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_partial_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_partial_fast_q15.c
emilmont 1:fdd22bb7aa52 9 *
emilmont 2:da51fb522205 10 * Description: Fast Q15 Partial 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 PartialConv
emilmont 1:fdd22bb7aa52 49 * @{
emilmont 1:fdd22bb7aa52 50 */
emilmont 1:fdd22bb7aa52 51
emilmont 1:fdd22bb7aa52 52 /**
emilmont 1:fdd22bb7aa52 53 * @brief Partial 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.
emilmont 1:fdd22bb7aa52 59 * @param[in] firstIndex is the first output sample to start with.
emilmont 1:fdd22bb7aa52 60 * @param[in] numPoints is the number of output points to be computed.
emilmont 1:fdd22bb7aa52 61 * @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 62 *
emilmont 1:fdd22bb7aa52 63 * 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 64 */
emilmont 1:fdd22bb7aa52 65
emilmont 1:fdd22bb7aa52 66
emilmont 1:fdd22bb7aa52 67 arm_status arm_conv_partial_fast_q15(
emilmont 1:fdd22bb7aa52 68 q15_t * pSrcA,
emilmont 1:fdd22bb7aa52 69 uint32_t srcALen,
emilmont 1:fdd22bb7aa52 70 q15_t * pSrcB,
emilmont 1:fdd22bb7aa52 71 uint32_t srcBLen,
emilmont 1:fdd22bb7aa52 72 q15_t * pDst,
emilmont 1:fdd22bb7aa52 73 uint32_t firstIndex,
emilmont 1:fdd22bb7aa52 74 uint32_t numPoints)
emilmont 1:fdd22bb7aa52 75 {
emilmont 1:fdd22bb7aa52 76 #ifndef UNALIGNED_SUPPORT_DISABLE
emilmont 1:fdd22bb7aa52 77
emilmont 1:fdd22bb7aa52 78 q15_t *pIn1; /* inputA pointer */
emilmont 1:fdd22bb7aa52 79 q15_t *pIn2; /* inputB pointer */
emilmont 1:fdd22bb7aa52 80 q15_t *pOut = pDst; /* output pointer */
emilmont 1:fdd22bb7aa52 81 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
emilmont 1:fdd22bb7aa52 82 q15_t *px; /* Intermediate inputA pointer */
emilmont 1:fdd22bb7aa52 83 q15_t *py; /* Intermediate inputB pointer */
emilmont 1:fdd22bb7aa52 84 q15_t *pSrc1, *pSrc2; /* Intermediate pointers */
emilmont 1:fdd22bb7aa52 85 q31_t x0, x1, x2, x3, c0;
emilmont 1:fdd22bb7aa52 86 uint32_t j, k, count, check, blkCnt;
emilmont 1:fdd22bb7aa52 87 int32_t blockSize1, blockSize2, blockSize3; /* loop counters */
emilmont 1:fdd22bb7aa52 88 arm_status status; /* status of Partial convolution */
emilmont 1:fdd22bb7aa52 89
emilmont 1:fdd22bb7aa52 90 /* Check for range of output samples to be calculated */
emilmont 1:fdd22bb7aa52 91 if((firstIndex + numPoints) > ((srcALen + (srcBLen - 1u))))
emilmont 1:fdd22bb7aa52 92 {
emilmont 1:fdd22bb7aa52 93 /* Set status as ARM_MATH_ARGUMENT_ERROR */
emilmont 1:fdd22bb7aa52 94 status = ARM_MATH_ARGUMENT_ERROR;
emilmont 1:fdd22bb7aa52 95 }
emilmont 1:fdd22bb7aa52 96 else
emilmont 1:fdd22bb7aa52 97 {
emilmont 1:fdd22bb7aa52 98
emilmont 1:fdd22bb7aa52 99 /* The algorithm implementation is based on the lengths of the inputs. */
emilmont 1:fdd22bb7aa52 100 /* srcB is always made to slide across srcA. */
emilmont 1:fdd22bb7aa52 101 /* So srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 102 if(srcALen >=srcBLen)
emilmont 1:fdd22bb7aa52 103 {
emilmont 1:fdd22bb7aa52 104 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 105 pIn1 = pSrcA;
emilmont 1:fdd22bb7aa52 106
emilmont 1:fdd22bb7aa52 107 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 108 pIn2 = pSrcB;
emilmont 1:fdd22bb7aa52 109 }
emilmont 1:fdd22bb7aa52 110 else
emilmont 1:fdd22bb7aa52 111 {
emilmont 1:fdd22bb7aa52 112 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 113 pIn1 = pSrcB;
emilmont 1:fdd22bb7aa52 114
emilmont 1:fdd22bb7aa52 115 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 116 pIn2 = pSrcA;
emilmont 1:fdd22bb7aa52 117
emilmont 1:fdd22bb7aa52 118 /* srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 119 j = srcBLen;
emilmont 1:fdd22bb7aa52 120 srcBLen = srcALen;
emilmont 1:fdd22bb7aa52 121 srcALen = j;
emilmont 1:fdd22bb7aa52 122 }
emilmont 1:fdd22bb7aa52 123
emilmont 1:fdd22bb7aa52 124 /* Conditions to check which loopCounter holds
emilmont 1:fdd22bb7aa52 125 * the first and last indices of the output samples to be calculated. */
emilmont 1:fdd22bb7aa52 126 check = firstIndex + numPoints;
emilmont 1:fdd22bb7aa52 127 blockSize3 = ((int32_t) check - (int32_t) srcALen);
emilmont 1:fdd22bb7aa52 128 blockSize3 = (blockSize3 > 0) ? blockSize3 : 0;
emilmont 1:fdd22bb7aa52 129 blockSize1 = (((int32_t) srcBLen - 1) - (int32_t) firstIndex);
emilmont 1:fdd22bb7aa52 130 blockSize1 = (blockSize1 > 0) ? ((check > (srcBLen - 1u)) ? blockSize1 :
emilmont 1:fdd22bb7aa52 131 (int32_t) numPoints) : 0;
emilmont 1:fdd22bb7aa52 132 blockSize2 = (int32_t) check - ((blockSize3 + blockSize1) +
emilmont 1:fdd22bb7aa52 133 (int32_t) firstIndex);
emilmont 1:fdd22bb7aa52 134 blockSize2 = (blockSize2 > 0) ? blockSize2 : 0;
emilmont 1:fdd22bb7aa52 135
emilmont 1:fdd22bb7aa52 136 /* 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 137 /* The function is internally
emilmont 1:fdd22bb7aa52 138 * divided into three stages according to the number of multiplications that has to be
emilmont 1:fdd22bb7aa52 139 * taken place between inputA samples and inputB samples. In the first stage of the
emilmont 1:fdd22bb7aa52 140 * algorithm, the multiplications increase by one for every iteration.
emilmont 1:fdd22bb7aa52 141 * In the second stage of the algorithm, srcBLen number of multiplications are done.
emilmont 1:fdd22bb7aa52 142 * In the third stage of the algorithm, the multiplications decrease by one
emilmont 1:fdd22bb7aa52 143 * for every iteration. */
emilmont 1:fdd22bb7aa52 144
emilmont 1:fdd22bb7aa52 145 /* Set the output pointer to point to the firstIndex
emilmont 1:fdd22bb7aa52 146 * of the output sample to be calculated. */
emilmont 1:fdd22bb7aa52 147 pOut = pDst + firstIndex;
emilmont 1:fdd22bb7aa52 148
emilmont 1:fdd22bb7aa52 149 /* --------------------------
emilmont 1:fdd22bb7aa52 150 * Initializations of stage1
emilmont 1:fdd22bb7aa52 151 * -------------------------*/
emilmont 1:fdd22bb7aa52 152
emilmont 1:fdd22bb7aa52 153 /* sum = x[0] * y[0]
emilmont 1:fdd22bb7aa52 154 * sum = x[0] * y[1] + x[1] * y[0]
emilmont 1:fdd22bb7aa52 155 * ....
emilmont 1:fdd22bb7aa52 156 * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
emilmont 1:fdd22bb7aa52 157 */
emilmont 1:fdd22bb7aa52 158
emilmont 1:fdd22bb7aa52 159 /* In this stage the MAC operations are increased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 160 The count variable holds the number of MAC operations performed.
