Eli Hughes / CMSIS_DSP_401

V4.0.1 of the ARM CMSIS DSP libraries. Note that arm_bitreversal2.s, arm_cfft_f32.c and arm_rfft_fast_f32.c had to be removed. arm_bitreversal2.s will not assemble with the online tools. So, the fast f32 FFT functions are not yet available. All the other FFT functions are available.

Dependents:   MPU9150_Example fir_f32 fir_f32 MPU9150_nucleo_noni2cdev ... more

FilteringFunctions/arm_conv_fast_q15.c@0:3d9c67d97d6f, 2014-07-28 (annotated)

Committer:
emh203
Date:
Mon Jul 28 15:03:15 2014 +0000
Revision:
0:3d9c67d97d6f
1st working commit.   Had to remove arm_bitreversal2.s     arm_cfft_f32.c and arm_rfft_fast_f32.c.    The .s will not assemble.      For now I removed these functions so we could at least have a library for the other functions.

Who changed what in which revision?

UserRevisionLine numberNew contents of line
emh203 0:3d9c67d97d6f 1 /* ----------------------------------------------------------------------
emh203 0:3d9c67d97d6f 2 * Copyright (C) 2010-2014 ARM Limited. All rights reserved.
emh203 0:3d9c67d97d6f 3 *
emh203 0:3d9c67d97d6f 4 * $Date: 12. March 2014
emh203 0:3d9c67d97d6f 5 * $Revision: V1.4.3
emh203 0:3d9c67d97d6f 6 *
emh203 0:3d9c67d97d6f 7 * Project: CMSIS DSP Library
emh203 0:3d9c67d97d6f 8 * Title: arm_conv_fast_q15.c
emh203 0:3d9c67d97d6f 9 *
emh203 0:3d9c67d97d6f 10 * Description: Fast Q15 Convolution.
emh203 0:3d9c67d97d6f 11 *
emh203 0:3d9c67d97d6f 12 * Target Processor: Cortex-M4/Cortex-M3
emh203 0:3d9c67d97d6f 13 *
emh203 0:3d9c67d97d6f 14 * Redistribution and use in source and binary forms, with or without
emh203 0:3d9c67d97d6f 15 * modification, are permitted provided that the following conditions
emh203 0:3d9c67d97d6f 16 * are met:
emh203 0:3d9c67d97d6f 17 * - Redistributions of source code must retain the above copyright
emh203 0:3d9c67d97d6f 18 * notice, this list of conditions and the following disclaimer.
emh203 0:3d9c67d97d6f 19 * - Redistributions in binary form must reproduce the above copyright
emh203 0:3d9c67d97d6f 20 * notice, this list of conditions and the following disclaimer in
emh203 0:3d9c67d97d6f 21 * the documentation and/or other materials provided with the
emh203 0:3d9c67d97d6f 22 * distribution.
emh203 0:3d9c67d97d6f 23 * - Neither the name of ARM LIMITED nor the names of its contributors
emh203 0:3d9c67d97d6f 24 * may be used to endorse or promote products derived from this
emh203 0:3d9c67d97d6f 25 * software without specific prior written permission.
emh203 0:3d9c67d97d6f 26 *
emh203 0:3d9c67d97d6f 27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
emh203 0:3d9c67d97d6f 28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
emh203 0:3d9c67d97d6f 29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
emh203 0:3d9c67d97d6f 30 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
emh203 0:3d9c67d97d6f 31 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
emh203 0:3d9c67d97d6f 32 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
emh203 0:3d9c67d97d6f 33 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
emh203 0:3d9c67d97d6f 34 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
emh203 0:3d9c67d97d6f 35 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
emh203 0:3d9c67d97d6f 36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
emh203 0:3d9c67d97d6f 37 * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
emh203 0:3d9c67d97d6f 38 * POSSIBILITY OF SUCH DAMAGE.
emh203 0:3d9c67d97d6f 39 * -------------------------------------------------------------------- */
emh203 0:3d9c67d97d6f 40
emh203 0:3d9c67d97d6f 41 #include "arm_math.h"
emh203 0:3d9c67d97d6f 42
emh203 0:3d9c67d97d6f 43 /**
emh203 0:3d9c67d97d6f 44 * @ingroup groupFilters
emh203 0:3d9c67d97d6f 45 */
emh203 0:3d9c67d97d6f 46
emh203 0:3d9c67d97d6f 47 /**
emh203 0:3d9c67d97d6f 48 * @addtogroup Conv
emh203 0:3d9c67d97d6f 49 * @{
emh203 0:3d9c67d97d6f 50 */
emh203 0:3d9c67d97d6f 51
emh203 0:3d9c67d97d6f 52 /**
emh203 0:3d9c67d97d6f 53 * @brief Convolution of Q15 sequences (fast version) for Cortex-M3 and Cortex-M4.
emh203 0:3d9c67d97d6f 54 * @param[in] *pSrcA points to the first input sequence.
emh203 0:3d9c67d97d6f 55 * @param[in] srcALen length of the first input sequence.
emh203 0:3d9c67d97d6f 56 * @param[in] *pSrcB points to the second input sequence.
emh203 0:3d9c67d97d6f 57 * @param[in] srcBLen length of the second input sequence.
emh203 0:3d9c67d97d6f 58 * @param[out] *pDst points to the location where the output result is written. Length srcALen+srcBLen-1.
emh203 0:3d9c67d97d6f 59 * @return none.
emh203 0:3d9c67d97d6f 60 *
emh203 0:3d9c67d97d6f 61 * <b>Scaling and Overflow Behavior:</b>
emh203 0:3d9c67d97d6f 62 *
emh203 0:3d9c67d97d6f 63 * \par
emh203 0:3d9c67d97d6f 64 * This fast version uses a 32-bit accumulator with 2.30 format.
emh203 0:3d9c67d97d6f 65 * The accumulator maintains full precision of the intermediate multiplication results
emh203 0:3d9c67d97d6f 66 * but provides only a single guard bit. There is no saturation on intermediate additions.
emh203 0:3d9c67d97d6f 67 * Thus, if the accumulator overflows it wraps around and distorts the result.
emh203 0:3d9c67d97d6f 68 * The input signals should be scaled down to avoid intermediate overflows.
emh203 0:3d9c67d97d6f 69 * Scale down the inputs by log2(min(srcALen, srcBLen)) (log2 is read as log to the base 2) times to avoid overflows,
emh203 0:3d9c67d97d6f 70 * as maximum of min(srcALen, srcBLen) number of additions are carried internally.
emh203 0:3d9c67d97d6f 71 * The 2.30 accumulator is right shifted by 15 bits and then saturated to 1.15 format to yield the final result.
emh203 0:3d9c67d97d6f 72 *
emh203 0:3d9c67d97d6f 73 * \par
emh203 0:3d9c67d97d6f 74 * See <code>arm_conv_q15()</code> for a slower implementation of this function which uses 64-bit accumulation to avoid wrap around distortion.
emh203 0:3d9c67d97d6f 75 */
emh203 0:3d9c67d97d6f 76
emh203 0:3d9c67d97d6f 77 void arm_conv_fast_q15(
emh203 0:3d9c67d97d6f 78 q15_t * pSrcA,
emh203 0:3d9c67d97d6f 79 uint32_t srcALen,
emh203 0:3d9c67d97d6f 80 q15_t * pSrcB,
emh203 0:3d9c67d97d6f 81 uint32_t srcBLen,
emh203 0:3d9c67d97d6f 82 q15_t * pDst)
emh203 0:3d9c67d97d6f 83 {
emh203 0:3d9c67d97d6f 84 #ifndef UNALIGNED_SUPPORT_DISABLE
emh203 0:3d9c67d97d6f 85 q15_t *pIn1; /* inputA pointer */
emh203 0:3d9c67d97d6f 86 q15_t *pIn2; /* inputB pointer */
emh203 0:3d9c67d97d6f 87 q15_t *pOut = pDst; /* output pointer */
emh203 0:3d9c67d97d6f 88 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
emh203 0:3d9c67d97d6f 89 q15_t *px; /* Intermediate inputA pointer */
emh203 0:3d9c67d97d6f 90 q15_t *py; /* Intermediate inputB pointer */
emh203 0:3d9c67d97d6f 91 q15_t *pSrc1, *pSrc2; /* Intermediate pointers */
emh203 0:3d9c67d97d6f 92 q31_t x0, x1, x2, x3, c0; /* Temporary variables to hold state and coefficient values */
emh203 0:3d9c67d97d6f 93 uint32_t blockSize1, blockSize2, blockSize3, j, k, count, blkCnt; /* loop counter */
emh203 0:3d9c67d97d6f 94
emh203 0:3d9c67d97d6f 95 /* The algorithm implementation is based on the lengths of the inputs. */
emh203 0:3d9c67d97d6f 96 /* srcB is always made to slide across srcA. */
emh203 0:3d9c67d97d6f 97 /* So srcBLen is always considered as shorter or equal to srcALen */
emh203 0:3d9c67d97d6f 98 if(srcALen >= srcBLen)
emh203 0:3d9c67d97d6f 99 {
emh203 0:3d9c67d97d6f 100 /* Initialization of inputA pointer */
emh203 0:3d9c67d97d6f 101 pIn1 = pSrcA;
emh203 0:3d9c67d97d6f 102
emh203 0:3d9c67d97d6f 103 /* Initialization of inputB pointer */
emh203 0:3d9c67d97d6f 104 pIn2 = pSrcB;
emh203 0:3d9c67d97d6f 105 }
emh203 0:3d9c67d97d6f 106 else
emh203 0:3d9c67d97d6f 107 {
emh203 0:3d9c67d97d6f 108 /* Initialization of inputA pointer */
emh203 0:3d9c67d97d6f 109 pIn1 = pSrcB;
emh203 0:3d9c67d97d6f 110
emh203 0:3d9c67d97d6f 111 /* Initialization of inputB pointer */
emh203 0:3d9c67d97d6f 112 pIn2 = pSrcA;
emh203 0:3d9c67d97d6f 113
emh203 0:3d9c67d97d6f 114 /* srcBLen is always considered as shorter or equal to srcALen */
emh203 0:3d9c67d97d6f 115 j = srcBLen;
emh203 0:3d9c67d97d6f 116 srcBLen = srcALen;
emh203 0:3d9c67d97d6f 117 srcALen = j;
emh203 0:3d9c67d97d6f 118 }
emh203 0:3d9c67d97d6f 119
emh203 0:3d9c67d97d6f 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] */
emh203 0:3d9c67d97d6f 121 /* The function is internally
emh203 0:3d9c67d97d6f 122 * divided into three stages according to the number of multiplications that has to be
emh203 0:3d9c67d97d6f 123 * taken place between inputA samples and inputB samples. In the first stage of the
emh203 0:3d9c67d97d6f 124 * algorithm, the multiplications increase by one for every iteration.
emh203 0:3d9c67d97d6f 125 * In the second stage of the algorithm, srcBLen number of multiplications are done.
emh203 0:3d9c67d97d6f 126 * In the third stage of the algorithm, the multiplications decrease by one
emh203 0:3d9c67d97d6f 127 * for every iteration. */
emh203 0:3d9c67d97d6f 128
emh203 0:3d9c67d97d6f 129 /* The algorithm is implemented in three stages.
