 337 c0 = *__SIMD32(py)--;
emilmont 1:fdd22bb7aa52 338
emilmont 1:fdd22bb7aa52 339 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 340 acc0 = __SMLADX(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 341
emilmont 1:fdd22bb7aa52 342 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 343 acc1 = __SMLADX(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 344
emilmont 1:fdd22bb7aa52 345 /* Read x[4], x[5] */
emilmont 1:fdd22bb7aa52 346 x0 = _SIMD32_OFFSET(px+2);
emilmont 1:fdd22bb7aa52 347
emilmont 1:fdd22bb7aa52 348 /* Read x[5], x[6] */
emilmont 1:fdd22bb7aa52 349 x1 = _SIMD32_OFFSET(px+3);
emilmont 2:da51fb522205 350 px += 4u;
emilmont 1:fdd22bb7aa52 351
emilmont 1:fdd22bb7aa52 352 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 353 acc2 = __SMLADX(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 354
emilmont 1:fdd22bb7aa52 355 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 356 acc3 = __SMLADX(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 357
emilmont 1:fdd22bb7aa52 358 } while(--k);
emilmont 1:fdd22bb7aa52 359
emilmont 1:fdd22bb7aa52 360 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 361 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 362
emilmont 1:fdd22bb7aa52 363 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 364 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 365 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 366
emilmont 1:fdd22bb7aa52 367 if(k == 1u)
emilmont 1:fdd22bb7aa52 368 {
emilmont 1:fdd22bb7aa52 369 /* Read y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 370 c0 = *(py+1);
emilmont 1:fdd22bb7aa52 371
emilmont 1:fdd22bb7aa52 372 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 373
emilmont 1:fdd22bb7aa52 374 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 375
emilmont 1:fdd22bb7aa52 376 #else
emilmont 1:fdd22bb7aa52 377
emilmont 1:fdd22bb7aa52 378 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 379
emilmont 1:fdd22bb7aa52 380 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 381
emilmont 1:fdd22bb7aa52 382 /* Read x[7] */
emilmont 1:fdd22bb7aa52 383 x3 = *__SIMD32(px);
emilmont 2:da51fb522205 384 px++;
emilmont 1:fdd22bb7aa52 385
emilmont 1:fdd22bb7aa52 386 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 387 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 388 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 389 acc2 = __SMLADX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 390 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 391 }
emilmont 1:fdd22bb7aa52 392
emilmont 1:fdd22bb7aa52 393 if(k == 2u)
emilmont 1:fdd22bb7aa52 394 {
emilmont 1:fdd22bb7aa52 395 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 396 c0 = _SIMD32_OFFSET(py);
emilmont 1:fdd22bb7aa52 397
emilmont 1:fdd22bb7aa52 398 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 399 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 400
emilmont 1:fdd22bb7aa52 401 /* Read x[9] */
emilmont 1:fdd22bb7aa52 402 x2 = _SIMD32_OFFSET(px+1);
emilmont 2:da51fb522205 403 px += 2u;
emilmont 1:fdd22bb7aa52 404
emilmont 1:fdd22bb7aa52 405 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 406 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 407 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 408 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 409 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 410 }
emilmont 1:fdd22bb7aa52 411
emilmont 1:fdd22bb7aa52 412 if(k == 3u)
emilmont 1:fdd22bb7aa52 413 {
emilmont 1:fdd22bb7aa52 414 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 415 c0 = _SIMD32_OFFSET(py);
emilmont 1:fdd22bb7aa52 416
emilmont 1:fdd22bb7aa52 417 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 418 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 419
emilmont 1:fdd22bb7aa52 420 /* Read x[9] */
emilmont 1:fdd22bb7aa52 421 x2 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 422
emilmont 1:fdd22bb7aa52 423 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 424 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 425 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 426 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 427 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 428
emilmont 1:fdd22bb7aa52 429 /* Read y[srcBLen - 7] */
emilmont 2:da51fb522205 430 c0 = *(py-1);
emilmont 1:fdd22bb7aa52 431 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 432
emilmont 1:fdd22bb7aa52 433 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 434 #else
emilmont 1:fdd22bb7aa52 435
emilmont 1:fdd22bb7aa52 436 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 437 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 438
emilmont 1:fdd22bb7aa52 439 /* Read x[10] */
emilmont 1:fdd22bb7aa52 440 x3 = _SIMD32_OFFSET(px+2);
emilmont 2:da51fb522205 441 px += 3u;
emilmont 1:fdd22bb7aa52 442
emilmont 1:fdd22bb7aa52 443 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 444 acc0 = __SMLADX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 445 acc1 = __SMLAD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 446 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 447 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 448 }
emilmont 1:fdd22bb7aa52 449
emilmont 1:fdd22bb7aa52 450 /* Store the results in the accumulators in the destination buffer. */
emilmont 1:fdd22bb7aa52 451 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 452
emilmont 1:fdd22bb7aa52 453 *__SIMD32(pOut)++ = __PKHBT((acc0 >> 15), (acc1 >> 15), 16);
emilmont 1:fdd22bb7aa52 454 *__SIMD32(pOut)++ = __PKHBT((acc2 >> 15), (acc3 >> 15), 16);
emilmont 1:fdd22bb7aa52 455
emilmont 1:fdd22bb7aa52 456 #else
emilmont 1:fdd22bb7aa52 457
emilmont 1:fdd22bb7aa52 458 *__SIMD32(pOut)++ = __PKHBT((acc1 >> 15), (acc0 >> 15), 16);
emilmont 1:fdd22bb7aa52 459 *__SIMD32(pOut)++ = __PKHBT((acc3 >> 15), (acc2 >> 15), 16);
emilmont 1:fdd22bb7aa52 460
emilmont 1:fdd22bb7aa52 461 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 462
emilmont 1:fdd22bb7aa52 463 /* Increment the pointer pIn1 index, count by 4 */
emilmont 1:fdd22bb7aa52 464 count += 4u;
emilmont 1:fdd22bb7aa52 465
emilmont 1:fdd22bb7aa52 466 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 467 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 468 py = pSrc2;
emilmont 1:fdd22bb7aa52 469
