The CMSIS DSP 5 library
Dependents: Nucleo-Heart-Rate ejercicioVrms2 PROYECTOFINAL ejercicioVrms ... more
functions/FilteringFunctions/arm_correlate_fast_q15.c@3:4098b9d3d571, 2018-06-21 (annotated)
- Committer:
- xorjoep
- Date:
- Thu Jun 21 11:56:27 2018 +0000
- Revision:
- 3:4098b9d3d571
- Parent:
- 1:24714b45cd1b
headers is a folder not a library
Who changed what in which revision?
User | Revision | Line number | New contents of line |
---|---|---|---|
xorjoep | 1:24714b45cd1b | 1 | /* ---------------------------------------------------------------------- |
xorjoep | 1:24714b45cd1b | 2 | * Project: CMSIS DSP Library |
xorjoep | 1:24714b45cd1b | 3 | * Title: arm_correlate_fast_q15.c |
xorjoep | 1:24714b45cd1b | 4 | * Description: Fast Q15 Correlation |
xorjoep | 1:24714b45cd1b | 5 | * |
xorjoep | 1:24714b45cd1b | 6 | * $Date: 27. January 2017 |
xorjoep | 1:24714b45cd1b | 7 | * $Revision: V.1.5.1 |
xorjoep | 1:24714b45cd1b | 8 | * |
xorjoep | 1:24714b45cd1b | 9 | * Target Processor: Cortex-M cores |
xorjoep | 1:24714b45cd1b | 10 | * -------------------------------------------------------------------- */ |
xorjoep | 1:24714b45cd1b | 11 | /* |
xorjoep | 1:24714b45cd1b | 12 | * Copyright (C) 2010-2017 ARM Limited or its affiliates. All rights reserved. |
xorjoep | 1:24714b45cd1b | 13 | * |
xorjoep | 1:24714b45cd1b | 14 | * SPDX-License-Identifier: Apache-2.0 |
xorjoep | 1:24714b45cd1b | 15 | * |
xorjoep | 1:24714b45cd1b | 16 | * Licensed under the Apache License, Version 2.0 (the License); you may |
xorjoep | 1:24714b45cd1b | 17 | * not use this file except in compliance with the License. |
xorjoep | 1:24714b45cd1b | 18 | * You may obtain a copy of the License at |
xorjoep | 1:24714b45cd1b | 19 | * |
xorjoep | 1:24714b45cd1b | 20 | * www.apache.org/licenses/LICENSE-2.0 |
xorjoep | 1:24714b45cd1b | 21 | * |
xorjoep | 1:24714b45cd1b | 22 | * Unless required by applicable law or agreed to in writing, software |
xorjoep | 1:24714b45cd1b | 23 | * distributed under the License is distributed on an AS IS BASIS, WITHOUT |
xorjoep | 1:24714b45cd1b | 24 | * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
xorjoep | 1:24714b45cd1b | 25 | * See the License for the specific language governing permissions and |
xorjoep | 1:24714b45cd1b | 26 | * limitations under the License. |
xorjoep | 1:24714b45cd1b | 27 | */ |
xorjoep | 1:24714b45cd1b | 28 | |
xorjoep | 1:24714b45cd1b | 29 | #include "arm_math.h" |
xorjoep | 1:24714b45cd1b | 30 | |
xorjoep | 1:24714b45cd1b | 31 | /** |
xorjoep | 1:24714b45cd1b | 32 | * @ingroup groupFilters |
xorjoep | 1:24714b45cd1b | 33 | */ |
xorjoep | 1:24714b45cd1b | 34 | |
xorjoep | 1:24714b45cd1b | 35 | /** |
xorjoep | 1:24714b45cd1b | 36 | * @addtogroup Corr |
xorjoep | 1:24714b45cd1b | 37 | * @{ |
xorjoep | 1:24714b45cd1b | 38 | */ |
xorjoep | 1:24714b45cd1b | 39 | |
xorjoep | 1:24714b45cd1b | 40 | /** |
xorjoep | 1:24714b45cd1b | 41 | * @brief Correlation of Q15 sequences (fast version) for Cortex-M3 and Cortex-M4. |
xorjoep | 1:24714b45cd1b | 42 | * @param[in] *pSrcA points to the first input sequence. |
xorjoep | 1:24714b45cd1b | 43 | * @param[in] srcALen length of the first input sequence. |
xorjoep | 1:24714b45cd1b | 44 | * @param[in] *pSrcB points to the second input sequence. |
xorjoep | 1:24714b45cd1b | 45 | * @param[in] srcBLen length of the second input sequence. |
xorjoep | 1:24714b45cd1b | 46 | * @param[out] *pDst points to the location where the output result is written. Length 2 * max(srcALen, srcBLen) - 1. |
xorjoep | 1:24714b45cd1b | 47 | * @return none. |
xorjoep | 1:24714b45cd1b | 48 | * |
xorjoep | 1:24714b45cd1b | 49 | * <b>Scaling and Overflow Behavior:</b> |
xorjoep | 1:24714b45cd1b | 50 | * |
xorjoep | 1:24714b45cd1b | 51 | * \par |
xorjoep | 1:24714b45cd1b | 52 | * This fast version uses a 32-bit accumulator with 2.30 format. |
xorjoep | 1:24714b45cd1b | 53 | * The accumulator maintains full precision of the intermediate multiplication results but provides only a single guard bit. |
xorjoep | 1:24714b45cd1b | 54 | * There is no saturation on intermediate additions. |
xorjoep | 1:24714b45cd1b | 55 | * Thus, if the accumulator overflows it wraps around and distorts the result. |
xorjoep | 1:24714b45cd1b | 56 | * The input signals should be scaled down to avoid intermediate overflows. |
xorjoep | 1:24714b45cd1b | 57 | * Scale down one of the inputs by 1/min(srcALen, srcBLen) to avoid overflow since a |
xorjoep | 1:24714b45cd1b | 58 | * maximum of min(srcALen, srcBLen) number of additions is carried internally. |
xorjoep | 1:24714b45cd1b | 59 | * The 2.30 accumulator is right shifted by 15 bits and then saturated to 1.15 format to yield the final result. |
xorjoep | 1:24714b45cd1b | 60 | * |
xorjoep | 1:24714b45cd1b | 61 | * \par |
xorjoep | 1:24714b45cd1b | 62 | * See <code>arm_correlate_q15()</code> for a slower implementation of this function which uses a 64-bit accumulator to avoid wrap around distortion. |
xorjoep | 1:24714b45cd1b | 63 | */ |
xorjoep | 1:24714b45cd1b | 64 | |
xorjoep | 1:24714b45cd1b | 65 | void arm_correlate_fast_q15( |
xorjoep | 1:24714b45cd1b | 66 | q15_t * pSrcA, |
xorjoep | 1:24714b45cd1b | 67 | uint32_t srcALen, |
xorjoep | 1:24714b45cd1b | 68 | q15_t * pSrcB, |
xorjoep | 1:24714b45cd1b | 69 | uint32_t srcBLen, |
xorjoep | 1:24714b45cd1b | 70 | q15_t * pDst) |
xorjoep | 1:24714b45cd1b | 71 | { |
xorjoep | 1:24714b45cd1b | 72 | #ifndef UNALIGNED_SUPPORT_DISABLE |
xorjoep | 1:24714b45cd1b | 73 | |
xorjoep | 1:24714b45cd1b | 74 | q15_t *pIn1; /* inputA pointer */ |
xorjoep | 1:24714b45cd1b | 75 | q15_t *pIn2; /* inputB pointer */ |
xorjoep | 1:24714b45cd1b | 76 | q15_t *pOut = pDst; /* output pointer */ |
xorjoep | 1:24714b45cd1b | 77 | q31_t sum, acc0, acc1, acc2, acc3; /* Accumulators */ |
xorjoep | 1:24714b45cd1b | 78 | q15_t *px; /* Intermediate inputA pointer */ |
xorjoep | 1:24714b45cd1b | 79 | q15_t *py; /* Intermediate inputB pointer */ |
xorjoep | 1:24714b45cd1b | 80 | q15_t *pSrc1; /* Intermediate pointers */ |
xorjoep | 1:24714b45cd1b | 81 | q31_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */ |
xorjoep | 1:24714b45cd1b | 82 | uint32_t j, k = 0U, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counter */ |
xorjoep | 1:24714b45cd1b | 83 | int32_t inc = 1; /* Destination address modifier */ |
xorjoep | 1:24714b45cd1b | 84 | |
xorjoep | 1:24714b45cd1b | 85 | |
xorjoep | 1:24714b45cd1b | 86 | /* The algorithm implementation is based on the lengths of the inputs. */ |
xorjoep | 1:24714b45cd1b | 87 | /* srcB is always made to slide across srcA. */ |
xorjoep | 1:24714b45cd1b | 88 | /* So srcBLen is always considered as shorter or equal to srcALen */ |
xorjoep | 1:24714b45cd1b | 89 | /* But CORR(x, y) is reverse of CORR(y, x) */ |
xorjoep | 1:24714b45cd1b | 90 | /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */ |
xorjoep | 1:24714b45cd1b | 91 | /* and the destination pointer modifier, inc is set to -1 */ |
xorjoep | 1:24714b45cd1b | 92 | /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */ |
xorjoep | 1:24714b45cd1b | 93 | /* But to improve the performance, |
xorjoep | 1:24714b45cd1b | 94 | * we include zeroes in the output instead of zero padding either of the the inputs*/ |
xorjoep | 1:24714b45cd1b | 95 | /* If srcALen > srcBLen, |
xorjoep | 1:24714b45cd1b | 96 | * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */ |
xorjoep | 1:24714b45cd1b | 97 | /* If srcALen < srcBLen, |
xorjoep | 1:24714b45cd1b | 98 | * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */ |
xorjoep | 1:24714b45cd1b | 99 | if (srcALen >= srcBLen) |
xorjoep | 1:24714b45cd1b | 100 | { |
xorjoep | 1:24714b45cd1b | 101 | /* Initialization of inputA pointer */ |
xorjoep | 1:24714b45cd1b | 102 | pIn1 = (pSrcA); |
xorjoep | 1:24714b45cd1b | 103 | |
xorjoep | 1:24714b45cd1b | 104 | /* Initialization of inputB pointer */ |
xorjoep | 1:24714b45cd1b | 105 | pIn2 = (pSrcB); |
xorjoep | 1:24714b45cd1b | 106 | |
xorjoep | 1:24714b45cd1b | 107 | /* Number of output samples is calculated */ |
xorjoep | 1:24714b45cd1b | 108 | outBlockSize = (2U * srcALen) - 1U; |
xorjoep | 1:24714b45cd1b | 109 | |
xorjoep | 1:24714b45cd1b | 110 | /* When srcALen > srcBLen, zero padding is done to srcB |
xorjoep | 1:24714b45cd1b | 111 | * to make their lengths equal. |
xorjoep | 1:24714b45cd1b | 112 | * Instead, (outBlockSize - (srcALen + srcBLen - 1)) |
xorjoep | 1:24714b45cd1b | 113 | * number of output samples are made zero */ |
xorjoep | 1:24714b45cd1b | 114 | j = outBlockSize - (srcALen + (srcBLen - 1U)); |
xorjoep | 1:24714b45cd1b | 115 | |
xorjoep | 1:24714b45cd1b | 116 | /* Updating the pointer position to non zero value */ |