emh203 0:3d9c67d97d6f 130 The loop counters of each stage is initiated here. */
emh203 0:3d9c67d97d6f 131 blockSize1 = srcBLen - 1u;
emh203 0:3d9c67d97d6f 132 blockSize2 = srcALen - (srcBLen - 1u);
emh203 0:3d9c67d97d6f 133 blockSize3 = blockSize1;
emh203 0:3d9c67d97d6f 134
emh203 0:3d9c67d97d6f 135 /* --------------------------
emh203 0:3d9c67d97d6f 136 * Initializations of stage1
emh203 0:3d9c67d97d6f 137 * -------------------------*/
emh203 0:3d9c67d97d6f 138
emh203 0:3d9c67d97d6f 139 /* sum = x[0] * y[0]
emh203 0:3d9c67d97d6f 140 * sum = x[0] * y[1] + x[1] * y[0]
emh203 0:3d9c67d97d6f 141 * ....
emh203 0:3d9c67d97d6f 142 * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
emh203 0:3d9c67d97d6f 143 */
emh203 0:3d9c67d97d6f 144
emh203 0:3d9c67d97d6f 145 /* In this stage the MAC operations are increased by 1 for every iteration.
emh203 0:3d9c67d97d6f 146 The count variable holds the number of MAC operations performed */
emh203 0:3d9c67d97d6f 147 count = 1u;
emh203 0:3d9c67d97d6f 148
emh203 0:3d9c67d97d6f 149 /* Working pointer of inputA */
emh203 0:3d9c67d97d6f 150 px = pIn1;
emh203 0:3d9c67d97d6f 151
emh203 0:3d9c67d97d6f 152 /* Working pointer of inputB */
emh203 0:3d9c67d97d6f 153 py = pIn2;
emh203 0:3d9c67d97d6f 154
emh203 0:3d9c67d97d6f 155
emh203 0:3d9c67d97d6f 156 /* ------------------------
emh203 0:3d9c67d97d6f 157 * Stage1 process
emh203 0:3d9c67d97d6f 158 * ----------------------*/
emh203 0:3d9c67d97d6f 159
emh203 0:3d9c67d97d6f 160 /* For loop unrolling by 4, this stage is divided into two. */
emh203 0:3d9c67d97d6f 161 /* First part of this stage computes the MAC operations less than 4 */
emh203 0:3d9c67d97d6f 162 /* Second part of this stage computes the MAC operations greater than or equal to 4 */
emh203 0:3d9c67d97d6f 163
emh203 0:3d9c67d97d6f 164 /* The first part of the stage starts here */
emh203 0:3d9c67d97d6f 165 while((count < 4u) && (blockSize1 > 0u))
emh203 0:3d9c67d97d6f 166 {
emh203 0:3d9c67d97d6f 167 /* Accumulator is made zero for every iteration */
emh203 0:3d9c67d97d6f 168 sum = 0;
emh203 0:3d9c67d97d6f 169
emh203 0:3d9c67d97d6f 170 /* Loop over number of MAC operations between
emh203 0:3d9c67d97d6f 171 * inputA samples and inputB samples */
emh203 0:3d9c67d97d6f 172 k = count;
emh203 0:3d9c67d97d6f 173
emh203 0:3d9c67d97d6f 174 while(k > 0u)
emh203 0:3d9c67d97d6f 175 {
emh203 0:3d9c67d97d6f 176 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 177 sum = __SMLAD(*px++, *py--, sum);
emh203 0:3d9c67d97d6f 178
emh203 0:3d9c67d97d6f 179 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 180 k--;
emh203 0:3d9c67d97d6f 181 }
emh203 0:3d9c67d97d6f 182
emh203 0:3d9c67d97d6f 183 /* Store the result in the accumulator in the destination buffer. */
emh203 0:3d9c67d97d6f 184 *pOut++ = (q15_t) (sum >> 15);
emh203 0:3d9c67d97d6f 185
emh203 0:3d9c67d97d6f 186 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 187 py = pIn2 + count;
emh203 0:3d9c67d97d6f 188 px = pIn1;
emh203 0:3d9c67d97d6f 189
emh203 0:3d9c67d97d6f 190 /* Increment the MAC count */
emh203 0:3d9c67d97d6f 191 count++;
emh203 0:3d9c67d97d6f 192
emh203 0:3d9c67d97d6f 193 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 194 blockSize1--;
emh203 0:3d9c67d97d6f 195 }
emh203 0:3d9c67d97d6f 196
emh203 0:3d9c67d97d6f 197 /* The second part of the stage starts here */
emh203 0:3d9c67d97d6f 198 /* The internal loop, over count, is unrolled by 4 */
emh203 0:3d9c67d97d6f 199 /* To, read the last two inputB samples using SIMD:
emh203 0:3d9c67d97d6f 200 * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */
emh203 0:3d9c67d97d6f 201 py = py - 1;
emh203 0:3d9c67d97d6f 202
emh203 0:3d9c67d97d6f 203 while(blockSize1 > 0u)
emh203 0:3d9c67d97d6f 204 {
emh203 0:3d9c67d97d6f 205 /* Accumulator is made zero for every iteration */
emh203 0:3d9c67d97d6f 206 sum = 0;
emh203 0:3d9c67d97d6f 207
emh203 0:3d9c67d97d6f 208 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emh203 0:3d9c67d97d6f 209 k = count >> 2u;
emh203 0:3d9c67d97d6f 210
emh203 0:3d9c67d97d6f 211 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emh203 0:3d9c67d97d6f 212 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emh203 0:3d9c67d97d6f 213 while(k > 0u)
emh203 0:3d9c67d97d6f 214 {
emh203 0:3d9c67d97d6f 215 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 216 /* x[0], x[1] are multiplied with y[srcBLen - 1], y[srcBLen - 2] respectively */
emh203 0:3d9c67d97d6f 217 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emh203 0:3d9c67d97d6f 218 /* x[2], x[3] are multiplied with y[srcBLen - 3], y[srcBLen - 4] respectively */
emh203 0:3d9c67d97d6f 219 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emh203 0:3d9c67d97d6f 220
emh203 0:3d9c67d97d6f 221 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 222 k--;
emh203 0:3d9c67d97d6f 223 }
emh203 0:3d9c67d97d6f 224
emh203 0:3d9c67d97d6f 225 /* For the next MAC operations, the pointer py is used without SIMD
emh203 0:3d9c67d97d6f 226 * So, py is incremented by 1 */
emh203 0:3d9c67d97d6f 227 py = py + 1u;
emh203 0:3d9c67d97d6f 228
emh203 0:3d9c67d97d6f 229 /* If the count is not a multiple of 4, compute any remaining MACs here.
emh203 0:3d9c67d97d6f 230 ** No loop unrolling is used. */
emh203 0:3d9c67d97d6f 231 k = count % 0x4u;
emh203 0:3d9c67d97d6f 232
emh203 0:3d9c67d97d6f 233 while(k > 0u)
emh203 0:3d9c67d97d6f 234 {
emh203 0:3d9c67d97d6f 235 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 236 sum = __SMLAD(*px++, *py--, sum);
emh203 0:3d9c67d97d6f 237
emh203 0:3d9c67d97d6f 238 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 239 k--;
emh203 0:3d9c67d97d6f 240 }
emh203 0:3d9c67d97d6f 241
emh203 0:3d9c67d97d6f 242 /* Store the result in the accumulator in the destination buffer. */
emh203 0:3d9c67d97d6f 243 *pOut++ = (q15_t) (sum >> 15);
emh203 0:3d9c67d97d6f 244
emh203 0:3d9c67d97d6f 245 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 246 py = pIn2 + (count - 1u);
emh203 0:3d9c67d97d6f 247 px = pIn1;
emh203 0:3d9c67d97d6f 248
emh203 0:3d9c67d97d6f 249 /* Increment the MAC count */
emh203 0:3d9c67d97d6f 250 count++;
emh203 0:3d9c67d97d6f 251
emh203 0:3d9c67d97d6f 252 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 253 blockSize1--;
emh203 0:3d9c67d97d6f 254 }
emh203 0:3d9c67d97d6f 255
emh203 0:3d9c67d97d6f 256 /* --------------------------
emh203 0:3d9c67d97d6f 257 * Initializations of stage2
emh203 0:3d9c67d97d6f 258 * ------------------------*/
emh203 0:3d9c67d97d6f 259
emh203 0:3d9c67d97d6f 260 /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
emh203 0:3d9c67d97d6f 261 * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
emh203 0:3d9c67d97d6f 262 * ....