emilmont 1:fdd22bb7aa52 161 Since the partial convolution starts from firstIndex
emilmont 1:fdd22bb7aa52 162 Number of Macs to be performed is firstIndex + 1 */
emilmont 1:fdd22bb7aa52 163 count = 1u + firstIndex;
emilmont 1:fdd22bb7aa52 164
emilmont 1:fdd22bb7aa52 165 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 166 px = pIn1;
emilmont 1:fdd22bb7aa52 167
emilmont 1:fdd22bb7aa52 168 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 169 pSrc2 = pIn2 + firstIndex;
emilmont 1:fdd22bb7aa52 170 py = pSrc2;
emilmont 1:fdd22bb7aa52 171
emilmont 1:fdd22bb7aa52 172 /* ------------------------
emilmont 1:fdd22bb7aa52 173 * Stage1 process
emilmont 1:fdd22bb7aa52 174 * ----------------------*/
emilmont 1:fdd22bb7aa52 175
emilmont 1:fdd22bb7aa52 176 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 177 /* First part of this stage computes the MAC operations less than 4 */
emilmont 1:fdd22bb7aa52 178 /* Second part of this stage computes the MAC operations greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 179
emilmont 1:fdd22bb7aa52 180 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 181 while((count < 4u) && (blockSize1 > 0))
emilmont 1:fdd22bb7aa52 182 {
emilmont 1:fdd22bb7aa52 183 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 184 sum = 0;
emilmont 1:fdd22bb7aa52 185
emilmont 1:fdd22bb7aa52 186 /* Loop over number of MAC operations between
emilmont 1:fdd22bb7aa52 187 * inputA samples and inputB samples */
emilmont 1:fdd22bb7aa52 188 k = count;
emilmont 1:fdd22bb7aa52 189
emilmont 1:fdd22bb7aa52 190 while(k > 0u)
emilmont 1:fdd22bb7aa52 191 {
emilmont 1:fdd22bb7aa52 192 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 193 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 194
emilmont 1:fdd22bb7aa52 195 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 196 k--;
emilmont 1:fdd22bb7aa52 197 }
emilmont 1:fdd22bb7aa52 198
emilmont 1:fdd22bb7aa52 199 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 200 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 201
emilmont 1:fdd22bb7aa52 202 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 203 py = ++pSrc2;
emilmont 1:fdd22bb7aa52 204 px = pIn1;
emilmont 1:fdd22bb7aa52 205
emilmont 1:fdd22bb7aa52 206 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 207 count++;
emilmont 1:fdd22bb7aa52 208
emilmont 1:fdd22bb7aa52 209 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 210 blockSize1--;
emilmont 1:fdd22bb7aa52 211 }
emilmont 1:fdd22bb7aa52 212
emilmont 1:fdd22bb7aa52 213 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 214 /* The internal loop, over count, is unrolled by 4 */
emilmont 1:fdd22bb7aa52 215 /* To, read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 216 * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */
emilmont 1:fdd22bb7aa52 217 py = py - 1;
emilmont 1:fdd22bb7aa52 218
emilmont 1:fdd22bb7aa52 219 while(blockSize1 > 0)
emilmont 1:fdd22bb7aa52 220 {
emilmont 1:fdd22bb7aa52 221 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 222 sum = 0;
emilmont 1:fdd22bb7aa52 223
emilmont 1:fdd22bb7aa52 224 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 225 k = count >> 2u;
emilmont 1:fdd22bb7aa52 226
emilmont 1:fdd22bb7aa52 227 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 228 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 229 while(k > 0u)
emilmont 1:fdd22bb7aa52 230 {
emilmont 1:fdd22bb7aa52 231 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 232 /* x[0], x[1] are multiplied with y[srcBLen - 1], y[srcBLen - 2] respectively */
emilmont 1:fdd22bb7aa52 233 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 234 /* x[2], x[3] are multiplied with y[srcBLen - 3], y[srcBLen - 4] respectively */
emilmont 1:fdd22bb7aa52 235 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 236
emilmont 1:fdd22bb7aa52 237 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 238 k--;
emilmont 1:fdd22bb7aa52 239 }
emilmont 1:fdd22bb7aa52 240
emilmont 1:fdd22bb7aa52 241 /* For the next MAC operations, the pointer py is used without SIMD
emilmont 1:fdd22bb7aa52 242 * So, py is incremented by 1 */
emilmont 1:fdd22bb7aa52 243 py = py + 1u;
emilmont 1:fdd22bb7aa52 244
emilmont 1:fdd22bb7aa52 245 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 246 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 247 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 248
emilmont 1:fdd22bb7aa52 249 while(k > 0u)
emilmont 1:fdd22bb7aa52 250 {
emilmont 1:fdd22bb7aa52 251 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 252 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 253
emilmont 1:fdd22bb7aa52 254 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 255 k--;
emilmont 1:fdd22bb7aa52 256 }
emilmont 1:fdd22bb7aa52 257
emilmont 1:fdd22bb7aa52 258 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 259 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 260
emilmont 1:fdd22bb7aa52 261 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 262 py = ++pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 263 px = pIn1;
emilmont 1:fdd22bb7aa52 264
emilmont 1:fdd22bb7aa52 265 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 266 count++;
emilmont 1:fdd22bb7aa52 267
emilmont 1:fdd22bb7aa52 268 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 269 blockSize1--;
emilmont 1:fdd22bb7aa52 270 }
emilmont 1:fdd22bb7aa52 271
emilmont 1:fdd22bb7aa52 272 /* --------------------------
emilmont 1:fdd22bb7aa52 273 * Initializations of stage2
emilmont 1:fdd22bb7aa52 274 * ------------------------*/
emilmont 1:fdd22bb7aa52 275
emilmont 1:fdd22bb7aa52 276 /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
emilmont 1:fdd22bb7aa52 277 * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
emilmont 1:fdd22bb7aa52 278 * ....
emilmont 1:fdd22bb7aa52 279 * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
emilmont 1:fdd22bb7aa52 280 */
emilmont 1:fdd22bb7aa52 281
emilmont 1:fdd22bb7aa52 282 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 283 px = pIn1;
emilmont 1:fdd22bb7aa52 284
emilmont 1:fdd22bb7aa52 285 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 286 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 287 py = pSrc2;
emilmont 1:fdd22bb7aa52 288
emilmont 1:fdd22bb7aa52 289 /* count is the index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 290 count = 0u;
emilmont 1:fdd22bb7aa52 291
emilmont 1:fdd22bb7aa52 292
emilmont 1:fdd22bb7aa52 293 /* --------------------
emilmont 1:fdd22bb7aa52 294 * Stage2 process
emilmont 1:fdd22bb7aa52 295 * -------------------*/
emilmont 1:fdd22bb7aa52 296
emilmont 1:fdd22bb7aa52 297 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 298 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 299 * srcBLen should be greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 300 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 301 {
emilmont 1:fdd22bb7aa52 302 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 303 blkCnt = ((uint32_t) blockSize2 >> 2u);
emilmont 1:fdd22bb7aa52 304
emilmont 1:fdd22bb7aa52 305 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 306 {
emilmont 1:fdd22bb7aa52 307 py = py - 1u;
emilmont 1:fdd22bb7aa52 308
emilmont 1:fdd22bb7aa52 309 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 310 acc0 = 0;
emilmont 1:fdd22bb7aa52 311 acc1 = 0;
emilmont 1:fdd22bb7aa52 312 acc2 = 0;
emilmont 1:fdd22bb7aa52 313 acc3 = 0;
emilmont 1:fdd22bb7aa52 314
emilmont 1:fdd22bb7aa52 315
emilmont 1:fdd22bb7aa52 316 /* read x[0], x[1] samples */
emilmont 1:fdd22bb7aa52 317 x0 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 318 /* read x[1], x[2] samples */
emilmont 1:fdd22bb7aa52 319 x1 = _SIMD32_OFFSET(px+1);
emilmont 2:da51fb522205 320 px+= 2u;
emilmont 1:fdd22bb7aa52 321
emilmont 1:fdd22bb7aa52 322
emilmont 1:fdd22bb7aa52 323 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 324 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 325
emilmont 1:fdd22bb7aa52 326 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 327 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 328 do
emilmont 1:fdd22bb7aa52 329 {
emilmont 1:fdd22bb7aa52 330 /* Read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 331 * y[srcBLen - 1] and y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 332 c0 = *__SIMD32(py)--;
emilmont 1:fdd22bb7aa52 333
emilmont 1:fdd22bb7aa52 334 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 335 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 336
emilmont 1:fdd22bb7aa52 337 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 338 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 339
emilmont 1:fdd22bb7aa52 340 /* Read x[2], x[3] */
emilmont 1:fdd22bb7aa52 341 x2 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 342
emilmont 1:fdd22bb7aa52 343 /* Read x[3], x[4] */
emilmont 1:fdd22bb7aa52 344 x3 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 345
emilmont 1:fdd22bb7aa52 346 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 347 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 348
emilmont 1:fdd22bb7aa52 349 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 350 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 351
emilmont 1:fdd22bb7aa52 352 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 353 c0 = *__SIMD32(py)--;
emilmont 1:fdd22bb7aa52 354
emilmont 1:fdd22bb7aa52 355 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 356 acc0 = __SMLADX(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 357
emilmont 1:fdd22bb7aa52 358 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 359 acc1 = __SMLADX(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 360
emilmont 1:fdd22bb7aa52 361 /* Read x[4], x[5] */
emilmont 1:fdd22bb7aa52 362 x0 = _SIMD32_OFFSET(px+2);
emilmont 1:fdd22bb7aa52 363
emilmont 1:fdd22bb7aa52 364 /* Read x[5], x[6] */
emilmont 1:fdd22bb7aa52 365 x1 = _SIMD32_OFFSET(px+3);
emilmont 2:da51fb522205 366 px += 4u;
emilmont 1:fdd22bb7aa52 367
emilmont 1:fdd22bb7aa52 368 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 369 acc2 = __SMLADX(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 370