emilmont 1:fdd22bb7aa52 470 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 471 blkCnt--;
emilmont 1:fdd22bb7aa52 472 }
emilmont 1:fdd22bb7aa52 473
emilmont 1:fdd22bb7aa52 474 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 475 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 476 blkCnt = blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 477
emilmont 1:fdd22bb7aa52 478 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 479 {
emilmont 1:fdd22bb7aa52 480 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 481 sum = 0;
emilmont 1:fdd22bb7aa52 482
emilmont 1:fdd22bb7aa52 483 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 484 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 485
emilmont 1:fdd22bb7aa52 486 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 487 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 488 while(k > 0u)
emilmont 1:fdd22bb7aa52 489 {
emilmont 1:fdd22bb7aa52 490 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 491 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 492 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 493 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 494 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 495
emilmont 1:fdd22bb7aa52 496 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 497 k--;
emilmont 1:fdd22bb7aa52 498 }
emilmont 1:fdd22bb7aa52 499
emilmont 1:fdd22bb7aa52 500 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 501 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 502 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 503
emilmont 1:fdd22bb7aa52 504 while(k > 0u)
emilmont 1:fdd22bb7aa52 505 {
emilmont 1:fdd22bb7aa52 506 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 507 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 508
emilmont 1:fdd22bb7aa52 509 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 510 k--;
emilmont 1:fdd22bb7aa52 511 }
emilmont 1:fdd22bb7aa52 512
emilmont 1:fdd22bb7aa52 513 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 514 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 515
emilmont 1:fdd22bb7aa52 516 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 517 count++;
emilmont 1:fdd22bb7aa52 518
emilmont 1:fdd22bb7aa52 519 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 520 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 521 py = pSrc2;
emilmont 1:fdd22bb7aa52 522
emilmont 1:fdd22bb7aa52 523 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 524 blkCnt--;
emilmont 1:fdd22bb7aa52 525 }
emilmont 1:fdd22bb7aa52 526 }
emilmont 1:fdd22bb7aa52 527 else
emilmont 1:fdd22bb7aa52 528 {
emilmont 1:fdd22bb7aa52 529 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 530 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 531 blkCnt = blockSize2;
emilmont 1:fdd22bb7aa52 532
emilmont 1:fdd22bb7aa52 533 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 534 {
emilmont 1:fdd22bb7aa52 535 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 536 sum = 0;
emilmont 1:fdd22bb7aa52 537
emilmont 1:fdd22bb7aa52 538 /* srcBLen number of MACS should be performed */
emilmont 1:fdd22bb7aa52 539 k = srcBLen;
emilmont 1:fdd22bb7aa52 540
emilmont 1:fdd22bb7aa52 541 while(k > 0u)
emilmont 1:fdd22bb7aa52 542 {
emilmont 1:fdd22bb7aa52 543 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 544 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 545
emilmont 1:fdd22bb7aa52 546 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 547 k--;
emilmont 1:fdd22bb7aa52 548 }
emilmont 1:fdd22bb7aa52 549
emilmont 1:fdd22bb7aa52 550 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 551 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 552
emilmont 1:fdd22bb7aa52 553 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 554 count++;
emilmont 1:fdd22bb7aa52 555
emilmont 1:fdd22bb7aa52 556 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 557 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 558 py = pSrc2;
emilmont 1:fdd22bb7aa52 559
emilmont 1:fdd22bb7aa52 560 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 561 blkCnt--;
emilmont 1:fdd22bb7aa52 562 }
emilmont 1:fdd22bb7aa52 563 }
emilmont 1:fdd22bb7aa52 564
emilmont 1:fdd22bb7aa52 565
emilmont 1:fdd22bb7aa52 566 /* --------------------------
emilmont 1:fdd22bb7aa52 567 * Initializations of stage3
emilmont 1:fdd22bb7aa52 568 * -------------------------*/
emilmont 1:fdd22bb7aa52 569
emilmont 1:fdd22bb7aa52 570 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 571 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emilmont 1:fdd22bb7aa52 572 * ....
emilmont 1:fdd22bb7aa52 573 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emilmont 1:fdd22bb7aa52 574 * sum += x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 575 */
emilmont 1:fdd22bb7aa52 576
emilmont 1:fdd22bb7aa52 577 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 578 The blockSize3 variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 579
emilmont 1:fdd22bb7aa52 580 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 581 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 582 px = pSrc1;
emilmont 1:fdd22bb7aa52 583
emilmont 1:fdd22bb7aa52 584 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 585 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 586 pIn2 = pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 587 py = pIn2;
emilmont 1:fdd22bb7aa52 588
emilmont 1:fdd22bb7aa52 589 /* -------------------
emilmont 1:fdd22bb7aa52 590 * Stage3 process
emilmont 1:fdd22bb7aa52 591 * ------------------*/
emilmont 1:fdd22bb7aa52 592
emilmont 1:fdd22bb7aa52 593 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 594 /* First part of this stage computes the MAC operations greater than 4 */
emilmont 1:fdd22bb7aa52 595 /* Second part of this stage computes the MAC operations less than or equal to 4 */
emilmont 1:fdd22bb7aa52 596
emilmont 1:fdd22bb7aa52 597 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 598 j = blockSize3 >> 2u;
emilmont 1:fdd22bb7aa52 599
emilmont 1:fdd22bb7aa52 600 while((j > 0u) && (blockSize3 > 0u))
emilmont 1:fdd22bb7aa52 601 {
emilmont 1:fdd22bb7aa52 602 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 603 sum = 0;
emilmont 1:fdd22bb7aa52 604
emilmont 1:fdd22bb7aa52 605 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 606 k = blockSize3 >> 2u;