xorjoep | 1:24714b45cd1b | 117 | pOut += j; |
xorjoep | 1:24714b45cd1b | 118 | |
xorjoep | 1:24714b45cd1b | 119 | } |
xorjoep | 1:24714b45cd1b | 120 | else |
xorjoep | 1:24714b45cd1b | 121 | { |
xorjoep | 1:24714b45cd1b | 122 | /* Initialization of inputA pointer */ |
xorjoep | 1:24714b45cd1b | 123 | pIn1 = (pSrcB); |
xorjoep | 1:24714b45cd1b | 124 | |
xorjoep | 1:24714b45cd1b | 125 | /* Initialization of inputB pointer */ |
xorjoep | 1:24714b45cd1b | 126 | pIn2 = (pSrcA); |
xorjoep | 1:24714b45cd1b | 127 | |
xorjoep | 1:24714b45cd1b | 128 | /* srcBLen is always considered as shorter or equal to srcALen */ |
xorjoep | 1:24714b45cd1b | 129 | j = srcBLen; |
xorjoep | 1:24714b45cd1b | 130 | srcBLen = srcALen; |
xorjoep | 1:24714b45cd1b | 131 | srcALen = j; |
xorjoep | 1:24714b45cd1b | 132 | |
xorjoep | 1:24714b45cd1b | 133 | /* CORR(x, y) = Reverse order(CORR(y, x)) */ |
xorjoep | 1:24714b45cd1b | 134 | /* Hence set the destination pointer to point to the last output sample */ |
xorjoep | 1:24714b45cd1b | 135 | pOut = pDst + ((srcALen + srcBLen) - 2U); |
xorjoep | 1:24714b45cd1b | 136 | |
xorjoep | 1:24714b45cd1b | 137 | /* Destination address modifier is set to -1 */ |
xorjoep | 1:24714b45cd1b | 138 | inc = -1; |
xorjoep | 1:24714b45cd1b | 139 | |
xorjoep | 1:24714b45cd1b | 140 | } |
xorjoep | 1:24714b45cd1b | 141 | |
xorjoep | 1:24714b45cd1b | 142 | /* The function is internally |
xorjoep | 1:24714b45cd1b | 143 | * divided into three parts according to the number of multiplications that has to be |
xorjoep | 1:24714b45cd1b | 144 | * taken place between inputA samples and inputB samples. In the first part of the |
xorjoep | 1:24714b45cd1b | 145 | * algorithm, the multiplications increase by one for every iteration. |
xorjoep | 1:24714b45cd1b | 146 | * In the second part of the algorithm, srcBLen number of multiplications are done. |
xorjoep | 1:24714b45cd1b | 147 | * In the third part of the algorithm, the multiplications decrease by one |
xorjoep | 1:24714b45cd1b | 148 | * for every iteration.*/ |
xorjoep | 1:24714b45cd1b | 149 | /* The algorithm is implemented in three stages. |
xorjoep | 1:24714b45cd1b | 150 | * The loop counters of each stage is initiated here. */ |
xorjoep | 1:24714b45cd1b | 151 | blockSize1 = srcBLen - 1U; |
xorjoep | 1:24714b45cd1b | 152 | blockSize2 = srcALen - (srcBLen - 1U); |
xorjoep | 1:24714b45cd1b | 153 | blockSize3 = blockSize1; |
xorjoep | 1:24714b45cd1b | 154 | |
xorjoep | 1:24714b45cd1b | 155 | /* -------------------------- |
xorjoep | 1:24714b45cd1b | 156 | * Initializations of stage1 |
xorjoep | 1:24714b45cd1b | 157 | * -------------------------*/ |
xorjoep | 1:24714b45cd1b | 158 | |
xorjoep | 1:24714b45cd1b | 159 | /* sum = x[0] * y[srcBlen - 1] |
xorjoep | 1:24714b45cd1b | 160 | * sum = x[0] * y[srcBlen - 2] + x[1] * y[srcBlen - 1] |
xorjoep | 1:24714b45cd1b | 161 | * .... |
xorjoep | 1:24714b45cd1b | 162 | * sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen - 1] * y[srcBLen - 1] |
xorjoep | 1:24714b45cd1b | 163 | */ |
xorjoep | 1:24714b45cd1b | 164 | |
xorjoep | 1:24714b45cd1b | 165 | /* In this stage the MAC operations are increased by 1 for every iteration. |
xorjoep | 1:24714b45cd1b | 166 | The count variable holds the number of MAC operations performed */ |
xorjoep | 1:24714b45cd1b | 167 | count = 1U; |
xorjoep | 1:24714b45cd1b | 168 | |
xorjoep | 1:24714b45cd1b | 169 | /* Working pointer of inputA */ |
xorjoep | 1:24714b45cd1b | 170 | px = pIn1; |
xorjoep | 1:24714b45cd1b | 171 | |
xorjoep | 1:24714b45cd1b | 172 | /* Working pointer of inputB */ |
xorjoep | 1:24714b45cd1b | 173 | pSrc1 = pIn2 + (srcBLen - 1U); |
xorjoep | 1:24714b45cd1b | 174 | py = pSrc1; |
xorjoep | 1:24714b45cd1b | 175 | |
xorjoep | 1:24714b45cd1b | 176 | /* ------------------------ |
xorjoep | 1:24714b45cd1b | 177 | * Stage1 process |
xorjoep | 1:24714b45cd1b | 178 | * ----------------------*/ |
xorjoep | 1:24714b45cd1b | 179 | |
xorjoep | 1:24714b45cd1b | 180 | /* The first loop starts here */ |
xorjoep | 1:24714b45cd1b | 181 | while (blockSize1 > 0U) |
xorjoep | 1:24714b45cd1b | 182 | { |
xorjoep | 1:24714b45cd1b | 183 | /* Accumulator is made zero for every iteration */ |
xorjoep | 1:24714b45cd1b | 184 | sum = 0; |
xorjoep | 1:24714b45cd1b | 185 | |
xorjoep | 1:24714b45cd1b | 186 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
xorjoep | 1:24714b45cd1b | 187 | k = count >> 2; |
xorjoep | 1:24714b45cd1b | 188 | |
xorjoep | 1:24714b45cd1b | 189 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
xorjoep | 1:24714b45cd1b | 190 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
xorjoep | 1:24714b45cd1b | 191 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 192 | { |
xorjoep | 1:24714b45cd1b | 193 | /* x[0] * y[srcBLen - 4] , x[1] * y[srcBLen - 3] */ |
xorjoep | 1:24714b45cd1b | 194 | sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); |
xorjoep | 1:24714b45cd1b | 195 | /* x[3] * y[srcBLen - 1] , x[2] * y[srcBLen - 2] */ |
xorjoep | 1:24714b45cd1b | 196 | sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); |
xorjoep | 1:24714b45cd1b | 197 | |
xorjoep | 1:24714b45cd1b | 198 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 199 | k--; |
xorjoep | 1:24714b45cd1b | 200 | } |
xorjoep | 1:24714b45cd1b | 201 | |
xorjoep | 1:24714b45cd1b | 202 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
xorjoep | 1:24714b45cd1b | 203 | ** No loop unrolling is used. */ |
xorjoep | 1:24714b45cd1b | 204 | k = count % 0x4U; |
xorjoep | 1:24714b45cd1b | 205 | |
xorjoep | 1:24714b45cd1b | 206 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 207 | { |
xorjoep | 1:24714b45cd1b | 208 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 209 | /* x[0] * y[srcBLen - 1] */ |
xorjoep | 1:24714b45cd1b | 210 | sum = __SMLAD(*px++, *py++, sum); |
xorjoep | 1:24714b45cd1b | 211 | |
xorjoep | 1:24714b45cd1b | 212 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 213 | k--; |
xorjoep | 1:24714b45cd1b | 214 | } |
xorjoep | 1:24714b45cd1b | 215 | |
xorjoep | 1:24714b45cd1b | 216 | /* Store the result in the accumulator in the destination buffer. */ |
xorjoep | 1:24714b45cd1b | 217 | *pOut = (q15_t) (sum >> 15); |
xorjoep | 1:24714b45cd1b | 218 | /* Destination pointer is updated according to the address modifier, inc */ |
xorjoep | 1:24714b45cd1b | 219 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 220 | |
xorjoep | 1:24714b45cd1b | 221 | /* Update the inputA and inputB pointers for next MAC calculation */ |
xorjoep | 1:24714b45cd1b | 222 | py = pSrc1 - count; |
xorjoep | 1:24714b45cd1b | 223 | px = pIn1; |
xorjoep | 1:24714b45cd1b | 224 | |
xorjoep | 1:24714b45cd1b | 225 | /* Increment the MAC count */ |
xorjoep | 1:24714b45cd1b | 226 | count++; |
xorjoep | 1:24714b45cd1b | 227 | |
xorjoep | 1:24714b45cd1b | 228 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 229 | blockSize1--; |
xorjoep | 1:24714b45cd1b | 230 | } |
xorjoep | 1:24714b45cd1b | 231 | |
xorjoep | 1:24714b45cd1b | 232 | /* -------------------------- |
xorjoep | 1:24714b45cd1b | 233 | * Initializations of stage2 |
xorjoep | 1:24714b45cd1b | 234 | * ------------------------*/ |
xorjoep | 1:24714b45cd1b | 235 | |
xorjoep | 1:24714b45cd1b | 236 | /* sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen-1] * y[srcBLen-1] |
xorjoep | 1:24714b45cd1b | 237 | * sum = x[1] * y[0] + x[2] * y[1] +...+ x[srcBLen] * y[srcBLen-1] |
xorjoep | 1:24714b45cd1b | 238 | * .... |
xorjoep | 1:24714b45cd1b | 239 | * sum = x[srcALen-srcBLen-2] * y[0] + x[srcALen-srcBLen-1] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
xorjoep | 1:24714b45cd1b | 240 | */ |
xorjoep | 1:24714b45cd1b | 241 | |
xorjoep | 1:24714b45cd1b | 242 | /* Working pointer of inputA */ |
xorjoep | 1:24714b45cd1b | 243 | px = pIn1; |
xorjoep | 1:24714b45cd1b | 244 | |
xorjoep | 1:24714b45cd1b | 245 | /* Working pointer of inputB */ |
xorjoep | 1:24714b45cd1b | 246 | py = pIn2; |
xorjoep | 1:24714b45cd1b | 247 | |
xorjoep | 1:24714b45cd1b | 248 | /* count is index by which the pointer pIn1 to be incremented */ |
xorjoep | 1:24714b45cd1b | 249 | count = 0U; |
xorjoep | 1:24714b45cd1b | 250 | |
xorjoep | 1:24714b45cd1b | 251 | /* ------------------- |
xorjoep | 1:24714b45cd1b | 252 | * Stage2 process |
xorjoep | 1:24714b45cd1b | 253 | * ------------------*/ |
xorjoep | 1:24714b45cd1b | 254 | |
xorjoep | 1:24714b45cd1b | 255 | /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed. |
xorjoep | 1:24714b45cd1b | 256 | * So, to loop unroll over blockSize2, |
xorjoep | 1:24714b45cd1b | 257 | * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */ |
