mbed official / mbed-dsp

CMSIS DSP library

Dependents:   performance_timer Surfboard_ gps2rtty Capstone ... more

Legacy Warning

This is an mbed 2 library. To learn more about mbed OS 5, visit the docs.

cmsis_dsp/FilteringFunctions/arm_conv_partial_fast_q15.c@5:3762170b6d4d, 2015-11-20 (annotated)

Committer:
mbed_official
Date:
Fri Nov 20 08:45:18 2015 +0000
Revision:
5:3762170b6d4d
Parent:
3:7a284390b0ce 
Synchronized with git revision 2eb940b9a73af188d3004a2575fdfbb05febe62b



Full URL: https://github.com/mbedmicro/mbed/commit/2eb940b9a73af188d3004a2575fdfbb05febe62b/



Added option to build rpc library. closes #1426

Who changed what in which revision?

UserRevisionLine numberNew contents of line
emilmont 1:fdd22bb7aa52 1 /* ----------------------------------------------------------------------
mbed_official 5:3762170b6d4d 2 * Copyright (C) 2010-2014 ARM Limited. All rights reserved.
emilmont 1:fdd22bb7aa52 3 *
mbed_official 5:3762170b6d4d 4 * $Date: 19. March 2015
mbed_official 5:3762170b6d4d 5 * $Revision: V.1.4.5
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;
mbed_official 5:3762170b6d4d 127 blockSize3 = ((int32_t)check > (int32_t)srcALen) ? (int32_t)check - (int32_t)srcALen : 0;
mbed_official 5:3762170b6d4d 128 blockSize3 = ((int32_t)firstIndex > (int32_t)srcALen - 1) ? blockSize3 - (int32_t)firstIndex + (int32_t)srcALen : blockSize3;
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 */
mbed_official 5:3762170b6d4d 283 if((int32_t)firstIndex - (int32_t)srcBLen + 1 > 0)
mbed_official 5:3762170b6d4d 284 {
mbed_official 5:3762170b6d4d 285 px = pIn1 + firstIndex - srcBLen + 1;
mbed_official 5:3762170b6d4d 286 }
mbed_official 5:3762170b6d4d 287 else
mbed_official 5:3762170b6d4d 288 {
mbed_official 5:3762170b6d4d 289 px = pIn1;
mbed_official 5:3762170b6d4d 290 }
emilmont 1:fdd22bb7aa52 291
emilmont 1:fdd22bb7aa52 292 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 293 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 294 py = pSrc2;
emilmont 1:fdd22bb7aa52 295
emilmont 1:fdd22bb7aa52 296 /* count is the index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 297 count = 0u;
emilmont 1:fdd22bb7aa52 298
emilmont 1:fdd22bb7aa52 299
emilmont 1:fdd22bb7aa52 300 /* --------------------
emilmont 1:fdd22bb7aa52 301 * Stage2 process
emilmont 1:fdd22bb7aa52 302 * -------------------*/
emilmont 1:fdd22bb7aa52 303
emilmont 1:fdd22bb7aa52 304 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 305 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 306 * srcBLen should be greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 307 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 308 {
emilmont 1:fdd22bb7aa52 309 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 310 blkCnt = ((uint32_t) blockSize2 >> 2u);
emilmont 1:fdd22bb7aa52 311
emilmont 1:fdd22bb7aa52 312 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 313 {
emilmont 1:fdd22bb7aa52 314 py = py - 1u;
emilmont 1:fdd22bb7aa52 315
emilmont 1:fdd22bb7aa52 316 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 317 acc0 = 0;
emilmont 1:fdd22bb7aa52 318 acc1 = 0;
emilmont 1:fdd22bb7aa52 319 acc2 = 0;
emilmont 1:fdd22bb7aa52 320 acc3 = 0;
emilmont 1:fdd22bb7aa52 321
emilmont 1:fdd22bb7aa52 322
emilmont 1:fdd22bb7aa52 323 /* read x[0], x[1] samples */
emilmont 1:fdd22bb7aa52 324 x0 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 325 /* read x[1], x[2] samples */
emilmont 1:fdd22bb7aa52 326 x1 = _SIMD32_OFFSET(px+1);
emilmont 2:da51fb522205 327 px+= 2u;
emilmont 1:fdd22bb7aa52 328
emilmont 1:fdd22bb7aa52 329
emilmont 1:fdd22bb7aa52 330 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 331 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 332
emilmont 1:fdd22bb7aa52 333 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 334 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 335 do
emilmont 1:fdd22bb7aa52 336 {
emilmont 1:fdd22bb7aa52 337 /* Read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 338 * y[srcBLen - 1] and y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 339 c0 = *__SIMD32(py)--;
emilmont 1:fdd22bb7aa52 340
emilmont 1:fdd22bb7aa52 341 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 342 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 343
emilmont 1:fdd22bb7aa52 344 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 345 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 346
emilmont 1:fdd22bb7aa52 347 /* Read x[2], x[3] */
emilmont 1:fdd22bb7aa52 348 x2 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 349
emilmont 1:fdd22bb7aa52 350 /* Read x[3], x[4] */
emilmont 1:fdd22bb7aa52 351 x3 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 352
emilmont 1:fdd22bb7aa52 353 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 354 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 355
emilmont 1:fdd22bb7aa52 356 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 357 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 358
emilmont 1:fdd22bb7aa52 359 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 360 c0 = *__SIMD32(py)--;
emilmont 1:fdd22bb7aa52 361
emilmont 1:fdd22bb7aa52 362 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 363 acc0 = __SMLADX(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 364
emilmont 1:fdd22bb7aa52 365 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 366 acc1 = __SMLADX(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 367
emilmont 1:fdd22bb7aa52 368 /* Read x[4], x[5] */
emilmont 1:fdd22bb7aa52 369 x0 = _SIMD32_OFFSET(px+2);
emilmont 1:fdd22bb7aa52 370
emilmont 1:fdd22bb7aa52 371 /* Read x[5], x[6] */
emilmont 1:fdd22bb7aa52 372 x1 = _SIMD32_OFFSET(px+3);
emilmont 2:da51fb522205 373 px += 4u;
emilmont 1:fdd22bb7aa52 374
emilmont 1:fdd22bb7aa52 375 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 376 acc2 = __SMLADX(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 377
emilmont 1:fdd22bb7aa52 378 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 379 acc3 = __SMLADX(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 380
emilmont 1:fdd22bb7aa52 381 } while(--k);
emilmont 1:fdd22bb7aa52 382
emilmont 1:fdd22bb7aa52 383 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 384 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 385
emilmont 1:fdd22bb7aa52 386 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 387 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 388 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 389
emilmont 1:fdd22bb7aa52 390 if(k == 1u)
emilmont 1:fdd22bb7aa52 391 {
emilmont 1:fdd22bb7aa52 392 /* Read y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 393 c0 = *(py+1);
emilmont 1:fdd22bb7aa52 394 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 395
emilmont 1:fdd22bb7aa52 396 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 397
emilmont 1:fdd22bb7aa52 398 #else
emilmont 1:fdd22bb7aa52 399
emilmont 1:fdd22bb7aa52 400 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 401
emilmont 1:fdd22bb7aa52 402 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 403
emilmont 1:fdd22bb7aa52 404 /* Read x[7] */
emilmont 1:fdd22bb7aa52 405 x3 = *__SIMD32(px);
emilmont 2:da51fb522205 406 px++;
emilmont 1:fdd22bb7aa52 407
emilmont 1:fdd22bb7aa52 408 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 409 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 410 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 411 acc2 = __SMLADX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 412 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 413 }
emilmont 1:fdd22bb7aa52 414
emilmont 1:fdd22bb7aa52 415 if(k == 2u)
emilmont 1:fdd22bb7aa52 416 {
emilmont 1:fdd22bb7aa52 417 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 418 c0 = _SIMD32_OFFSET(py);
emilmont 1:fdd22bb7aa52 419
emilmont 1:fdd22bb7aa52 420 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 421 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 422