emilmont 1:fdd22bb7aa52 607
emilmont 1:fdd22bb7aa52 608 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 609 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 610 while(k > 0u)
emilmont 1:fdd22bb7aa52 611 {
emilmont 1:fdd22bb7aa52 612 /* x[srcALen - srcBLen + 1], x[srcALen - srcBLen + 2] are multiplied
emilmont 1:fdd22bb7aa52 613 * with y[srcBLen - 1], y[srcBLen - 2] respectively */
emilmont 1:fdd22bb7aa52 614 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 615 /* x[srcALen - srcBLen + 3], x[srcALen - srcBLen + 4] are multiplied
emilmont 1:fdd22bb7aa52 616 * with y[srcBLen - 3], y[srcBLen - 4] respectively */
emilmont 1:fdd22bb7aa52 617 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 618
emilmont 1:fdd22bb7aa52 619 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 620 k--;
emilmont 1:fdd22bb7aa52 621 }
emilmont 1:fdd22bb7aa52 622
emilmont 1:fdd22bb7aa52 623 /* For the next MAC operations, the pointer py is used without SIMD
emilmont 1:fdd22bb7aa52 624 * So, py is incremented by 1 */
emilmont 1:fdd22bb7aa52 625 py = py + 1u;
emilmont 1:fdd22bb7aa52 626
emilmont 1:fdd22bb7aa52 627 /* If the blockSize3 is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 628 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 629 k = blockSize3 % 0x4u;
emilmont 1:fdd22bb7aa52 630
emilmont 1:fdd22bb7aa52 631 while(k > 0u)
emilmont 1:fdd22bb7aa52 632 {
emilmont 1:fdd22bb7aa52 633 /* sum += x[srcALen - srcBLen + 5] * y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 634 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 635
emilmont 1:fdd22bb7aa52 636 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 637 k--;
emilmont 1:fdd22bb7aa52 638 }
emilmont 1:fdd22bb7aa52 639
emilmont 1:fdd22bb7aa52 640 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 641 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 642
emilmont 1:fdd22bb7aa52 643 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 644 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 645 py = pIn2;
emilmont 1:fdd22bb7aa52 646
emilmont 1:fdd22bb7aa52 647 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 648 blockSize3--;
emilmont 1:fdd22bb7aa52 649
emilmont 1:fdd22bb7aa52 650 j--;
emilmont 1:fdd22bb7aa52 651 }
emilmont 1:fdd22bb7aa52 652
emilmont 1:fdd22bb7aa52 653 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 654 /* SIMD is not used for the next MAC operations,
emilmont 1:fdd22bb7aa52 655 * so pointer py is updated to read only one sample at a time */
emilmont 1:fdd22bb7aa52 656 py = py + 1u;
emilmont 1:fdd22bb7aa52 657
emilmont 1:fdd22bb7aa52 658 while(blockSize3 > 0u)
emilmont 1:fdd22bb7aa52 659 {
emilmont 1:fdd22bb7aa52 660 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 661 sum = 0;
emilmont 1:fdd22bb7aa52 662
emilmont 1:fdd22bb7aa52 663 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 664 k = blockSize3;
emilmont 1:fdd22bb7aa52 665
emilmont 1:fdd22bb7aa52 666 while(k > 0u)
emilmont 1:fdd22bb7aa52 667 {
emilmont 1:fdd22bb7aa52 668 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 669 /* sum += x[srcALen-1] * y[srcBLen-1] */
emilmont 1:fdd22bb7aa52 670 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 671
emilmont 1:fdd22bb7aa52 672 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 673 k--;
emilmont 1:fdd22bb7aa52 674 }
emilmont 1:fdd22bb7aa52 675
emilmont 1:fdd22bb7aa52 676 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 677 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 678
emilmont 1:fdd22bb7aa52 679 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 680 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 681 py = pSrc2;
emilmont 1:fdd22bb7aa52 682
emilmont 1:fdd22bb7aa52 683 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 684 blockSize3--;
emilmont 1:fdd22bb7aa52 685 }
emilmont 1:fdd22bb7aa52 686
emilmont 1:fdd22bb7aa52 687 #else
emilmont 1:fdd22bb7aa52 688 q15_t *pIn1; /* inputA pointer */
emilmont 1:fdd22bb7aa52 689 q15_t *pIn2; /* inputB pointer */
emilmont 1:fdd22bb7aa52 690 q15_t *pOut = pDst; /* output pointer */
emilmont 1:fdd22bb7aa52 691 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
emilmont 1:fdd22bb7aa52 692 q15_t *px; /* Intermediate inputA pointer */
emilmont 1:fdd22bb7aa52 693 q15_t *py; /* Intermediate inputB pointer */
emilmont 1:fdd22bb7aa52 694 q15_t *pSrc1, *pSrc2; /* Intermediate pointers */
emilmont 1:fdd22bb7aa52 695 q31_t x0, x1, x2, x3, c0; /* Temporary variables to hold state and coefficient values */
emilmont 1:fdd22bb7aa52 696 uint32_t blockSize1, blockSize2, blockSize3, j, k, count, blkCnt; /* loop counter */
emilmont 1:fdd22bb7aa52 697 q15_t a, b;
emilmont 1:fdd22bb7aa52 698
emilmont 1:fdd22bb7aa52 699 /* The algorithm implementation is based on the lengths of the inputs. */
emilmont 1:fdd22bb7aa52 700 /* srcB is always made to slide across srcA. */
emilmont 1:fdd22bb7aa52 701 /* So srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 702 if(srcALen >= srcBLen)
emilmont 1:fdd22bb7aa52 703 {
emilmont 1:fdd22bb7aa52 704 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 705 pIn1 = pSrcA;
emilmont 1:fdd22bb7aa52 706
emilmont 1:fdd22bb7aa52 707 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 708 pIn2 = pSrcB;
emilmont 1:fdd22bb7aa52 709 }
emilmont 1:fdd22bb7aa52 710 else
emilmont 1:fdd22bb7aa52 711 {
emilmont 1:fdd22bb7aa52 712 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 713 pIn1 = pSrcB;
emilmont 1:fdd22bb7aa52 714
emilmont 1:fdd22bb7aa52 715 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 716 pIn2 = pSrcA;
emilmont 1:fdd22bb7aa52 717
emilmont 1:fdd22bb7aa52 718 /* srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 719 j = srcBLen;
emilmont 1:fdd22bb7aa52 720 srcBLen = srcALen;
emilmont 1:fdd22bb7aa52 721 srcALen = j;
emilmont 1:fdd22bb7aa52 722 }
emilmont 1:fdd22bb7aa52 723
emilmont 1:fdd22bb7aa52 724 /* 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 725 /* The function is internally
emilmont 1:fdd22bb7aa52 726 * divided into three stages according to the number of multiplications that has to be
emilmont 1:fdd22bb7aa52 727 * taken place between inputA samples and inputB samples. In the first stage of the
emilmont 1:fdd22bb7aa52 728 * algorithm, the multiplications increase by one for every iteration.