xorjoep | 1:24714b45cd1b | 258 | if (srcBLen >= 4U) |
xorjoep | 1:24714b45cd1b | 259 | { |
xorjoep | 1:24714b45cd1b | 260 | /* Loop unroll over blockSize2, by 4 */ |
xorjoep | 1:24714b45cd1b | 261 | blkCnt = blockSize2 >> 2U; |
xorjoep | 1:24714b45cd1b | 262 | |
xorjoep | 1:24714b45cd1b | 263 | while (blkCnt > 0U) |
xorjoep | 1:24714b45cd1b | 264 | { |
xorjoep | 1:24714b45cd1b | 265 | /* Set all accumulators to zero */ |
xorjoep | 1:24714b45cd1b | 266 | acc0 = 0; |
xorjoep | 1:24714b45cd1b | 267 | acc1 = 0; |
xorjoep | 1:24714b45cd1b | 268 | acc2 = 0; |
xorjoep | 1:24714b45cd1b | 269 | acc3 = 0; |
xorjoep | 1:24714b45cd1b | 270 | |
xorjoep | 1:24714b45cd1b | 271 | /* read x[0], x[1] samples */ |
xorjoep | 1:24714b45cd1b | 272 | x0 = *__SIMD32(px); |
xorjoep | 1:24714b45cd1b | 273 | /* read x[1], x[2] samples */ |
xorjoep | 1:24714b45cd1b | 274 | x1 = _SIMD32_OFFSET(px + 1); |
xorjoep | 1:24714b45cd1b | 275 | px += 2U; |
xorjoep | 1:24714b45cd1b | 276 | |
xorjoep | 1:24714b45cd1b | 277 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
xorjoep | 1:24714b45cd1b | 278 | k = srcBLen >> 2U; |
xorjoep | 1:24714b45cd1b | 279 | |
xorjoep | 1:24714b45cd1b | 280 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
xorjoep | 1:24714b45cd1b | 281 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
xorjoep | 1:24714b45cd1b | 282 | do |
xorjoep | 1:24714b45cd1b | 283 | { |
xorjoep | 1:24714b45cd1b | 284 | /* Read the first two inputB samples using SIMD: |
xorjoep | 1:24714b45cd1b | 285 | * y[0] and y[1] */ |
xorjoep | 1:24714b45cd1b | 286 | c0 = *__SIMD32(py)++; |
xorjoep | 1:24714b45cd1b | 287 | |
xorjoep | 1:24714b45cd1b | 288 | /* acc0 += x[0] * y[0] + x[1] * y[1] */ |
xorjoep | 1:24714b45cd1b | 289 | acc0 = __SMLAD(x0, c0, acc0); |
xorjoep | 1:24714b45cd1b | 290 | |
xorjoep | 1:24714b45cd1b | 291 | /* acc1 += x[1] * y[0] + x[2] * y[1] */ |
xorjoep | 1:24714b45cd1b | 292 | acc1 = __SMLAD(x1, c0, acc1); |
xorjoep | 1:24714b45cd1b | 293 | |
xorjoep | 1:24714b45cd1b | 294 | /* Read x[2], x[3] */ |
xorjoep | 1:24714b45cd1b | 295 | x2 = *__SIMD32(px); |
xorjoep | 1:24714b45cd1b | 296 | |
xorjoep | 1:24714b45cd1b | 297 | /* Read x[3], x[4] */ |
xorjoep | 1:24714b45cd1b | 298 | x3 = _SIMD32_OFFSET(px + 1); |
xorjoep | 1:24714b45cd1b | 299 | |
xorjoep | 1:24714b45cd1b | 300 | /* acc2 += x[2] * y[0] + x[3] * y[1] */ |
xorjoep | 1:24714b45cd1b | 301 | acc2 = __SMLAD(x2, c0, acc2); |
xorjoep | 1:24714b45cd1b | 302 | |
xorjoep | 1:24714b45cd1b | 303 | /* acc3 += x[3] * y[0] + x[4] * y[1] */ |
xorjoep | 1:24714b45cd1b | 304 | acc3 = __SMLAD(x3, c0, acc3); |
xorjoep | 1:24714b45cd1b | 305 | |
xorjoep | 1:24714b45cd1b | 306 | /* Read y[2] and y[3] */ |
xorjoep | 1:24714b45cd1b | 307 | c0 = *__SIMD32(py)++; |
xorjoep | 1:24714b45cd1b | 308 | |
xorjoep | 1:24714b45cd1b | 309 | /* acc0 += x[2] * y[2] + x[3] * y[3] */ |
xorjoep | 1:24714b45cd1b | 310 | acc0 = __SMLAD(x2, c0, acc0); |
xorjoep | 1:24714b45cd1b | 311 | |
xorjoep | 1:24714b45cd1b | 312 | /* acc1 += x[3] * y[2] + x[4] * y[3] */ |
xorjoep | 1:24714b45cd1b | 313 | acc1 = __SMLAD(x3, c0, acc1); |
xorjoep | 1:24714b45cd1b | 314 | |
xorjoep | 1:24714b45cd1b | 315 | /* Read x[4], x[5] */ |
xorjoep | 1:24714b45cd1b | 316 | x0 = _SIMD32_OFFSET(px + 2); |
xorjoep | 1:24714b45cd1b | 317 | |
xorjoep | 1:24714b45cd1b | 318 | /* Read x[5], x[6] */ |
xorjoep | 1:24714b45cd1b | 319 | x1 = _SIMD32_OFFSET(px + 3); |
xorjoep | 1:24714b45cd1b | 320 | px += 4U; |
xorjoep | 1:24714b45cd1b | 321 | |
xorjoep | 1:24714b45cd1b | 322 | /* acc2 += x[4] * y[2] + x[5] * y[3] */ |
xorjoep | 1:24714b45cd1b | 323 | acc2 = __SMLAD(x0, c0, acc2); |
xorjoep | 1:24714b45cd1b | 324 | |
xorjoep | 1:24714b45cd1b | 325 | /* acc3 += x[5] * y[2] + x[6] * y[3] */ |
xorjoep | 1:24714b45cd1b | 326 | acc3 = __SMLAD(x1, c0, acc3); |
xorjoep | 1:24714b45cd1b | 327 | |
xorjoep | 1:24714b45cd1b | 328 | } while (--k); |
xorjoep | 1:24714b45cd1b | 329 | |
xorjoep | 1:24714b45cd1b | 330 | /* For the next MAC operations, SIMD is not used |
xorjoep | 1:24714b45cd1b | 331 | * So, the 16 bit pointer if inputB, py is updated */ |
xorjoep | 1:24714b45cd1b | 332 | |
xorjoep | 1:24714b45cd1b | 333 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
xorjoep | 1:24714b45cd1b | 334 | ** No loop unrolling is used. */ |
xorjoep | 1:24714b45cd1b | 335 | k = srcBLen % 0x4U; |
xorjoep | 1:24714b45cd1b | 336 | |
xorjoep | 1:24714b45cd1b | 337 | if (k == 1U) |
xorjoep | 1:24714b45cd1b | 338 | { |
xorjoep | 1:24714b45cd1b | 339 | /* Read y[4] */ |
xorjoep | 1:24714b45cd1b | 340 | c0 = *py; |
xorjoep | 1:24714b45cd1b | 341 | #ifdef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 342 | |
xorjoep | 1:24714b45cd1b | 343 | c0 = c0 << 16U; |
xorjoep | 1:24714b45cd1b | 344 | |
xorjoep | 1:24714b45cd1b | 345 | #else |
xorjoep | 1:24714b45cd1b | 346 | |
xorjoep | 1:24714b45cd1b | 347 | c0 = c0 & 0x0000FFFF; |
xorjoep | 1:24714b45cd1b | 348 | |
xorjoep | 1:24714b45cd1b | 349 | #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 350 | |
xorjoep | 1:24714b45cd1b | 351 | /* Read x[7] */ |
xorjoep | 1:24714b45cd1b | 352 | x3 = *__SIMD32(px); |
xorjoep | 1:24714b45cd1b | 353 | px++; |
xorjoep | 1:24714b45cd1b | 354 | |
xorjoep | 1:24714b45cd1b | 355 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 356 | acc0 = __SMLAD(x0, c0, acc0); |
xorjoep | 1:24714b45cd1b | 357 | acc1 = __SMLAD(x1, c0, acc1); |
xorjoep | 1:24714b45cd1b | 358 | acc2 = __SMLADX(x1, c0, acc2); |
xorjoep | 1:24714b45cd1b | 359 | acc3 = __SMLADX(x3, c0, acc3); |
xorjoep | 1:24714b45cd1b | 360 | } |
xorjoep | 1:24714b45cd1b | 361 | |
xorjoep | 1:24714b45cd1b | 362 | if (k == 2U) |
xorjoep | 1:24714b45cd1b | 363 | { |
xorjoep | 1:24714b45cd1b | 364 | /* Read y[4], y[5] */ |
xorjoep | 1:24714b45cd1b | 365 | c0 = *__SIMD32(py); |
xorjoep | 1:24714b45cd1b | 366 | |
xorjoep | 1:24714b45cd1b | 367 | /* Read x[7], x[8] */ |
xorjoep | 1:24714b45cd1b | 368 | x3 = *__SIMD32(px); |
xorjoep | 1:24714b45cd1b | 369 | |
xorjoep | 1:24714b45cd1b | 370 | /* Read x[9] */ |
xorjoep | 1:24714b45cd1b | 371 | x2 = _SIMD32_OFFSET(px + 1); |
xorjoep | 1:24714b45cd1b | 372 | px += 2U; |
xorjoep | 1:24714b45cd1b | 373 | |
xorjoep | 1:24714b45cd1b | 374 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 375 | acc0 = __SMLAD(x0, c0, acc0); |
xorjoep | 1:24714b45cd1b | 376 | acc1 = __SMLAD(x1, c0, acc1); |
xorjoep | 1:24714b45cd1b | 377 | acc2 = __SMLAD(x3, c0, acc2); |
xorjoep | 1:24714b45cd1b | 378 | acc3 = __SMLAD(x2, c0, acc3); |
xorjoep | 1:24714b45cd1b | 379 | } |
xorjoep | 1:24714b45cd1b | 380 | |
xorjoep | 1:24714b45cd1b | 381 | if (k == 3U) |
xorjoep | 1:24714b45cd1b | 382 | { |
xorjoep | 1:24714b45cd1b | 383 | /* Read y[4], y[5] */ |
xorjoep | 1:24714b45cd1b | 384 | c0 = *__SIMD32(py)++; |
xorjoep | 1:24714b45cd1b | 385 | |
xorjoep | 1:24714b45cd1b | 386 | /* Read x[7], x[8] */ |
xorjoep | 1:24714b45cd1b | 387 | x3 = *__SIMD32(px); |
xorjoep | 1:24714b45cd1b | 388 | |
xorjoep | 1:24714b45cd1b | 389 | /* Read x[9] */ |
xorjoep | 1:24714b45cd1b | 390 | x2 = _SIMD32_OFFSET(px + 1); |
xorjoep | 1:24714b45cd1b | 391 | |
xorjoep | 1:24714b45cd1b | 392 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 393 | acc0 = __SMLAD(x0, c0, acc0); |
xorjoep | 1:24714b45cd1b | 394 | acc1 = __SMLAD(x1, c0, acc1); |
xorjoep | 1:24714b45cd1b | 395 | acc2 = __SMLAD(x3, c0, acc2); |
xorjoep | 1:24714b45cd1b | 396 | acc3 = __SMLAD(x2, c0, acc3); |
xorjoep | 1:24714b45cd1b | 397 | |
xorjoep | 1:24714b45cd1b | 398 | c0 = (*py); |
xorjoep | 1:24714b45cd1b | 399 | /* Read y[6] */ |
xorjoep | 1:24714b45cd1b | 400 | #ifdef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 401 | |
xorjoep | 1:24714b45cd1b | 402 | c0 = c0 << 16U; |
xorjoep | 1:24714b45cd1b | 403 | #else |
xorjoep | 1:24714b45cd1b | 404 | |
xorjoep | 1:24714b45cd1b | 405 | c0 = c0 & 0x0000FFFF; |
xorjoep | 1:24714b45cd1b | 406 | #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 407 | |
xorjoep | 1:24714b45cd1b | 408 | /* Read x[10] */ |
xorjoep | 1:24714b45cd1b | 409 | x3 = _SIMD32_OFFSET(px + 2); |
xorjoep | 1:24714b45cd1b | 410 | px += 3U; |
xorjoep | 1:24714b45cd1b | 411 | |
xorjoep | 1:24714b45cd1b | 412 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 413 | acc0 = __SMLADX(x1, c0, acc0); |
xorjoep | 1:24714b45cd1b | 414 | acc1 = __SMLAD(x2, c0, acc1); |
xorjoep | 1:24714b45cd1b | 415 | acc2 = __SMLADX(x2, c0, acc2); |
xorjoep | 1:24714b45cd1b | 416 | acc3 = __SMLADX(x3, c0, acc3); |
xorjoep | 1:24714b45cd1b | 417 | } |
xorjoep | 1:24714b45cd1b | 418 | |
xorjoep | 1:24714b45cd1b | 419 | /* Store the result in the accumulator in the destination buffer. */ |