emilmont 1:fdd22bb7aa52 423 /* Read x[9] */
emilmont 1:fdd22bb7aa52 424 x2 = _SIMD32_OFFSET(px+1);
emilmont 2:da51fb522205 425 px += 2u;
emilmont 1:fdd22bb7aa52 426
emilmont 1:fdd22bb7aa52 427 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 428 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 429 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 430 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 431 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 432 }
emilmont 1:fdd22bb7aa52 433
emilmont 1:fdd22bb7aa52 434 if(k == 3u)
emilmont 1:fdd22bb7aa52 435 {
emilmont 1:fdd22bb7aa52 436 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 1:fdd22bb7aa52 437 c0 = _SIMD32_OFFSET(py);
emilmont 1:fdd22bb7aa52 438
emilmont 1:fdd22bb7aa52 439 /* Read x[7], x[8] */
emilmont 1:fdd22bb7aa52 440 x3 = *__SIMD32(px);
emilmont 1:fdd22bb7aa52 441
emilmont 1:fdd22bb7aa52 442 /* Read x[9] */
emilmont 1:fdd22bb7aa52 443 x2 = _SIMD32_OFFSET(px+1);
emilmont 1:fdd22bb7aa52 444
emilmont 1:fdd22bb7aa52 445 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 446 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 447 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 448 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 449 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 450
emilmont 2:da51fb522205 451 c0 = *(py-1);
emilmont 1:fdd22bb7aa52 452 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 453
emilmont 1:fdd22bb7aa52 454 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 455 #else
emilmont 1:fdd22bb7aa52 456
emilmont 1:fdd22bb7aa52 457 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 458 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 459
emilmont 1:fdd22bb7aa52 460 /* Read x[10] */
emilmont 1:fdd22bb7aa52 461 x3 = _SIMD32_OFFSET(px+2);
emilmont 2:da51fb522205 462 px += 3u;
emilmont 1:fdd22bb7aa52 463
emilmont 1:fdd22bb7aa52 464 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 465 acc0 = __SMLADX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 466 acc1 = __SMLAD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 467 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 468 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 469 }
emilmont 1:fdd22bb7aa52 470
emilmont 1:fdd22bb7aa52 471 /* Store the results in the accumulators in the destination buffer. */
emilmont 1:fdd22bb7aa52 472 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 473
emilmont 1:fdd22bb7aa52 474 *__SIMD32(pOut)++ = __PKHBT(acc0 >> 15, acc1 >> 15, 16);
emilmont 1:fdd22bb7aa52 475 *__SIMD32(pOut)++ = __PKHBT(acc2 >> 15, acc3 >> 15, 16);
emilmont 1:fdd22bb7aa52 476
emilmont 1:fdd22bb7aa52 477 #else
emilmont 1:fdd22bb7aa52 478
emilmont 1:fdd22bb7aa52 479 *__SIMD32(pOut)++ = __PKHBT(acc1 >> 15, acc0 >> 15, 16);
emilmont 1:fdd22bb7aa52 480 *__SIMD32(pOut)++ = __PKHBT(acc3 >> 15, acc2 >> 15, 16);
emilmont 1:fdd22bb7aa52 481
emilmont 1:fdd22bb7aa52 482 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 483
emilmont 1:fdd22bb7aa52 484 /* Increment the pointer pIn1 index, count by 4 */
emilmont 1:fdd22bb7aa52 485 count += 4u;
emilmont 1:fdd22bb7aa52 486
emilmont 1:fdd22bb7aa52 487 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 488 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 489 py = pSrc2;
emilmont 1:fdd22bb7aa52 490
emilmont 1:fdd22bb7aa52 491 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 492 blkCnt--;
emilmont 1:fdd22bb7aa52 493 }
emilmont 1:fdd22bb7aa52 494
emilmont 1:fdd22bb7aa52 495 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 496 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 497 blkCnt = (uint32_t) blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 498
emilmont 1:fdd22bb7aa52 499 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 500 {
emilmont 1:fdd22bb7aa52 501 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 502 sum = 0;
emilmont 1:fdd22bb7aa52 503
emilmont 1:fdd22bb7aa52 504 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 505 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 506
emilmont 1:fdd22bb7aa52 507 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 508 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 509 while(k > 0u)
emilmont 1:fdd22bb7aa52 510 {
emilmont 1:fdd22bb7aa52 511 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 512 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 513 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 514 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 515 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 516
emilmont 1:fdd22bb7aa52 517 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 518 k--;
emilmont 1:fdd22bb7aa52 519 }
emilmont 1:fdd22bb7aa52 520
emilmont 1:fdd22bb7aa52 521 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 522 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 523 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 524
emilmont 1:fdd22bb7aa52 525 while(k > 0u)
emilmont 1:fdd22bb7aa52 526 {
emilmont 1:fdd22bb7aa52 527 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 528 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 529
emilmont 1:fdd22bb7aa52 530 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 531 k--;
emilmont 1:fdd22bb7aa52 532 }
emilmont 1:fdd22bb7aa52 533
emilmont 1:fdd22bb7aa52 534 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 535 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 536
emilmont 1:fdd22bb7aa52 537 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 538 count++;
emilmont 1:fdd22bb7aa52 539
emilmont 1:fdd22bb7aa52 540 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 541 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 542 py = pSrc2;
emilmont 1:fdd22bb7aa52 543
emilmont 1:fdd22bb7aa52 544 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 545 blkCnt--;
emilmont 1:fdd22bb7aa52 546 }
emilmont 1:fdd22bb7aa52 547 }
emilmont 1:fdd22bb7aa52 548 else
emilmont 1:fdd22bb7aa52 549 {
emilmont 1:fdd22bb7aa52 550 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 551 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 552 blkCnt = (uint32_t) blockSize2;
emilmont 1:fdd22bb7aa52 553
emilmont 1:fdd22bb7aa52 554 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 555 {
emilmont 1:fdd22bb7aa52 556 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 557 sum = 0;
emilmont 1:fdd22bb7aa52 558
emilmont 1:fdd22bb7aa52 559 /* srcBLen number of MACS should be performed */
emilmont 1:fdd22bb7aa52 560 k = srcBLen;
emilmont 1:fdd22bb7aa52 561
emilmont 1:fdd22bb7aa52 562 while(k > 0u)
emilmont 1:fdd22bb7aa52 563 {
emilmont 1:fdd22bb7aa52 564 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 565 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 566
emilmont 1:fdd22bb7aa52 567 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 568 k--;
emilmont 1:fdd22bb7aa52 569 }
emilmont 1:fdd22bb7aa52 570
emilmont 1:fdd22bb7aa52 571 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 572 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 573
emilmont 1:fdd22bb7aa52 574 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 575 count++;
emilmont 1:fdd22bb7aa52 576
emilmont 1:fdd22bb7aa52 577 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 578 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 579 py = pSrc2;
emilmont 1:fdd22bb7aa52 580
emilmont 1:fdd22bb7aa52 581 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 582 blkCnt--;
emilmont 1:fdd22bb7aa52 583 }
emilmont 1:fdd22bb7aa52 584 }
emilmont 1:fdd22bb7aa52 585
emilmont 1:fdd22bb7aa52 586
emilmont 1:fdd22bb7aa52 587 /* --------------------------
emilmont 1:fdd22bb7aa52 588 * Initializations of stage3
emilmont 1:fdd22bb7aa52 589 * -------------------------*/
emilmont 1:fdd22bb7aa52 590
emilmont 1:fdd22bb7aa52 591 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 592 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emilmont 1:fdd22bb7aa52 593 * ....