emilmont 1:fdd22bb7aa52 729 * In the second stage of the algorithm, srcBLen number of multiplications are done.
emilmont 1:fdd22bb7aa52 730 * In the third stage of the algorithm, the multiplications decrease by one
emilmont 1:fdd22bb7aa52 731 * for every iteration. */
emilmont 1:fdd22bb7aa52 732
emilmont 1:fdd22bb7aa52 733 /* The algorithm is implemented in three stages.
emilmont 1:fdd22bb7aa52 734 The loop counters of each stage is initiated here. */
emilmont 1:fdd22bb7aa52 735 blockSize1 = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 736 blockSize2 = srcALen - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 737 blockSize3 = blockSize1;
emilmont 1:fdd22bb7aa52 738
emilmont 1:fdd22bb7aa52 739 /* --------------------------
emilmont 1:fdd22bb7aa52 740 * Initializations of stage1
emilmont 1:fdd22bb7aa52 741 * -------------------------*/
emilmont 1:fdd22bb7aa52 742
emilmont 1:fdd22bb7aa52 743 /* sum = x[0] * y[0]
emilmont 1:fdd22bb7aa52 744 * sum = x[0] * y[1] + x[1] * y[0]
emilmont 1:fdd22bb7aa52 745 * ....
emilmont 1:fdd22bb7aa52 746 * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
emilmont 1:fdd22bb7aa52 747 */
emilmont 1:fdd22bb7aa52 748
emilmont 1:fdd22bb7aa52 749 /* In this stage the MAC operations are increased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 750 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 751 count = 1u;
emilmont 1:fdd22bb7aa52 752
emilmont 1:fdd22bb7aa52 753 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 754 px = pIn1;
emilmont 1:fdd22bb7aa52 755
emilmont 1:fdd22bb7aa52 756 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 757 py = pIn2;
emilmont 1:fdd22bb7aa52 758
emilmont 1:fdd22bb7aa52 759
emilmont 1:fdd22bb7aa52 760 /* ------------------------
emilmont 1:fdd22bb7aa52 761 * Stage1 process
emilmont 1:fdd22bb7aa52 762 * ----------------------*/
emilmont 1:fdd22bb7aa52 763
emilmont 1:fdd22bb7aa52 764 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 765 /* First part of this stage computes the MAC operations less than 4 */
emilmont 1:fdd22bb7aa52 766 /* Second part of this stage computes the MAC operations greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 767
emilmont 1:fdd22bb7aa52 768 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 769 while((count < 4u) && (blockSize1 > 0u))
emilmont 1:fdd22bb7aa52 770 {
emilmont 1:fdd22bb7aa52 771 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 772 sum = 0;
emilmont 1:fdd22bb7aa52 773
emilmont 1:fdd22bb7aa52 774 /* Loop over number of MAC operations between
emilmont 1:fdd22bb7aa52 775 * inputA samples and inputB samples */
emilmont 1:fdd22bb7aa52 776 k = count;
emilmont 1:fdd22bb7aa52 777
emilmont 1:fdd22bb7aa52 778 while(k > 0u)
emilmont 1:fdd22bb7aa52 779 {
emilmont 1:fdd22bb7aa52 780 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 781 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 782
emilmont 1:fdd22bb7aa52 783 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 784 k--;
emilmont 1:fdd22bb7aa52 785 }
emilmont 1:fdd22bb7aa52 786
emilmont 1:fdd22bb7aa52 787 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 788 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 789
emilmont 1:fdd22bb7aa52 790 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 791 py = pIn2 + count;
emilmont 1:fdd22bb7aa52 792 px = pIn1;
emilmont 1:fdd22bb7aa52 793
emilmont 1:fdd22bb7aa52 794 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 795 count++;
emilmont 1:fdd22bb7aa52 796
emilmont 1:fdd22bb7aa52 797 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 798 blockSize1--;
emilmont 1:fdd22bb7aa52 799 }
emilmont 1:fdd22bb7aa52 800
emilmont 1:fdd22bb7aa52 801 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 802 /* The internal loop, over count, is unrolled by 4 */
emilmont 1:fdd22bb7aa52 803 /* To, read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 804 * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */
emilmont 1:fdd22bb7aa52 805 py = py - 1;
emilmont 1:fdd22bb7aa52 806
emilmont 1:fdd22bb7aa52 807 while(blockSize1 > 0u)
emilmont 1:fdd22bb7aa52 808 {
emilmont 1:fdd22bb7aa52 809 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 810 sum = 0;
emilmont 1:fdd22bb7aa52 811
emilmont 1:fdd22bb7aa52 812 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 813 k = count >> 2u;
emilmont 1:fdd22bb7aa52 814
emilmont 1:fdd22bb7aa52 815 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 816 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 2:da51fb522205 817 py++;
emilmont 1:fdd22bb7aa52 818
emilmont 1:fdd22bb7aa52 819 while(k > 0u)
emilmont 1:fdd22bb7aa52 820 {
emilmont 1:fdd22bb7aa52 821 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 822 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 823 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 824 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 825 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 826
emilmont 1:fdd22bb7aa52 827 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 828 k--;
emilmont 1:fdd22bb7aa52 829 }
emilmont 1:fdd22bb7aa52 830
emilmont 1:fdd22bb7aa52 831 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 832 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 833 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 834
emilmont 1:fdd22bb7aa52 835 while(k > 0u)
emilmont 1:fdd22bb7aa52 836 {
emilmont 1:fdd22bb7aa52 837 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 838 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 839
emilmont 1:fdd22bb7aa52 840 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 841 k--;
emilmont 1:fdd22bb7aa52 842 }
emilmont 1:fdd22bb7aa52 843
emilmont 1:fdd22bb7aa52 844 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 845 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 846
emilmont 1:fdd22bb7aa52 847 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 848 py = pIn2 + (count - 1u);
emilmont 1:fdd22bb7aa52 849 px = pIn1;
emilmont 1:fdd22bb7aa52 850
emilmont 1:fdd22bb7aa52 851 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 852 count++;
emilmont 1:fdd22bb7aa52 853
emilmont 1:fdd22bb7aa52 854 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 855 blockSize1--;
emilmont 1:fdd22bb7aa52 856 }
emilmont 1:fdd22bb7aa52 857
emilmont 1:fdd22bb7aa52 858 /* --------------------------
emilmont 1:fdd22bb7aa52 859 * Initializations of stage2
emilmont 1:fdd22bb7aa52 860 * ------------------------*/
emilmont 1:fdd22bb7aa52 861
emilmont 1:fdd22bb7aa52 862 /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
emilmont 1:fdd22bb7aa52 863 * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
emilmont 1:fdd22bb7aa52 864 * ....