emilmont 1:fdd22bb7aa52 371 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 372 acc3 = __SMLADX(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 373
emilmont 1:fdd22bb7aa52 374 } while(--k);
emilmont 1:fdd22bb7aa52 375
emilmont 1:fdd22bb7aa52 376 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 377 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 378
emilmont 1:fdd22bb7aa52 379 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 380 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 381 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 382
emilmont 1:fdd22bb7aa52 383 if(k == 1u)
emilmont 1:fdd22bb7aa52 384 {
emilmont 1:fdd22bb7aa52 385 /* Read y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 386 c0 = *(py+1);
emilmont 1:fdd22bb7aa52 387 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 388
emilmont 1:fdd22bb7aa52 389 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 390
emilmont 1:fdd22bb7aa52 391 #else
emilmont 1:fdd22bb7aa52 392
emilmont 1:fdd22bb7aa52 393 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 394
emilmont 1:fdd22bb7aa52 395 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 396
emilmont 1:fdd22bb7aa52 397 /* Read x[7] */
emilmont 1:fdd22bb7aa52 398 x3 = *__SIMD32(px);
emilmont 2:da51fb522205 399 px++;
emilmont 1:fdd22bb7aa52 400
emilmont 1:fdd22bb7aa52 401 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 402 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 403 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 404 acc2 = __SMLADX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 405 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 406 }
emilmont 1:fdd22bb7aa52 407
emilmont 1:fdd22bb7aa52 408 if(k == 2u)
emilmont 1:fdd22bb7aa52 409 {
emilmont 1:fdd22bb7aa52 410 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 411 c0 = _SIMD32_OFFSET(py);
emilmont 1:fdd22bb7aa52 412
emilmont 1:fdd22bb7aa52 413 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 414 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 415
emilmont 1:fdd22bb7aa52 416 /* Read x[9] */
emilmont 1:fdd22bb7aa52 417 x2 = _SIMD32_OFFSET(px+1);
emilmont 2:da51fb522205 418 px += 2u;
emilmont 1:fdd22bb7aa52 419
emilmont 1:fdd22bb7aa52 420 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 421 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 422 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 423 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 424 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 425 }
emilmont 1:fdd22bb7aa52 426
emilmont 1:fdd22bb7aa52 427 if(k == 3u)
emilmont 1:fdd22bb7aa52 428 {
emilmont 1:fdd22bb7aa52 429 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 430 c0 = _SIMD32_OFFSET(py);
emilmont 1:fdd22bb7aa52 431
emilmont 1:fdd22bb7aa52 432 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 433 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 434
emilmont 1:fdd22bb7aa52 435 /* Read x[9] */
emilmont 1:fdd22bb7aa52 436 x2 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 437
emilmont 1:fdd22bb7aa52 438 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 439 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 440 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 441 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 442 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 443
emilmont 2:da51fb522205 444 c0 = *(py-1);
emilmont 1:fdd22bb7aa52 445 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 446
emilmont 1:fdd22bb7aa52 447 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 448 #else
emilmont 1:fdd22bb7aa52 449
emilmont 1:fdd22bb7aa52 450 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 451 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 452
emilmont 1:fdd22bb7aa52 453 /* Read x[10] */
emilmont 1:fdd22bb7aa52 454 x3 = _SIMD32_OFFSET(px+2);
emilmont 2:da51fb522205 455 px += 3u;
emilmont 1:fdd22bb7aa52 456
emilmont 1:fdd22bb7aa52 457 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 458 acc0 = __SMLADX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 459 acc1 = __SMLAD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 460 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 461 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 462 }
emilmont 1:fdd22bb7aa52 463
emilmont 1:fdd22bb7aa52 464 /* Store the results in the accumulators in the destination buffer. */
emilmont 1:fdd22bb7aa52 465 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 466
emilmont 1:fdd22bb7aa52 467 *__SIMD32(pOut)++ = __PKHBT(acc0 >> 15, acc1 >> 15, 16);
emilmont 1:fdd22bb7aa52 468 *__SIMD32(pOut)++ = __PKHBT(acc2 >> 15, acc3 >> 15, 16);
emilmont 1:fdd22bb7aa52 469
emilmont 1:fdd22bb7aa52 470 #else
emilmont 1:fdd22bb7aa52 471
emilmont 1:fdd22bb7aa52 472 *__SIMD32(pOut)++ = __PKHBT(acc1 >> 15, acc0 >> 15, 16);
emilmont 1:fdd22bb7aa52 473 *__SIMD32(pOut)++ = __PKHBT(acc3 >> 15, acc2 >> 15, 16);
emilmont 1:fdd22bb7aa52 474
emilmont 1:fdd22bb7aa52 475 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 476
emilmont 1:fdd22bb7aa52 477 /* Increment the pointer pIn1 index, count by 4 */
emilmont 1:fdd22bb7aa52 478 count += 4u;
emilmont 1:fdd22bb7aa52 479
emilmont 1:fdd22bb7aa52 480 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 481 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 482 py = pSrc2;
emilmont 1:fdd22bb7aa52 483
emilmont 1:fdd22bb7aa52 484 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 485 blkCnt--;
emilmont 1:fdd22bb7aa52 486 }
emilmont 1:fdd22bb7aa52 487
emilmont 1:fdd22bb7aa52 488 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 489 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 490 blkCnt = (uint32_t) blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 491
emilmont 1:fdd22bb7aa52 492 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 493 {
emilmont 1:fdd22bb7aa52 494 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 495 sum = 0;
emilmont 1:fdd22bb7aa52 496
emilmont 1:fdd22bb7aa52 497 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 498 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 499
emilmont 1:fdd22bb7aa52 500 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 501 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 502 while(k > 0u)
emilmont 1:fdd22bb7aa52 503 {
emilmont 1:fdd22bb7aa52 504 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 505 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 506 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 507 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 508 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 509
emilmont 1:fdd22bb7aa52 510 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 511 k--;
emilmont 1:fdd22bb7aa52 512 }
emilmont 1:fdd22bb7aa52 513
emilmont 1:fdd22bb7aa52 514 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 515 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 516 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 517
emilmont 1:fdd22bb7aa52 518 while(k > 0u)
emilmont 1:fdd22bb7aa52 519 {
emilmont 1:fdd22bb7aa52 520 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 521 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 522
emilmont 1:fdd22bb7aa52 523 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 524 k--;
emilmont 1:fdd22bb7aa52 525 }
emilmont 1:fdd22bb7aa52 526
emilmont 1:fdd22bb7aa52 527 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 528 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 529
emilmont 1:fdd22bb7aa52 530 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 531 count++;
emilmont 1:fdd22bb7aa52 532
emilmont 1:fdd22bb7aa52 533 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 534 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 535 py = pSrc2;
emilmont 1:fdd22bb7aa52 536
emilmont 1:fdd22bb7aa52 537 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 538 blkCnt--;
emilmont 1:fdd22bb7aa52 539 }
emilmont 1:fdd22bb7aa52 540 }
emilmont 1:fdd22bb7aa52 541 else
emilmont 1:fdd22bb7aa52 542 {
emilmont 1:fdd22bb7aa52 543 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 544 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 545 blkCnt = (uint32_t) blockSize2;
emilmont 1:fdd22bb7aa52 546
emilmont 1:fdd22bb7aa52 547 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 548 {
emilmont 1:fdd22bb7aa52 549 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 550 sum = 0;
emilmont 1:fdd22bb7aa52 551
emilmont 1:fdd22bb7aa52 552 /* srcBLen number of MACS should be performed */
emilmont 1:fdd22bb7aa52 553 k = srcBLen;
emilmont 1:fdd22bb7aa52 554
emilmont 1:fdd22bb7aa52 555 while(k > 0u)
emilmont 1:fdd22bb7aa52 556 {
emilmont 1:fdd22bb7aa52 557 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 558 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 559
emilmont 1:fdd22bb7aa52 560 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 561 k--;
emilmont 1:fdd22bb7aa52 562 }
emilmont 1:fdd22bb7aa52 563
emilmont 1:fdd22bb7aa52 564 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 565 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 566
emilmont 1:fdd22bb7aa52 567 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 568 count++;
emilmont 1:fdd22bb7aa52 569
emilmont 1:fdd22bb7aa52 570 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 571 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 572 py = pSrc2;
emilmont 1:fdd22bb7aa52 573
emilmont 1:fdd22bb7aa52 574 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 575 blkCnt--;
emilmont 1:fdd22bb7aa52 576 }
emilmont 1:fdd22bb7aa52 577 }
emilmont 1:fdd22bb7aa52 578
emilmont 1:fdd22bb7aa52 579
emilmont 1:fdd22bb7aa52 580 /* --------------------------
emilmont 1:fdd22bb7aa52 581 * Initializations of stage3
emilmont 1:fdd22bb7aa52 582 * -------------------------*/
emilmont 1:fdd22bb7aa52 583
emilmont 1:fdd22bb7aa52 584 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 585 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emilmont 1:fdd22bb7aa52 586 * ....