emilmont 1:fdd22bb7aa52 594 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emilmont 1:fdd22bb7aa52 595 * sum += x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 596 */
emilmont 1:fdd22bb7aa52 597
emilmont 1:fdd22bb7aa52 598 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 599 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 600 count = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 601
emilmont 1:fdd22bb7aa52 602 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 603 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 604 px = pSrc1;
emilmont 1:fdd22bb7aa52 605
emilmont 1:fdd22bb7aa52 606 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 607 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 608 pIn2 = pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 609 py = pIn2;
emilmont 1:fdd22bb7aa52 610
emilmont 1:fdd22bb7aa52 611 /* -------------------
emilmont 1:fdd22bb7aa52 612 * Stage3 process
emilmont 1:fdd22bb7aa52 613 * ------------------*/
emilmont 1:fdd22bb7aa52 614
emilmont 1:fdd22bb7aa52 615 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 616 /* First part of this stage computes the MAC operations greater than 4 */
emilmont 1:fdd22bb7aa52 617 /* Second part of this stage computes the MAC operations less than or equal to 4 */
emilmont 1:fdd22bb7aa52 618
emilmont 1:fdd22bb7aa52 619 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 620 j = count >> 2u;
emilmont 1:fdd22bb7aa52 621
emilmont 1:fdd22bb7aa52 622 while((j > 0u) && (blockSize3 > 0))
emilmont 1:fdd22bb7aa52 623 {
emilmont 1:fdd22bb7aa52 624 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 625 sum = 0;
emilmont 1:fdd22bb7aa52 626
emilmont 1:fdd22bb7aa52 627 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 628 k = count >> 2u;
emilmont 1:fdd22bb7aa52 629
emilmont 1:fdd22bb7aa52 630 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 631 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 632 while(k > 0u)
emilmont 1:fdd22bb7aa52 633 {
emilmont 1:fdd22bb7aa52 634 /* x[srcALen - srcBLen + 1], x[srcALen - srcBLen + 2] are multiplied
emilmont 1:fdd22bb7aa52 635 * with y[srcBLen - 1], y[srcBLen - 2] respectively */
emilmont 1:fdd22bb7aa52 636 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 637 /* x[srcALen - srcBLen + 3], x[srcALen - srcBLen + 4] are multiplied
emilmont 1:fdd22bb7aa52 638 * with y[srcBLen - 3], y[srcBLen - 4] respectively */
emilmont 1:fdd22bb7aa52 639 sum = __SMLADX(*__SIMD32(px)++, *__SIMD32(py)--, sum);
emilmont 1:fdd22bb7aa52 640
emilmont 1:fdd22bb7aa52 641 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 642 k--;
emilmont 1:fdd22bb7aa52 643 }
emilmont 1:fdd22bb7aa52 644
emilmont 1:fdd22bb7aa52 645 /* For the next MAC operations, the pointer py is used without SIMD
emilmont 1:fdd22bb7aa52 646 * So, py is incremented by 1 */
emilmont 1:fdd22bb7aa52 647 py = py + 1u;
emilmont 1:fdd22bb7aa52 648
emilmont 1:fdd22bb7aa52 649 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 650 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 651 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 652
emilmont 1:fdd22bb7aa52 653 while(k > 0u)
emilmont 1:fdd22bb7aa52 654 {
emilmont 1:fdd22bb7aa52 655 /* sum += x[srcALen - srcBLen + 5] * y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 656 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 657
emilmont 1:fdd22bb7aa52 658 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 659 k--;
emilmont 1:fdd22bb7aa52 660 }
emilmont 1:fdd22bb7aa52 661
emilmont 1:fdd22bb7aa52 662 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 663 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 664
emilmont 1:fdd22bb7aa52 665 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 666 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 667 py = pIn2;
emilmont 1:fdd22bb7aa52 668
emilmont 1:fdd22bb7aa52 669 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 670 count--;
emilmont 1:fdd22bb7aa52 671
emilmont 1:fdd22bb7aa52 672 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 673 blockSize3--;
emilmont 1:fdd22bb7aa52 674
emilmont 1:fdd22bb7aa52 675 j--;
emilmont 1:fdd22bb7aa52 676 }
emilmont 1:fdd22bb7aa52 677
emilmont 1:fdd22bb7aa52 678 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 679 /* SIMD is not used for the next MAC operations,
emilmont 1:fdd22bb7aa52 680 * so pointer py is updated to read only one sample at a time */
emilmont 1:fdd22bb7aa52 681 py = py + 1u;
emilmont 1:fdd22bb7aa52 682
emilmont 1:fdd22bb7aa52 683 while(blockSize3 > 0)
emilmont 1:fdd22bb7aa52 684 {
emilmont 1:fdd22bb7aa52 685 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 686 sum = 0;
emilmont 1:fdd22bb7aa52 687
emilmont 1:fdd22bb7aa52 688 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 689 k = count;
emilmont 1:fdd22bb7aa52 690
emilmont 1:fdd22bb7aa52 691 while(k > 0u)
emilmont 1:fdd22bb7aa52 692 {
emilmont 1:fdd22bb7aa52 693 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 694 /* sum += x[srcALen-1] * y[srcBLen-1] */
emilmont 1:fdd22bb7aa52 695 sum = __SMLAD(*px++, *py--, sum);
emilmont 1:fdd22bb7aa52 696
emilmont 1:fdd22bb7aa52 697 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 698 k--;
emilmont 1:fdd22bb7aa52 699 }
emilmont 1:fdd22bb7aa52 700
emilmont 1:fdd22bb7aa52 701 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 702 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 703
emilmont 1:fdd22bb7aa52 704 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 705 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 706 py = pSrc2;
emilmont 1:fdd22bb7aa52 707
emilmont 1:fdd22bb7aa52 708 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 709 count--;
emilmont 1:fdd22bb7aa52 710
emilmont 1:fdd22bb7aa52 711 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 712 blockSize3--;
emilmont 1:fdd22bb7aa52 713 }
emilmont 1:fdd22bb7aa52 714
emilmont 1:fdd22bb7aa52 715 /* set status as ARM_MATH_SUCCESS */
emilmont 1:fdd22bb7aa52 716 status = ARM_MATH_SUCCESS;
emilmont 1:fdd22bb7aa52 717 }
emilmont 1:fdd22bb7aa52 718
emilmont 1:fdd22bb7aa52 719 /* Return to application */
emilmont 1:fdd22bb7aa52 720 return (status);
emilmont 1:fdd22bb7aa52 721
emilmont 1:fdd22bb7aa52 722 #else
emilmont 1:fdd22bb7aa52 723
emilmont 1:fdd22bb7aa52 724 q15_t *pIn1; /* inputA pointer */
emilmont 1:fdd22bb7aa52 725 q15_t *pIn2; /* inputB pointer */
emilmont 1:fdd22bb7aa52 726 q15_t *pOut = pDst; /* output pointer */
emilmont 1:fdd22bb7aa52 727 q31_t sum, acc0, acc1, acc2, acc3; /* Accumulator */
emilmont 1:fdd22bb7aa52 728 q15_t *px; /* Intermediate inputA pointer */
emilmont 1:fdd22bb7aa52 729 q15_t *py; /* Intermediate inputB pointer */
emilmont 1:fdd22bb7aa52 730 q15_t *pSrc1, *pSrc2; /* Intermediate pointers */
emilmont 1:fdd22bb7aa52 731 q31_t x0, x1, x2, x3, c0;
emilmont 1:fdd22bb7aa52 732 uint32_t j, k, count, check, blkCnt;
emilmont 1:fdd22bb7aa52 733 int32_t blockSize1, blockSize2, blockSize3; /* loop counters */
emilmont 1:fdd22bb7aa52 734 arm_status status; /* status of Partial convolution */
emilmont 1:fdd22bb7aa52 735 q15_t a, b;
emilmont 1:fdd22bb7aa52 736
emilmont 1:fdd22bb7aa52 737 /* Check for range of output samples to be calculated */
emilmont 1:fdd22bb7aa52 738 if((firstIndex + numPoints) > ((srcALen + (srcBLen - 1u))))
emilmont 1:fdd22bb7aa52 739 {
emilmont 1:fdd22bb7aa52 740 /* Set status as ARM_MATH_ARGUMENT_ERROR */
emilmont 1:fdd22bb7aa52 741 status = ARM_MATH_ARGUMENT_ERROR;
emilmont 1:fdd22bb7aa52 742 }
emilmont 1:fdd22bb7aa52 743 else
emilmont 1:fdd22bb7aa52 744 {
emilmont 1:fdd22bb7aa52 745
emilmont 1:fdd22bb7aa52 746 /* The algorithm implementation is based on the lengths of the inputs. */
emilmont 1:fdd22bb7aa52 747 /* srcB is always made to slide across srcA. */
emilmont 1:fdd22bb7aa52 748 /* So srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 749 if(srcALen >=srcBLen)
emilmont 1:fdd22bb7aa52 750 {
emilmont 1:fdd22bb7aa52 751 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 752 pIn1 = pSrcA;
emilmont 1:fdd22bb7aa52 753
emilmont 1:fdd22bb7aa52 754 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 755 pIn2 = pSrcB;
emilmont 1:fdd22bb7aa52 756 }
emilmont 1:fdd22bb7aa52 757 else
emilmont 1:fdd22bb7aa52 758 {
emilmont 1:fdd22bb7aa52 759 /* Initialization of inputA pointer */
emilmont 1:fdd22bb7aa52 760 pIn1 = pSrcB;
emilmont 1:fdd22bb7aa52 761
emilmont 1:fdd22bb7aa52 762 /* Initialization of inputB pointer */
emilmont 1:fdd22bb7aa52 763 pIn2 = pSrcA;
emilmont 1:fdd22bb7aa52 764