emh203 0:3d9c67d97d6f 263 * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
emh203 0:3d9c67d97d6f 264 */
emh203 0:3d9c67d97d6f 265
emh203 0:3d9c67d97d6f 266 /* Working pointer of inputA */
emh203 0:3d9c67d97d6f 267 px = pIn1;
emh203 0:3d9c67d97d6f 268
emh203 0:3d9c67d97d6f 269 /* Working pointer of inputB */
emh203 0:3d9c67d97d6f 270 pSrc2 = pIn2 + (srcBLen - 1u);
emh203 0:3d9c67d97d6f 271 py = pSrc2;
emh203 0:3d9c67d97d6f 272
emh203 0:3d9c67d97d6f 273 /* count is the index by which the pointer pIn1 to be incremented */
emh203 0:3d9c67d97d6f 274 count = 0u;
emh203 0:3d9c67d97d6f 275
emh203 0:3d9c67d97d6f 276
emh203 0:3d9c67d97d6f 277 /* --------------------
emh203 0:3d9c67d97d6f 278 * Stage2 process
emh203 0:3d9c67d97d6f 279 * -------------------*/
emh203 0:3d9c67d97d6f 280
emh203 0:3d9c67d97d6f 281 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emh203 0:3d9c67d97d6f 282 * So, to loop unroll over blockSize2,
emh203 0:3d9c67d97d6f 283 * srcBLen should be greater than or equal to 4 */
emh203 0:3d9c67d97d6f 284 if(srcBLen >= 4u)
emh203 0:3d9c67d97d6f 285 {
emh203 0:3d9c67d97d6f 286 /* Loop unroll over blockSize2, by 4 */
emh203 0:3d9c67d97d6f 287 blkCnt = blockSize2 >> 2u;
emh203 0:3d9c67d97d6f 288
emh203 0:3d9c67d97d6f 289 while(blkCnt > 0u)
emh203 0:3d9c67d97d6f 290 {
emh203 0:3d9c67d97d6f 291 py = py - 1u;
emh203 0:3d9c67d97d6f 292
emh203 0:3d9c67d97d6f 293 /* Set all accumulators to zero */
emh203 0:3d9c67d97d6f 294 acc0 = 0;
emh203 0:3d9c67d97d6f 295 acc1 = 0;
emh203 0:3d9c67d97d6f 296 acc2 = 0;
emh203 0:3d9c67d97d6f 297 acc3 = 0;
emh203 0:3d9c67d97d6f 298
emh203 0:3d9c67d97d6f 299
emh203 0:3d9c67d97d6f 300 /* read x[0], x[1] samples */
emh203 0:3d9c67d97d6f 301 x0 = *__SIMD32(px);
emh203 0:3d9c67d97d6f 302 /* read x[1], x[2] samples */
emh203 0:3d9c67d97d6f 303 x1 = _SIMD32_OFFSET(px+1);
emh203 0:3d9c67d97d6f 304 px+= 2u;
emh203 0:3d9c67d97d6f 305
emh203 0:3d9c67d97d6f 306
emh203 0:3d9c67d97d6f 307 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emh203 0:3d9c67d97d6f 308 k = srcBLen >> 2u;
emh203 0:3d9c67d97d6f 309
emh203 0:3d9c67d97d6f 310 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emh203 0:3d9c67d97d6f 311 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emh203 0:3d9c67d97d6f 312 do
emh203 0:3d9c67d97d6f 313 {
emh203 0:3d9c67d97d6f 314 /* Read the last two inputB samples using SIMD:
emh203 0:3d9c67d97d6f 315 * y[srcBLen - 1] and y[srcBLen - 2] */
emh203 0:3d9c67d97d6f 316 c0 = *__SIMD32(py)--;
emh203 0:3d9c67d97d6f 317
emh203 0:3d9c67d97d6f 318 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emh203 0:3d9c67d97d6f 319 acc0 = __SMLADX(x0, c0, acc0);
emh203 0:3d9c67d97d6f 320
emh203 0:3d9c67d97d6f 321 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emh203 0:3d9c67d97d6f 322 acc1 = __SMLADX(x1, c0, acc1);
emh203 0:3d9c67d97d6f 323
emh203 0:3d9c67d97d6f 324 /* Read x[2], x[3] */
emh203 0:3d9c67d97d6f 325 x2 = *__SIMD32(px);
emh203 0:3d9c67d97d6f 326
emh203 0:3d9c67d97d6f 327 /* Read x[3], x[4] */
emh203 0:3d9c67d97d6f 328 x3 = _SIMD32_OFFSET(px+1);
emh203 0:3d9c67d97d6f 329
emh203 0:3d9c67d97d6f 330 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emh203 0:3d9c67d97d6f 331 acc2 = __SMLADX(x2, c0, acc2);
emh203 0:3d9c67d97d6f 332
emh203 0:3d9c67d97d6f 333 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emh203 0:3d9c67d97d6f 334 acc3 = __SMLADX(x3, c0, acc3);
emh203 0:3d9c67d97d6f 335
emh203 0:3d9c67d97d6f 336 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emh203 0:3d9c67d97d6f 337 c0 = *__SIMD32(py)--;
emh203 0:3d9c67d97d6f 338
emh203 0:3d9c67d97d6f 339 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emh203 0:3d9c67d97d6f 340 acc0 = __SMLADX(x2, c0, acc0);
emh203 0:3d9c67d97d6f 341
emh203 0:3d9c67d97d6f 342 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emh203 0:3d9c67d97d6f 343 acc1 = __SMLADX(x3, c0, acc1);
emh203 0:3d9c67d97d6f 344
emh203 0:3d9c67d97d6f 345 /* Read x[4], x[5] */
emh203 0:3d9c67d97d6f 346 x0 = _SIMD32_OFFSET(px+2);
emh203 0:3d9c67d97d6f 347
emh203 0:3d9c67d97d6f 348 /* Read x[5], x[6] */
emh203 0:3d9c67d97d6f 349 x1 = _SIMD32_OFFSET(px+3);
emh203 0:3d9c67d97d6f 350 px += 4u;
emh203 0:3d9c67d97d6f 351
emh203 0:3d9c67d97d6f 352 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emh203 0:3d9c67d97d6f 353 acc2 = __SMLADX(x0, c0, acc2);
emh203 0:3d9c67d97d6f 354
emh203 0:3d9c67d97d6f 355 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emh203 0:3d9c67d97d6f 356 acc3 = __SMLADX(x1, c0, acc3);
emh203 0:3d9c67d97d6f 357
emh203 0:3d9c67d97d6f 358 } while(--k);
emh203 0:3d9c67d97d6f 359
emh203 0:3d9c67d97d6f 360 /* For the next MAC operations, SIMD is not used
emh203 0:3d9c67d97d6f 361 * So, the 16 bit pointer if inputB, py is updated */
emh203 0:3d9c67d97d6f 362
emh203 0:3d9c67d97d6f 363 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emh203 0:3d9c67d97d6f 364 ** No loop unrolling is used. */
emh203 0:3d9c67d97d6f 365 k = srcBLen % 0x4u;
emh203 0:3d9c67d97d6f 366
emh203 0:3d9c67d97d6f 367 if(k == 1u)
emh203 0:3d9c67d97d6f 368 {
emh203 0:3d9c67d97d6f 369 /* Read y[srcBLen - 5] */
emh203 0:3d9c67d97d6f 370 c0 = *(py+1);
emh203 0:3d9c67d97d6f 371
emh203 0:3d9c67d97d6f 372 #ifdef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 373
emh203 0:3d9c67d97d6f 374 c0 = c0 << 16u;
emh203 0:3d9c67d97d6f 375
emh203 0:3d9c67d97d6f 376 #else
emh203 0:3d9c67d97d6f 377
emh203 0:3d9c67d97d6f 378 c0 = c0 & 0x0000FFFF;
emh203 0:3d9c67d97d6f 379
emh203 0:3d9c67d97d6f 380 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 381
emh203 0:3d9c67d97d6f 382 /* Read x[7] */
emh203 0:3d9c67d97d6f 383 x3 = *__SIMD32(px);
emh203 0:3d9c67d97d6f 384 px++;
emh203 0:3d9c67d97d6f 385
emh203 0:3d9c67d97d6f 386 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 387 acc0 = __SMLAD(x0, c0, acc0);
emh203 0:3d9c67d97d6f 388 acc1 = __SMLAD(x1, c0, acc1);
emh203 0:3d9c67d97d6f 389 acc2 = __SMLADX(x1, c0, acc2);
emh203 0:3d9c67d97d6f 390 acc3 = __SMLADX(x3, c0, acc3);
emh203 0:3d9c67d97d6f 391 }
emh203 0:3d9c67d97d6f 392
emh203 0:3d9c67d97d6f 393 if(k == 2u)
emh203 0:3d9c67d97d6f 394 {
emh203 0:3d9c67d97d6f 395 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emh203 0:3d9c67d97d6f 396 c0 = _SIMD32_OFFSET(py);
emh203 0:3d9c67d97d6f 397
emh203 0:3d9c67d97d6f 398 /* Read x[7], x[8] */
emh203 0:3d9c67d97d6f 399 x3 = *__SIMD32(px);
emh203 0:3d9c67d97d6f 400
emh203 0:3d9c67d97d6f 401 /* Read x[9] */
emh203 0:3d9c67d97d6f 402 x2 = _SIMD32_OFFSET(px+1);
emh203 0:3d9c67d97d6f 403 px += 2u;
emh203 0:3d9c67d97d6f 404
emh203 0:3d9c67d97d6f 405 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 406 acc0 = __SMLADX(x0, c0, acc0);
emh203 0:3d9c67d97d6f 407 acc1 = __SMLADX(x1, c0, acc1);
emh203 0:3d9c67d97d6f 408 acc2 = __SMLADX(x3, c0, acc2);
emh203 0:3d9c67d97d6f 409 acc3 = __SMLADX(x2, c0, acc3);
emh203 0:3d9c67d97d6f 410 }
emh203 0:3d9c67d97d6f 411
emh203 0:3d9c67d97d6f 412 if(k == 3u)
emh203 0:3d9c67d97d6f 413 {
emh203 0:3d9c67d97d6f 414 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emh203 0:3d9c67d97d6f 415 c0 = _SIMD32_OFFSET(py);
emh203 0:3d9c67d97d6f 416
emh203 0:3d9c67d97d6f 417 /* Read x[7], x[8] */
emh203 0:3d9c67d97d6f 418 x3 = *__SIMD32(px);
emh203 0:3d9c67d97d6f 419
emh203 0:3d9c67d97d6f 420 /* Read x[9] */
emh203 0:3d9c67d97d6f 421 x2 = _SIMD32_OFFSET(px+1);
emh203 0:3d9c67d97d6f 422
emh203 0:3d9c67d97d6f 423 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 424 acc0 = __SMLADX(x0, c0, acc0);
emh203 0:3d9c67d97d6f 425 acc1 = __SMLADX(x1, c0, acc1);
emh203 0:3d9c67d97d6f 426 acc2 = __SMLADX(x3, c0, acc2);
emh203 0:3d9c67d97d6f 427 acc3 = __SMLADX(x2, c0, acc3);
emh203 0:3d9c67d97d6f 428
emh203 0:3d9c67d97d6f 429 /* Read y[srcBLen - 7] */
emh203 0:3d9c67d97d6f 430 c0 = *(py-1);
emh203 0:3d9c67d97d6f 431 #ifdef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 432
emh203 0:3d9c67d97d6f 433 c0 = c0 << 16u;
emh203 0:3d9c67d97d6f 434 #else
emh203 0:3d9c67d97d6f 435
emh203 0:3d9c67d97d6f 436 c0 = c0 & 0x0000FFFF;
emh203 0:3d9c67d97d6f 437 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 438
emh203 0:3d9c67d97d6f 439 /* Read x[10] */
emh203 0:3d9c67d97d6f 440 x3 = _SIMD32_OFFSET(px+2);
emh203 0:3d9c67d97d6f 441 px += 3u;
emh203 0:3d9c67d97d6f 442
emh203 0:3d9c67d97d6f 443 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 444 acc0 = __SMLADX(x1, c0, acc0);
emh203 0:3d9c67d97d6f 445 acc1 = __SMLAD(x2, c0, acc1);
emh203 0:3d9c67d97d6f 446 acc2 = __SMLADX(x2, c0, acc2);
emh203 0:3d9c67d97d6f 447 acc3 = __SMLADX(x3, c0, acc3);
emh203 0:3d9c67d97d6f 448 }
emh203 0:3d9c67d97d6f 449
emh203 0:3d9c67d97d6f 450 /* Store the results in the accumulators in the destination buffer. */