emilmont 1:fdd22bb7aa52 865 * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
emilmont 1:fdd22bb7aa52 866 */
emilmont 1:fdd22bb7aa52 867
emilmont 1:fdd22bb7aa52 868 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 869 px = pIn1;
emilmont 1:fdd22bb7aa52 870
emilmont 1:fdd22bb7aa52 871 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 872 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 873 py = pSrc2;
emilmont 1:fdd22bb7aa52 874
emilmont 1:fdd22bb7aa52 875 /* count is the index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 876 count = 0u;
emilmont 1:fdd22bb7aa52 877
emilmont 1:fdd22bb7aa52 878
emilmont 1:fdd22bb7aa52 879 /* --------------------
emilmont 1:fdd22bb7aa52 880 * Stage2 process
emilmont 1:fdd22bb7aa52 881 * -------------------*/
emilmont 1:fdd22bb7aa52 882
emilmont 1:fdd22bb7aa52 883 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 884 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 885 * srcBLen should be greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 886 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 887 {
emilmont 1:fdd22bb7aa52 888 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 889 blkCnt = blockSize2 >> 2u;
emilmont 1:fdd22bb7aa52 890
emilmont 1:fdd22bb7aa52 891 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 892 {
emilmont 1:fdd22bb7aa52 893 py = py - 1u;
emilmont 1:fdd22bb7aa52 894
emilmont 1:fdd22bb7aa52 895 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 896 acc0 = 0;
emilmont 1:fdd22bb7aa52 897 acc1 = 0;
emilmont 1:fdd22bb7aa52 898 acc2 = 0;
emilmont 2:da51fb522205 899 acc3 = 0;
emilmont 1:fdd22bb7aa52 900
emilmont 1:fdd22bb7aa52 901 /* read x[0], x[1] samples */
emilmont 2:da51fb522205 902 a = *px++;
emilmont 2:da51fb522205 903 b = *px++;
emilmont 1:fdd22bb7aa52 904
emilmont 1:fdd22bb7aa52 905 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 906
emilmont 2:da51fb522205 907 x0 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 908 a = *px;
emilmont 2:da51fb522205 909 x1 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 910
emilmont 1:fdd22bb7aa52 911 #else
emilmont 1:fdd22bb7aa52 912
emilmont 2:da51fb522205 913 x0 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 914 a = *px;
emilmont 2:da51fb522205 915 x1 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 916
emilmont 2:da51fb522205 917 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 918
emilmont 1:fdd22bb7aa52 919 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 920 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 921
emilmont 1:fdd22bb7aa52 922 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 923 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 924 do
emilmont 1:fdd22bb7aa52 925 {
emilmont 1:fdd22bb7aa52 926 /* Read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 927 * y[srcBLen - 1] and y[srcBLen - 2] */
emilmont 2:da51fb522205 928 a = *py;
emilmont 2:da51fb522205 929 b = *(py+1);
emilmont 2:da51fb522205 930 py -= 2;
emilmont 1:fdd22bb7aa52 931
emilmont 1:fdd22bb7aa52 932 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 933
emilmont 2:da51fb522205 934 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 935
emilmont 1:fdd22bb7aa52 936 #else
emilmont 1:fdd22bb7aa52 937
emilmont 2:da51fb522205 938 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 939
emilmont 2:da51fb522205 940 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 941
emilmont 1:fdd22bb7aa52 942 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 943 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 944
emilmont 1:fdd22bb7aa52 945 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 946 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 947
emilmont 2:da51fb522205 948 a = *px;
emilmont 2:da51fb522205 949 b = *(px + 1);
emilmont 1:fdd22bb7aa52 950
emilmont 1:fdd22bb7aa52 951 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 952
emilmont 2:da51fb522205 953 x2 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 954 a = *(px + 2);
emilmont 2:da51fb522205 955 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 956
emilmont 1:fdd22bb7aa52 957 #else
emilmont 1:fdd22bb7aa52 958
emilmont 2:da51fb522205 959 x2 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 960 a = *(px + 2);
emilmont 2:da51fb522205 961 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 962
emilmont 2:da51fb522205 963 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 964
emilmont 1:fdd22bb7aa52 965 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 966 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 967
emilmont 1:fdd22bb7aa52 968 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 969 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 970
emilmont 1:fdd22bb7aa52 971 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emilmont 2:da51fb522205 972 a = *py;
emilmont 2:da51fb522205 973 b = *(py+1);
emilmont 2:da51fb522205 974 py -= 2;
emilmont 1:fdd22bb7aa52 975
emilmont 1:fdd22bb7aa52 976 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 977
emilmont 2:da51fb522205 978 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 979
emilmont 1:fdd22bb7aa52 980 #else
emilmont 1:fdd22bb7aa52 981
emilmont 2:da51fb522205 982 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 983
emilmont 2:da51fb522205 984 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 985
emilmont 1:fdd22bb7aa52 986 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 987 acc0 = __SMLADX(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 988
emilmont 1:fdd22bb7aa52 989 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 990 acc1 = __SMLADX(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 991
emilmont 1:fdd22bb7aa52 992 /* Read x[4], x[5], x[6] */
emilmont 2:da51fb522205 993 a = *(px + 2);
emilmont 2:da51fb522205 994 b = *(px + 3);
emilmont 1:fdd22bb7aa52 995
emilmont 1:fdd22bb7aa52 996 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 997
emilmont 2:da51fb522205 998 x0 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 999 a = *(px + 4);
emilmont 2:da51fb522205 1000 x1 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1001
emilmont 1:fdd22bb7aa52 1002 #else
emilmont 1:fdd22bb7aa52 1003
emilmont 2:da51fb522205 1004 x0 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 1005 a = *(px + 4);
emilmont 2:da51fb522205 1006 x1 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1007
emilmont 2:da51fb522205 1008 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1009