emilmont 1:fdd22bb7aa52 587 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emilmont 1:fdd22bb7aa52 588 * sum += x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 589 */
emilmont 1:fdd22bb7aa52 590
emilmont 1:fdd22bb7aa52 591 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 592 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 593 count = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 594
emilmont 1:fdd22bb7aa52 595 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 596 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 597 px = pSrc1;
emilmont 1:fdd22bb7aa52 598
emilmont 1:fdd22bb7aa52 599 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 600 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 601 pIn2 = pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 602 py = pIn2;
emilmont 1:fdd22bb7aa52 603
emilmont 1:fdd22bb7aa52 604 /* -------------------
emilmont 1:fdd22bb7aa52 605 * Stage3 process
emilmont 1:fdd22bb7aa52 606 * ------------------*/
emilmont 1:fdd22bb7aa52 607
emilmont 1:fdd22bb7aa52 608 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 609 /* First part of this stage computes the MAC operations greater than 4 */
emilmont 1:fdd22bb7aa52 610 /* Second part of this stage computes the MAC operations less than or equal to 4 */
emilmont 1:fdd22bb7aa52 611
emilmont 1:fdd22bb7aa52 612 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 613 j = count >> 2u;
emilmont 1:fdd22bb7aa52 614
emilmont 1:fdd22bb7aa52 615 while((j > 0u) && (blockSize3 > 0))
emilmont 1:fdd22bb7aa52 616 {
emilmont 1:fdd22bb7aa52 617 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 618 sum = 0;
emilmont 1:fdd22bb7aa52 619
emilmont 1:fdd22bb7aa52 620 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 621 k = count >> 2u;
emilmont 1:fdd22bb7aa52 622
emilmont 1:fdd22bb7aa52 623 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 624 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 625 while(k > 0u)
emilmont 1:fdd22bb7aa52 626 {
emilmont 1:fdd22bb7aa52 627 /* x[srcALen - srcBLen + 1], x[srcALen - srcBLen + 2] are multiplied
emilmont 1:fdd22bb7aa52 628 * with y[srcBLen - 1], y[srcBLen - 2] respectively */
emilmont 1:fdd22bb7aa52 629 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 630 /* x[srcALen - srcBLen + 3], x[srcALen - srcBLen + 4] are multiplied
emilmont 1:fdd22bb7aa52 631 * with y[srcBLen - 3], y[srcBLen - 4] respectively */
emilmont 1:fdd22bb7aa52 632 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 633
emilmont 1:fdd22bb7aa52 634 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 635 k--;
emilmont 1:fdd22bb7aa52 636 }
emilmont 1:fdd22bb7aa52 637
emilmont 1:fdd22bb7aa52 638 /* For the next MAC operations, the pointer py is used without SIMD
emilmont 1:fdd22bb7aa52 639 * So, py is incremented by 1 */
emilmont 1:fdd22bb7aa52 640 py = py + 1u;
emilmont 1:fdd22bb7aa52 641
emilmont 1:fdd22bb7aa52 642 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 643 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 644 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 645
emilmont 1:fdd22bb7aa52 646 while(k > 0u)
emilmont 1:fdd22bb7aa52 647 {
emilmont 1:fdd22bb7aa52 648 /* sum += x[srcALen - srcBLen + 5] * y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 649 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 650
emilmont 1:fdd22bb7aa52 651 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 652 k--;
emilmont 1:fdd22bb7aa52 653 }
emilmont 1:fdd22bb7aa52 654
emilmont 1:fdd22bb7aa52 655 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 656 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 657
emilmont 1:fdd22bb7aa52 658 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 659 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 660 py = pIn2;
emilmont 1:fdd22bb7aa52 661
emilmont 1:fdd22bb7aa52 662 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 663 count--;
emilmont 1:fdd22bb7aa52 664
emilmont 1:fdd22bb7aa52 665 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 666 blockSize3--;
emilmont 1:fdd22bb7aa52 667
emilmont 1:fdd22bb7aa52 668 j--;
emilmont 1:fdd22bb7aa52 669 }
emilmont 1:fdd22bb7aa52 670
emilmont 1:fdd22bb7aa52 671 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 672 /* SIMD is not used for the next MAC operations,
emilmont 1:fdd22bb7aa52 673 * so pointer py is updated to read only one sample at a time */
emilmont 1:fdd22bb7aa52 674 py = py + 1u;
emilmont 1:fdd22bb7aa52 675
emilmont 1:fdd22bb7aa52 676 while(blockSize3 > 0)
emilmont 1:fdd22bb7aa52 677 {
emilmont 1:fdd22bb7aa52 678 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 679 sum = 0;
emilmont 1:fdd22bb7aa52 680
emilmont 1:fdd22bb7aa52 681 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 682 k = count;
emilmont 1:fdd22bb7aa52 683
emilmont 1:fdd22bb7aa52 684 while(k > 0u)
emilmont 1:fdd22bb7aa52 685 {
emilmont 1:fdd22bb7aa52 686 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 687 /* sum += x[srcALen-1] * y[srcBLen-1] */
emilmont 1:fdd22bb7aa52 688 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 689
emilmont 1:fdd22bb7aa52 690 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 691 k--;
emilmont 1:fdd22bb7aa52 692 }
emilmont 1:fdd22bb7aa52 693
emilmont 1:fdd22bb7aa52 694 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 695 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 696
emilmont 1:fdd22bb7aa52 697 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 698 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 699 py = pSrc2;
emilmont 1:fdd22bb7aa52 700
emilmont 1:fdd22bb7aa52 701 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 702 count--;
emilmont 1:fdd22bb7aa52 703
emilmont 1:fdd22bb7aa52 704 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 705 blockSize3--;
emilmont 1:fdd22bb7aa52 706 }
emilmont 1:fdd22bb7aa52 707
emilmont 1:fdd22bb7aa52 708 /* set status as ARM_MATH_SUCCESS */
emilmont 1:fdd22bb7aa52 709 status = ARM_MATH_SUCCESS;
emilmont 1:fdd22bb7aa52 710 }
emilmont 1:fdd22bb7aa52 711
emilmont 1:fdd22bb7aa52 712 /* Return to application */
emilmont 1:fdd22bb7aa52 713 return (status);
emilmont 1:fdd22bb7aa52 714
emilmont 1:fdd22bb7aa52 715 #else
emilmont 1:fdd22bb7aa52 716
emilmont 1:fdd22bb7aa52 717 q15_t *pIn1; /* inputA pointer */
emilmont 1:fdd22bb7aa52 718 q15_t *pIn2; /* inputB pointer */
emilmont 1:fdd22bb7aa52 719 q15_t *pOut = pDst; /* output pointer */
emilmont 1:fdd22bb7aa52 720 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
emilmont 1:fdd22bb7aa52 721 q15_t *px; /* Intermediate inputA pointer */
emilmont 1:fdd22bb7aa52 722 q15_t *py; /* Intermediate inputB pointer */
emilmont 1:fdd22bb7aa52 723 q15_t *pSrc1, *pSrc2; /* Intermediate pointers */
emilmont 1:fdd22bb7aa52 724 q31_t x0, x1, x2, x3, c0;
emilmont 1:fdd22bb7aa52 725 uint32_t j, k, count, check, blkCnt;
emilmont 1:fdd22bb7aa52 726 int32_t blockSize1, blockSize2, blockSize3; /* loop counters */
emilmont 1:fdd22bb7aa52 727 arm_status status; /* status of Partial convolution */
emilmont 1:fdd22bb7aa52 728 q15_t a, b;
emilmont 1:fdd22bb7aa52 729
emilmont 1:fdd22bb7aa52 730 /* Check for range of output samples to be calculated */
emilmont 1:fdd22bb7aa52 731 if((firstIndex + numPoints) > ((srcALen + (srcBLen - 1u))))
emilmont 1:fdd22bb7aa52 732 {
emilmont 1:fdd22bb7aa52 733 /* Set status as ARM_MATH_ARGUMENT_ERROR */
emilmont 1:fdd22bb7aa52 734 status = ARM_MATH_ARGUMENT_ERROR;
emilmont 1:fdd22bb7aa52 735 }
emilmont 1:fdd22bb7aa52 736 else
emilmont 1:fdd22bb7aa52 737 {
emilmont 1:fdd22bb7aa52 738
emilmont 1:fdd22bb7aa52 739 /* The algorithm implementation is based on the lengths of the inputs. */
emilmont 1:fdd22bb7aa52 740 /* srcB is always made to slide across srcA. */
emilmont 1:fdd22bb7aa52 741 /* So srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 742 if(srcALen >=srcBLen)
emilmont 1:fdd22bb7aa52 743 {
emilmont 1:fdd22bb7aa52 744 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 745 pIn1 = pSrcA;
emilmont 1:fdd22bb7aa52 746
emilmont 1:fdd22bb7aa52 747 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 748 pIn2 = pSrcB;
emilmont 1:fdd22bb7aa52 749 }
emilmont 1:fdd22bb7aa52 750 else
emilmont 1:fdd22bb7aa52 751 {
emilmont 1:fdd22bb7aa52 752 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 753 pIn1 = pSrcB;
emilmont 1:fdd22bb7aa52 754
emilmont 1:fdd22bb7aa52 755 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 756 pIn2 = pSrcA;