emilmont 1:fdd22bb7aa52 765 /* srcBLen is always considered as shorter or equal to srcALen */
emilmont 1:fdd22bb7aa52 766 j = srcBLen;
emilmont 1:fdd22bb7aa52 767 srcBLen = srcALen;
emilmont 1:fdd22bb7aa52 768 srcALen = j;
emilmont 1:fdd22bb7aa52 769 }
emilmont 1:fdd22bb7aa52 770
emilmont 1:fdd22bb7aa52 771 /* Conditions to check which loopCounter holds
emilmont 1:fdd22bb7aa52 772 * the first and last indices of the output samples to be calculated. */
emilmont 1:fdd22bb7aa52 773 check = firstIndex + numPoints;
mbed_official 5:3762170b6d4d 774 blockSize3 = ((int32_t)check > (int32_t)srcALen) ? (int32_t)check - (int32_t)srcALen : 0;
mbed_official 5:3762170b6d4d 775 blockSize3 = ((int32_t)firstIndex > (int32_t)srcALen - 1) ? blockSize3 - (int32_t)firstIndex + (int32_t)srcALen : blockSize3;
mbed_official 5:3762170b6d4d 776 blockSize1 = ((int32_t) srcBLen - 1) - (int32_t) firstIndex;
emilmont 1:fdd22bb7aa52 777 blockSize1 = (blockSize1 > 0) ? ((check > (srcBLen - 1u)) ? blockSize1 :
emilmont 1:fdd22bb7aa52 778 (int32_t) numPoints) : 0;
mbed_official 5:3762170b6d4d 779 blockSize2 = ((int32_t) check - blockSize3) -
mbed_official 5:3762170b6d4d 780 (blockSize1 + (int32_t) firstIndex);
emilmont 1:fdd22bb7aa52 781 blockSize2 = (blockSize2 > 0) ? blockSize2 : 0;
emilmont 1:fdd22bb7aa52 782
emilmont 1:fdd22bb7aa52 783 /* 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 784 /* The function is internally
emilmont 1:fdd22bb7aa52 785 * divided into three stages according to the number of multiplications that has to be
emilmont 1:fdd22bb7aa52 786 * taken place between inputA samples and inputB samples. In the first stage of the
emilmont 1:fdd22bb7aa52 787 * algorithm, the multiplications increase by one for every iteration.
emilmont 1:fdd22bb7aa52 788 * In the second stage of the algorithm, srcBLen number of multiplications are done.
emilmont 1:fdd22bb7aa52 789 * In the third stage of the algorithm, the multiplications decrease by one
emilmont 1:fdd22bb7aa52 790 * for every iteration. */
emilmont 1:fdd22bb7aa52 791
emilmont 1:fdd22bb7aa52 792 /* Set the output pointer to point to the firstIndex
emilmont 1:fdd22bb7aa52 793 * of the output sample to be calculated. */
emilmont 1:fdd22bb7aa52 794 pOut = pDst + firstIndex;
emilmont 1:fdd22bb7aa52 795
emilmont 1:fdd22bb7aa52 796 /* --------------------------
emilmont 1:fdd22bb7aa52 797 * Initializations of stage1
emilmont 1:fdd22bb7aa52 798 * -------------------------*/
emilmont 1:fdd22bb7aa52 799
emilmont 1:fdd22bb7aa52 800 /* sum = x[0] * y[0]
emilmont 1:fdd22bb7aa52 801 * sum = x[0] * y[1] + x[1] * y[0]
emilmont 1:fdd22bb7aa52 802 * ....
emilmont 1:fdd22bb7aa52 803 * sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
emilmont 1:fdd22bb7aa52 804 */
emilmont 1:fdd22bb7aa52 805
emilmont 1:fdd22bb7aa52 806 /* In this stage the MAC operations are increased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 807 The count variable holds the number of MAC operations performed.
emilmont 1:fdd22bb7aa52 808 Since the partial convolution starts from firstIndex
emilmont 1:fdd22bb7aa52 809 Number of Macs to be performed is firstIndex + 1 */
emilmont 1:fdd22bb7aa52 810 count = 1u + firstIndex;
emilmont 1:fdd22bb7aa52 811
emilmont 1:fdd22bb7aa52 812 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 813 px = pIn1;
emilmont 1:fdd22bb7aa52 814
emilmont 1:fdd22bb7aa52 815 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 816 pSrc2 = pIn2 + firstIndex;
emilmont 1:fdd22bb7aa52 817 py = pSrc2;
emilmont 1:fdd22bb7aa52 818
emilmont 1:fdd22bb7aa52 819 /* ------------------------
emilmont 1:fdd22bb7aa52 820 * Stage1 process
emilmont 1:fdd22bb7aa52 821 * ----------------------*/
emilmont 1:fdd22bb7aa52 822
emilmont 1:fdd22bb7aa52 823 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 824 /* First part of this stage computes the MAC operations less than 4 */
emilmont 1:fdd22bb7aa52 825 /* Second part of this stage computes the MAC operations greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 826
emilmont 1:fdd22bb7aa52 827 /* The first part of the stage starts here */
mbed_official 5:3762170b6d4d 828 while((count < 4u) && (blockSize1 > 0))
emilmont 1:fdd22bb7aa52 829 {
emilmont 1:fdd22bb7aa52 830 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 831 sum = 0;
emilmont 1:fdd22bb7aa52 832
emilmont 1:fdd22bb7aa52 833 /* Loop over number of MAC operations between
emilmont 1:fdd22bb7aa52 834 * inputA samples and inputB samples */
emilmont 1:fdd22bb7aa52 835 k = count;
emilmont 1:fdd22bb7aa52 836
emilmont 1:fdd22bb7aa52 837 while(k > 0u)
emilmont 1:fdd22bb7aa52 838 {
emilmont 1:fdd22bb7aa52 839 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 840 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 841
emilmont 1:fdd22bb7aa52 842 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 843 k--;
emilmont 1:fdd22bb7aa52 844 }
emilmont 1:fdd22bb7aa52 845
emilmont 1:fdd22bb7aa52 846 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 847 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 848
emilmont 1:fdd22bb7aa52 849 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 850 py = ++pSrc2;
emilmont 1:fdd22bb7aa52 851 px = pIn1;
emilmont 1:fdd22bb7aa52 852
emilmont 1:fdd22bb7aa52 853 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 854 count++;
emilmont 1:fdd22bb7aa52 855
emilmont 1:fdd22bb7aa52 856 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 857 blockSize1--;
emilmont 1:fdd22bb7aa52 858 }
emilmont 1:fdd22bb7aa52 859
emilmont 1:fdd22bb7aa52 860 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 861 /* The internal loop, over count, is unrolled by 4 */
emilmont 1:fdd22bb7aa52 862 /* To, read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 863 * y[srcBLen] and y[srcBLen-1] coefficients, py is decremented by 1 */
emilmont 1:fdd22bb7aa52 864 py = py - 1;
emilmont 1:fdd22bb7aa52 865
mbed_official 5:3762170b6d4d 866 while(blockSize1 > 0)
emilmont 1:fdd22bb7aa52 867 {
emilmont 1:fdd22bb7aa52 868 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 869 sum = 0;
emilmont 1:fdd22bb7aa52 870
emilmont 1:fdd22bb7aa52 871 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 872 k = count >> 2u;
emilmont 1:fdd22bb7aa52 873
emilmont 1:fdd22bb7aa52 874 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 875 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 2:da51fb522205 876 py++;
emilmont 1:fdd22bb7aa52 877
emilmont 1:fdd22bb7aa52 878 while(k > 0u)
emilmont 1:fdd22bb7aa52 879 {
emilmont 1:fdd22bb7aa52 880 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 881 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 882 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 883 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 884 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 885
emilmont 1:fdd22bb7aa52 886 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 887 k--;
emilmont 1:fdd22bb7aa52 888 }
emilmont 1:fdd22bb7aa52 889
emilmont 1:fdd22bb7aa52 890 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 891 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 892 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 893
emilmont 1:fdd22bb7aa52 894 while(k > 0u)
emilmont 1:fdd22bb7aa52 895 {
emilmont 1:fdd22bb7aa52 896 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 897 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 898
emilmont 1:fdd22bb7aa52 899 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 900 k--;
emilmont 1:fdd22bb7aa52 901 }
emilmont 1:fdd22bb7aa52 902
emilmont 1:fdd22bb7aa52 903 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 904 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 905
emilmont 1:fdd22bb7aa52 906 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 907 py = ++pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 908 px = pIn1;
emilmont 1:fdd22bb7aa52 909
emilmont 1:fdd22bb7aa52 910 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 911 count++;
emilmont 1:fdd22bb7aa52 912
emilmont 1:fdd22bb7aa52 913 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 914 blockSize1--;
emilmont 1:fdd22bb7aa52 915 }
emilmont 1:fdd22bb7aa52 916
emilmont 1:fdd22bb7aa52 917 /* --------------------------
emilmont 1:fdd22bb7aa52 918 * Initializations of stage2
emilmont 1:fdd22bb7aa52 919 * ------------------------*/
emilmont 1:fdd22bb7aa52 920
emilmont 1:fdd22bb7aa52 921 /* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
emilmont 1:fdd22bb7aa52 922 * sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
emilmont 1:fdd22bb7aa52 923 * ....