xorjoep | 1:24714b45cd1b | 420 | *pOut = (q15_t) (acc0 >> 15); |
xorjoep | 1:24714b45cd1b | 421 | /* Destination pointer is updated according to the address modifier, inc */ |
xorjoep | 1:24714b45cd1b | 422 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 423 | |
xorjoep | 1:24714b45cd1b | 424 | *pOut = (q15_t) (acc1 >> 15); |
xorjoep | 1:24714b45cd1b | 425 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 426 | |
xorjoep | 1:24714b45cd1b | 427 | *pOut = (q15_t) (acc2 >> 15); |
xorjoep | 1:24714b45cd1b | 428 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 429 | |
xorjoep | 1:24714b45cd1b | 430 | *pOut = (q15_t) (acc3 >> 15); |
xorjoep | 1:24714b45cd1b | 431 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 432 | |
xorjoep | 1:24714b45cd1b | 433 | /* Increment the pointer pIn1 index, count by 1 */ |
xorjoep | 1:24714b45cd1b | 434 | count += 4U; |
xorjoep | 1:24714b45cd1b | 435 | |
xorjoep | 1:24714b45cd1b | 436 | /* Update the inputA and inputB pointers for next MAC calculation */ |
xorjoep | 1:24714b45cd1b | 437 | px = pIn1 + count; |
xorjoep | 1:24714b45cd1b | 438 | py = pIn2; |
xorjoep | 1:24714b45cd1b | 439 | |
xorjoep | 1:24714b45cd1b | 440 | |
xorjoep | 1:24714b45cd1b | 441 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 442 | blkCnt--; |
xorjoep | 1:24714b45cd1b | 443 | } |
xorjoep | 1:24714b45cd1b | 444 | |
xorjoep | 1:24714b45cd1b | 445 | /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here. |
xorjoep | 1:24714b45cd1b | 446 | ** No loop unrolling is used. */ |
xorjoep | 1:24714b45cd1b | 447 | blkCnt = blockSize2 % 0x4U; |
xorjoep | 1:24714b45cd1b | 448 | |
xorjoep | 1:24714b45cd1b | 449 | while (blkCnt > 0U) |
xorjoep | 1:24714b45cd1b | 450 | { |
xorjoep | 1:24714b45cd1b | 451 | /* Accumulator is made zero for every iteration */ |
xorjoep | 1:24714b45cd1b | 452 | sum = 0; |
xorjoep | 1:24714b45cd1b | 453 | |
xorjoep | 1:24714b45cd1b | 454 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
xorjoep | 1:24714b45cd1b | 455 | k = srcBLen >> 2U; |
xorjoep | 1:24714b45cd1b | 456 | |
xorjoep | 1:24714b45cd1b | 457 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
xorjoep | 1:24714b45cd1b | 458 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
xorjoep | 1:24714b45cd1b | 459 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 460 | { |
xorjoep | 1:24714b45cd1b | 461 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 462 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 463 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 464 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 465 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 466 | |
xorjoep | 1:24714b45cd1b | 467 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 468 | k--; |
xorjoep | 1:24714b45cd1b | 469 | } |
xorjoep | 1:24714b45cd1b | 470 | |
xorjoep | 1:24714b45cd1b | 471 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
xorjoep | 1:24714b45cd1b | 472 | ** No loop unrolling is used. */ |
xorjoep | 1:24714b45cd1b | 473 | k = srcBLen % 0x4U; |
xorjoep | 1:24714b45cd1b | 474 | |
xorjoep | 1:24714b45cd1b | 475 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 476 | { |
xorjoep | 1:24714b45cd1b | 477 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 478 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 479 | |
xorjoep | 1:24714b45cd1b | 480 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 481 | k--; |
xorjoep | 1:24714b45cd1b | 482 | } |
xorjoep | 1:24714b45cd1b | 483 | |
xorjoep | 1:24714b45cd1b | 484 | /* Store the result in the accumulator in the destination buffer. */ |
xorjoep | 1:24714b45cd1b | 485 | *pOut = (q15_t) (sum >> 15); |
xorjoep | 1:24714b45cd1b | 486 | /* Destination pointer is updated according to the address modifier, inc */ |
xorjoep | 1:24714b45cd1b | 487 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 488 | |
xorjoep | 1:24714b45cd1b | 489 | /* Increment the pointer pIn1 index, count by 1 */ |
xorjoep | 1:24714b45cd1b | 490 | count++; |
xorjoep | 1:24714b45cd1b | 491 | |
xorjoep | 1:24714b45cd1b | 492 | /* Update the inputA and inputB pointers for next MAC calculation */ |
xorjoep | 1:24714b45cd1b | 493 | px = pIn1 + count; |
xorjoep | 1:24714b45cd1b | 494 | py = pIn2; |
xorjoep | 1:24714b45cd1b | 495 | |
xorjoep | 1:24714b45cd1b | 496 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 497 | blkCnt--; |
xorjoep | 1:24714b45cd1b | 498 | } |
xorjoep | 1:24714b45cd1b | 499 | } |
xorjoep | 1:24714b45cd1b | 500 | else |
xorjoep | 1:24714b45cd1b | 501 | { |
xorjoep | 1:24714b45cd1b | 502 | /* If the srcBLen is not a multiple of 4, |
xorjoep | 1:24714b45cd1b | 503 | * the blockSize2 loop cannot be unrolled by 4 */ |
xorjoep | 1:24714b45cd1b | 504 | blkCnt = blockSize2; |
xorjoep | 1:24714b45cd1b | 505 | |
xorjoep | 1:24714b45cd1b | 506 | while (blkCnt > 0U) |
xorjoep | 1:24714b45cd1b | 507 | { |
xorjoep | 1:24714b45cd1b | 508 | /* Accumulator is made zero for every iteration */ |
xorjoep | 1:24714b45cd1b | 509 | sum = 0; |
xorjoep | 1:24714b45cd1b | 510 | |
xorjoep | 1:24714b45cd1b | 511 | /* Loop over srcBLen */ |
xorjoep | 1:24714b45cd1b | 512 | k = srcBLen; |
xorjoep | 1:24714b45cd1b | 513 | |
xorjoep | 1:24714b45cd1b | 514 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 515 | { |
xorjoep | 1:24714b45cd1b | 516 | /* Perform the multiply-accumulate */ |
xorjoep | 1:24714b45cd1b | 517 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 518 | |
xorjoep | 1:24714b45cd1b | 519 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 520 | k--; |
xorjoep | 1:24714b45cd1b | 521 | } |
xorjoep | 1:24714b45cd1b | 522 | |
xorjoep | 1:24714b45cd1b | 523 | /* Store the result in the accumulator in the destination buffer. */ |
xorjoep | 1:24714b45cd1b | 524 | *pOut = (q15_t) (sum >> 15); |
xorjoep | 1:24714b45cd1b | 525 | /* Destination pointer is updated according to the address modifier, inc */ |
xorjoep | 1:24714b45cd1b | 526 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 527 | |
xorjoep | 1:24714b45cd1b | 528 | /* Increment the MAC count */ |
xorjoep | 1:24714b45cd1b | 529 | count++; |
xorjoep | 1:24714b45cd1b | 530 | |
xorjoep | 1:24714b45cd1b | 531 | /* Update the inputA and inputB pointers for next MAC calculation */ |
xorjoep | 1:24714b45cd1b | 532 | px = pIn1 + count; |
xorjoep | 1:24714b45cd1b | 533 | py = pIn2; |
xorjoep | 1:24714b45cd1b | 534 | |
xorjoep | 1:24714b45cd1b | 535 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 536 | blkCnt--; |
xorjoep | 1:24714b45cd1b | 537 | } |
xorjoep | 1:24714b45cd1b | 538 | } |
xorjoep | 1:24714b45cd1b | 539 | |
xorjoep | 1:24714b45cd1b | 540 | /* -------------------------- |
xorjoep | 1:24714b45cd1b | 541 | * Initializations of stage3 |
xorjoep | 1:24714b45cd1b | 542 | * -------------------------*/ |
xorjoep | 1:24714b45cd1b | 543 | |
xorjoep | 1:24714b45cd1b | 544 | /* sum += x[srcALen-srcBLen+1] * y[0] + x[srcALen-srcBLen+2] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
xorjoep | 1:24714b45cd1b | 545 | * sum += x[srcALen-srcBLen+2] * y[0] + x[srcALen-srcBLen+3] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
xorjoep | 1:24714b45cd1b | 546 | * .... |
xorjoep | 1:24714b45cd1b | 547 | * sum += x[srcALen-2] * y[0] + x[srcALen-1] * y[1] |
xorjoep | 1:24714b45cd1b | 548 | * sum += x[srcALen-1] * y[0] |
xorjoep | 1:24714b45cd1b | 549 | */ |
xorjoep | 1:24714b45cd1b | 550 | |
xorjoep | 1:24714b45cd1b | 551 | /* In this stage the MAC operations are decreased by 1 for every iteration. |
xorjoep | 1:24714b45cd1b | 552 | The count variable holds the number of MAC operations performed */ |
xorjoep | 1:24714b45cd1b | 553 | count = srcBLen - 1U; |
xorjoep | 1:24714b45cd1b | 554 | |
xorjoep | 1:24714b45cd1b | 555 | /* Working pointer of inputA */ |
xorjoep | 1:24714b45cd1b | 556 | pSrc1 = (pIn1 + srcALen) - (srcBLen - 1U); |
xorjoep | 1:24714b45cd1b | 557 | px = pSrc1; |
xorjoep | 1:24714b45cd1b | 558 | |
xorjoep | 1:24714b45cd1b | 559 | /* Working pointer of inputB */ |
xorjoep | 1:24714b45cd1b | 560 | py = pIn2; |
xorjoep | 1:24714b45cd1b | 561 | |
xorjoep | 1:24714b45cd1b | 562 | /* ------------------- |
xorjoep | 1:24714b45cd1b | 563 | * Stage3 process |
xorjoep | 1:24714b45cd1b | 564 | * ------------------*/ |
xorjoep | 1:24714b45cd1b | 565 | |
xorjoep | 1:24714b45cd1b | 566 | while (blockSize3 > 0U) |
xorjoep | 1:24714b45cd1b | 567 | { |
xorjoep | 1:24714b45cd1b | 568 | /* Accumulator is made zero for every iteration */ |
xorjoep | 1:24714b45cd1b | 569 | sum = 0; |
xorjoep | 1:24714b45cd1b | 570 | |
xorjoep | 1:24714b45cd1b | 571 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
xorjoep | 1:24714b45cd1b | 572 | k = count >> 2U; |