emilmont 2:da51fb522205 1010 px += 4u;
emilmont 1:fdd22bb7aa52 1011
emilmont 1:fdd22bb7aa52 1012 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1013 acc2 = __SMLADX(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 1014
emilmont 1:fdd22bb7aa52 1015 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1016 acc3 = __SMLADX(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 1017
emilmont 1:fdd22bb7aa52 1018 } while(--k);
emilmont 1:fdd22bb7aa52 1019
emilmont 1:fdd22bb7aa52 1020 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 1021 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 1022
emilmont 1:fdd22bb7aa52 1023 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1024 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1025 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 1026
emilmont 1:fdd22bb7aa52 1027 if(k == 1u)
emilmont 1:fdd22bb7aa52 1028 {
emilmont 1:fdd22bb7aa52 1029 /* Read y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 1030 c0 = *(py+1);
emilmont 1:fdd22bb7aa52 1031
emilmont 1:fdd22bb7aa52 1032 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1033
emilmont 1:fdd22bb7aa52 1034 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 1035
emilmont 1:fdd22bb7aa52 1036 #else
emilmont 1:fdd22bb7aa52 1037
emilmont 1:fdd22bb7aa52 1038 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 1039
emilmont 1:fdd22bb7aa52 1040 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1041
emilmont 1:fdd22bb7aa52 1042 /* Read x[7] */
emilmont 2:da51fb522205 1043 a = *px;
emilmont 2:da51fb522205 1044 b = *(px+1);
emilmont 2:da51fb522205 1045 px++;
emilmont 1:fdd22bb7aa52 1046
emilmont 1:fdd22bb7aa52 1047 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1048
emilmont 2:da51fb522205 1049 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1050
emilmont 1:fdd22bb7aa52 1051 #else
emilmont 1:fdd22bb7aa52 1052
emilmont 2:da51fb522205 1053 x3 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1054
emilmont 2:da51fb522205 1055 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1056
emilmont 1:fdd22bb7aa52 1057
emilmont 1:fdd22bb7aa52 1058 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1059 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1060 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1061 acc2 = __SMLADX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 1062 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1063 }
emilmont 1:fdd22bb7aa52 1064
emilmont 1:fdd22bb7aa52 1065 if(k == 2u)
emilmont 1:fdd22bb7aa52 1066 {
emilmont 1:fdd22bb7aa52 1067 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 2:da51fb522205 1068 a = *py;
emilmont 2:da51fb522205 1069 b = *(py+1);
emilmont 1:fdd22bb7aa52 1070
emilmont 1:fdd22bb7aa52 1071 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1072
emilmont 2:da51fb522205 1073 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1074
emilmont 1:fdd22bb7aa52 1075 #else
emilmont 1:fdd22bb7aa52 1076
emilmont 2:da51fb522205 1077 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1078
emilmont 2:da51fb522205 1079 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1080
emilmont 1:fdd22bb7aa52 1081 /* Read x[7], x[8], x[9] */
emilmont 2:da51fb522205 1082 a = *px;
emilmont 2:da51fb522205 1083 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1084
emilmont 1:fdd22bb7aa52 1085 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 1086
emilmont 2:da51fb522205 1087 x3 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 1088 a = *(px + 2);
emilmont 2:da51fb522205 1089 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1090
emilmont 1:fdd22bb7aa52 1091 #else
emilmont 1:fdd22bb7aa52 1092
emilmont 2:da51fb522205 1093 x3 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 1094 a = *(px + 2);
emilmont 2:da51fb522205 1095 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1096
emilmont 2:da51fb522205 1097 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 2:da51fb522205 1098 px += 2u;
emilmont 1:fdd22bb7aa52 1099
emilmont 1:fdd22bb7aa52 1100 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1101 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1102 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1103 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 1104 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 1105 }
emilmont 1:fdd22bb7aa52 1106
emilmont 1:fdd22bb7aa52 1107 if(k == 3u)
emilmont 1:fdd22bb7aa52 1108 {
emilmont 1:fdd22bb7aa52 1109 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 2:da51fb522205 1110 a = *py;
emilmont 2:da51fb522205 1111 b = *(py+1);
emilmont 1:fdd22bb7aa52 1112
emilmont 1:fdd22bb7aa52 1113 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1114
emilmont 2:da51fb522205 1115 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1116
emilmont 1:fdd22bb7aa52 1117 #else
emilmont 1:fdd22bb7aa52 1118
emilmont 2:da51fb522205 1119 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1120
emilmont 2:da51fb522205 1121 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1122
emilmont 1:fdd22bb7aa52 1123 /* Read x[7], x[8], x[9] */
emilmont 2:da51fb522205 1124 a = *px;
emilmont 2:da51fb522205 1125 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1126
emilmont 1:fdd22bb7aa52 1127 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 1128
emilmont 2:da51fb522205 1129 x3 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 1130 a = *(px + 2);
emilmont 2:da51fb522205 1131 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1132
emilmont 1:fdd22bb7aa52 1133 #else
emilmont 1:fdd22bb7aa52 1134
emilmont 2:da51fb522205 1135 x3 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 1136 a = *(px + 2);
emilmont 2:da51fb522205 1137 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1138
emilmont 2:da51fb522205 1139 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1140
emilmont 1:fdd22bb7aa52 1141 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1142 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1143 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1144 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 1145 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 1146
emilmont 1:fdd22bb7aa52 1147 /* Read y[srcBLen - 7] */
emilmont 2:da51fb522205 1148 c0 = *(py-1);
emilmont 1:fdd22bb7aa52 1149 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1150
emilmont 1:fdd22bb7aa52 1151 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 1152 #else
emilmont 1:fdd22bb7aa52 1153