emh203 0:3d9c67d97d6f 451 #ifndef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 452
emh203 0:3d9c67d97d6f 453 *__SIMD32(pOut)++ = __PKHBT((acc0 >> 15), (acc1 >> 15), 16);
emh203 0:3d9c67d97d6f 454 *__SIMD32(pOut)++ = __PKHBT((acc2 >> 15), (acc3 >> 15), 16);
emh203 0:3d9c67d97d6f 455
emh203 0:3d9c67d97d6f 456 #else
emh203 0:3d9c67d97d6f 457
emh203 0:3d9c67d97d6f 458 *__SIMD32(pOut)++ = __PKHBT((acc1 >> 15), (acc0 >> 15), 16);
emh203 0:3d9c67d97d6f 459 *__SIMD32(pOut)++ = __PKHBT((acc3 >> 15), (acc2 >> 15), 16);
emh203 0:3d9c67d97d6f 460
emh203 0:3d9c67d97d6f 461 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 462
emh203 0:3d9c67d97d6f 463 /* Increment the pointer pIn1 index, count by 4 */
emh203 0:3d9c67d97d6f 464 count += 4u;
emh203 0:3d9c67d97d6f 465
emh203 0:3d9c67d97d6f 466 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 467 px = pIn1 + count;
emh203 0:3d9c67d97d6f 468 py = pSrc2;
emh203 0:3d9c67d97d6f 469
emh203 0:3d9c67d97d6f 470 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 471 blkCnt--;
emh203 0:3d9c67d97d6f 472 }
emh203 0:3d9c67d97d6f 473
emh203 0:3d9c67d97d6f 474 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emh203 0:3d9c67d97d6f 475 ** No loop unrolling is used. */
emh203 0:3d9c67d97d6f 476 blkCnt = blockSize2 % 0x4u;
emh203 0:3d9c67d97d6f 477
emh203 0:3d9c67d97d6f 478 while(blkCnt > 0u)
emh203 0:3d9c67d97d6f 479 {
emh203 0:3d9c67d97d6f 480 /* Accumulator is made zero for every iteration */
emh203 0:3d9c67d97d6f 481 sum = 0;
emh203 0:3d9c67d97d6f 482
emh203 0:3d9c67d97d6f 483 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emh203 0:3d9c67d97d6f 484 k = srcBLen >> 2u;
emh203 0:3d9c67d97d6f 485
emh203 0:3d9c67d97d6f 486 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emh203 0:3d9c67d97d6f 487 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emh203 0:3d9c67d97d6f 488 while(k > 0u)
emh203 0:3d9c67d97d6f 489 {
emh203 0:3d9c67d97d6f 490 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 491 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 492 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 493 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 494 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 495
emh203 0:3d9c67d97d6f 496 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 497 k--;
emh203 0:3d9c67d97d6f 498 }
emh203 0:3d9c67d97d6f 499
emh203 0:3d9c67d97d6f 500 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emh203 0:3d9c67d97d6f 501 ** No loop unrolling is used. */
emh203 0:3d9c67d97d6f 502 k = srcBLen % 0x4u;
emh203 0:3d9c67d97d6f 503
emh203 0:3d9c67d97d6f 504 while(k > 0u)
emh203 0:3d9c67d97d6f 505 {
emh203 0:3d9c67d97d6f 506 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 507 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 508
emh203 0:3d9c67d97d6f 509 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 510 k--;
emh203 0:3d9c67d97d6f 511 }
emh203 0:3d9c67d97d6f 512
emh203 0:3d9c67d97d6f 513 /* Store the result in the accumulator in the destination buffer. */
emh203 0:3d9c67d97d6f 514 *pOut++ = (q15_t) (sum >> 15);
emh203 0:3d9c67d97d6f 515
emh203 0:3d9c67d97d6f 516 /* Increment the pointer pIn1 index, count by 1 */
emh203 0:3d9c67d97d6f 517 count++;
emh203 0:3d9c67d97d6f 518
emh203 0:3d9c67d97d6f 519 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 520 px = pIn1 + count;
emh203 0:3d9c67d97d6f 521 py = pSrc2;
emh203 0:3d9c67d97d6f 522
emh203 0:3d9c67d97d6f 523 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 524 blkCnt--;
emh203 0:3d9c67d97d6f 525 }
emh203 0:3d9c67d97d6f 526 }
emh203 0:3d9c67d97d6f 527 else
emh203 0:3d9c67d97d6f 528 {
emh203 0:3d9c67d97d6f 529 /* If the srcBLen is not a multiple of 4,
emh203 0:3d9c67d97d6f 530 * the blockSize2 loop cannot be unrolled by 4 */
emh203 0:3d9c67d97d6f 531 blkCnt = blockSize2;
emh203 0:3d9c67d97d6f 532
emh203 0:3d9c67d97d6f 533 while(blkCnt > 0u)
emh203 0:3d9c67d97d6f 534 {
emh203 0:3d9c67d97d6f 535 /* Accumulator is made zero for every iteration */
emh203 0:3d9c67d97d6f 536 sum = 0;
emh203 0:3d9c67d97d6f 537
emh203 0:3d9c67d97d6f 538 /* srcBLen number of MACS should be performed */
emh203 0:3d9c67d97d6f 539 k = srcBLen;
emh203 0:3d9c67d97d6f 540
emh203 0:3d9c67d97d6f 541 while(k > 0u)
emh203 0:3d9c67d97d6f 542 {
emh203 0:3d9c67d97d6f 543 /* Perform the multiply-accumulate */
emh203 0:3d9c67d97d6f 544 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 545
emh203 0:3d9c67d97d6f 546 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 547 k--;
emh203 0:3d9c67d97d6f 548 }
emh203 0:3d9c67d97d6f 549
emh203 0:3d9c67d97d6f 550 /* Store the result in the accumulator in the destination buffer. */
emh203 0:3d9c67d97d6f 551 *pOut++ = (q15_t) (sum >> 15);
emh203 0:3d9c67d97d6f 552
emh203 0:3d9c67d97d6f 553 /* Increment the MAC count */
emh203 0:3d9c67d97d6f 554 count++;
emh203 0:3d9c67d97d6f 555
emh203 0:3d9c67d97d6f 556 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 557 px = pIn1 + count;
emh203 0:3d9c67d97d6f 558 py = pSrc2;
emh203 0:3d9c67d97d6f 559
emh203 0:3d9c67d97d6f 560 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 561 blkCnt--;
emh203 0:3d9c67d97d6f 562 }
emh203 0:3d9c67d97d6f 563 }
emh203 0:3d9c67d97d6f 564
emh203 0:3d9c67d97d6f 565
emh203 0:3d9c67d97d6f 566 /* --------------------------
emh203 0:3d9c67d97d6f 567 * Initializations of stage3
emh203 0:3d9c67d97d6f 568 * -------------------------*/
emh203 0:3d9c67d97d6f 569
emh203 0:3d9c67d97d6f 570 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emh203 0:3d9c67d97d6f 571 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emh203 0:3d9c67d97d6f 572 * ....
emh203 0:3d9c67d97d6f 573 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emh203 0:3d9c67d97d6f 574 * sum += x[srcALen-1] * y[srcBLen-1]
emh203 0:3d9c67d97d6f 575 */
emh203 0:3d9c67d97d6f 576
emh203 0:3d9c67d97d6f 577 /* In this stage the MAC operations are decreased by 1 for every iteration.
emh203 0:3d9c67d97d6f 578 The blockSize3 variable holds the number of MAC operations performed */
emh203 0:3d9c67d97d6f 579
emh203 0:3d9c67d97d6f 580 /* Working pointer of inputA */
emh203 0:3d9c67d97d6f 581 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emh203 0:3d9c67d97d6f 582 px = pSrc1;
emh203 0:3d9c67d97d6f 583
emh203 0:3d9c67d97d6f 584 /* Working pointer of inputB */
emh203 0:3d9c67d97d6f 585 pSrc2 = pIn2 + (srcBLen - 1u);
emh203 0:3d9c67d97d6f 586 pIn2 = pSrc2 - 1u;
emh203 0:3d9c67d97d6f 587 py = pIn2;
emh203 0:3d9c67d97d6f 588
emh203 0:3d9c67d97d6f 589 /* -------------------
emh203 0:3d9c67d97d6f 590 * Stage3 process
emh203 0:3d9c67d97d6f 591 * ------------------*/
emh203 0:3d9c67d97d6f 592
emh203 0:3d9c67d97d6f 593 /* For loop unrolling by 4, this stage is divided into two. */
emh203 0:3d9c67d97d6f 594 /* First part of this stage computes the MAC operations greater than 4 */
emh203 0:3d9c67d97d6f 595 /* Second part of this stage computes the MAC operations less than or equal to 4 */
emh203 0:3d9c67d97d6f 596
emh203 0:3d9c67d97d6f 597 /* The first part of the stage starts here */
emh203 0:3d9c67d97d6f 598 j = blockSize3 >> 2u;
emh203 0:3d9c67d97d6f 599
emh203 0:3d9c67d97d6f 600 while((j > 0u) && (blockSize3 > 0u))
emh203 0:3d9c67d97d6f 601 {
emh203 0:3d9c67d97d6f 602 /* Accumulator is made zero for every iteration */
emh203 0:3d9c67d97d6f 603 sum = 0;
emh203 0:3d9c67d97d6f 604
emh203 0:3d9c67d97d6f 605 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emh203 0:3d9c67d97d6f 606 k = blockSize3 >> 2u;
emh203 0:3d9c67d97d6f 607
emh203 0:3d9c67d97d6f 608 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emh203 0:3d9c67d97d6f 609 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emh203 0:3d9c67d97d6f 610 while(k > 0u)
emh203 0:3d9c67d97d6f 611 {
emh203 0:3d9c67d97d6f 612 /* x[srcALen - srcBLen + 1], x[srcALen - srcBLen + 2] are multiplied
emh203 0:3d9c67d97d6f 613 * with y[srcBLen - 1], y[srcBLen - 2] respectively */
emh203 0:3d9c67d97d6f 614 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emh203 0:3d9c67d97d6f 615 /* x[srcALen - srcBLen + 3], x[srcALen - srcBLen + 4] are multiplied
emh203 0:3d9c67d97d6f 616 * with y[srcBLen - 3], y[srcBLen - 4] respectively */
emh203 0:3d9c67d97d6f 617 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emh203 0:3d9c67d97d6f 618
emh203 0:3d9c67d97d6f 619 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 620 k--;
emh203 0:3d9c67d97d6f 621 }
emh203 0:3d9c67d97d6f 622
emh203 0:3d9c67d97d6f 623 /* For the next MAC operations, the pointer py is used without SIMD
emh203 0:3d9c67d97d6f 624 * So, py is incremented by 1 */
emh203 0:3d9c67d97d6f 625 py = py + 1u;
emh203 0:3d9c67d97d6f 626
emh203 0:3d9c67d97d6f 627 /* If the blockSize3 is not a multiple of 4, compute any remaining MACs here.