emilmont 1:fdd22bb7aa52 757
emilmont 1:fdd22bb7aa52 758 /* srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 759 j = srcBLen;
emilmont 1:fdd22bb7aa52 760 srcBLen = srcALen;
emilmont 1:fdd22bb7aa52 761 srcALen = j;
emilmont 1:fdd22bb7aa52 762 }
emilmont 1:fdd22bb7aa52 763
emilmont 1:fdd22bb7aa52 764 /* Conditions to check which loopCounter holds
emilmont 1:fdd22bb7aa52 765 * the first and last indices of the output samples to be calculated. */
emilmont 1:fdd22bb7aa52 766 check = firstIndex + numPoints;
emilmont 1:fdd22bb7aa52 767 blockSize3 = ((int32_t) check - (int32_t) srcALen);
emilmont 1:fdd22bb7aa52 768 blockSize3 = (blockSize3 > 0) ? blockSize3 : 0;
emilmont 1:fdd22bb7aa52 769 blockSize1 = (((int32_t) srcBLen - 1) - (int32_t) firstIndex);
emilmont 1:fdd22bb7aa52 770 blockSize1 = (blockSize1 > 0) ? ((check > (srcBLen - 1u)) ? blockSize1 :
emilmont 1:fdd22bb7aa52 771 (int32_t) numPoints) : 0;
emilmont 1:fdd22bb7aa52 772 blockSize2 = (int32_t) check - ((blockSize3 + blockSize1) +
emilmont 1:fdd22bb7aa52 773 (int32_t) firstIndex);
emilmont 1:fdd22bb7aa52 774 blockSize2 = (blockSize2 > 0) ? blockSize2 : 0;
emilmont 1:fdd22bb7aa52 775
emilmont 1:fdd22bb7aa52 776 /* 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 777 /* The function is internally
emilmont 1:fdd22bb7aa52 778 * divided into three stages according to the number of multiplications that has to be
emilmont 1:fdd22bb7aa52 779 * taken place between inputA samples and inputB samples. In the first stage of the
emilmont 1:fdd22bb7aa52 780 * algorithm, the multiplications increase by one for every iteration.
emilmont 1:fdd22bb7aa52 781 * In the second stage of the algorithm, srcBLen number of multiplications are done.
emilmont 1:fdd22bb7aa52 782 * In the third stage of the algorithm, the multiplications decrease by one
emilmont 1:fdd22bb7aa52 783 * for every iteration. */
emilmont 1:fdd22bb7aa52 784
emilmont 1:fdd22bb7aa52 785 /* Set the output pointer to point to the firstIndex
emilmont 1:fdd22bb7aa52 786 * of the output sample to be calculated. */
emilmont 1:fdd22bb7aa52 787 pOut = pDst + firstIndex;
emilmont 1:fdd22bb7aa52 788
emilmont 1:fdd22bb7aa52 789 /* --------------------------
emilmont 1:fdd22bb7aa52 790 * Initializations of stage1
emilmont 1:fdd22bb7aa52 791 * -------------------------*/
emilmont 1:fdd22bb7aa52 792
emilmont 1:fdd22bb7aa52 793 /* sum = x[0] * y[0]
emilmont 1:fdd22bb7aa52 794 * sum = x[0] * y[1] + x[1] * y[0]
emilmont 1:fdd22bb7aa52 795 * ....
emilmont 1:fdd22bb7aa52 796 * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
emilmont 1:fdd22bb7aa52 797 */
emilmont 1:fdd22bb7aa52 798
emilmont 1:fdd22bb7aa52 799 /* In this stage the MAC operations are increased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 800 The count variable holds the number of MAC operations performed.
emilmont 1:fdd22bb7aa52 801 Since the partial convolution starts from firstIndex
emilmont 1:fdd22bb7aa52 802 Number of Macs to be performed is firstIndex + 1 */
emilmont 1:fdd22bb7aa52 803 count = 1u + firstIndex;
emilmont 1:fdd22bb7aa52 804
emilmont 1:fdd22bb7aa52 805 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 806 px = pIn1;
emilmont 1:fdd22bb7aa52 807
emilmont 1:fdd22bb7aa52 808 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 809 pSrc2 = pIn2 + firstIndex;
emilmont 1:fdd22bb7aa52 810 py = pSrc2;
emilmont 1:fdd22bb7aa52 811
emilmont 1:fdd22bb7aa52 812 /* ------------------------
emilmont 1:fdd22bb7aa52 813 * Stage1 process
emilmont 1:fdd22bb7aa52 814 * ----------------------*/
emilmont 1:fdd22bb7aa52 815
emilmont 1:fdd22bb7aa52 816 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 817 /* First part of this stage computes the MAC operations less than 4 */
emilmont 1:fdd22bb7aa52 818 /* Second part of this stage computes the MAC operations greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 819
emilmont 1:fdd22bb7aa52 820 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 821 while((count < 4u) && (blockSize1 > 0u))
emilmont 1:fdd22bb7aa52 822 {
emilmont 1:fdd22bb7aa52 823 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 824 sum = 0;
emilmont 1:fdd22bb7aa52 825
emilmont 1:fdd22bb7aa52 826 /* Loop over number of MAC operations between
emilmont 1:fdd22bb7aa52 827 * inputA samples and inputB samples */
emilmont 1:fdd22bb7aa52 828 k = count;
emilmont 1:fdd22bb7aa52 829
emilmont 1:fdd22bb7aa52 830 while(k > 0u)
emilmont 1:fdd22bb7aa52 831 {
emilmont 1:fdd22bb7aa52 832 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 833 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 834
emilmont 1:fdd22bb7aa52 835 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 836 k--;
emilmont 1:fdd22bb7aa52 837 }
emilmont 1:fdd22bb7aa52 838
emilmont 1:fdd22bb7aa52 839 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 840 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 841
emilmont 1:fdd22bb7aa52 842 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 843 py = ++pSrc2;
emilmont 1:fdd22bb7aa52 844 px = pIn1;
emilmont 1:fdd22bb7aa52 845
emilmont 1:fdd22bb7aa52 846 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 847 count++;
emilmont 1:fdd22bb7aa52 848
emilmont 1:fdd22bb7aa52 849 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 850 blockSize1--;
emilmont 1:fdd22bb7aa52 851 }
emilmont 1:fdd22bb7aa52 852
emilmont 1:fdd22bb7aa52 853 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 854 /* The internal loop, over count, is unrolled by 4 */
emilmont 1:fdd22bb7aa52 855 /* To, read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 856 * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */
emilmont 1:fdd22bb7aa52 857 py = py - 1;
emilmont 1:fdd22bb7aa52 858
emilmont 1:fdd22bb7aa52 859 while(blockSize1 > 0u)
emilmont 1:fdd22bb7aa52 860 {
emilmont 1:fdd22bb7aa52 861 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 862 sum = 0;
emilmont 1:fdd22bb7aa52 863
emilmont 1:fdd22bb7aa52 864 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 865 k = count >> 2u;
emilmont 1:fdd22bb7aa52 866
emilmont 1:fdd22bb7aa52 867 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 868 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 2:da51fb522205 869 py++;
emilmont 1:fdd22bb7aa52 870
emilmont 1:fdd22bb7aa52 871 while(k > 0u)
emilmont 1:fdd22bb7aa52 872 {
emilmont 1:fdd22bb7aa52 873 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 874 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 875 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 876 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 877 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 878
emilmont 1:fdd22bb7aa52 879 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 880 k--;
emilmont 1:fdd22bb7aa52 881 }
emilmont 1:fdd22bb7aa52 882
emilmont 1:fdd22bb7aa52 883 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 884 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 885 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 886
emilmont 1:fdd22bb7aa52 887 while(k > 0u)
emilmont 1:fdd22bb7aa52 888 {
emilmont 1:fdd22bb7aa52 889 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 890 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 891
emilmont 1:fdd22bb7aa52 892 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 893 k--;
emilmont 1:fdd22bb7aa52 894 }
emilmont 1:fdd22bb7aa52 895
emilmont 1:fdd22bb7aa52 896 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 897 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 898
emilmont 1:fdd22bb7aa52 899 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 900 py = ++pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 901 px = pIn1;
emilmont 1:fdd22bb7aa52 902
emilmont 1:fdd22bb7aa52 903 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 904 count++;
emilmont 1:fdd22bb7aa52 905
emilmont 1:fdd22bb7aa52 906 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 907 blockSize1--;
emilmont 1:fdd22bb7aa52 908 }
emilmont 1:fdd22bb7aa52 909
emilmont 1:fdd22bb7aa52 910 /* --------------------------
emilmont 1:fdd22bb7aa52 911 * Initializations of stage2
emilmont 1:fdd22bb7aa52 912 * ------------------------*/
emilmont 1:fdd22bb7aa52 913
emilmont 1:fdd22bb7aa52 914 /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
emilmont 1:fdd22bb7aa52 915 * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
emilmont 1:fdd22bb7aa52 916 * ....