xorjoep | 1:24714b45cd1b | 573 | |
xorjoep | 1:24714b45cd1b | 574 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
xorjoep | 1:24714b45cd1b | 575 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
xorjoep | 1:24714b45cd1b | 576 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 577 | { |
xorjoep | 1:24714b45cd1b | 578 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 579 | /* sum += x[srcALen - srcBLen + 4] * y[3] , sum += x[srcALen - srcBLen + 3] * y[2] */ |
xorjoep | 1:24714b45cd1b | 580 | sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); |
xorjoep | 1:24714b45cd1b | 581 | /* sum += x[srcALen - srcBLen + 2] * y[1] , sum += x[srcALen - srcBLen + 1] * y[0] */ |
xorjoep | 1:24714b45cd1b | 582 | sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); |
xorjoep | 1:24714b45cd1b | 583 | |
xorjoep | 1:24714b45cd1b | 584 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 585 | k--; |
xorjoep | 1:24714b45cd1b | 586 | } |
xorjoep | 1:24714b45cd1b | 587 | |
xorjoep | 1:24714b45cd1b | 588 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
xorjoep | 1:24714b45cd1b | 589 | ** No loop unrolling is used. */ |
xorjoep | 1:24714b45cd1b | 590 | k = count % 0x4U; |
xorjoep | 1:24714b45cd1b | 591 | |
xorjoep | 1:24714b45cd1b | 592 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 593 | { |
xorjoep | 1:24714b45cd1b | 594 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 595 | sum = __SMLAD(*px++, *py++, sum); |
xorjoep | 1:24714b45cd1b | 596 | |
xorjoep | 1:24714b45cd1b | 597 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 598 | k--; |
xorjoep | 1:24714b45cd1b | 599 | } |
xorjoep | 1:24714b45cd1b | 600 | |
xorjoep | 1:24714b45cd1b | 601 | /* Store the result in the accumulator in the destination buffer. */ |
xorjoep | 1:24714b45cd1b | 602 | *pOut = (q15_t) (sum >> 15); |
xorjoep | 1:24714b45cd1b | 603 | /* Destination pointer is updated according to the address modifier, inc */ |
xorjoep | 1:24714b45cd1b | 604 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 605 | |
xorjoep | 1:24714b45cd1b | 606 | /* Update the inputA and inputB pointers for next MAC calculation */ |
xorjoep | 1:24714b45cd1b | 607 | px = ++pSrc1; |
xorjoep | 1:24714b45cd1b | 608 | py = pIn2; |
xorjoep | 1:24714b45cd1b | 609 | |
xorjoep | 1:24714b45cd1b | 610 | /* Decrement the MAC count */ |
xorjoep | 1:24714b45cd1b | 611 | count--; |
xorjoep | 1:24714b45cd1b | 612 | |
xorjoep | 1:24714b45cd1b | 613 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 614 | blockSize3--; |
xorjoep | 1:24714b45cd1b | 615 | } |
xorjoep | 1:24714b45cd1b | 616 | |
xorjoep | 1:24714b45cd1b | 617 | #else |
xorjoep | 1:24714b45cd1b | 618 | |
xorjoep | 1:24714b45cd1b | 619 | q15_t *pIn1; /* inputA pointer */ |
xorjoep | 1:24714b45cd1b | 620 | q15_t *pIn2; /* inputB pointer */ |
xorjoep | 1:24714b45cd1b | 621 | q15_t *pOut = pDst; /* output pointer */ |
xorjoep | 1:24714b45cd1b | 622 | q31_t sum, acc0, acc1, acc2, acc3; /* Accumulators */ |
xorjoep | 1:24714b45cd1b | 623 | q15_t *px; /* Intermediate inputA pointer */ |
xorjoep | 1:24714b45cd1b | 624 | q15_t *py; /* Intermediate inputB pointer */ |
xorjoep | 1:24714b45cd1b | 625 | q15_t *pSrc1; /* Intermediate pointers */ |
xorjoep | 1:24714b45cd1b | 626 | q31_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */ |
xorjoep | 1:24714b45cd1b | 627 | uint32_t j, k = 0U, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counter */ |
xorjoep | 1:24714b45cd1b | 628 | int32_t inc = 1; /* Destination address modifier */ |
xorjoep | 1:24714b45cd1b | 629 | q15_t a, b; |
xorjoep | 1:24714b45cd1b | 630 | |
xorjoep | 1:24714b45cd1b | 631 | |
xorjoep | 1:24714b45cd1b | 632 | /* The algorithm implementation is based on the lengths of the inputs. */ |
xorjoep | 1:24714b45cd1b | 633 | /* srcB is always made to slide across srcA. */ |
xorjoep | 1:24714b45cd1b | 634 | /* So srcBLen is always considered as shorter or equal to srcALen */ |
xorjoep | 1:24714b45cd1b | 635 | /* But CORR(x, y) is reverse of CORR(y, x) */ |
xorjoep | 1:24714b45cd1b | 636 | /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */ |
xorjoep | 1:24714b45cd1b | 637 | /* and the destination pointer modifier, inc is set to -1 */ |
xorjoep | 1:24714b45cd1b | 638 | /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */ |
xorjoep | 1:24714b45cd1b | 639 | /* But to improve the performance, |
xorjoep | 1:24714b45cd1b | 640 | * we include zeroes in the output instead of zero padding either of the the inputs*/ |
xorjoep | 1:24714b45cd1b | 641 | /* If srcALen > srcBLen, |
xorjoep | 1:24714b45cd1b | 642 | * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */ |
xorjoep | 1:24714b45cd1b | 643 | /* If srcALen < srcBLen, |
xorjoep | 1:24714b45cd1b | 644 | * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */ |
xorjoep | 1:24714b45cd1b | 645 | if (srcALen >= srcBLen) |
xorjoep | 1:24714b45cd1b | 646 | { |
xorjoep | 1:24714b45cd1b | 647 | /* Initialization of inputA pointer */ |
xorjoep | 1:24714b45cd1b | 648 | pIn1 = (pSrcA); |
xorjoep | 1:24714b45cd1b | 649 | |
xorjoep | 1:24714b45cd1b | 650 | /* Initialization of inputB pointer */ |
xorjoep | 1:24714b45cd1b | 651 | pIn2 = (pSrcB); |
xorjoep | 1:24714b45cd1b | 652 | |
xorjoep | 1:24714b45cd1b | 653 | /* Number of output samples is calculated */ |
xorjoep | 1:24714b45cd1b | 654 | outBlockSize = (2U * srcALen) - 1U; |
xorjoep | 1:24714b45cd1b | 655 | |
xorjoep | 1:24714b45cd1b | 656 | /* When srcALen > srcBLen, zero padding is done to srcB |
xorjoep | 1:24714b45cd1b | 657 | * to make their lengths equal. |
xorjoep | 1:24714b45cd1b | 658 | * Instead, (outBlockSize - (srcALen + srcBLen - 1)) |
xorjoep | 1:24714b45cd1b | 659 | * number of output samples are made zero */ |
xorjoep | 1:24714b45cd1b | 660 | j = outBlockSize - (srcALen + (srcBLen - 1U)); |
xorjoep | 1:24714b45cd1b | 661 | |
xorjoep | 1:24714b45cd1b | 662 | /* Updating the pointer position to non zero value */ |
xorjoep | 1:24714b45cd1b | 663 | pOut += j; |
xorjoep | 1:24714b45cd1b | 664 | |
xorjoep | 1:24714b45cd1b | 665 | } |
xorjoep | 1:24714b45cd1b | 666 | else |
xorjoep | 1:24714b45cd1b | 667 | { |
xorjoep | 1:24714b45cd1b | 668 | /* Initialization of inputA pointer */ |
xorjoep | 1:24714b45cd1b | 669 | pIn1 = (pSrcB); |
xorjoep | 1:24714b45cd1b | 670 | |
xorjoep | 1:24714b45cd1b | 671 | /* Initialization of inputB pointer */ |
xorjoep | 1:24714b45cd1b | 672 | pIn2 = (pSrcA); |
xorjoep | 1:24714b45cd1b | 673 | |
xorjoep | 1:24714b45cd1b | 674 | /* srcBLen is always considered as shorter or equal to srcALen */ |
xorjoep | 1:24714b45cd1b | 675 | j = srcBLen; |
xorjoep | 1:24714b45cd1b | 676 | srcBLen = srcALen; |
xorjoep | 1:24714b45cd1b | 677 | srcALen = j; |
xorjoep | 1:24714b45cd1b | 678 | |
xorjoep | 1:24714b45cd1b | 679 | /* CORR(x, y) = Reverse order(CORR(y, x)) */ |
xorjoep | 1:24714b45cd1b | 680 | /* Hence set the destination pointer to point to the last output sample */ |
xorjoep | 1:24714b45cd1b | 681 | pOut = pDst + ((srcALen + srcBLen) - 2U); |
xorjoep | 1:24714b45cd1b | 682 | |
xorjoep | 1:24714b45cd1b | 683 | /* Destination address modifier is set to -1 */ |
xorjoep | 1:24714b45cd1b | 684 | inc = -1; |
xorjoep | 1:24714b45cd1b | 685 | |
xorjoep | 1:24714b45cd1b | 686 | } |
xorjoep | 1:24714b45cd1b | 687 | |
xorjoep | 1:24714b45cd1b | 688 | /* The function is internally |
xorjoep | 1:24714b45cd1b | 689 | * divided into three parts according to the number of multiplications that has to be |
xorjoep | 1:24714b45cd1b | 690 | * taken place between inputA samples and inputB samples. In the first part of the |
xorjoep | 1:24714b45cd1b | 691 | * algorithm, the multiplications increase by one for every iteration. |
xorjoep | 1:24714b45cd1b | 692 | * In the second part of the algorithm, srcBLen number of multiplications are done. |
xorjoep | 1:24714b45cd1b | 693 | * In the third part of the algorithm, the multiplications decrease by one |
xorjoep | 1:24714b45cd1b | 694 | * for every iteration.*/ |
xorjoep | 1:24714b45cd1b | 695 | /* The algorithm is implemented in three stages. |
xorjoep | 1:24714b45cd1b | 696 | * The loop counters of each stage is initiated here. */ |
xorjoep | 1:24714b45cd1b | 697 | blockSize1 = srcBLen - 1U; |
xorjoep | 1:24714b45cd1b | 698 | blockSize2 = srcALen - (srcBLen - 1U); |
xorjoep | 1:24714b45cd1b | 699 | blockSize3 = blockSize1; |
xorjoep | 1:24714b45cd1b | 700 | |
xorjoep | 1:24714b45cd1b | 701 | /* -------------------------- |
xorjoep | 1:24714b45cd1b | 702 | * Initializations of stage1 |
xorjoep | 1:24714b45cd1b | 703 | * -------------------------*/ |
xorjoep | 1:24714b45cd1b | 704 | |