emh203 0:3d9c67d97d6f 628 ** No loop unrolling is used. */
emh203 0:3d9c67d97d6f 629 k = blockSize3 % 0x4u;
emh203 0:3d9c67d97d6f 630
emh203 0:3d9c67d97d6f 631 while(k > 0u)
emh203 0:3d9c67d97d6f 632 {
emh203 0:3d9c67d97d6f 633 /* sum += x[srcALen - srcBLen + 5] * y[srcBLen - 5] */
emh203 0:3d9c67d97d6f 634 sum = __SMLAD(*px++, *py--, sum);
emh203 0:3d9c67d97d6f 635
emh203 0:3d9c67d97d6f 636 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 637 k--;
emh203 0:3d9c67d97d6f 638 }
emh203 0:3d9c67d97d6f 639
emh203 0:3d9c67d97d6f 640 /* Store the result in the accumulator in the destination buffer. */
emh203 0:3d9c67d97d6f 641 *pOut++ = (q15_t) (sum >> 15);
emh203 0:3d9c67d97d6f 642
emh203 0:3d9c67d97d6f 643 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 644 px = ++pSrc1;
emh203 0:3d9c67d97d6f 645 py = pIn2;
emh203 0:3d9c67d97d6f 646
emh203 0:3d9c67d97d6f 647 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 648 blockSize3--;
emh203 0:3d9c67d97d6f 649
emh203 0:3d9c67d97d6f 650 j--;
emh203 0:3d9c67d97d6f 651 }
emh203 0:3d9c67d97d6f 652
emh203 0:3d9c67d97d6f 653 /* The second part of the stage starts here */
emh203 0:3d9c67d97d6f 654 /* SIMD is not used for the next MAC operations,
emh203 0:3d9c67d97d6f 655 * so pointer py is updated to read only one sample at a time */
emh203 0:3d9c67d97d6f 656 py = py + 1u;
emh203 0:3d9c67d97d6f 657
emh203 0:3d9c67d97d6f 658 while(blockSize3 > 0u)
emh203 0:3d9c67d97d6f 659 {
emh203 0:3d9c67d97d6f 660 /* Accumulator is made zero for every iteration */
emh203 0:3d9c67d97d6f 661 sum = 0;
emh203 0:3d9c67d97d6f 662
emh203 0:3d9c67d97d6f 663 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emh203 0:3d9c67d97d6f 664 k = blockSize3;
emh203 0:3d9c67d97d6f 665
emh203 0:3d9c67d97d6f 666 while(k > 0u)
emh203 0:3d9c67d97d6f 667 {
emh203 0:3d9c67d97d6f 668 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 669 /* sum += x[srcALen-1] * y[srcBLen-1] */
emh203 0:3d9c67d97d6f 670 sum = __SMLAD(*px++, *py--, sum);
emh203 0:3d9c67d97d6f 671
emh203 0:3d9c67d97d6f 672 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 673 k--;
emh203 0:3d9c67d97d6f 674 }
emh203 0:3d9c67d97d6f 675
emh203 0:3d9c67d97d6f 676 /* Store the result in the accumulator in the destination buffer. */
emh203 0:3d9c67d97d6f 677 *pOut++ = (q15_t) (sum >> 15);
emh203 0:3d9c67d97d6f 678
emh203 0:3d9c67d97d6f 679 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 680 px = ++pSrc1;
emh203 0:3d9c67d97d6f 681 py = pSrc2;
emh203 0:3d9c67d97d6f 682
emh203 0:3d9c67d97d6f 683 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 684 blockSize3--;
emh203 0:3d9c67d97d6f 685 }
emh203 0:3d9c67d97d6f 686
emh203 0:3d9c67d97d6f 687 #else
emh203 0:3d9c67d97d6f 688 q15_t *pIn1; /* inputA pointer */
emh203 0:3d9c67d97d6f 689 q15_t *pIn2; /* inputB pointer */
emh203 0:3d9c67d97d6f 690 q15_t *pOut = pDst; /* output pointer */
emh203 0:3d9c67d97d6f 691 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
emh203 0:3d9c67d97d6f 692 q15_t *px; /* Intermediate inputA pointer */
emh203 0:3d9c67d97d6f 693 q15_t *py; /* Intermediate inputB pointer */
emh203 0:3d9c67d97d6f 694 q15_t *pSrc1, *pSrc2; /* Intermediate pointers */
emh203 0:3d9c67d97d6f 695 q31_t x0, x1, x2, x3, c0; /* Temporary variables to hold state and coefficient values */
emh203 0:3d9c67d97d6f 696 uint32_t blockSize1, blockSize2, blockSize3, j, k, count, blkCnt; /* loop counter */
emh203 0:3d9c67d97d6f 697 q15_t a, b;
emh203 0:3d9c67d97d6f 698
emh203 0:3d9c67d97d6f 699 /* The algorithm implementation is based on the lengths of the inputs. */
emh203 0:3d9c67d97d6f 700 /* srcB is always made to slide across srcA. */
emh203 0:3d9c67d97d6f 701 /* So srcBLen is always considered as shorter or equal to srcALen */
emh203 0:3d9c67d97d6f 702 if(srcALen >= srcBLen)
emh203 0:3d9c67d97d6f 703 {
emh203 0:3d9c67d97d6f 704 /* Initialization of inputA pointer */
emh203 0:3d9c67d97d6f 705 pIn1 = pSrcA;
emh203 0:3d9c67d97d6f 706
emh203 0:3d9c67d97d6f 707 /* Initialization of inputB pointer */
emh203 0:3d9c67d97d6f 708 pIn2 = pSrcB;
emh203 0:3d9c67d97d6f 709 }
emh203 0:3d9c67d97d6f 710 else
emh203 0:3d9c67d97d6f 711 {
emh203 0:3d9c67d97d6f 712 /* Initialization of inputA pointer */
emh203 0:3d9c67d97d6f 713 pIn1 = pSrcB;
emh203 0:3d9c67d97d6f 714
emh203 0:3d9c67d97d6f 715 /* Initialization of inputB pointer */
emh203 0:3d9c67d97d6f 716 pIn2 = pSrcA;
emh203 0:3d9c67d97d6f 717
emh203 0:3d9c67d97d6f 718 /* srcBLen is always considered as shorter or equal to srcALen */
emh203 0:3d9c67d97d6f 719 j = srcBLen;
emh203 0:3d9c67d97d6f 720 srcBLen = srcALen;
emh203 0:3d9c67d97d6f 721 srcALen = j;
emh203 0:3d9c67d97d6f 722 }
emh203 0:3d9c67d97d6f 723
emh203 0:3d9c67d97d6f 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] */
emh203 0:3d9c67d97d6f 725 /* The function is internally
emh203 0:3d9c67d97d6f 726 * divided into three stages according to the number of multiplications that has to be
emh203 0:3d9c67d97d6f 727 * taken place between inputA samples and inputB samples. In the first stage of the
emh203 0:3d9c67d97d6f 728 * algorithm, the multiplications increase by one for every iteration.
emh203 0:3d9c67d97d6f 729 * In the second stage of the algorithm, srcBLen number of multiplications are done.
emh203 0:3d9c67d97d6f 730 * In the third stage of the algorithm, the multiplications decrease by one
emh203 0:3d9c67d97d6f 731 * for every iteration. */
emh203 0:3d9c67d97d6f 732
emh203 0:3d9c67d97d6f 733 /* The algorithm is implemented in three stages.
emh203 0:3d9c67d97d6f 734 The loop counters of each stage is initiated here. */
emh203 0:3d9c67d97d6f 735 blockSize1 = srcBLen - 1u;
emh203 0:3d9c67d97d6f 736 blockSize2 = srcALen - (srcBLen - 1u);
emh203 0:3d9c67d97d6f 737 blockSize3 = blockSize1;
emh203 0:3d9c67d97d6f 738
emh203 0:3d9c67d97d6f 739 /* --------------------------
emh203 0:3d9c67d97d6f 740 * Initializations of stage1
emh203 0:3d9c67d97d6f 741 * -------------------------*/
emh203 0:3d9c67d97d6f 742
emh203 0:3d9c67d97d6f 743 /* sum = x[0] * y[0]
emh203 0:3d9c67d97d6f 744 * sum = x[0] * y[1] + x[1] * y[0]
emh203 0:3d9c67d97d6f 745 * ....
emh203 0:3d9c67d97d6f 746 * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
emh203 0:3d9c67d97d6f 747 */
emh203 0:3d9c67d97d6f 748
emh203 0:3d9c67d97d6f 749 /* In this stage the MAC operations are increased by 1 for every iteration.
emh203 0:3d9c67d97d6f 750 The count variable holds the number of MAC operations performed */
emh203 0:3d9c67d97d6f 751 count = 1u;
emh203 0:3d9c67d97d6f 752
emh203 0:3d9c67d97d6f 753 /* Working pointer of inputA */
emh203 0:3d9c67d97d6f 754 px = pIn1;
emh203 0:3d9c67d97d6f 755
emh203 0:3d9c67d97d6f 756 /* Working pointer of inputB */
emh203 0:3d9c67d97d6f 757 py = pIn2;
emh203 0:3d9c67d97d6f 758
emh203 0:3d9c67d97d6f 759
emh203 0:3d9c67d97d6f 760 /* ------------------------
emh203 0:3d9c67d97d6f 761 * Stage1 process
emh203 0:3d9c67d97d6f 762 * ----------------------*/
emh203 0:3d9c67d97d6f 763
emh203 0:3d9c67d97d6f 764 /* For loop unrolling by 4, this stage is divided into two. */
emh203 0:3d9c67d97d6f 765 /* First part of this stage computes the MAC operations less than 4 */
emh203 0:3d9c67d97d6f 766 /* Second part of this stage computes the MAC operations greater than or equal to 4 */
emh203 0:3d9c67d97d6f 767
emh203 0:3d9c67d97d6f 768 /* The first part of the stage starts here */
emh203 0:3d9c67d97d6f 769 while((count < 4u) && (blockSize1 > 0u))
emh203 0:3d9c67d97d6f 770 {
emh203 0:3d9c67d97d6f 771 /* Accumulator is made zero for every iteration */
emh203 0:3d9c67d97d6f 772 sum = 0;
emh203 0:3d9c67d97d6f 773
emh203 0:3d9c67d97d6f 774 /* Loop over number of MAC operations between
emh203 0:3d9c67d97d6f 775 * inputA samples and inputB samples */
emh203 0:3d9c67d97d6f 776 k = count;
emh203 0:3d9c67d97d6f 777
emh203 0:3d9c67d97d6f 778 while(k > 0u)
emh203 0:3d9c67d97d6f 779 {
emh203 0:3d9c67d97d6f 780 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 781 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 782
emh203 0:3d9c67d97d6f 783 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 784 k--;
emh203 0:3d9c67d97d6f 785 }
emh203 0:3d9c67d97d6f 786
emh203 0:3d9c67d97d6f 787 /* Store the result in the accumulator in the destination buffer. */
emh203 0:3d9c67d97d6f 788 *pOut++ = (q15_t) (sum >> 15);
emh203 0:3d9c67d97d6f 789
emh203 0:3d9c67d97d6f 790 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 791 py = pIn2 + count;
emh203 0:3d9c67d97d6f 792 px = pIn1;
emh203 0:3d9c67d97d6f 793
emh203 0:3d9c67d97d6f 794 /* Increment the MAC count */
emh203 0:3d9c67d97d6f 795 count++;
emh203 0:3d9c67d97d6f 796
emh203 0:3d9c67d97d6f 797 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 798 blockSize1--;
emh203 0:3d9c67d97d6f 799 }
emh203 0:3d9c67d97d6f 800
emh203 0:3d9c67d97d6f 801 /* The second part of the stage starts here */
emh203 0:3d9c67d97d6f 802 /* The internal loop, over count, is unrolled by 4 */
emh203 0:3d9c67d97d6f 803 /* To, read the last two inputB samples using SIMD:
emh203 0:3d9c67d97d6f 804 * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */
emh203 0:3d9c67d97d6f 805 py = py - 1;
emh203 0:3d9c67d97d6f 806
emh203 0:3d9c67d97d6f 807 while(blockSize1 > 0u)
emh203 0:3d9c67d97d6f 808 {
emh203 0:3d9c67d97d6f 809 /* Accumulator is made zero for every iteration */
emh203 0:3d9c67d97d6f 810 sum = 0;
emh203 0:3d9c67d97d6f 811
emh203 0:3d9c67d97d6f 812 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emh203 0:3d9c67d97d6f 813 k = count >> 2u;
emh203 0:3d9c67d97d6f 814
emh203 0:3d9c67d97d6f 815 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emh203 0:3d9c67d97d6f 816 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emh203 0:3d9c67d97d6f 817 py++;
emh203 0:3d9c67d97d6f 818
emh203 0:3d9c67d97d6f 819 while(k > 0u)
emh203 0:3d9c67d97d6f 820 {
emh203 0:3d9c67d97d6f 821 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 822 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 823 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 824 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 825 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 826
emh203 0:3d9c67d97d6f 827 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 828 k--;
emh203 0:3d9c67d97d6f 829 }
emh203 0:3d9c67d97d6f 830
emh203 0:3d9c67d97d6f 831 /* If the count is not a multiple of 4, compute any remaining MACs here.