emilmont 1:fdd22bb7aa52 924 * sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] * y[0]
emilmont 1:fdd22bb7aa52 925 */
emilmont 1:fdd22bb7aa52 926
emilmont 1:fdd22bb7aa52 927 /* Working pointer of inputA */
mbed_official 5:3762170b6d4d 928 if((int32_t)firstIndex - (int32_t)srcBLen + 1 > 0)
mbed_official 5:3762170b6d4d 929 {
mbed_official 5:3762170b6d4d 930 px = pIn1 + firstIndex - srcBLen + 1;
mbed_official 5:3762170b6d4d 931 }
mbed_official 5:3762170b6d4d 932 else
mbed_official 5:3762170b6d4d 933 {
mbed_official 5:3762170b6d4d 934 px = pIn1;
mbed_official 5:3762170b6d4d 935 }
emilmont 1:fdd22bb7aa52 936
emilmont 1:fdd22bb7aa52 937 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 938 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 939 py = pSrc2;
emilmont 1:fdd22bb7aa52 940
emilmont 1:fdd22bb7aa52 941 /* count is the index by which the pointer pIn1 to be incremented */
emilmont 1:fdd22bb7aa52 942 count = 0u;
emilmont 1:fdd22bb7aa52 943
emilmont 1:fdd22bb7aa52 944
emilmont 1:fdd22bb7aa52 945 /* --------------------
emilmont 1:fdd22bb7aa52 946 * Stage2 process
emilmont 1:fdd22bb7aa52 947 * -------------------*/
emilmont 1:fdd22bb7aa52 948
emilmont 1:fdd22bb7aa52 949 /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
emilmont 1:fdd22bb7aa52 950 * So, to loop unroll over blockSize2,
emilmont 1:fdd22bb7aa52 951 * srcBLen should be greater than or equal to 4 */
emilmont 1:fdd22bb7aa52 952 if(srcBLen >= 4u)
emilmont 1:fdd22bb7aa52 953 {
emilmont 1:fdd22bb7aa52 954 /* Loop unroll over blockSize2, by 4 */
emilmont 1:fdd22bb7aa52 955 blkCnt = ((uint32_t) blockSize2 >> 2u);
emilmont 1:fdd22bb7aa52 956
emilmont 1:fdd22bb7aa52 957 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 958 {
emilmont 1:fdd22bb7aa52 959 py = py - 1u;
emilmont 1:fdd22bb7aa52 960
emilmont 1:fdd22bb7aa52 961 /* Set all accumulators to zero */
emilmont 1:fdd22bb7aa52 962 acc0 = 0;
emilmont 1:fdd22bb7aa52 963 acc1 = 0;
emilmont 1:fdd22bb7aa52 964 acc2 = 0;
emilmont 1:fdd22bb7aa52 965 acc3 = 0;
emilmont 1:fdd22bb7aa52 966
emilmont 1:fdd22bb7aa52 967 /* read x[0], x[1] samples */
emilmont 2:da51fb522205 968 a = *px++;
emilmont 2:da51fb522205 969 b = *px++;
emilmont 1:fdd22bb7aa52 970
emilmont 1:fdd22bb7aa52 971 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 972
emilmont 2:da51fb522205 973 x0 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 974 a = *px;
emilmont 2:da51fb522205 975 x1 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 976
emilmont 1:fdd22bb7aa52 977 #else
emilmont 1:fdd22bb7aa52 978
emilmont 2:da51fb522205 979 x0 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 980 a = *px;
emilmont 2:da51fb522205 981 x1 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 982
emilmont 2:da51fb522205 983 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 984
emilmont 1:fdd22bb7aa52 985 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 986 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 987
emilmont 1:fdd22bb7aa52 988 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 989 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 990 do
emilmont 1:fdd22bb7aa52 991 {
emilmont 1:fdd22bb7aa52 992 /* Read the last two inputB samples using SIMD:
emilmont 1:fdd22bb7aa52 993 * y[srcBLen - 1] and y[srcBLen - 2] */
emilmont 2:da51fb522205 994 a = *py;
emilmont 2:da51fb522205 995 b = *(py+1);
emilmont 2:da51fb522205 996 py -= 2;
emilmont 1:fdd22bb7aa52 997
emilmont 1:fdd22bb7aa52 998 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 999
emilmont 2:da51fb522205 1000 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1001
emilmont 1:fdd22bb7aa52 1002 #else
emilmont 1:fdd22bb7aa52 1003
emilmont 2:da51fb522205 1004 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1005
emilmont 2:da51fb522205 1006 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1007
emilmont 1:fdd22bb7aa52 1008 /* acc0 += x[0] * y[srcBLen - 1] + x[1] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 1009 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1010
emilmont 1:fdd22bb7aa52 1011 /* acc1 += x[1] * y[srcBLen - 1] + x[2] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 1012 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1013
emilmont 2:da51fb522205 1014 a = *px;
emilmont 2:da51fb522205 1015 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1016
emilmont 1:fdd22bb7aa52 1017 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 1018
emilmont 2:da51fb522205 1019 x2 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 1020 a = *(px + 2);
emilmont 2:da51fb522205 1021 x3 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1022
emilmont 1:fdd22bb7aa52 1023 #else
emilmont 1:fdd22bb7aa52 1024
emilmont 2:da51fb522205 1025 x2 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 1026 a = *(px + 2);
emilmont 2:da51fb522205 1027 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1028
emilmont 2:da51fb522205 1029 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1030
emilmont 1:fdd22bb7aa52 1031 /* acc2 += x[2] * y[srcBLen - 1] + x[3] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 1032 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 1033
emilmont 1:fdd22bb7aa52 1034 /* acc3 += x[3] * y[srcBLen - 1] + x[4] * y[srcBLen - 2] */
emilmont 1:fdd22bb7aa52 1035 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1036
emilmont 1:fdd22bb7aa52 1037 /* Read y[srcBLen - 3] and y[srcBLen - 4] */
emilmont 2:da51fb522205 1038 a = *py;
emilmont 2:da51fb522205 1039 b = *(py+1);
emilmont 2:da51fb522205 1040 py -= 2;
emilmont 1:fdd22bb7aa52 1041
emilmont 1:fdd22bb7aa52 1042 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1043
emilmont 2:da51fb522205 1044 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1045
emilmont 1:fdd22bb7aa52 1046 #else
emilmont 1:fdd22bb7aa52 1047
emilmont 2:da51fb522205 1048 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1049
emilmont 2:da51fb522205 1050 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1051
emilmont 1:fdd22bb7aa52 1052 /* acc0 += x[2] * y[srcBLen - 3] + x[3] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1053 acc0 = __SMLADX(x2, c0, acc0);
emilmont 1:fdd22bb7aa52 1054
emilmont 1:fdd22bb7aa52 1055 /* acc1 += x[3] * y[srcBLen - 3] + x[4] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1056 acc1 = __SMLADX(x3, c0, acc1);
emilmont 1:fdd22bb7aa52 1057
emilmont 1:fdd22bb7aa52 1058 /* Read x[4], x[5], x[6] */
emilmont 2:da51fb522205 1059 a = *(px + 2);
emilmont 2:da51fb522205 1060 b = *(px + 3);
emilmont 1:fdd22bb7aa52 1061
emilmont 1:fdd22bb7aa52 1062 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 1063
emilmont 2:da51fb522205 1064 x0 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 1065 a = *(px + 4);
emilmont 2:da51fb522205 1066 x1 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1067