emilmont 1:fdd22bb7aa52 917 * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
emilmont 1:fdd22bb7aa52 918 */
emilmont 1:fdd22bb7aa52 919
emilmont 1:fdd22bb7aa52 920 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 921 px = pIn1;
emilmont 1:fdd22bb7aa52 922
emilmont 1:fdd22bb7aa52 923 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 924 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 925 py = pSrc2;
emilmont 1:fdd22bb7aa52 926
emilmont 1:fdd22bb7aa52 927 /* count is the index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 928 count = 0u;
emilmont 1:fdd22bb7aa52 929
emilmont 1:fdd22bb7aa52 930
emilmont 1:fdd22bb7aa52 931 /* --------------------
emilmont 1:fdd22bb7aa52 932 * Stage2 process
emilmont 1:fdd22bb7aa52 933 * -------------------*/
emilmont 1:fdd22bb7aa52 934
emilmont 1:fdd22bb7aa52 935 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 936 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 937 * srcBLen should be greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 938 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 939 {
emilmont 1:fdd22bb7aa52 940 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 941 blkCnt = ((uint32_t) blockSize2 >> 2u);
emilmont 1:fdd22bb7aa52 942
emilmont 1:fdd22bb7aa52 943 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 944 {
emilmont 1:fdd22bb7aa52 945 py = py - 1u;
emilmont 1:fdd22bb7aa52 946
emilmont 1:fdd22bb7aa52 947 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 948 acc0 = 0;
emilmont 1:fdd22bb7aa52 949 acc1 = 0;
emilmont 1:fdd22bb7aa52 950 acc2 = 0;
emilmont 1:fdd22bb7aa52 951 acc3 = 0;
emilmont 1:fdd22bb7aa52 952
emilmont 1:fdd22bb7aa52 953 /* read x[0], x[1] samples */
emilmont 2:da51fb522205 954 a = *px++;
emilmont 2:da51fb522205 955 b = *px++;
emilmont 1:fdd22bb7aa52 956
emilmont 1:fdd22bb7aa52 957 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 958
emilmont 2:da51fb522205 959 x0 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 960 a = *px;
emilmont 2:da51fb522205 961 x1 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 962
emilmont 1:fdd22bb7aa52 963 #else
emilmont 1:fdd22bb7aa52 964
emilmont 2:da51fb522205 965 x0 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 966 a = *px;
emilmont 2:da51fb522205 967 x1 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 968
emilmont 2:da51fb522205 969 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 970
emilmont 1:fdd22bb7aa52 971 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 972 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 973
emilmont 1:fdd22bb7aa52 974 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 975 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 976 do
emilmont 1:fdd22bb7aa52 977 {
emilmont 1:fdd22bb7aa52 978 /* Read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 979 * y[srcBLen - 1] and y[srcBLen - 2] */
emilmont 2:da51fb522205 980 a = *py;
emilmont 2:da51fb522205 981 b = *(py+1);
emilmont 2:da51fb522205 982 py -= 2;
emilmont 1:fdd22bb7aa52 983
emilmont 1:fdd22bb7aa52 984 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 985
emilmont 2:da51fb522205 986 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 987
emilmont 1:fdd22bb7aa52 988 #else
emilmont 1:fdd22bb7aa52 989
emilmont 2:da51fb522205 990 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 991
emilmont 2:da51fb522205 992 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 993
emilmont 1:fdd22bb7aa52 994 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 995 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 996
emilmont 1:fdd22bb7aa52 997 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 998 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 999
emilmont 2:da51fb522205 1000 a = *px;
emilmont 2:da51fb522205 1001 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1002
emilmont 1:fdd22bb7aa52 1003 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 1004
emilmont 2:da51fb522205 1005 x2 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 1006 a = *(px + 2);
emilmont 2:da51fb522205 1007 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1008
emilmont 1:fdd22bb7aa52 1009 #else
emilmont 1:fdd22bb7aa52 1010
emilmont 2:da51fb522205 1011 x2 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 1012 a = *(px + 2);
emilmont 2:da51fb522205 1013 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1014
emilmont 2:da51fb522205 1015 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1016
emilmont 1:fdd22bb7aa52 1017 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 1018 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 1019
emilmont 1:fdd22bb7aa52 1020 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 1021 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1022
emilmont 1:fdd22bb7aa52 1023 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emilmont 2:da51fb522205 1024 a = *py;
emilmont 2:da51fb522205 1025 b = *(py+1);
emilmont 2:da51fb522205 1026 py -= 2;
emilmont 1:fdd22bb7aa52 1027
emilmont 1:fdd22bb7aa52 1028 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1029
emilmont 2:da51fb522205 1030 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1031
emilmont 1:fdd22bb7aa52 1032 #else
emilmont 1:fdd22bb7aa52 1033
emilmont 2:da51fb522205 1034 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1035
emilmont 2:da51fb522205 1036 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1037
emilmont 1:fdd22bb7aa52 1038 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1039 acc0 = __SMLADX(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 1040
emilmont 1:fdd22bb7aa52 1041 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1042 acc1 = __SMLADX(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 1043
emilmont 1:fdd22bb7aa52 1044 /* Read x[4], x[5], x[6] */
emilmont 2:da51fb522205 1045 a = *(px + 2);
emilmont 2:da51fb522205 1046 b = *(px + 3);
emilmont 1:fdd22bb7aa52 1047
emilmont 1:fdd22bb7aa52 1048 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 1049
emilmont 2:da51fb522205 1050 x0 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 1051 a = *(px + 4);
emilmont 2:da51fb522205 1052 x1 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1053
emilmont 1:fdd22bb7aa52 1054 #else
emilmont 1:fdd22bb7aa52 1055
emilmont 2:da51fb522205 1056 x0 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 1057 a = *(px + 4);
emilmont 2:da51fb522205 1058 x1 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1059
emilmont 2:da51fb522205 1060 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1061
emilmont 2:da51fb522205 1062 px += 4u;
emilmont 1:fdd22bb7aa52 1063
emilmont 1:fdd22bb7aa52 1064 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1065 acc2 = __SMLADX(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 1066
emilmont 1:fdd22bb7aa52 1067 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1068 acc3 = __SMLADX(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 1069
emilmont 1:fdd22bb7aa52 1070 } while(--k);
emilmont 1:fdd22bb7aa52 1071
emilmont 1:fdd22bb7aa52 1072 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 1073 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 1074
emilmont 1:fdd22bb7aa52 1075 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1076 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1077 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 1078
emilmont 1:fdd22bb7aa52 1079 if(k == 1u)
emilmont 1:fdd22bb7aa52 1080 {
emilmont 1:fdd22bb7aa52 1081 /* Read y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 1082 c0 = *(py+1);
emilmont 1:fdd22bb7aa52 1083
emilmont 1:fdd22bb7aa52 1084 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1085
emilmont 1:fdd22bb7aa52 1086 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 1087
emilmont 1:fdd22bb7aa52 1088 #else
emilmont 1:fdd22bb7aa52 1089
emilmont 1:fdd22bb7aa52 1090 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 1091
emilmont 1:fdd22bb7aa52 1092 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1093
emilmont 1:fdd22bb7aa52 1094 /* Read x[7] */
emilmont 2:da51fb522205 1095 a = *px;
emilmont 2:da51fb522205 1096 b = *(px+1);
emilmont 2:da51fb522205 1097 px++;
emilmont 1:fdd22bb7aa52 1098
emilmont 1:fdd22bb7aa52 1099 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1100
emilmont 2:da51fb522205 1101 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1102
emilmont 1:fdd22bb7aa52 1103 #else
emilmont 1:fdd22bb7aa52 1104
emilmont 2:da51fb522205 1105 x3 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1106