emh203 0:3d9c67d97d6f 832 ** No loop unrolling is used. */
emh203 0:3d9c67d97d6f 833 k = count % 0x4u;
emh203 0:3d9c67d97d6f 834
emh203 0:3d9c67d97d6f 835 while(k > 0u)
emh203 0:3d9c67d97d6f 836 {
emh203 0:3d9c67d97d6f 837 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 838 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 839
emh203 0:3d9c67d97d6f 840 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 841 k--;
emh203 0:3d9c67d97d6f 842 }
emh203 0:3d9c67d97d6f 843
emh203 0:3d9c67d97d6f 844 /* Store the result in the accumulator in the destination buffer. */
emh203 0:3d9c67d97d6f 845 *pOut++ = (q15_t) (sum >> 15);
emh203 0:3d9c67d97d6f 846
emh203 0:3d9c67d97d6f 847 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 848 py = pIn2 + (count - 1u);
emh203 0:3d9c67d97d6f 849 px = pIn1;
emh203 0:3d9c67d97d6f 850
emh203 0:3d9c67d97d6f 851 /* Increment the MAC count */
emh203 0:3d9c67d97d6f 852 count++;
emh203 0:3d9c67d97d6f 853
emh203 0:3d9c67d97d6f 854 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 855 blockSize1--;
emh203 0:3d9c67d97d6f 856 }
emh203 0:3d9c67d97d6f 857
emh203 0:3d9c67d97d6f 858 /* --------------------------
emh203 0:3d9c67d97d6f 859 * Initializations of stage2
emh203 0:3d9c67d97d6f 860 * ------------------------*/
emh203 0:3d9c67d97d6f 861
emh203 0:3d9c67d97d6f 862 /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
emh203 0:3d9c67d97d6f 863 * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
emh203 0:3d9c67d97d6f 864 * ....
emh203 0:3d9c67d97d6f 865 * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
emh203 0:3d9c67d97d6f 866 */
emh203 0:3d9c67d97d6f 867
emh203 0:3d9c67d97d6f 868 /* Working pointer of inputA */
emh203 0:3d9c67d97d6f 869 px = pIn1;
emh203 0:3d9c67d97d6f 870
emh203 0:3d9c67d97d6f 871 /* Working pointer of inputB */
emh203 0:3d9c67d97d6f 872 pSrc2 = pIn2 + (srcBLen - 1u);
emh203 0:3d9c67d97d6f 873 py = pSrc2;
emh203 0:3d9c67d97d6f 874
emh203 0:3d9c67d97d6f 875 /* count is the index by which the pointer pIn1 to be incremented */
emh203 0:3d9c67d97d6f 876 count = 0u;
emh203 0:3d9c67d97d6f 877
emh203 0:3d9c67d97d6f 878
emh203 0:3d9c67d97d6f 879 /* --------------------
emh203 0:3d9c67d97d6f 880 * Stage2 process
emh203 0:3d9c67d97d6f 881 * -------------------*/
emh203 0:3d9c67d97d6f 882
emh203 0:3d9c67d97d6f 883 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emh203 0:3d9c67d97d6f 884 * So, to loop unroll over blockSize2,
emh203 0:3d9c67d97d6f 885 * srcBLen should be greater than or equal to 4 */
emh203 0:3d9c67d97d6f 886 if(srcBLen >= 4u)
emh203 0:3d9c67d97d6f 887 {
emh203 0:3d9c67d97d6f 888 /* Loop unroll over blockSize2, by 4 */
emh203 0:3d9c67d97d6f 889 blkCnt = blockSize2 >> 2u;
emh203 0:3d9c67d97d6f 890
emh203 0:3d9c67d97d6f 891 while(blkCnt > 0u)
emh203 0:3d9c67d97d6f 892 {
emh203 0:3d9c67d97d6f 893 py = py - 1u;
emh203 0:3d9c67d97d6f 894
emh203 0:3d9c67d97d6f 895 /* Set all accumulators to zero */
emh203 0:3d9c67d97d6f 896 acc0 = 0;
emh203 0:3d9c67d97d6f 897 acc1 = 0;
emh203 0:3d9c67d97d6f 898 acc2 = 0;
emh203 0:3d9c67d97d6f 899 acc3 = 0;
emh203 0:3d9c67d97d6f 900
emh203 0:3d9c67d97d6f 901 /* read x[0], x[1] samples */
emh203 0:3d9c67d97d6f 902 a = *px++;
emh203 0:3d9c67d97d6f 903 b = *px++;
emh203 0:3d9c67d97d6f 904
emh203 0:3d9c67d97d6f 905 #ifndef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 906
emh203 0:3d9c67d97d6f 907 x0 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 908 a = *px;
emh203 0:3d9c67d97d6f 909 x1 = __PKHBT(b, a, 16);
emh203 0:3d9c67d97d6f 910
emh203 0:3d9c67d97d6f 911 #else
emh203 0:3d9c67d97d6f 912
emh203 0:3d9c67d97d6f 913 x0 = __PKHBT(b, a, 16);
emh203 0:3d9c67d97d6f 914 a = *px;
emh203 0:3d9c67d97d6f 915 x1 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 916
emh203 0:3d9c67d97d6f 917 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 918
emh203 0:3d9c67d97d6f 919 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emh203 0:3d9c67d97d6f 920 k = srcBLen >> 2u;
emh203 0:3d9c67d97d6f 921
emh203 0:3d9c67d97d6f 922 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emh203 0:3d9c67d97d6f 923 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emh203 0:3d9c67d97d6f 924 do
emh203 0:3d9c67d97d6f 925 {
emh203 0:3d9c67d97d6f 926 /* Read the last two inputB samples using SIMD:
emh203 0:3d9c67d97d6f 927 * y[srcBLen - 1] and y[srcBLen - 2] */
emh203 0:3d9c67d97d6f 928 a = *py;
emh203 0:3d9c67d97d6f 929 b = *(py+1);
emh203 0:3d9c67d97d6f 930 py -= 2;
emh203 0:3d9c67d97d6f 931
emh203 0:3d9c67d97d6f 932 #ifndef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 933
emh203 0:3d9c67d97d6f 934 c0 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 935
emh203 0:3d9c67d97d6f 936 #else
emh203 0:3d9c67d97d6f 937
emh203 0:3d9c67d97d6f 938 c0 = __PKHBT(b, a, 16);;
emh203 0:3d9c67d97d6f 939
emh203 0:3d9c67d97d6f 940 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 941
emh203 0:3d9c67d97d6f 942 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emh203 0:3d9c67d97d6f 943 acc0 = __SMLADX(x0, c0, acc0);
emh203 0:3d9c67d97d6f 944
emh203 0:3d9c67d97d6f 945 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emh203 0:3d9c67d97d6f 946 acc1 = __SMLADX(x1, c0, acc1);
emh203 0:3d9c67d97d6f 947
emh203 0:3d9c67d97d6f 948 a = *px;
emh203 0:3d9c67d97d6f 949 b = *(px + 1);
emh203 0:3d9c67d97d6f 950
emh203 0:3d9c67d97d6f 951 #ifndef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 952
emh203 0:3d9c67d97d6f 953 x2 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 954 a = *(px + 2);
emh203 0:3d9c67d97d6f 955 x3 = __PKHBT(b, a, 16);
emh203 0:3d9c67d97d6f 956
emh203 0:3d9c67d97d6f 957 #else
emh203 0:3d9c67d97d6f 958
emh203 0:3d9c67d97d6f 959 x2 = __PKHBT(b, a, 16);
emh203 0:3d9c67d97d6f 960 a = *(px + 2);
emh203 0:3d9c67d97d6f 961 x3 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 962
emh203 0:3d9c67d97d6f 963 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 964
emh203 0:3d9c67d97d6f 965 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emh203 0:3d9c67d97d6f 966 acc2 = __SMLADX(x2, c0, acc2);
emh203 0:3d9c67d97d6f 967
emh203 0:3d9c67d97d6f 968 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emh203 0:3d9c67d97d6f 969 acc3 = __SMLADX(x3, c0, acc3);
emh203 0:3d9c67d97d6f 970
emh203 0:3d9c67d97d6f 971 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emh203 0:3d9c67d97d6f 972 a = *py;
emh203 0:3d9c67d97d6f 973 b = *(py+1);
emh203 0:3d9c67d97d6f 974 py -= 2;
emh203 0:3d9c67d97d6f 975
emh203 0:3d9c67d97d6f 976 #ifndef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 977
emh203 0:3d9c67d97d6f 978 c0 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 979
emh203 0:3d9c67d97d6f 980 #else
emh203 0:3d9c67d97d6f 981
emh203 0:3d9c67d97d6f 982 c0 = __PKHBT(b, a, 16);;
emh203 0:3d9c67d97d6f 983
emh203 0:3d9c67d97d6f 984 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 985
emh203 0:3d9c67d97d6f 986 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emh203 0:3d9c67d97d6f 987 acc0 = __SMLADX(x2, c0, acc0);
emh203 0:3d9c67d97d6f 988
emh203 0:3d9c67d97d6f 989 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emh203 0:3d9c67d97d6f 990 acc1 = __SMLADX(x3, c0, acc1);
emh203 0:3d9c67d97d6f 991
emh203 0:3d9c67d97d6f 992 /* Read x[4], x[5], x[6] */
emh203 0:3d9c67d97d6f 993 a = *(px + 2);
emh203 0:3d9c67d97d6f 994 b = *(px + 3);
emh203 0:3d9c67d97d6f 995
emh203 0:3d9c67d97d6f 996 #ifndef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 997
emh203 0:3d9c67d97d6f 998 x0 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 999 a = *(px + 4);
emh203 0:3d9c67d97d6f 1000 x1 = __PKHBT(b, a, 16);
emh203 0:3d9c67d97d6f 1001
emh203 0:3d9c67d97d6f 1002 #else
emh203 0:3d9c67d97d6f 1003
emh203 0:3d9c67d97d6f 1004 x0 = __PKHBT(b, a, 16);
emh203 0:3d9c67d97d6f 1005 a = *(px + 4);
emh203 0:3d9c67d97d6f 1006 x1 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 1007
emh203 0:3d9c67d97d6f 1008 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 1009
emh203 0:3d9c67d97d6f 1010 px += 4u;
emh203 0:3d9c67d97d6f 1011
emh203 0:3d9c67d97d6f 1012 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emh203 0:3d9c67d97d6f 1013 acc2 = __SMLADX(x0, c0, acc2);
emh203 0:3d9c67d97d6f 1014
emh203 0:3d9c67d97d6f 1015 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emh203 0:3d9c67d97d6f 1016 acc3 = __SMLADX(x1, c0, acc3);
emh203 0:3d9c67d97d6f 1017
emh203 0:3d9c67d97d6f 1018 } while(--k);
emh203 0:3d9c67d97d6f 1019
emh203 0:3d9c67d97d6f 1020 /* For the next MAC operations, SIMD is not used
emh203 0:3d9c67d97d6f 1021 * So, the 16 bit pointer if inputB, py is updated */
emh203 0:3d9c67d97d6f 1022
emh203 0:3d9c67d97d6f 1023 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emh203 0:3d9c67d97d6f 1024 ** No loop unrolling is used. */
emh203 0:3d9c67d97d6f 1025 k = srcBLen % 0x4u;
emh203 0:3d9c67d97d6f 1026