xorjoep | 1:24714b45cd1b | 705 | /* sum = x[0] * y[srcBlen - 1] |
xorjoep | 1:24714b45cd1b | 706 | * sum = x[0] * y[srcBlen - 2] + x[1] * y[srcBlen - 1] |
xorjoep | 1:24714b45cd1b | 707 | * .... |
xorjoep | 1:24714b45cd1b | 708 | * sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen - 1] * y[srcBLen - 1] |
xorjoep | 1:24714b45cd1b | 709 | */ |
xorjoep | 1:24714b45cd1b | 710 | |
xorjoep | 1:24714b45cd1b | 711 | /* In this stage the MAC operations are increased by 1 for every iteration. |
xorjoep | 1:24714b45cd1b | 712 | The count variable holds the number of MAC operations performed */ |
xorjoep | 1:24714b45cd1b | 713 | count = 1U; |
xorjoep | 1:24714b45cd1b | 714 | |
xorjoep | 1:24714b45cd1b | 715 | /* Working pointer of inputA */ |
xorjoep | 1:24714b45cd1b | 716 | px = pIn1; |
xorjoep | 1:24714b45cd1b | 717 | |
xorjoep | 1:24714b45cd1b | 718 | /* Working pointer of inputB */ |
xorjoep | 1:24714b45cd1b | 719 | pSrc1 = pIn2 + (srcBLen - 1U); |
xorjoep | 1:24714b45cd1b | 720 | py = pSrc1; |
xorjoep | 1:24714b45cd1b | 721 | |
xorjoep | 1:24714b45cd1b | 722 | /* ------------------------ |
xorjoep | 1:24714b45cd1b | 723 | * Stage1 process |
xorjoep | 1:24714b45cd1b | 724 | * ----------------------*/ |
xorjoep | 1:24714b45cd1b | 725 | |
xorjoep | 1:24714b45cd1b | 726 | /* The first loop starts here */ |
xorjoep | 1:24714b45cd1b | 727 | while (blockSize1 > 0U) |
xorjoep | 1:24714b45cd1b | 728 | { |
xorjoep | 1:24714b45cd1b | 729 | /* Accumulator is made zero for every iteration */ |
xorjoep | 1:24714b45cd1b | 730 | sum = 0; |
xorjoep | 1:24714b45cd1b | 731 | |
xorjoep | 1:24714b45cd1b | 732 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
xorjoep | 1:24714b45cd1b | 733 | k = count >> 2; |
xorjoep | 1:24714b45cd1b | 734 | |
xorjoep | 1:24714b45cd1b | 735 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
xorjoep | 1:24714b45cd1b | 736 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
xorjoep | 1:24714b45cd1b | 737 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 738 | { |
xorjoep | 1:24714b45cd1b | 739 | /* x[0] * y[srcBLen - 4] , x[1] * y[srcBLen - 3] */ |
xorjoep | 1:24714b45cd1b | 740 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 741 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 742 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 743 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 744 | |
xorjoep | 1:24714b45cd1b | 745 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 746 | k--; |
xorjoep | 1:24714b45cd1b | 747 | } |
xorjoep | 1:24714b45cd1b | 748 | |
xorjoep | 1:24714b45cd1b | 749 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
xorjoep | 1:24714b45cd1b | 750 | ** No loop unrolling is used. */ |
xorjoep | 1:24714b45cd1b | 751 | k = count % 0x4U; |
xorjoep | 1:24714b45cd1b | 752 | |
xorjoep | 1:24714b45cd1b | 753 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 754 | { |
xorjoep | 1:24714b45cd1b | 755 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 756 | /* x[0] * y[srcBLen - 1] */ |
xorjoep | 1:24714b45cd1b | 757 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 758 | |
xorjoep | 1:24714b45cd1b | 759 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 760 | k--; |
xorjoep | 1:24714b45cd1b | 761 | } |
xorjoep | 1:24714b45cd1b | 762 | |
xorjoep | 1:24714b45cd1b | 763 | /* Store the result in the accumulator in the destination buffer. */ |
xorjoep | 1:24714b45cd1b | 764 | *pOut = (q15_t) (sum >> 15); |
xorjoep | 1:24714b45cd1b | 765 | /* Destination pointer is updated according to the address modifier, inc */ |
xorjoep | 1:24714b45cd1b | 766 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 767 | |
xorjoep | 1:24714b45cd1b | 768 | /* Update the inputA and inputB pointers for next MAC calculation */ |
xorjoep | 1:24714b45cd1b | 769 | py = pSrc1 - count; |
xorjoep | 1:24714b45cd1b | 770 | px = pIn1; |
xorjoep | 1:24714b45cd1b | 771 | |
xorjoep | 1:24714b45cd1b | 772 | /* Increment the MAC count */ |
xorjoep | 1:24714b45cd1b | 773 | count++; |
xorjoep | 1:24714b45cd1b | 774 | |
xorjoep | 1:24714b45cd1b | 775 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 776 | blockSize1--; |
xorjoep | 1:24714b45cd1b | 777 | } |
xorjoep | 1:24714b45cd1b | 778 | |
xorjoep | 1:24714b45cd1b | 779 | /* -------------------------- |
xorjoep | 1:24714b45cd1b | 780 | * Initializations of stage2 |
xorjoep | 1:24714b45cd1b | 781 | * ------------------------*/ |
xorjoep | 1:24714b45cd1b | 782 | |
xorjoep | 1:24714b45cd1b | 783 | /* sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen-1] * y[srcBLen-1] |
xorjoep | 1:24714b45cd1b | 784 | * sum = x[1] * y[0] + x[2] * y[1] +...+ x[srcBLen] * y[srcBLen-1] |
xorjoep | 1:24714b45cd1b | 785 | * .... |
xorjoep | 1:24714b45cd1b | 786 | * sum = x[srcALen-srcBLen-2] * y[0] + x[srcALen-srcBLen-1] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
xorjoep | 1:24714b45cd1b | 787 | */ |
xorjoep | 1:24714b45cd1b | 788 | |
xorjoep | 1:24714b45cd1b | 789 | /* Working pointer of inputA */ |
xorjoep | 1:24714b45cd1b | 790 | px = pIn1; |
xorjoep | 1:24714b45cd1b | 791 | |
xorjoep | 1:24714b45cd1b | 792 | /* Working pointer of inputB */ |
xorjoep | 1:24714b45cd1b | 793 | py = pIn2; |
xorjoep | 1:24714b45cd1b | 794 | |
xorjoep | 1:24714b45cd1b | 795 | /* count is index by which the pointer pIn1 to be incremented */ |
xorjoep | 1:24714b45cd1b | 796 | count = 0U; |
xorjoep | 1:24714b45cd1b | 797 | |
xorjoep | 1:24714b45cd1b | 798 | /* ------------------- |
xorjoep | 1:24714b45cd1b | 799 | * Stage2 process |
xorjoep | 1:24714b45cd1b | 800 | * ------------------*/ |
xorjoep | 1:24714b45cd1b | 801 | |
xorjoep | 1:24714b45cd1b | 802 | /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed. |
xorjoep | 1:24714b45cd1b | 803 | * So, to loop unroll over blockSize2, |
xorjoep | 1:24714b45cd1b | 804 | * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */ |
xorjoep | 1:24714b45cd1b | 805 | if (srcBLen >= 4U) |
xorjoep | 1:24714b45cd1b | 806 | { |
xorjoep | 1:24714b45cd1b | 807 | /* Loop unroll over blockSize2, by 4 */ |
xorjoep | 1:24714b45cd1b | 808 | blkCnt = blockSize2 >> 2U; |
xorjoep | 1:24714b45cd1b | 809 | |
xorjoep | 1:24714b45cd1b | 810 | while (blkCnt > 0U) |
xorjoep | 1:24714b45cd1b | 811 | { |
xorjoep | 1:24714b45cd1b | 812 | /* Set all accumulators to zero */ |
xorjoep | 1:24714b45cd1b | 813 | acc0 = 0; |
xorjoep | 1:24714b45cd1b | 814 | acc1 = 0; |
xorjoep | 1:24714b45cd1b | 815 | acc2 = 0; |
xorjoep | 1:24714b45cd1b | 816 | acc3 = 0; |
xorjoep | 1:24714b45cd1b | 817 | |
xorjoep | 1:24714b45cd1b | 818 | /* read x[0], x[1], x[2] samples */ |
xorjoep | 1:24714b45cd1b | 819 | a = *px; |
xorjoep | 1:24714b45cd1b | 820 | b = *(px + 1); |
xorjoep | 1:24714b45cd1b | 821 | |
xorjoep | 1:24714b45cd1b | 822 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 823 | |
xorjoep | 1:24714b45cd1b | 824 | x0 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 825 | a = *(px + 2); |
xorjoep | 1:24714b45cd1b | 826 | x1 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 827 | |
xorjoep | 1:24714b45cd1b | 828 | #else |
xorjoep | 1:24714b45cd1b | 829 | |
xorjoep | 1:24714b45cd1b | 830 | x0 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 831 | a = *(px + 2); |
xorjoep | 1:24714b45cd1b | 832 | x1 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 833 | |
xorjoep | 1:24714b45cd1b | 834 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 835 | |
xorjoep | 1:24714b45cd1b | 836 | px += 2U; |
xorjoep | 1:24714b45cd1b | 837 | |
xorjoep | 1:24714b45cd1b | 838 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
xorjoep | 1:24714b45cd1b | 839 | k = srcBLen >> 2U; |
xorjoep | 1:24714b45cd1b | 840 | |
xorjoep | 1:24714b45cd1b | 841 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
xorjoep | 1:24714b45cd1b | 842 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
xorjoep | 1:24714b45cd1b | 843 | do |
xorjoep | 1:24714b45cd1b | 844 | { |
xorjoep | 1:24714b45cd1b | 845 | /* Read the first two inputB samples using SIMD: |
xorjoep | 1:24714b45cd1b | 846 | * y[0] and y[1] */ |
xorjoep | 1:24714b45cd1b | 847 | a = *py; |
xorjoep | 1:24714b45cd1b | 848 | b = *(py + 1); |
xorjoep | 1:24714b45cd1b | 849 | |
xorjoep | 1:24714b45cd1b | 850 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 851 | |
xorjoep | 1:24714b45cd1b | 852 | c0 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 853 | |
xorjoep | 1:24714b45cd1b | 854 | #else |
xorjoep | 1:24714b45cd1b | 855 | |
xorjoep | 1:24714b45cd1b | 856 | c0 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 857 | |
xorjoep | 1:24714b45cd1b | 858 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 859 | |
xorjoep | 1:24714b45cd1b | 860 | /* acc0 += x[0] * y[0] + x[1] * y[1] */ |