emilmont 1:fdd22bb7aa52 1068 #else
emilmont 1:fdd22bb7aa52 1069
emilmont 2:da51fb522205 1070 x0 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 1071 a = *(px + 4);
emilmont 2:da51fb522205 1072 x1 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1073
emilmont 2:da51fb522205 1074 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1075
emilmont 2:da51fb522205 1076 px += 4u;
emilmont 1:fdd22bb7aa52 1077
emilmont 1:fdd22bb7aa52 1078 /* acc2 += x[4] * y[srcBLen - 3] + x[5] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1079 acc2 = __SMLADX(x0, c0, acc2);
emilmont 1:fdd22bb7aa52 1080
emilmont 1:fdd22bb7aa52 1081 /* acc3 += x[5] * y[srcBLen - 3] + x[6] * y[srcBLen - 4] */
emilmont 1:fdd22bb7aa52 1082 acc3 = __SMLADX(x1, c0, acc3);
emilmont 1:fdd22bb7aa52 1083
emilmont 1:fdd22bb7aa52 1084 } while(--k);
emilmont 1:fdd22bb7aa52 1085
emilmont 1:fdd22bb7aa52 1086 /* For the next MAC operations, SIMD is not used
emilmont 1:fdd22bb7aa52 1087 * So, the 16 bit pointer if inputB, py is updated */
emilmont 1:fdd22bb7aa52 1088
emilmont 1:fdd22bb7aa52 1089 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1090 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1091 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 1092
emilmont 1:fdd22bb7aa52 1093 if(k == 1u)
emilmont 1:fdd22bb7aa52 1094 {
emilmont 1:fdd22bb7aa52 1095 /* Read y[srcBLen - 5] */
emilmont 1:fdd22bb7aa52 1096 c0 = *(py+1);
emilmont 1:fdd22bb7aa52 1097
emilmont 1:fdd22bb7aa52 1098 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1099
emilmont 1:fdd22bb7aa52 1100 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 1101
emilmont 1:fdd22bb7aa52 1102 #else
emilmont 1:fdd22bb7aa52 1103
emilmont 1:fdd22bb7aa52 1104 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 1105
emilmont 1:fdd22bb7aa52 1106 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1107
emilmont 1:fdd22bb7aa52 1108 /* Read x[7] */
emilmont 2:da51fb522205 1109 a = *px;
emilmont 2:da51fb522205 1110 b = *(px+1);
emilmont 2:da51fb522205 1111 px++;
emilmont 1:fdd22bb7aa52 1112
emilmont 1:fdd22bb7aa52 1113 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1114
emilmont 2:da51fb522205 1115 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1116
emilmont 1:fdd22bb7aa52 1117 #else
emilmont 1:fdd22bb7aa52 1118
emilmont 2:da51fb522205 1119 x3 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1120
emilmont 2:da51fb522205 1121 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1122
emilmont 1:fdd22bb7aa52 1123
emilmont 1:fdd22bb7aa52 1124 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1125 acc0 = __SMLAD(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1126 acc1 = __SMLAD(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1127 acc2 = __SMLADX(x1, c0, acc2);
emilmont 1:fdd22bb7aa52 1128 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1129 }
emilmont 1:fdd22bb7aa52 1130
emilmont 1:fdd22bb7aa52 1131 if(k == 2u)
emilmont 1:fdd22bb7aa52 1132 {
emilmont 1:fdd22bb7aa52 1133 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 2:da51fb522205 1134 a = *py;
emilmont 2:da51fb522205 1135 b = *(py+1);
emilmont 1:fdd22bb7aa52 1136
emilmont 1:fdd22bb7aa52 1137 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1138
emilmont 2:da51fb522205 1139 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1140
emilmont 1:fdd22bb7aa52 1141 #else
emilmont 1:fdd22bb7aa52 1142
emilmont 2:da51fb522205 1143 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1144
emilmont 2:da51fb522205 1145 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1146
emilmont 1:fdd22bb7aa52 1147 /* Read x[7], x[8], x[9] */
emilmont 2:da51fb522205 1148 a = *px;
emilmont 2:da51fb522205 1149 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1150
emilmont 1:fdd22bb7aa52 1151 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 1152
emilmont 2:da51fb522205 1153 x3 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 1154 a = *(px + 2);
emilmont 2:da51fb522205 1155 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1156
emilmont 1:fdd22bb7aa52 1157 #else
emilmont 1:fdd22bb7aa52 1158
emilmont 2:da51fb522205 1159 x3 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 1160 a = *(px + 2);
emilmont 2:da51fb522205 1161 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1162
emilmont 2:da51fb522205 1163 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 2:da51fb522205 1164 px += 2u;
emilmont 1:fdd22bb7aa52 1165
emilmont 1:fdd22bb7aa52 1166 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1167 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1168 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1169 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 1170 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 1171 }
emilmont 1:fdd22bb7aa52 1172
emilmont 1:fdd22bb7aa52 1173 if(k == 3u)
emilmont 1:fdd22bb7aa52 1174 {
emilmont 1:fdd22bb7aa52 1175 /* Read y[srcBLen - 5], y[srcBLen - 6] */
emilmont 2:da51fb522205 1176 a = *py;
emilmont 2:da51fb522205 1177 b = *(py+1);
emilmont 1:fdd22bb7aa52 1178
emilmont 1:fdd22bb7aa52 1179 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1180
emilmont 2:da51fb522205 1181 c0 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1182
emilmont 1:fdd22bb7aa52 1183 #else
emilmont 1:fdd22bb7aa52 1184
emilmont 2:da51fb522205 1185 c0 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1186
emilmont 2:da51fb522205 1187 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1188
emilmont 1:fdd22bb7aa52 1189 /* Read x[7], x[8], x[9] */
emilmont 2:da51fb522205 1190 a = *px;
emilmont 2:da51fb522205 1191 b = *(px + 1);
emilmont 1:fdd22bb7aa52 1192
emilmont 1:fdd22bb7aa52 1193 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 2:da51fb522205 1194
emilmont 2:da51fb522205 1195 x3 = __PKHBT(a, b, 16);
emilmont 2:da51fb522205 1196 a = *(px + 2);
emilmont 2:da51fb522205 1197 x2 = __PKHBT(b, a, 16);
emilmont 1:fdd22bb7aa52 1198
emilmont 1:fdd22bb7aa52 1199 #else
emilmont 1:fdd22bb7aa52 1200
emilmont 2:da51fb522205 1201 x3 = __PKHBT(b, a, 16);
emilmont 2:da51fb522205 1202 a = *(px + 2);
emilmont 2:da51fb522205 1203 x2 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1204
emilmont 2:da51fb522205 1205 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1206
emilmont 1:fdd22bb7aa52 1207 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1208 acc0 = __SMLADX(x0, c0, acc0);
emilmont 1:fdd22bb7aa52 1209 acc1 = __SMLADX(x1, c0, acc1);
emilmont 1:fdd22bb7aa52 1210 acc2 = __SMLADX(x3, c0, acc2);
emilmont 1:fdd22bb7aa52 1211 acc3 = __SMLADX(x2, c0, acc3);
emilmont 1:fdd22bb7aa52 1212
emilmont 1:fdd22bb7aa52 1213 /* Read y[srcBLen - 7] */
emilmont 2:da51fb522205 1214 c0 = *(py-1);
emilmont 1:fdd22bb7aa52 1215 #ifdef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1216