emh203 0:3d9c67d97d6f 1027 if(k == 1u)
emh203 0:3d9c67d97d6f 1028 {
emh203 0:3d9c67d97d6f 1029 /* Read y[srcBLen - 5] */
emh203 0:3d9c67d97d6f 1030 c0 = *(py+1);
emh203 0:3d9c67d97d6f 1031
emh203 0:3d9c67d97d6f 1032 #ifdef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 1033
emh203 0:3d9c67d97d6f 1034 c0 = c0 << 16u;
emh203 0:3d9c67d97d6f 1035
emh203 0:3d9c67d97d6f 1036 #else
emh203 0:3d9c67d97d6f 1037
emh203 0:3d9c67d97d6f 1038 c0 = c0 & 0x0000FFFF;
emh203 0:3d9c67d97d6f 1039
emh203 0:3d9c67d97d6f 1040 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 1041
emh203 0:3d9c67d97d6f 1042 /* Read x[7] */
emh203 0:3d9c67d97d6f 1043 a = *px;
emh203 0:3d9c67d97d6f 1044 b = *(px+1);
emh203 0:3d9c67d97d6f 1045 px++;
emh203 0:3d9c67d97d6f 1046
emh203 0:3d9c67d97d6f 1047 #ifndef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 1048
emh203 0:3d9c67d97d6f 1049 x3 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 1050
emh203 0:3d9c67d97d6f 1051 #else
emh203 0:3d9c67d97d6f 1052
emh203 0:3d9c67d97d6f 1053 x3 = __PKHBT(b, a, 16);;
emh203 0:3d9c67d97d6f 1054
emh203 0:3d9c67d97d6f 1055 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 1056
emh203 0:3d9c67d97d6f 1057
emh203 0:3d9c67d97d6f 1058 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 1059 acc0 = __SMLAD(x0, c0, acc0);
emh203 0:3d9c67d97d6f 1060 acc1 = __SMLAD(x1, c0, acc1);
emh203 0:3d9c67d97d6f 1061 acc2 = __SMLADX(x1, c0, acc2);
emh203 0:3d9c67d97d6f 1062 acc3 = __SMLADX(x3, c0, acc3);
emh203 0:3d9c67d97d6f 1063 }
emh203 0:3d9c67d97d6f 1064
emh203 0:3d9c67d97d6f 1065 if(k == 2u)
emh203 0:3d9c67d97d6f 1066 {
emh203 0:3d9c67d97d6f 1067 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emh203 0:3d9c67d97d6f 1068 a = *py;
emh203 0:3d9c67d97d6f 1069 b = *(py+1);
emh203 0:3d9c67d97d6f 1070
emh203 0:3d9c67d97d6f 1071 #ifndef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 1072
emh203 0:3d9c67d97d6f 1073 c0 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 1074
emh203 0:3d9c67d97d6f 1075 #else
emh203 0:3d9c67d97d6f 1076
emh203 0:3d9c67d97d6f 1077 c0 = __PKHBT(b, a, 16);;
emh203 0:3d9c67d97d6f 1078
emh203 0:3d9c67d97d6f 1079 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 1080
emh203 0:3d9c67d97d6f 1081 /* Read x[7], x[8], x[9] */
emh203 0:3d9c67d97d6f 1082 a = *px;
emh203 0:3d9c67d97d6f 1083 b = *(px + 1);
emh203 0:3d9c67d97d6f 1084
emh203 0:3d9c67d97d6f 1085 #ifndef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 1086
emh203 0:3d9c67d97d6f 1087 x3 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 1088 a = *(px + 2);
emh203 0:3d9c67d97d6f 1089 x2 = __PKHBT(b, a, 16);
emh203 0:3d9c67d97d6f 1090
emh203 0:3d9c67d97d6f 1091 #else
emh203 0:3d9c67d97d6f 1092
emh203 0:3d9c67d97d6f 1093 x3 = __PKHBT(b, a, 16);
emh203 0:3d9c67d97d6f 1094 a = *(px + 2);
emh203 0:3d9c67d97d6f 1095 x2 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 1096
emh203 0:3d9c67d97d6f 1097 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 1098 px += 2u;
emh203 0:3d9c67d97d6f 1099
emh203 0:3d9c67d97d6f 1100 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 1101 acc0 = __SMLADX(x0, c0, acc0);
emh203 0:3d9c67d97d6f 1102 acc1 = __SMLADX(x1, c0, acc1);
emh203 0:3d9c67d97d6f 1103 acc2 = __SMLADX(x3, c0, acc2);
emh203 0:3d9c67d97d6f 1104 acc3 = __SMLADX(x2, c0, acc3);
emh203 0:3d9c67d97d6f 1105 }
emh203 0:3d9c67d97d6f 1106
emh203 0:3d9c67d97d6f 1107 if(k == 3u)
emh203 0:3d9c67d97d6f 1108 {
emh203 0:3d9c67d97d6f 1109 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emh203 0:3d9c67d97d6f 1110 a = *py;
emh203 0:3d9c67d97d6f 1111 b = *(py+1);
emh203 0:3d9c67d97d6f 1112
emh203 0:3d9c67d97d6f 1113 #ifndef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 1114
emh203 0:3d9c67d97d6f 1115 c0 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 1116
emh203 0:3d9c67d97d6f 1117 #else
emh203 0:3d9c67d97d6f 1118
emh203 0:3d9c67d97d6f 1119 c0 = __PKHBT(b, a, 16);;
emh203 0:3d9c67d97d6f 1120
emh203 0:3d9c67d97d6f 1121 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 1122
emh203 0:3d9c67d97d6f 1123 /* Read x[7], x[8], x[9] */
emh203 0:3d9c67d97d6f 1124 a = *px;
emh203 0:3d9c67d97d6f 1125 b = *(px + 1);
emh203 0:3d9c67d97d6f 1126
emh203 0:3d9c67d97d6f 1127 #ifndef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 1128
emh203 0:3d9c67d97d6f 1129 x3 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 1130 a = *(px + 2);
emh203 0:3d9c67d97d6f 1131 x2 = __PKHBT(b, a, 16);
emh203 0:3d9c67d97d6f 1132
emh203 0:3d9c67d97d6f 1133 #else
emh203 0:3d9c67d97d6f 1134
emh203 0:3d9c67d97d6f 1135 x3 = __PKHBT(b, a, 16);
emh203 0:3d9c67d97d6f 1136 a = *(px + 2);
emh203 0:3d9c67d97d6f 1137 x2 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 1138
emh203 0:3d9c67d97d6f 1139 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 1140
emh203 0:3d9c67d97d6f 1141 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 1142 acc0 = __SMLADX(x0, c0, acc0);
emh203 0:3d9c67d97d6f 1143 acc1 = __SMLADX(x1, c0, acc1);
emh203 0:3d9c67d97d6f 1144 acc2 = __SMLADX(x3, c0, acc2);
emh203 0:3d9c67d97d6f 1145 acc3 = __SMLADX(x2, c0, acc3);
emh203 0:3d9c67d97d6f 1146
emh203 0:3d9c67d97d6f 1147 /* Read y[srcBLen - 7] */
emh203 0:3d9c67d97d6f 1148 c0 = *(py-1);
emh203 0:3d9c67d97d6f 1149 #ifdef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 1150
emh203 0:3d9c67d97d6f 1151 c0 = c0 << 16u;
emh203 0:3d9c67d97d6f 1152 #else
emh203 0:3d9c67d97d6f 1153
emh203 0:3d9c67d97d6f 1154 c0 = c0 & 0x0000FFFF;
emh203 0:3d9c67d97d6f 1155 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 1156
emh203 0:3d9c67d97d6f 1157 /* Read x[10] */
emh203 0:3d9c67d97d6f 1158 a = *(px+2);
emh203 0:3d9c67d97d6f 1159 b = *(px+3);
emh203 0:3d9c67d97d6f 1160
emh203 0:3d9c67d97d6f 1161 #ifndef ARM_MATH_BIG_ENDIAN
emh203 0:3d9c67d97d6f 1162
emh203 0:3d9c67d97d6f 1163 x3 = __PKHBT(a, b, 16);
emh203 0:3d9c67d97d6f 1164
emh203 0:3d9c67d97d6f 1165 #else
emh203 0:3d9c67d97d6f 1166
emh203 0:3d9c67d97d6f 1167 x3 = __PKHBT(b, a, 16);;
emh203 0:3d9c67d97d6f 1168
emh203 0:3d9c67d97d6f 1169 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emh203 0:3d9c67d97d6f 1170
emh203 0:3d9c67d97d6f 1171 px += 3u;
emh203 0:3d9c67d97d6f 1172
emh203 0:3d9c67d97d6f 1173 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 1174 acc0 = __SMLADX(x1, c0, acc0);
emh203 0:3d9c67d97d6f 1175 acc1 = __SMLAD(x2, c0, acc1);
emh203 0:3d9c67d97d6f 1176 acc2 = __SMLADX(x2, c0, acc2);
emh203 0:3d9c67d97d6f 1177 acc3 = __SMLADX(x3, c0, acc3);
emh203 0:3d9c67d97d6f 1178 }
emh203 0:3d9c67d97d6f 1179
emh203 0:3d9c67d97d6f 1180 /* Store the results in the accumulators in the destination buffer. */
emh203 0:3d9c67d97d6f 1181 *pOut++ = (q15_t)(acc0 >> 15);
emh203 0:3d9c67d97d6f 1182 *pOut++ = (q15_t)(acc1 >> 15);
emh203 0:3d9c67d97d6f 1183 *pOut++ = (q15_t)(acc2 >> 15);
emh203 0:3d9c67d97d6f 1184 *pOut++ = (q15_t)(acc3 >> 15);
emh203 0:3d9c67d97d6f 1185
emh203 0:3d9c67d97d6f 1186 /* Increment the pointer pIn1 index, count by 4 */
emh203 0:3d9c67d97d6f 1187 count += 4u;
emh203 0:3d9c67d97d6f 1188
emh203 0:3d9c67d97d6f 1189 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 1190 px = pIn1 + count;
emh203 0:3d9c67d97d6f 1191 py = pSrc2;
emh203 0:3d9c67d97d6f 1192
emh203 0:3d9c67d97d6f 1193 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 1194 blkCnt--;
emh203 0:3d9c67d97d6f 1195 }
emh203 0:3d9c67d97d6f 1196
emh203 0:3d9c67d97d6f 1197 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emh203 0:3d9c67d97d6f 1198 ** No loop unrolling is used. */
emh203 0:3d9c67d97d6f 1199 blkCnt = blockSize2 % 0x4u;
emh203 0:3d9c67d97d6f 1200
emh203 0:3d9c67d97d6f 1201 while(blkCnt > 0u)
emh203 0:3d9c67d97d6f 1202 {
emh203 0:3d9c67d97d6f 1203 /* Accumulator is made zero for every iteration */
emh203 0:3d9c67d97d6f 1204 sum = 0;
emh203 0:3d9c67d97d6f 1205
emh203 0:3d9c67d97d6f 1206 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emh203 0:3d9c67d97d6f 1207 k = srcBLen >> 2u;
emh203 0:3d9c67d97d6f 1208
emh203 0:3d9c67d97d6f 1209 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emh203 0:3d9c67d97d6f 1210 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emh203 0:3d9c67d97d6f 1211 while(k > 0u)
emh203 0:3d9c67d97d6f 1212 {
emh203 0:3d9c67d97d6f 1213 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 1214 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 1215 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 1216 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 1217 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 1218
emh203 0:3d9c67d97d6f 1219 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 1220 k--;
emh203 0:3d9c67d97d6f 1221 }
emh203 0:3d9c67d97d6f 1222
emh203 0:3d9c67d97d6f 1223 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emh203 0:3d9c67d97d6f 1224 ** No loop unrolling is used. */
emh203 0:3d9c67d97d6f 1225 k = srcBLen % 0x4u;
emh203 0:3d9c67d97d6f 1226