emilmont 1:fdd22bb7aa52 1154 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 1155 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1156
emilmont 1:fdd22bb7aa52 1157 /* Read x[10] */
emilmont 2:da51fb522205 1158 a = *(px+2);
emilmont 2:da51fb522205 1159 b = *(px+3);
emilmont 1:fdd22bb7aa52 1160
emilmont 1:fdd22bb7aa52 1161 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1162
emilmont 2:da51fb522205 1163 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1164
emilmont 1:fdd22bb7aa52 1165 #else
emilmont 1:fdd22bb7aa52 1166
emilmont 2:da51fb522205 1167 x3 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1168
emilmont 2:da51fb522205 1169 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1170
emilmont 2:da51fb522205 1171 px += 3u;
emilmont 1:fdd22bb7aa52 1172
emilmont 1:fdd22bb7aa52 1173 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1174 acc0 = __SMLADX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 1175 acc1 = __SMLAD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 1176 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 1177 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1178 }
emilmont 1:fdd22bb7aa52 1179
emilmont 1:fdd22bb7aa52 1180 /* Store the results in the accumulators in the destination buffer. */
emilmont 2:da51fb522205 1181 *pOut++ = (q15_t)(acc0 >> 15);
emilmont 2:da51fb522205 1182 *pOut++ = (q15_t)(acc1 >> 15);
emilmont 2:da51fb522205 1183 *pOut++ = (q15_t)(acc2 >> 15);
emilmont 2:da51fb522205 1184 *pOut++ = (q15_t)(acc3 >> 15);
emilmont 1:fdd22bb7aa52 1185
emilmont 1:fdd22bb7aa52 1186 /* Increment the pointer pIn1 index, count by 4 */
emilmont 1:fdd22bb7aa52 1187 count += 4u;
emilmont 1:fdd22bb7aa52 1188
emilmont 1:fdd22bb7aa52 1189 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1190 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1191 py = pSrc2;
emilmont 1:fdd22bb7aa52 1192
emilmont 1:fdd22bb7aa52 1193 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1194 blkCnt--;
emilmont 1:fdd22bb7aa52 1195 }
emilmont 1:fdd22bb7aa52 1196
emilmont 1:fdd22bb7aa52 1197 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 1198 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1199 blkCnt = blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 1200
emilmont 1:fdd22bb7aa52 1201 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 1202 {
emilmont 1:fdd22bb7aa52 1203 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1204 sum = 0;
emilmont 1:fdd22bb7aa52 1205
emilmont 1:fdd22bb7aa52 1206 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1207 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 1208
emilmont 1:fdd22bb7aa52 1209 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 1210 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 1211 while(k > 0u)
emilmont 1:fdd22bb7aa52 1212 {
emilmont 1:fdd22bb7aa52 1213 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1214 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1215 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1216 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1217 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1218
emilmont 1:fdd22bb7aa52 1219 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1220 k--;
emilmont 1:fdd22bb7aa52 1221 }
emilmont 1:fdd22bb7aa52 1222
emilmont 1:fdd22bb7aa52 1223 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1224 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1225 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 1226
emilmont 1:fdd22bb7aa52 1227 while(k > 0u)
emilmont 1:fdd22bb7aa52 1228 {
emilmont 1:fdd22bb7aa52 1229 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1230 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1231
emilmont 1:fdd22bb7aa52 1232 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1233 k--;
emilmont 1:fdd22bb7aa52 1234 }
emilmont 1:fdd22bb7aa52 1235
emilmont 1:fdd22bb7aa52 1236 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1237 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1238
emilmont 1:fdd22bb7aa52 1239 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 1240 count++;
emilmont 1:fdd22bb7aa52 1241
emilmont 1:fdd22bb7aa52 1242 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1243 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1244 py = pSrc2;
emilmont 1:fdd22bb7aa52 1245
emilmont 1:fdd22bb7aa52 1246 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1247 blkCnt--;
emilmont 1:fdd22bb7aa52 1248 }
emilmont 1:fdd22bb7aa52 1249 }
emilmont 1:fdd22bb7aa52 1250 else
emilmont 1:fdd22bb7aa52 1251 {
emilmont 1:fdd22bb7aa52 1252 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 1253 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 1254 blkCnt = blockSize2;
emilmont 1:fdd22bb7aa52 1255
emilmont 1:fdd22bb7aa52 1256 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 1257 {
emilmont 1:fdd22bb7aa52 1258 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1259 sum = 0;
emilmont 1:fdd22bb7aa52 1260
emilmont 1:fdd22bb7aa52 1261 /* srcBLen number of MACS should be performed */
emilmont 1:fdd22bb7aa52 1262 k = srcBLen;
emilmont 1:fdd22bb7aa52 1263
emilmont 1:fdd22bb7aa52 1264 while(k > 0u)
emilmont 1:fdd22bb7aa52 1265 {
emilmont 1:fdd22bb7aa52 1266 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 1267 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1268
emilmont 1:fdd22bb7aa52 1269 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1270 k--;
emilmont 1:fdd22bb7aa52 1271 }
emilmont 1:fdd22bb7aa52 1272
emilmont 1:fdd22bb7aa52 1273 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1274 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1275
emilmont 1:fdd22bb7aa52 1276 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 1277 count++;
emilmont 1:fdd22bb7aa52 1278
emilmont 1:fdd22bb7aa52 1279 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1280 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1281 py = pSrc2;
emilmont 1:fdd22bb7aa52 1282
emilmont 1:fdd22bb7aa52 1283 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1284 blkCnt--;
emilmont 1:fdd22bb7aa52 1285 }
emilmont 1:fdd22bb7aa52 1286 }
emilmont 1:fdd22bb7aa52 1287
emilmont 1:fdd22bb7aa52 1288
emilmont 1:fdd22bb7aa52 1289 /* --------------------------
emilmont 1:fdd22bb7aa52 1290 * Initializations of stage3
emilmont 1:fdd22bb7aa52 1291 * -------------------------*/
emilmont 1:fdd22bb7aa52 1292
emilmont 1:fdd22bb7aa52 1293 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 1294 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emilmont 1:fdd22bb7aa52 1295 * ....