emilmont 2:da51fb522205 1107 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1108
emilmont 1:fdd22bb7aa52 1109
emilmont 1:fdd22bb7aa52 1110 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1111 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1112 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1113 acc2 = __SMLADX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 1114 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1115 }
emilmont 1:fdd22bb7aa52 1116
emilmont 1:fdd22bb7aa52 1117 if(k == 2u)
emilmont 1:fdd22bb7aa52 1118 {
emilmont 1:fdd22bb7aa52 1119 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 2:da51fb522205 1120 a = *py;
emilmont 2:da51fb522205 1121 b = *(py+1);
emilmont 1:fdd22bb7aa52 1122
emilmont 1:fdd22bb7aa52 1123 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1124
emilmont 2:da51fb522205 1125 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1126
emilmont 1:fdd22bb7aa52 1127 #else
emilmont 1:fdd22bb7aa52 1128
emilmont 2:da51fb522205 1129 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1130
emilmont 2:da51fb522205 1131 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1132
emilmont 1:fdd22bb7aa52 1133 /* Read x[7], x[8], x[9] */
emilmont 2:da51fb522205 1134 a = *px;
emilmont 2:da51fb522205 1135 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1136
emilmont 1:fdd22bb7aa52 1137 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 1138
emilmont 2:da51fb522205 1139 x3 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 1140 a = *(px + 2);
emilmont 2:da51fb522205 1141 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1142
emilmont 1:fdd22bb7aa52 1143 #else
emilmont 1:fdd22bb7aa52 1144
emilmont 2:da51fb522205 1145 x3 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 1146 a = *(px + 2);
emilmont 2:da51fb522205 1147 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1148
emilmont 2:da51fb522205 1149 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 2:da51fb522205 1150 px += 2u;
emilmont 1:fdd22bb7aa52 1151
emilmont 1:fdd22bb7aa52 1152 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1153 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1154 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1155 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 1156 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 1157 }
emilmont 1:fdd22bb7aa52 1158
emilmont 1:fdd22bb7aa52 1159 if(k == 3u)
emilmont 1:fdd22bb7aa52 1160 {
emilmont 1:fdd22bb7aa52 1161 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 2:da51fb522205 1162 a = *py;
emilmont 2:da51fb522205 1163 b = *(py+1);
emilmont 1:fdd22bb7aa52 1164
emilmont 1:fdd22bb7aa52 1165 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1166
emilmont 2:da51fb522205 1167 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1168
emilmont 1:fdd22bb7aa52 1169 #else
emilmont 1:fdd22bb7aa52 1170
emilmont 2:da51fb522205 1171 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1172
emilmont 2:da51fb522205 1173 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1174
emilmont 1:fdd22bb7aa52 1175 /* Read x[7], x[8], x[9] */
emilmont 2:da51fb522205 1176 a = *px;
emilmont 2:da51fb522205 1177 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1178
emilmont 1:fdd22bb7aa52 1179 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 1180
emilmont 2:da51fb522205 1181 x3 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 1182 a = *(px + 2);
emilmont 2:da51fb522205 1183 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1184
emilmont 1:fdd22bb7aa52 1185 #else
emilmont 1:fdd22bb7aa52 1186
emilmont 2:da51fb522205 1187 x3 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 1188 a = *(px + 2);
emilmont 2:da51fb522205 1189 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1190
emilmont 2:da51fb522205 1191 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1192
emilmont 1:fdd22bb7aa52 1193 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1194 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1195 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1196 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 1197 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 1198
emilmont 1:fdd22bb7aa52 1199 /* Read y[srcBLen - 7] */
emilmont 2:da51fb522205 1200 c0 = *(py-1);
emilmont 1:fdd22bb7aa52 1201 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1202
emilmont 1:fdd22bb7aa52 1203 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 1204 #else
emilmont 1:fdd22bb7aa52 1205
emilmont 1:fdd22bb7aa52 1206 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 1207 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1208
emilmont 1:fdd22bb7aa52 1209 /* Read x[10] */
emilmont 2:da51fb522205 1210 a = *(px+2);
emilmont 2:da51fb522205 1211 b = *(px+3);
emilmont 1:fdd22bb7aa52 1212
emilmont 1:fdd22bb7aa52 1213 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1214
emilmont 2:da51fb522205 1215 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1216
emilmont 1:fdd22bb7aa52 1217 #else
emilmont 1:fdd22bb7aa52 1218
emilmont 2:da51fb522205 1219 x3 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1220
emilmont 2:da51fb522205 1221 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1222
emilmont 2:da51fb522205 1223 px += 3u;
emilmont 1:fdd22bb7aa52 1224
emilmont 1:fdd22bb7aa52 1225 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1226 acc0 = __SMLADX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 1227 acc1 = __SMLAD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 1228 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 1229 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1230 }
emilmont 1:fdd22bb7aa52 1231
emilmont 1:fdd22bb7aa52 1232 /* Store the results in the accumulators in the destination buffer. */
emilmont 2:da51fb522205 1233 *pOut++ = (q15_t)(acc0 >> 15);
emilmont 2:da51fb522205 1234 *pOut++ = (q15_t)(acc1 >> 15);
emilmont 2:da51fb522205 1235 *pOut++ = (q15_t)(acc2 >> 15);
emilmont 2:da51fb522205 1236 *pOut++ = (q15_t)(acc3 >> 15);
emilmont 1:fdd22bb7aa52 1237
emilmont 1:fdd22bb7aa52 1238 /* Increment the pointer pIn1 index, count by 4 */
emilmont 1:fdd22bb7aa52 1239 count += 4u;
emilmont 1:fdd22bb7aa52 1240
emilmont 1:fdd22bb7aa52 1241 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1242 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1243 py = pSrc2;
emilmont 1:fdd22bb7aa52 1244
emilmont 1:fdd22bb7aa52 1245 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1246 blkCnt--;
emilmont 1:fdd22bb7aa52 1247 }
emilmont 1:fdd22bb7aa52 1248
emilmont 1:fdd22bb7aa52 1249 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 1250 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1251 blkCnt = (uint32_t) blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 1252
emilmont 1:fdd22bb7aa52 1253 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 1254 {
emilmont 1:fdd22bb7aa52 1255 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1256 sum = 0;
emilmont 1:fdd22bb7aa52 1257
emilmont 1:fdd22bb7aa52 1258 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1259 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 1260
emilmont 1:fdd22bb7aa52 1261 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 1262 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 1263 while(k > 0u)
emilmont 1:fdd22bb7aa52 1264 {
emilmont 1:fdd22bb7aa52 1265 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1266 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1267 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1268 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1269 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1270
emilmont 1:fdd22bb7aa52 1271 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1272 k--;
emilmont 1:fdd22bb7aa52 1273 }
emilmont 1:fdd22bb7aa52 1274
emilmont 1:fdd22bb7aa52 1275 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1276 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1277 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 1278
emilmont 1:fdd22bb7aa52 1279 while(k > 0u)
emilmont 1:fdd22bb7aa52 1280 {
emilmont 1:fdd22bb7aa52 1281 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1282 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1283
emilmont 1:fdd22bb7aa52 1284 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1285 k--;
emilmont 1:fdd22bb7aa52 1286 }
emilmont 1:fdd22bb7aa52 1287
emilmont 1:fdd22bb7aa52 1288 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1289 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1290