xorjoep | 1:24714b45cd1b | 861 | acc0 = __SMLAD(x0, c0, acc0); |
xorjoep | 1:24714b45cd1b | 862 | |
xorjoep | 1:24714b45cd1b | 863 | /* acc1 += x[1] * y[0] + x[2] * y[1] */ |
xorjoep | 1:24714b45cd1b | 864 | acc1 = __SMLAD(x1, c0, acc1); |
xorjoep | 1:24714b45cd1b | 865 | |
xorjoep | 1:24714b45cd1b | 866 | /* Read x[2], x[3], x[4] */ |
xorjoep | 1:24714b45cd1b | 867 | a = *px; |
xorjoep | 1:24714b45cd1b | 868 | b = *(px + 1); |
xorjoep | 1:24714b45cd1b | 869 | |
xorjoep | 1:24714b45cd1b | 870 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 871 | |
xorjoep | 1:24714b45cd1b | 872 | x2 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 873 | a = *(px + 2); |
xorjoep | 1:24714b45cd1b | 874 | x3 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 875 | |
xorjoep | 1:24714b45cd1b | 876 | #else |
xorjoep | 1:24714b45cd1b | 877 | |
xorjoep | 1:24714b45cd1b | 878 | x2 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 879 | a = *(px + 2); |
xorjoep | 1:24714b45cd1b | 880 | x3 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 881 | |
xorjoep | 1:24714b45cd1b | 882 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 883 | |
xorjoep | 1:24714b45cd1b | 884 | /* acc2 += x[2] * y[0] + x[3] * y[1] */ |
xorjoep | 1:24714b45cd1b | 885 | acc2 = __SMLAD(x2, c0, acc2); |
xorjoep | 1:24714b45cd1b | 886 | |
xorjoep | 1:24714b45cd1b | 887 | /* acc3 += x[3] * y[0] + x[4] * y[1] */ |
xorjoep | 1:24714b45cd1b | 888 | acc3 = __SMLAD(x3, c0, acc3); |
xorjoep | 1:24714b45cd1b | 889 | |
xorjoep | 1:24714b45cd1b | 890 | /* Read y[2] and y[3] */ |
xorjoep | 1:24714b45cd1b | 891 | a = *(py + 2); |
xorjoep | 1:24714b45cd1b | 892 | b = *(py + 3); |
xorjoep | 1:24714b45cd1b | 893 | |
xorjoep | 1:24714b45cd1b | 894 | py += 4U; |
xorjoep | 1:24714b45cd1b | 895 | |
xorjoep | 1:24714b45cd1b | 896 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 897 | |
xorjoep | 1:24714b45cd1b | 898 | c0 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 899 | |
xorjoep | 1:24714b45cd1b | 900 | #else |
xorjoep | 1:24714b45cd1b | 901 | |
xorjoep | 1:24714b45cd1b | 902 | c0 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 903 | |
xorjoep | 1:24714b45cd1b | 904 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 905 | |
xorjoep | 1:24714b45cd1b | 906 | /* acc0 += x[2] * y[2] + x[3] * y[3] */ |
xorjoep | 1:24714b45cd1b | 907 | acc0 = __SMLAD(x2, c0, acc0); |
xorjoep | 1:24714b45cd1b | 908 | |
xorjoep | 1:24714b45cd1b | 909 | /* acc1 += x[3] * y[2] + x[4] * y[3] */ |
xorjoep | 1:24714b45cd1b | 910 | acc1 = __SMLAD(x3, c0, acc1); |
xorjoep | 1:24714b45cd1b | 911 | |
xorjoep | 1:24714b45cd1b | 912 | /* Read x[4], x[5], x[6] */ |
xorjoep | 1:24714b45cd1b | 913 | a = *(px + 2); |
xorjoep | 1:24714b45cd1b | 914 | b = *(px + 3); |
xorjoep | 1:24714b45cd1b | 915 | |
xorjoep | 1:24714b45cd1b | 916 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 917 | |
xorjoep | 1:24714b45cd1b | 918 | x0 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 919 | a = *(px + 4); |
xorjoep | 1:24714b45cd1b | 920 | x1 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 921 | |
xorjoep | 1:24714b45cd1b | 922 | #else |
xorjoep | 1:24714b45cd1b | 923 | |
xorjoep | 1:24714b45cd1b | 924 | x0 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 925 | a = *(px + 4); |
xorjoep | 1:24714b45cd1b | 926 | x1 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 927 | |
xorjoep | 1:24714b45cd1b | 928 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 929 | |
xorjoep | 1:24714b45cd1b | 930 | px += 4U; |
xorjoep | 1:24714b45cd1b | 931 | |
xorjoep | 1:24714b45cd1b | 932 | /* acc2 += x[4] * y[2] + x[5] * y[3] */ |
xorjoep | 1:24714b45cd1b | 933 | acc2 = __SMLAD(x0, c0, acc2); |
xorjoep | 1:24714b45cd1b | 934 | |
xorjoep | 1:24714b45cd1b | 935 | /* acc3 += x[5] * y[2] + x[6] * y[3] */ |
xorjoep | 1:24714b45cd1b | 936 | acc3 = __SMLAD(x1, c0, acc3); |
xorjoep | 1:24714b45cd1b | 937 | |
xorjoep | 1:24714b45cd1b | 938 | } while (--k); |
xorjoep | 1:24714b45cd1b | 939 | |
xorjoep | 1:24714b45cd1b | 940 | /* For the next MAC operations, SIMD is not used |
xorjoep | 1:24714b45cd1b | 941 | * So, the 16 bit pointer if inputB, py is updated */ |
xorjoep | 1:24714b45cd1b | 942 | |
xorjoep | 1:24714b45cd1b | 943 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
xorjoep | 1:24714b45cd1b | 944 | ** No loop unrolling is used. */ |
xorjoep | 1:24714b45cd1b | 945 | k = srcBLen % 0x4U; |
xorjoep | 1:24714b45cd1b | 946 | |
xorjoep | 1:24714b45cd1b | 947 | if (k == 1U) |
xorjoep | 1:24714b45cd1b | 948 | { |
xorjoep | 1:24714b45cd1b | 949 | /* Read y[4] */ |
xorjoep | 1:24714b45cd1b | 950 | c0 = *py; |
xorjoep | 1:24714b45cd1b | 951 | #ifdef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 952 | |
xorjoep | 1:24714b45cd1b | 953 | c0 = c0 << 16U; |
xorjoep | 1:24714b45cd1b | 954 | |
xorjoep | 1:24714b45cd1b | 955 | #else |
xorjoep | 1:24714b45cd1b | 956 | |
xorjoep | 1:24714b45cd1b | 957 | c0 = c0 & 0x0000FFFF; |
xorjoep | 1:24714b45cd1b | 958 | |
xorjoep | 1:24714b45cd1b | 959 | #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 960 | |
xorjoep | 1:24714b45cd1b | 961 | /* Read x[7] */ |
xorjoep | 1:24714b45cd1b | 962 | a = *px; |
xorjoep | 1:24714b45cd1b | 963 | b = *(px + 1); |
xorjoep | 1:24714b45cd1b | 964 | |
xorjoep | 1:24714b45cd1b | 965 | px++;; |
xorjoep | 1:24714b45cd1b | 966 | |
xorjoep | 1:24714b45cd1b | 967 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 968 | |
xorjoep | 1:24714b45cd1b | 969 | x3 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 970 | |
xorjoep | 1:24714b45cd1b | 971 | #else |
xorjoep | 1:24714b45cd1b | 972 | |
xorjoep | 1:24714b45cd1b | 973 | x3 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 974 | |
xorjoep | 1:24714b45cd1b | 975 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 976 | |
xorjoep | 1:24714b45cd1b | 977 | px++; |
xorjoep | 1:24714b45cd1b | 978 | |
xorjoep | 1:24714b45cd1b | 979 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 980 | acc0 = __SMLAD(x0, c0, acc0); |
xorjoep | 1:24714b45cd1b | 981 | acc1 = __SMLAD(x1, c0, acc1); |
xorjoep | 1:24714b45cd1b | 982 | acc2 = __SMLADX(x1, c0, acc2); |
xorjoep | 1:24714b45cd1b | 983 | acc3 = __SMLADX(x3, c0, acc3); |
xorjoep | 1:24714b45cd1b | 984 | } |
xorjoep | 1:24714b45cd1b | 985 | |
xorjoep | 1:24714b45cd1b | 986 | if (k == 2U) |
xorjoep | 1:24714b45cd1b | 987 | { |
xorjoep | 1:24714b45cd1b | 988 | /* Read y[4], y[5] */ |
xorjoep | 1:24714b45cd1b | 989 | a = *py; |
xorjoep | 1:24714b45cd1b | 990 | b = *(py + 1); |
xorjoep | 1:24714b45cd1b | 991 | |
xorjoep | 1:24714b45cd1b | 992 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 993 | |
xorjoep | 1:24714b45cd1b | 994 | c0 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 995 | |
xorjoep | 1:24714b45cd1b | 996 | #else |
xorjoep | 1:24714b45cd1b | 997 | |
xorjoep | 1:24714b45cd1b | 998 | c0 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 999 | |
xorjoep | 1:24714b45cd1b | 1000 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 1001 | |
xorjoep | 1:24714b45cd1b | 1002 | /* Read x[7], x[8], x[9] */ |
xorjoep | 1:24714b45cd1b | 1003 | a = *px; |
xorjoep | 1:24714b45cd1b | 1004 | b = *(px + 1); |
xorjoep | 1:24714b45cd1b | 1005 | |
xorjoep | 1:24714b45cd1b | 1006 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 1007 | |
xorjoep | 1:24714b45cd1b | 1008 | x3 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 1009 | a = *(px + 2); |
xorjoep | 1:24714b45cd1b | 1010 | x2 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 1011 | |
xorjoep | 1:24714b45cd1b | 1012 | #else |
xorjoep | 1:24714b45cd1b | 1013 | |
xorjoep | 1:24714b45cd1b | 1014 | x3 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 1015 | a = *(px + 2); |
xorjoep | 1:24714b45cd1b | 1016 | x2 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 1017 | |
xorjoep | 1:24714b45cd1b | 1018 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 1019 | |
xorjoep | 1:24714b45cd1b | 1020 | px += 2U; |
xorjoep | 1:24714b45cd1b | 1021 | |
xorjoep | 1:24714b45cd1b | 1022 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 1023 | acc0 = __SMLAD(x0, c0, acc0); |
xorjoep | 1:24714b45cd1b | 1024 | acc1 = __SMLAD(x1, c0, acc1); |
xorjoep | 1:24714b45cd1b | 1025 | acc2 = __SMLAD(x3, c0, acc2); |
xorjoep | 1:24714b45cd1b | 1026 | acc3 = __SMLAD(x2, c0, acc3); |
xorjoep | 1:24714b45cd1b | 1027 | } |
xorjoep | 1:24714b45cd1b | 1028 | |
xorjoep | 1:24714b45cd1b | 1029 | if (k == 3U) |
xorjoep | 1:24714b45cd1b | 1030 | { |
xorjoep | 1:24714b45cd1b | 1031 | /* Read y[4], y[5] */ |
xorjoep | 1:24714b45cd1b | 1032 | a = *py; |