emilmont 1:fdd22bb7aa52 1296 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emilmont 1:fdd22bb7aa52 1297 * sum += x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 1298 */
emilmont 1:fdd22bb7aa52 1299
emilmont 1:fdd22bb7aa52 1300 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 1301 The blockSize3 variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 1302
emilmont 1:fdd22bb7aa52 1303 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 1304 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 1305 px = pSrc1;
emilmont 1:fdd22bb7aa52 1306
emilmont 1:fdd22bb7aa52 1307 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 1308 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 1309 pIn2 = pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 1310 py = pIn2;
emilmont 1:fdd22bb7aa52 1311
emilmont 1:fdd22bb7aa52 1312 /* -------------------
emilmont 1:fdd22bb7aa52 1313 * Stage3 process
emilmont 1:fdd22bb7aa52 1314 * ------------------*/
emilmont 1:fdd22bb7aa52 1315
emilmont 1:fdd22bb7aa52 1316 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 1317 /* First part of this stage computes the MAC operations greater than 4 */
emilmont 1:fdd22bb7aa52 1318 /* Second part of this stage computes the MAC operations less than or equal to 4 */
emilmont 1:fdd22bb7aa52 1319
emilmont 1:fdd22bb7aa52 1320 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 1321 j = blockSize3 >> 2u;
emilmont 1:fdd22bb7aa52 1322
emilmont 1:fdd22bb7aa52 1323 while((j > 0u) && (blockSize3 > 0u))
emilmont 1:fdd22bb7aa52 1324 {
emilmont 1:fdd22bb7aa52 1325 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1326 sum = 0;
emilmont 1:fdd22bb7aa52 1327
emilmont 1:fdd22bb7aa52 1328 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1329 k = blockSize3 >> 2u;
emilmont 1:fdd22bb7aa52 1330
emilmont 1:fdd22bb7aa52 1331 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 1332 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 2:da51fb522205 1333 py++;
emilmont 1:fdd22bb7aa52 1334
emilmont 1:fdd22bb7aa52 1335 while(k > 0u)
emilmont 2:da51fb522205 1336 {
emilmont 1:fdd22bb7aa52 1337 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1338 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1339 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1340 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1341 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1342 k--;
emilmont 1:fdd22bb7aa52 1343 }
emilmont 1:fdd22bb7aa52 1344
emilmont 1:fdd22bb7aa52 1345 /* If the blockSize3 is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1346 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1347 k = blockSize3 % 0x4u;
emilmont 1:fdd22bb7aa52 1348
emilmont 1:fdd22bb7aa52 1349 while(k > 0u)
emilmont 1:fdd22bb7aa52 1350 {
emilmont 1:fdd22bb7aa52 1351 /* sum += x[srcALen - srcBLen + 5] * y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 1352 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1353
emilmont 1:fdd22bb7aa52 1354 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1355 k--;
emilmont 1:fdd22bb7aa52 1356 }
emilmont 1:fdd22bb7aa52 1357
emilmont 1:fdd22bb7aa52 1358 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1359 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1360
emilmont 1:fdd22bb7aa52 1361 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1362 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 1363 py = pIn2;
emilmont 1:fdd22bb7aa52 1364
emilmont 1:fdd22bb7aa52 1365 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1366 blockSize3--;
emilmont 1:fdd22bb7aa52 1367
emilmont 1:fdd22bb7aa52 1368 j--;
emilmont 1:fdd22bb7aa52 1369 }
emilmont 1:fdd22bb7aa52 1370
emilmont 1:fdd22bb7aa52 1371 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 1372 /* SIMD is not used for the next MAC operations,
emilmont 1:fdd22bb7aa52 1373 * so pointer py is updated to read only one sample at a time */
emilmont 1:fdd22bb7aa52 1374 py = py + 1u;
emilmont 1:fdd22bb7aa52 1375
emilmont 1:fdd22bb7aa52 1376 while(blockSize3 > 0u)
emilmont 1:fdd22bb7aa52 1377 {
emilmont 1:fdd22bb7aa52 1378 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1379 sum = 0;
emilmont 1:fdd22bb7aa52 1380
emilmont 1:fdd22bb7aa52 1381 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1382 k = blockSize3;
emilmont 1:fdd22bb7aa52 1383
emilmont 1:fdd22bb7aa52 1384 while(k > 0u)
emilmont 1:fdd22bb7aa52 1385 {
emilmont 1:fdd22bb7aa52 1386 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1387 /* sum += x[srcALen-1] * y[srcBLen-1] */
emilmont 1:fdd22bb7aa52 1388 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1389
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 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1395 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1396
emilmont 1:fdd22bb7aa52 1397 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1398 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 1399 py = pSrc2;
emilmont 1:fdd22bb7aa52 1400
emilmont 1:fdd22bb7aa52 1401 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1402 blockSize3--;
emilmont 1:fdd22bb7aa52 1403 }
emilmont 1:fdd22bb7aa52 1404
emilmont 2:da51fb522205 1405 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
emilmont 1:fdd22bb7aa52 1406 }
emilmont 1:fdd22bb7aa52 1407
emilmont 1:fdd22bb7aa52 1408 /**
emilmont 1:fdd22bb7aa52 1409 * @} end of Conv group
emilmont 1:fdd22bb7aa52 1410 */