emilmont 1:fdd22bb7aa52 1217 c0 = c0 << 16u;
emilmont 1:fdd22bb7aa52 1218 #else
emilmont 1:fdd22bb7aa52 1219
emilmont 1:fdd22bb7aa52 1220 c0 = c0 & 0x0000FFFF;
emilmont 1:fdd22bb7aa52 1221 #endif /* #ifdef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1222
emilmont 1:fdd22bb7aa52 1223 /* Read x[10] */
emilmont 2:da51fb522205 1224 a = *(px+2);
emilmont 2:da51fb522205 1225 b = *(px+3);
emilmont 1:fdd22bb7aa52 1226
emilmont 1:fdd22bb7aa52 1227 #ifndef ARM_MATH_BIG_ENDIAN
emilmont 1:fdd22bb7aa52 1228
emilmont 2:da51fb522205 1229 x3 = __PKHBT(a, b, 16);
emilmont 1:fdd22bb7aa52 1230
emilmont 1:fdd22bb7aa52 1231 #else
emilmont 1:fdd22bb7aa52 1232
emilmont 2:da51fb522205 1233 x3 = __PKHBT(b, a, 16);;
emilmont 1:fdd22bb7aa52 1234
emilmont 2:da51fb522205 1235 #endif /* #ifndef ARM_MATH_BIG_ENDIAN */
emilmont 1:fdd22bb7aa52 1236
emilmont 2:da51fb522205 1237 px += 3u;
emilmont 1:fdd22bb7aa52 1238
emilmont 1:fdd22bb7aa52 1239 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1240 acc0 = __SMLADX(x1, c0, acc0);
emilmont 1:fdd22bb7aa52 1241 acc1 = __SMLAD(x2, c0, acc1);
emilmont 1:fdd22bb7aa52 1242 acc2 = __SMLADX(x2, c0, acc2);
emilmont 1:fdd22bb7aa52 1243 acc3 = __SMLADX(x3, c0, acc3);
emilmont 1:fdd22bb7aa52 1244 }
emilmont 1:fdd22bb7aa52 1245
emilmont 1:fdd22bb7aa52 1246 /* Store the results in the accumulators in the destination buffer. */
emilmont 2:da51fb522205 1247 *pOut++ = (q15_t)(acc0 >> 15);
emilmont 2:da51fb522205 1248 *pOut++ = (q15_t)(acc1 >> 15);
emilmont 2:da51fb522205 1249 *pOut++ = (q15_t)(acc2 >> 15);
emilmont 2:da51fb522205 1250 *pOut++ = (q15_t)(acc3 >> 15);
emilmont 1:fdd22bb7aa52 1251
emilmont 1:fdd22bb7aa52 1252 /* Increment the pointer pIn1 index, count by 4 */
emilmont 1:fdd22bb7aa52 1253 count += 4u;
emilmont 1:fdd22bb7aa52 1254
emilmont 1:fdd22bb7aa52 1255 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1256 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1257 py = pSrc2;
emilmont 1:fdd22bb7aa52 1258
emilmont 1:fdd22bb7aa52 1259 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1260 blkCnt--;
emilmont 1:fdd22bb7aa52 1261 }
emilmont 1:fdd22bb7aa52 1262
emilmont 1:fdd22bb7aa52 1263 /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
emilmont 1:fdd22bb7aa52 1264 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1265 blkCnt = (uint32_t) blockSize2 % 0x4u;
emilmont 1:fdd22bb7aa52 1266
emilmont 1:fdd22bb7aa52 1267 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 1268 {
emilmont 1:fdd22bb7aa52 1269 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1270 sum = 0;
emilmont 1:fdd22bb7aa52 1271
emilmont 1:fdd22bb7aa52 1272 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1273 k = srcBLen >> 2u;
emilmont 1:fdd22bb7aa52 1274
emilmont 1:fdd22bb7aa52 1275 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 1276 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 1:fdd22bb7aa52 1277 while(k > 0u)
emilmont 1:fdd22bb7aa52 1278 {
emilmont 1:fdd22bb7aa52 1279 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1280 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1281 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1282 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1283 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1284
emilmont 1:fdd22bb7aa52 1285 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1286 k--;
emilmont 1:fdd22bb7aa52 1287 }
emilmont 1:fdd22bb7aa52 1288
emilmont 1:fdd22bb7aa52 1289 /* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1290 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1291 k = srcBLen % 0x4u;
emilmont 1:fdd22bb7aa52 1292
emilmont 1:fdd22bb7aa52 1293 while(k > 0u)
emilmont 1:fdd22bb7aa52 1294 {
emilmont 1:fdd22bb7aa52 1295 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1296 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1297
emilmont 1:fdd22bb7aa52 1298 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1299 k--;
emilmont 1:fdd22bb7aa52 1300 }
emilmont 1:fdd22bb7aa52 1301
emilmont 1:fdd22bb7aa52 1302 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1303 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1304
emilmont 1:fdd22bb7aa52 1305 /* Increment the pointer pIn1 index, count by 1 */
emilmont 1:fdd22bb7aa52 1306 count++;
emilmont 1:fdd22bb7aa52 1307
emilmont 1:fdd22bb7aa52 1308 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1309 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1310 py = pSrc2;
emilmont 1:fdd22bb7aa52 1311
emilmont 1:fdd22bb7aa52 1312 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1313 blkCnt--;
emilmont 1:fdd22bb7aa52 1314 }
emilmont 1:fdd22bb7aa52 1315 }
emilmont 1:fdd22bb7aa52 1316 else
emilmont 1:fdd22bb7aa52 1317 {
emilmont 1:fdd22bb7aa52 1318 /* If the srcBLen is not a multiple of 4,
emilmont 1:fdd22bb7aa52 1319 * the blockSize2 loop cannot be unrolled by 4 */
emilmont 1:fdd22bb7aa52 1320 blkCnt = (uint32_t) blockSize2;
emilmont 1:fdd22bb7aa52 1321
emilmont 1:fdd22bb7aa52 1322 while(blkCnt > 0u)
emilmont 1:fdd22bb7aa52 1323 {
emilmont 1:fdd22bb7aa52 1324 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1325 sum = 0;
emilmont 1:fdd22bb7aa52 1326
emilmont 1:fdd22bb7aa52 1327 /* srcBLen number of MACS should be performed */
emilmont 1:fdd22bb7aa52 1328 k = srcBLen;
emilmont 1:fdd22bb7aa52 1329
emilmont 1:fdd22bb7aa52 1330 while(k > 0u)
emilmont 1:fdd22bb7aa52 1331 {
emilmont 1:fdd22bb7aa52 1332 /* Perform the multiply-accumulate */
emilmont 1:fdd22bb7aa52 1333 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1334
emilmont 1:fdd22bb7aa52 1335 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1336 k--;
emilmont 1:fdd22bb7aa52 1337 }
emilmont 1:fdd22bb7aa52 1338
emilmont 1:fdd22bb7aa52 1339 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1340 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1341
emilmont 1:fdd22bb7aa52 1342 /* Increment the MAC count */
emilmont 1:fdd22bb7aa52 1343 count++;
emilmont 1:fdd22bb7aa52 1344
emilmont 1:fdd22bb7aa52 1345 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1346 px = pIn1 + count;
emilmont 1:fdd22bb7aa52 1347 py = pSrc2;
emilmont 1:fdd22bb7aa52 1348
emilmont 1:fdd22bb7aa52 1349 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1350 blkCnt--;
emilmont 1:fdd22bb7aa52 1351 }
emilmont 1:fdd22bb7aa52 1352 }
emilmont 1:fdd22bb7aa52 1353
emilmont 1:fdd22bb7aa52 1354
emilmont 1:fdd22bb7aa52 1355 /* --------------------------
emilmont 1:fdd22bb7aa52 1356 * Initializations of stage3
emilmont 1:fdd22bb7aa52 1357 * -------------------------*/
emilmont 1:fdd22bb7aa52 1358
emilmont 1:fdd22bb7aa52 1359 /* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+ x[srcALen-1] * y[1]
emilmont 1:fdd22bb7aa52 1360 * sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] * y[srcBLen-2] +...+ x[srcALen-1] * y[2]
emilmont 1:fdd22bb7aa52 1361 * ....