emilmont 1:fdd22bb7aa52 1291 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 1292 count++;
emilmont 1:fdd22bb7aa52 1293
emilmont 1:fdd22bb7aa52 1294 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1295 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1296 py = pSrc2;
emilmont 1:fdd22bb7aa52 1297
emilmont 1:fdd22bb7aa52 1298 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1299 blkCnt--;
emilmont 1:fdd22bb7aa52 1300 }
emilmont 1:fdd22bb7aa52 1301 }
emilmont 1:fdd22bb7aa52 1302 else
emilmont 1:fdd22bb7aa52 1303 {
emilmont 1:fdd22bb7aa52 1304 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 1305 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 1306 blkCnt = (uint32_t) blockSize2;
emilmont 1:fdd22bb7aa52 1307
emilmont 1:fdd22bb7aa52 1308 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 1309 {
emilmont 1:fdd22bb7aa52 1310 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1311 sum = 0;
emilmont 1:fdd22bb7aa52 1312
emilmont 1:fdd22bb7aa52 1313 /* srcBLen number of MACS should be performed */
emilmont 1:fdd22bb7aa52 1314 k = srcBLen;
emilmont 1:fdd22bb7aa52 1315
emilmont 1:fdd22bb7aa52 1316 while(k > 0u)
emilmont 1:fdd22bb7aa52 1317 {
emilmont 1:fdd22bb7aa52 1318 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 1319 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1320
emilmont 1:fdd22bb7aa52 1321 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1322 k--;
emilmont 1:fdd22bb7aa52 1323 }
emilmont 1:fdd22bb7aa52 1324
emilmont 1:fdd22bb7aa52 1325 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1326 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1327
emilmont 1:fdd22bb7aa52 1328 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 1329 count++;
emilmont 1:fdd22bb7aa52 1330
emilmont 1:fdd22bb7aa52 1331 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1332 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1333 py = pSrc2;
emilmont 1:fdd22bb7aa52 1334
emilmont 1:fdd22bb7aa52 1335 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1336 blkCnt--;
emilmont 1:fdd22bb7aa52 1337 }
emilmont 1:fdd22bb7aa52 1338 }
emilmont 1:fdd22bb7aa52 1339
emilmont 1:fdd22bb7aa52 1340
emilmont 1:fdd22bb7aa52 1341 /* --------------------------
emilmont 1:fdd22bb7aa52 1342 * Initializations of stage3
emilmont 1:fdd22bb7aa52 1343 * -------------------------*/
emilmont 1:fdd22bb7aa52 1344
emilmont 1:fdd22bb7aa52 1345 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 1346 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emilmont 1:fdd22bb7aa52 1347 * ....
emilmont 1:fdd22bb7aa52 1348 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emilmont 1:fdd22bb7aa52 1349 * sum += x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 1350 */
emilmont 1:fdd22bb7aa52 1351
emilmont 1:fdd22bb7aa52 1352 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 1353 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 1354 count = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 1355
emilmont 1:fdd22bb7aa52 1356 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 1357 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 1358 px = pSrc1;
emilmont 1:fdd22bb7aa52 1359
emilmont 1:fdd22bb7aa52 1360 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 1361 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 1362 pIn2 = pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 1363 py = pIn2;
emilmont 1:fdd22bb7aa52 1364
emilmont 1:fdd22bb7aa52 1365 /* -------------------
emilmont 1:fdd22bb7aa52 1366 * Stage3 process
emilmont 1:fdd22bb7aa52 1367 * ------------------*/
emilmont 1:fdd22bb7aa52 1368
emilmont 1:fdd22bb7aa52 1369 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 1370 /* First part of this stage computes the MAC operations greater than 4 */
emilmont 1:fdd22bb7aa52 1371 /* Second part of this stage computes the MAC operations less than or equal to 4 */
emilmont 1:fdd22bb7aa52 1372
emilmont 1:fdd22bb7aa52 1373 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 1374 j = count >> 2u;
emilmont 1:fdd22bb7aa52 1375
emilmont 1:fdd22bb7aa52 1376 while((j > 0u) && (blockSize3 > 0))
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 = count >> 2u;
emilmont 1:fdd22bb7aa52 1383
emilmont 1:fdd22bb7aa52 1384 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 1385 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 2:da51fb522205 1386 py++;
emilmont 1:fdd22bb7aa52 1387
emilmont 1:fdd22bb7aa52 1388 while(k > 0u)
emilmont 2:da51fb522205 1389 {
emilmont 1:fdd22bb7aa52 1390 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1391 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1392 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1393 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1394 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1395 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1396 k--;
emilmont 1:fdd22bb7aa52 1397 }
emilmont 1:fdd22bb7aa52 1398
emilmont 1:fdd22bb7aa52 1399
emilmont 1:fdd22bb7aa52 1400 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1401 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1402 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 1403
emilmont 1:fdd22bb7aa52 1404 while(k > 0u)
emilmont 1:fdd22bb7aa52 1405 {
emilmont 1:fdd22bb7aa52 1406 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1407 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1408
emilmont 1:fdd22bb7aa52 1409 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1410 k--;
emilmont 1:fdd22bb7aa52 1411 }
emilmont 1:fdd22bb7aa52 1412
emilmont 1:fdd22bb7aa52 1413 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1414 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1415
emilmont 1:fdd22bb7aa52 1416 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1417 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 1418 py = pIn2;
emilmont 1:fdd22bb7aa52 1419
emilmont 1:fdd22bb7aa52 1420 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 1421 count--;
emilmont 1:fdd22bb7aa52 1422
emilmont 1:fdd22bb7aa52 1423 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1424 blockSize3--;
emilmont 1:fdd22bb7aa52 1425
emilmont 1:fdd22bb7aa52 1426 j--;
emilmont 1:fdd22bb7aa52 1427 }
emilmont 1:fdd22bb7aa52 1428
emilmont 1:fdd22bb7aa52 1429 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 1430 /* SIMD is not used for the next MAC operations,
emilmont 1:fdd22bb7aa52 1431 * so pointer py is updated to read only one sample at a time */
emilmont 1:fdd22bb7aa52 1432 py = py + 1u;
emilmont 1:fdd22bb7aa52 1433
emilmont 1:fdd22bb7aa52 1434 while(blockSize3 > 0u)
emilmont 1:fdd22bb7aa52 1435 {
emilmont 1:fdd22bb7aa52 1436 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1437 sum = 0;
emilmont 1:fdd22bb7aa52 1438
emilmont 1:fdd22bb7aa52 1439 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1440 k = count;
emilmont 1:fdd22bb7aa52 1441
emilmont 1:fdd22bb7aa52 1442 while(k > 0u)
emilmont 1:fdd22bb7aa52 1443 {
emilmont 1:fdd22bb7aa52 1444 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1445 /* sum += x[srcALen-1] * y[srcBLen-1] */
emilmont 1:fdd22bb7aa52 1446 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1447
emilmont 1:fdd22bb7aa52 1448 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1449 k--;
emilmont 1:fdd22bb7aa52 1450 }
emilmont 1:fdd22bb7aa52 1451
emilmont 1:fdd22bb7aa52 1452 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1453 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1454
emilmont 1:fdd22bb7aa52 1455 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1456 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 1457 py = pSrc2;
emilmont 1:fdd22bb7aa52 1458
emilmont 1:fdd22bb7aa52 1459 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 1460 count--;
emilmont 1:fdd22bb7aa52 1461
emilmont 1:fdd22bb7aa52 1462 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1463 blockSize3--;
emilmont 1:fdd22bb7aa52 1464 }
emilmont 1:fdd22bb7aa52 1465
emilmont 1:fdd22bb7aa52 1466 /* set status as ARM_MATH_SUCCESS */
emilmont 1:fdd22bb7aa52 1467 status = ARM_MATH_SUCCESS;
emilmont 1:fdd22bb7aa52 1468 }
emilmont 1:fdd22bb7aa52 1469
emilmont 1:fdd22bb7aa52 1470 /* Return to application */
emilmont 1:fdd22bb7aa52 1471 return (status);
emilmont 1:fdd22bb7aa52 1472
emilmont 1:fdd22bb7aa52 1473 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
emilmont 1:fdd22bb7aa52 1474 }
emilmont 1:fdd22bb7aa52 1475
emilmont 1:fdd22bb7aa52 1476 /**
emilmont 1:fdd22bb7aa52 1477 * @} end of PartialConv group
emilmont 1:fdd22bb7aa52 1478 */