emh203 0:3d9c67d97d6f 1227 while(k > 0u)
emh203 0:3d9c67d97d6f 1228 {
emh203 0:3d9c67d97d6f 1229 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 1230 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 1231
emh203 0:3d9c67d97d6f 1232 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 1233 k--;
emh203 0:3d9c67d97d6f 1234 }
emh203 0:3d9c67d97d6f 1235
emh203 0:3d9c67d97d6f 1236 /* Store the result in the accumulator in the destination buffer. */
emh203 0:3d9c67d97d6f 1237 *pOut++ = (q15_t) (sum >> 15);
emh203 0:3d9c67d97d6f 1238
emh203 0:3d9c67d97d6f 1239 /* Increment the pointer pIn1 index, count by 1 */
emh203 0:3d9c67d97d6f 1240 count++;
emh203 0:3d9c67d97d6f 1241
emh203 0:3d9c67d97d6f 1242 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 1243 px = pIn1 + count;
emh203 0:3d9c67d97d6f 1244 py = pSrc2;
emh203 0:3d9c67d97d6f 1245
emh203 0:3d9c67d97d6f 1246 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 1247 blkCnt--;
emh203 0:3d9c67d97d6f 1248 }
emh203 0:3d9c67d97d6f 1249 }
emh203 0:3d9c67d97d6f 1250 else
emh203 0:3d9c67d97d6f 1251 {
emh203 0:3d9c67d97d6f 1252 /* If the srcBLen is not a multiple of 4,
emh203 0:3d9c67d97d6f 1253 * the blockSize2 loop cannot be unrolled by 4 */
emh203 0:3d9c67d97d6f 1254 blkCnt = blockSize2;
emh203 0:3d9c67d97d6f 1255
emh203 0:3d9c67d97d6f 1256 while(blkCnt > 0u)
emh203 0:3d9c67d97d6f 1257 {
emh203 0:3d9c67d97d6f 1258 /* Accumulator is made zero for every iteration */
emh203 0:3d9c67d97d6f 1259 sum = 0;
emh203 0:3d9c67d97d6f 1260
emh203 0:3d9c67d97d6f 1261 /* srcBLen number of MACS should be performed */
emh203 0:3d9c67d97d6f 1262 k = srcBLen;
emh203 0:3d9c67d97d6f 1263
emh203 0:3d9c67d97d6f 1264 while(k > 0u)
emh203 0:3d9c67d97d6f 1265 {
emh203 0:3d9c67d97d6f 1266 /* Perform the multiply-accumulate */
emh203 0:3d9c67d97d6f 1267 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 1268
emh203 0:3d9c67d97d6f 1269 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 1270 k--;
emh203 0:3d9c67d97d6f 1271 }
emh203 0:3d9c67d97d6f 1272
emh203 0:3d9c67d97d6f 1273 /* Store the result in the accumulator in the destination buffer. */
emh203 0:3d9c67d97d6f 1274 *pOut++ = (q15_t) (sum >> 15);
emh203 0:3d9c67d97d6f 1275
emh203 0:3d9c67d97d6f 1276 /* Increment the MAC count */
emh203 0:3d9c67d97d6f 1277 count++;
emh203 0:3d9c67d97d6f 1278
emh203 0:3d9c67d97d6f 1279 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 1280 px = pIn1 + count;
emh203 0:3d9c67d97d6f 1281 py = pSrc2;
emh203 0:3d9c67d97d6f 1282
emh203 0:3d9c67d97d6f 1283 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 1284 blkCnt--;
emh203 0:3d9c67d97d6f 1285 }
emh203 0:3d9c67d97d6f 1286 }
emh203 0:3d9c67d97d6f 1287
emh203 0:3d9c67d97d6f 1288
emh203 0:3d9c67d97d6f 1289 /* --------------------------
emh203 0:3d9c67d97d6f 1290 * Initializations of stage3
emh203 0:3d9c67d97d6f 1291 * -------------------------*/
emh203 0:3d9c67d97d6f 1292
emh203 0:3d9c67d97d6f 1293 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emh203 0:3d9c67d97d6f 1294 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emh203 0:3d9c67d97d6f 1295 * ....
emh203 0:3d9c67d97d6f 1296 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emh203 0:3d9c67d97d6f 1297 * sum += x[srcALen-1] * y[srcBLen-1]
emh203 0:3d9c67d97d6f 1298 */
emh203 0:3d9c67d97d6f 1299
emh203 0:3d9c67d97d6f 1300 /* In this stage the MAC operations are decreased by 1 for every iteration.
emh203 0:3d9c67d97d6f 1301 The blockSize3 variable holds the number of MAC operations performed */
emh203 0:3d9c67d97d6f 1302
emh203 0:3d9c67d97d6f 1303 /* Working pointer of inputA */
emh203 0:3d9c67d97d6f 1304 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emh203 0:3d9c67d97d6f 1305 px = pSrc1;
emh203 0:3d9c67d97d6f 1306
emh203 0:3d9c67d97d6f 1307 /* Working pointer of inputB */
emh203 0:3d9c67d97d6f 1308 pSrc2 = pIn2 + (srcBLen - 1u);
emh203 0:3d9c67d97d6f 1309 pIn2 = pSrc2 - 1u;
emh203 0:3d9c67d97d6f 1310 py = pIn2;
emh203 0:3d9c67d97d6f 1311
emh203 0:3d9c67d97d6f 1312 /* -------------------
emh203 0:3d9c67d97d6f 1313 * Stage3 process
emh203 0:3d9c67d97d6f 1314 * ------------------*/
emh203 0:3d9c67d97d6f 1315
emh203 0:3d9c67d97d6f 1316 /* For loop unrolling by 4, this stage is divided into two. */
emh203 0:3d9c67d97d6f 1317 /* First part of this stage computes the MAC operations greater than 4 */
emh203 0:3d9c67d97d6f 1318 /* Second part of this stage computes the MAC operations less than or equal to 4 */
emh203 0:3d9c67d97d6f 1319
emh203 0:3d9c67d97d6f 1320 /* The first part of the stage starts here */
emh203 0:3d9c67d97d6f 1321 j = blockSize3 >> 2u;
emh203 0:3d9c67d97d6f 1322
emh203 0:3d9c67d97d6f 1323 while((j > 0u) && (blockSize3 > 0u))
emh203 0:3d9c67d97d6f 1324 {
emh203 0:3d9c67d97d6f 1325 /* Accumulator is made zero for every iteration */
emh203 0:3d9c67d97d6f 1326 sum = 0;
emh203 0:3d9c67d97d6f 1327
emh203 0:3d9c67d97d6f 1328 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emh203 0:3d9c67d97d6f 1329 k = blockSize3 >> 2u;
emh203 0:3d9c67d97d6f 1330
emh203 0:3d9c67d97d6f 1331 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emh203 0:3d9c67d97d6f 1332 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emh203 0:3d9c67d97d6f 1333 py++;
emh203 0:3d9c67d97d6f 1334
emh203 0:3d9c67d97d6f 1335 while(k > 0u)
emh203 0:3d9c67d97d6f 1336 {
emh203 0:3d9c67d97d6f 1337 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 1338 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 1339 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 1340 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 1341 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 1342 k--;
emh203 0:3d9c67d97d6f 1343 }
emh203 0:3d9c67d97d6f 1344
emh203 0:3d9c67d97d6f 1345 /* If the blockSize3 is not a multiple of 4, compute any remaining MACs here.
emh203 0:3d9c67d97d6f 1346 ** No loop unrolling is used. */
emh203 0:3d9c67d97d6f 1347 k = blockSize3 % 0x4u;
emh203 0:3d9c67d97d6f 1348
emh203 0:3d9c67d97d6f 1349 while(k > 0u)
emh203 0:3d9c67d97d6f 1350 {
emh203 0:3d9c67d97d6f 1351 /* sum += x[srcALen - srcBLen + 5] * y[srcBLen - 5] */
emh203 0:3d9c67d97d6f 1352 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 1353
emh203 0:3d9c67d97d6f 1354 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 1355 k--;
emh203 0:3d9c67d97d6f 1356 }
emh203 0:3d9c67d97d6f 1357
emh203 0:3d9c67d97d6f 1358 /* Store the result in the accumulator in the destination buffer. */
emh203 0:3d9c67d97d6f 1359 *pOut++ = (q15_t) (sum >> 15);
emh203 0:3d9c67d97d6f 1360
emh203 0:3d9c67d97d6f 1361 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 1362 px = ++pSrc1;
emh203 0:3d9c67d97d6f 1363 py = pIn2;
emh203 0:3d9c67d97d6f 1364
emh203 0:3d9c67d97d6f 1365 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 1366 blockSize3--;
emh203 0:3d9c67d97d6f 1367
emh203 0:3d9c67d97d6f 1368 j--;
emh203 0:3d9c67d97d6f 1369 }
emh203 0:3d9c67d97d6f 1370
emh203 0:3d9c67d97d6f 1371 /* The second part of the stage starts here */
emh203 0:3d9c67d97d6f 1372 /* SIMD is not used for the next MAC operations,
emh203 0:3d9c67d97d6f 1373 * so pointer py is updated to read only one sample at a time */
emh203 0:3d9c67d97d6f 1374 py = py + 1u;
emh203 0:3d9c67d97d6f 1375
emh203 0:3d9c67d97d6f 1376 while(blockSize3 > 0u)
emh203 0:3d9c67d97d6f 1377 {
emh203 0:3d9c67d97d6f 1378 /* Accumulator is made zero for every iteration */
emh203 0:3d9c67d97d6f 1379 sum = 0;
emh203 0:3d9c67d97d6f 1380
emh203 0:3d9c67d97d6f 1381 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emh203 0:3d9c67d97d6f 1382 k = blockSize3;
emh203 0:3d9c67d97d6f 1383
emh203 0:3d9c67d97d6f 1384 while(k > 0u)
emh203 0:3d9c67d97d6f 1385 {
emh203 0:3d9c67d97d6f 1386 /* Perform the multiply-accumulates */
emh203 0:3d9c67d97d6f 1387 /* sum += x[srcALen-1] * y[srcBLen-1] */
emh203 0:3d9c67d97d6f 1388 sum += ((q31_t) * px++ * *py--);
emh203 0:3d9c67d97d6f 1389
emh203 0:3d9c67d97d6f 1390 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 1391 k--;
emh203 0:3d9c67d97d6f 1392 }
emh203 0:3d9c67d97d6f 1393
emh203 0:3d9c67d97d6f 1394 /* Store the result in the accumulator in the destination buffer. */
emh203 0:3d9c67d97d6f 1395 *pOut++ = (q15_t) (sum >> 15);
emh203 0:3d9c67d97d6f 1396
emh203 0:3d9c67d97d6f 1397 /* Update the inputA and inputB pointers for next MAC calculation */
emh203 0:3d9c67d97d6f 1398 px = ++pSrc1;
emh203 0:3d9c67d97d6f 1399 py = pSrc2;
emh203 0:3d9c67d97d6f 1400
emh203 0:3d9c67d97d6f 1401 /* Decrement the loop counter */
emh203 0:3d9c67d97d6f 1402 blockSize3--;
emh203 0:3d9c67d97d6f 1403 }
emh203 0:3d9c67d97d6f 1404
emh203 0:3d9c67d97d6f 1405 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
emh203 0:3d9c67d97d6f 1406 }
emh203 0:3d9c67d97d6f 1407
emh203 0:3d9c67d97d6f 1408 /**
emh203 0:3d9c67d97d6f 1409 * @} end of Conv group
emh203 0:3d9c67d97d6f 1410 */