emilmont 1:fdd22bb7aa52 1362 * sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
emilmont 1:fdd22bb7aa52 1363 * sum += x[srcALen-1] * y[srcBLen-1]
emilmont 1:fdd22bb7aa52 1364 */
emilmont 1:fdd22bb7aa52 1365
emilmont 1:fdd22bb7aa52 1366 /* In this stage the MAC operations are decreased by 1 for every iteration.
emilmont 1:fdd22bb7aa52 1367 The count variable holds the number of MAC operations performed */
emilmont 1:fdd22bb7aa52 1368 count = srcBLen - 1u;
emilmont 1:fdd22bb7aa52 1369
emilmont 1:fdd22bb7aa52 1370 /* Working pointer of inputA */
emilmont 1:fdd22bb7aa52 1371 pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 1372 px = pSrc1;
emilmont 1:fdd22bb7aa52 1373
emilmont 1:fdd22bb7aa52 1374 /* Working pointer of inputB */
emilmont 1:fdd22bb7aa52 1375 pSrc2 = pIn2 + (srcBLen - 1u);
emilmont 1:fdd22bb7aa52 1376 pIn2 = pSrc2 - 1u;
emilmont 1:fdd22bb7aa52 1377 py = pIn2;
emilmont 1:fdd22bb7aa52 1378
emilmont 1:fdd22bb7aa52 1379 /* -------------------
emilmont 1:fdd22bb7aa52 1380 * Stage3 process
emilmont 1:fdd22bb7aa52 1381 * ------------------*/
emilmont 1:fdd22bb7aa52 1382
emilmont 1:fdd22bb7aa52 1383 /* For loop unrolling by 4, this stage is divided into two. */
emilmont 1:fdd22bb7aa52 1384 /* First part of this stage computes the MAC operations greater than 4 */
emilmont 1:fdd22bb7aa52 1385 /* Second part of this stage computes the MAC operations less than or equal to 4 */
emilmont 1:fdd22bb7aa52 1386
emilmont 1:fdd22bb7aa52 1387 /* The first part of the stage starts here */
emilmont 1:fdd22bb7aa52 1388 j = count >> 2u;
emilmont 1:fdd22bb7aa52 1389
emilmont 1:fdd22bb7aa52 1390 while((j > 0u) && (blockSize3 > 0))
emilmont 1:fdd22bb7aa52 1391 {
emilmont 1:fdd22bb7aa52 1392 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1393 sum = 0;
emilmont 1:fdd22bb7aa52 1394
emilmont 1:fdd22bb7aa52 1395 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1396 k = count >> 2u;
emilmont 1:fdd22bb7aa52 1397
emilmont 1:fdd22bb7aa52 1398 /* First part of the processing with loop unrolling. Compute 4 MACs at a time.
emilmont 1:fdd22bb7aa52 1399 ** a second loop below computes MACs for the remaining 1 to 3 samples. */
emilmont 2:da51fb522205 1400 py++;
emilmont 1:fdd22bb7aa52 1401
emilmont 1:fdd22bb7aa52 1402 while(k > 0u)
emilmont 2:da51fb522205 1403 {
emilmont 1:fdd22bb7aa52 1404 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1405 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1406 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1407 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1408 sum += ((q31_t) * px++ * *py--);
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
emilmont 1:fdd22bb7aa52 1414 /* If the count is not a multiple of 4, compute any remaining MACs here.
emilmont 1:fdd22bb7aa52 1415 ** No loop unrolling is used. */
emilmont 1:fdd22bb7aa52 1416 k = count % 0x4u;
emilmont 1:fdd22bb7aa52 1417
emilmont 1:fdd22bb7aa52 1418 while(k > 0u)
emilmont 1:fdd22bb7aa52 1419 {
emilmont 1:fdd22bb7aa52 1420 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1421 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1422
emilmont 1:fdd22bb7aa52 1423 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1424 k--;
emilmont 1:fdd22bb7aa52 1425 }
emilmont 1:fdd22bb7aa52 1426
emilmont 1:fdd22bb7aa52 1427 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1428 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1429
emilmont 1:fdd22bb7aa52 1430 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1431 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 1432 py = pIn2;
emilmont 1:fdd22bb7aa52 1433
emilmont 1:fdd22bb7aa52 1434 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 1435 count--;
emilmont 1:fdd22bb7aa52 1436
emilmont 1:fdd22bb7aa52 1437 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1438 blockSize3--;
emilmont 1:fdd22bb7aa52 1439
emilmont 1:fdd22bb7aa52 1440 j--;
emilmont 1:fdd22bb7aa52 1441 }
emilmont 1:fdd22bb7aa52 1442
emilmont 1:fdd22bb7aa52 1443 /* The second part of the stage starts here */
emilmont 1:fdd22bb7aa52 1444 /* SIMD is not used for the next MAC operations,
emilmont 1:fdd22bb7aa52 1445 * so pointer py is updated to read only one sample at a time */
emilmont 1:fdd22bb7aa52 1446 py = py + 1u;
emilmont 1:fdd22bb7aa52 1447
mbed_official 5:3762170b6d4d 1448 while(blockSize3 > 0)
emilmont 1:fdd22bb7aa52 1449 {
emilmont 1:fdd22bb7aa52 1450 /* Accumulator is made zero for every iteration */
emilmont 1:fdd22bb7aa52 1451 sum = 0;
emilmont 1:fdd22bb7aa52 1452
emilmont 1:fdd22bb7aa52 1453 /* Apply loop unrolling and compute 4 MACs simultaneously. */
emilmont 1:fdd22bb7aa52 1454 k = count;
emilmont 1:fdd22bb7aa52 1455
emilmont 1:fdd22bb7aa52 1456 while(k > 0u)
emilmont 1:fdd22bb7aa52 1457 {
emilmont 1:fdd22bb7aa52 1458 /* Perform the multiply-accumulates */
emilmont 1:fdd22bb7aa52 1459 /* sum += x[srcALen-1] * y[srcBLen-1] */
emilmont 1:fdd22bb7aa52 1460 sum += ((q31_t) * px++ * *py--);
emilmont 1:fdd22bb7aa52 1461
emilmont 1:fdd22bb7aa52 1462 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1463 k--;
emilmont 1:fdd22bb7aa52 1464 }
emilmont 1:fdd22bb7aa52 1465
emilmont 1:fdd22bb7aa52 1466 /* Store the result in the accumulator in the destination buffer. */
emilmont 1:fdd22bb7aa52 1467 *pOut++ = (q15_t) (sum >> 15);
emilmont 1:fdd22bb7aa52 1468
emilmont 1:fdd22bb7aa52 1469 /* Update the inputA and inputB pointers for next MAC calculation */
emilmont 1:fdd22bb7aa52 1470 px = ++pSrc1;
emilmont 1:fdd22bb7aa52 1471 py = pSrc2;
emilmont 1:fdd22bb7aa52 1472
emilmont 1:fdd22bb7aa52 1473 /* Decrement the MAC count */
emilmont 1:fdd22bb7aa52 1474 count--;
emilmont 1:fdd22bb7aa52 1475
emilmont 1:fdd22bb7aa52 1476 /* Decrement the loop counter */
emilmont 1:fdd22bb7aa52 1477 blockSize3--;
emilmont 1:fdd22bb7aa52 1478 }
emilmont 1:fdd22bb7aa52 1479
emilmont 1:fdd22bb7aa52 1480 /* set status as ARM_MATH_SUCCESS */
emilmont 1:fdd22bb7aa52 1481 status = ARM_MATH_SUCCESS;
emilmont 1:fdd22bb7aa52 1482 }
emilmont 1:fdd22bb7aa52 1483
emilmont 1:fdd22bb7aa52 1484 /* Return to application */
emilmont 1:fdd22bb7aa52 1485 return (status);
emilmont 1:fdd22bb7aa52 1486
emilmont 1:fdd22bb7aa52 1487 #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */
emilmont 1:fdd22bb7aa52 1488 }
emilmont 1:fdd22bb7aa52 1489
emilmont 1:fdd22bb7aa52 1490 /**
emilmont 1:fdd22bb7aa52 1491 * @} end of PartialConv group
emilmont 1:fdd22bb7aa52 1492 */