xorjoep | 1:24714b45cd1b | 1033 | b = *(py + 1); |
xorjoep | 1:24714b45cd1b | 1034 | |
xorjoep | 1:24714b45cd1b | 1035 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 1036 | |
xorjoep | 1:24714b45cd1b | 1037 | c0 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 1038 | |
xorjoep | 1:24714b45cd1b | 1039 | #else |
xorjoep | 1:24714b45cd1b | 1040 | |
xorjoep | 1:24714b45cd1b | 1041 | c0 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 1042 | |
xorjoep | 1:24714b45cd1b | 1043 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 1044 | |
xorjoep | 1:24714b45cd1b | 1045 | py += 2U; |
xorjoep | 1:24714b45cd1b | 1046 | |
xorjoep | 1:24714b45cd1b | 1047 | /* Read x[7], x[8], x[9] */ |
xorjoep | 1:24714b45cd1b | 1048 | a = *px; |
xorjoep | 1:24714b45cd1b | 1049 | b = *(px + 1); |
xorjoep | 1:24714b45cd1b | 1050 | |
xorjoep | 1:24714b45cd1b | 1051 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 1052 | |
xorjoep | 1:24714b45cd1b | 1053 | x3 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 1054 | a = *(px + 2); |
xorjoep | 1:24714b45cd1b | 1055 | x2 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 1056 | |
xorjoep | 1:24714b45cd1b | 1057 | #else |
xorjoep | 1:24714b45cd1b | 1058 | |
xorjoep | 1:24714b45cd1b | 1059 | x3 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 1060 | a = *(px + 2); |
xorjoep | 1:24714b45cd1b | 1061 | x2 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 1062 | |
xorjoep | 1:24714b45cd1b | 1063 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 1064 | |
xorjoep | 1:24714b45cd1b | 1065 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 1066 | acc0 = __SMLAD(x0, c0, acc0); |
xorjoep | 1:24714b45cd1b | 1067 | acc1 = __SMLAD(x1, c0, acc1); |
xorjoep | 1:24714b45cd1b | 1068 | acc2 = __SMLAD(x3, c0, acc2); |
xorjoep | 1:24714b45cd1b | 1069 | acc3 = __SMLAD(x2, c0, acc3); |
xorjoep | 1:24714b45cd1b | 1070 | |
xorjoep | 1:24714b45cd1b | 1071 | c0 = (*py); |
xorjoep | 1:24714b45cd1b | 1072 | /* Read y[6] */ |
xorjoep | 1:24714b45cd1b | 1073 | #ifdef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 1074 | |
xorjoep | 1:24714b45cd1b | 1075 | c0 = c0 << 16U; |
xorjoep | 1:24714b45cd1b | 1076 | #else |
xorjoep | 1:24714b45cd1b | 1077 | |
xorjoep | 1:24714b45cd1b | 1078 | c0 = c0 & 0x0000FFFF; |
xorjoep | 1:24714b45cd1b | 1079 | #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 1080 | |
xorjoep | 1:24714b45cd1b | 1081 | /* Read x[10] */ |
xorjoep | 1:24714b45cd1b | 1082 | b = *(px + 3); |
xorjoep | 1:24714b45cd1b | 1083 | |
xorjoep | 1:24714b45cd1b | 1084 | #ifndef ARM_MATH_BIG_ENDIAN |
xorjoep | 1:24714b45cd1b | 1085 | |
xorjoep | 1:24714b45cd1b | 1086 | x3 = __PKHBT(a, b, 16); |
xorjoep | 1:24714b45cd1b | 1087 | |
xorjoep | 1:24714b45cd1b | 1088 | #else |
xorjoep | 1:24714b45cd1b | 1089 | |
xorjoep | 1:24714b45cd1b | 1090 | x3 = __PKHBT(b, a, 16); |
xorjoep | 1:24714b45cd1b | 1091 | |
xorjoep | 1:24714b45cd1b | 1092 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
xorjoep | 1:24714b45cd1b | 1093 | |
xorjoep | 1:24714b45cd1b | 1094 | px += 3U; |
xorjoep | 1:24714b45cd1b | 1095 | |
xorjoep | 1:24714b45cd1b | 1096 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 1097 | acc0 = __SMLADX(x1, c0, acc0); |
xorjoep | 1:24714b45cd1b | 1098 | acc1 = __SMLAD(x2, c0, acc1); |
xorjoep | 1:24714b45cd1b | 1099 | acc2 = __SMLADX(x2, c0, acc2); |
xorjoep | 1:24714b45cd1b | 1100 | acc3 = __SMLADX(x3, c0, acc3); |
xorjoep | 1:24714b45cd1b | 1101 | } |
xorjoep | 1:24714b45cd1b | 1102 | |
xorjoep | 1:24714b45cd1b | 1103 | /* Store the result in the accumulator in the destination buffer. */ |
xorjoep | 1:24714b45cd1b | 1104 | *pOut = (q15_t) (acc0 >> 15); |
xorjoep | 1:24714b45cd1b | 1105 | /* Destination pointer is updated according to the address modifier, inc */ |
xorjoep | 1:24714b45cd1b | 1106 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 1107 | |
xorjoep | 1:24714b45cd1b | 1108 | *pOut = (q15_t) (acc1 >> 15); |
xorjoep | 1:24714b45cd1b | 1109 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 1110 | |
xorjoep | 1:24714b45cd1b | 1111 | *pOut = (q15_t) (acc2 >> 15); |
xorjoep | 1:24714b45cd1b | 1112 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 1113 | |
xorjoep | 1:24714b45cd1b | 1114 | *pOut = (q15_t) (acc3 >> 15); |
xorjoep | 1:24714b45cd1b | 1115 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 1116 | |
xorjoep | 1:24714b45cd1b | 1117 | /* Increment the pointer pIn1 index, count by 1 */ |
xorjoep | 1:24714b45cd1b | 1118 | count += 4U; |
xorjoep | 1:24714b45cd1b | 1119 | |
xorjoep | 1:24714b45cd1b | 1120 | /* Update the inputA and inputB pointers for next MAC calculation */ |
xorjoep | 1:24714b45cd1b | 1121 | px = pIn1 + count; |
xorjoep | 1:24714b45cd1b | 1122 | py = pIn2; |
xorjoep | 1:24714b45cd1b | 1123 | |
xorjoep | 1:24714b45cd1b | 1124 | |
xorjoep | 1:24714b45cd1b | 1125 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 1126 | blkCnt--; |
xorjoep | 1:24714b45cd1b | 1127 | } |
xorjoep | 1:24714b45cd1b | 1128 | |
xorjoep | 1:24714b45cd1b | 1129 | /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here. |
xorjoep | 1:24714b45cd1b | 1130 | ** No loop unrolling is used. */ |
xorjoep | 1:24714b45cd1b | 1131 | blkCnt = blockSize2 % 0x4U; |
xorjoep | 1:24714b45cd1b | 1132 | |
xorjoep | 1:24714b45cd1b | 1133 | while (blkCnt > 0U) |
xorjoep | 1:24714b45cd1b | 1134 | { |
xorjoep | 1:24714b45cd1b | 1135 | /* Accumulator is made zero for every iteration */ |
xorjoep | 1:24714b45cd1b | 1136 | sum = 0; |
xorjoep | 1:24714b45cd1b | 1137 | |
xorjoep | 1:24714b45cd1b | 1138 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
xorjoep | 1:24714b45cd1b | 1139 | k = srcBLen >> 2U; |
xorjoep | 1:24714b45cd1b | 1140 | |
xorjoep | 1:24714b45cd1b | 1141 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
xorjoep | 1:24714b45cd1b | 1142 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
xorjoep | 1:24714b45cd1b | 1143 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 1144 | { |
xorjoep | 1:24714b45cd1b | 1145 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 1146 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 1147 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 1148 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 1149 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 1150 | |
xorjoep | 1:24714b45cd1b | 1151 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 1152 | k--; |
xorjoep | 1:24714b45cd1b | 1153 | } |
xorjoep | 1:24714b45cd1b | 1154 | |
xorjoep | 1:24714b45cd1b | 1155 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
xorjoep | 1:24714b45cd1b | 1156 | ** No loop unrolling is used. */ |
xorjoep | 1:24714b45cd1b | 1157 | k = srcBLen % 0x4U; |
xorjoep | 1:24714b45cd1b | 1158 | |
xorjoep | 1:24714b45cd1b | 1159 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 1160 | { |
xorjoep | 1:24714b45cd1b | 1161 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 1162 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 1163 | |
xorjoep | 1:24714b45cd1b | 1164 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 1165 | k--; |
xorjoep | 1:24714b45cd1b | 1166 | } |
xorjoep | 1:24714b45cd1b | 1167 | |
xorjoep | 1:24714b45cd1b | 1168 | /* Store the result in the accumulator in the destination buffer. */ |
xorjoep | 1:24714b45cd1b | 1169 | *pOut = (q15_t) (sum >> 15); |
xorjoep | 1:24714b45cd1b | 1170 | /* Destination pointer is updated according to the address modifier, inc */ |
xorjoep | 1:24714b45cd1b | 1171 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 1172 | |
xorjoep | 1:24714b45cd1b | 1173 | /* Increment the pointer pIn1 index, count by 1 */ |
xorjoep | 1:24714b45cd1b | 1174 | count++; |
xorjoep | 1:24714b45cd1b | 1175 | |
xorjoep | 1:24714b45cd1b | 1176 | /* Update the inputA and inputB pointers for next MAC calculation */ |
xorjoep | 1:24714b45cd1b | 1177 | px = pIn1 + count; |
xorjoep | 1:24714b45cd1b | 1178 | py = pIn2; |
xorjoep | 1:24714b45cd1b | 1179 | |
xorjoep | 1:24714b45cd1b | 1180 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 1181 | blkCnt--; |
xorjoep | 1:24714b45cd1b | 1182 | } |
xorjoep | 1:24714b45cd1b | 1183 | } |
xorjoep | 1:24714b45cd1b | 1184 | else |
xorjoep | 1:24714b45cd1b | 1185 | { |
xorjoep | 1:24714b45cd1b | 1186 | /* If the srcBLen is not a multiple of 4, |
xorjoep | 1:24714b45cd1b | 1187 | * the blockSize2 loop cannot be unrolled by 4 */ |
xorjoep | 1:24714b45cd1b | 1188 | blkCnt = blockSize2; |
xorjoep | 1:24714b45cd1b | 1189 | |
xorjoep | 1:24714b45cd1b | 1190 | while (blkCnt > 0U) |
xorjoep | 1:24714b45cd1b | 1191 | { |
xorjoep | 1:24714b45cd1b | 1192 | /* Accumulator is made zero for every iteration */ |
xorjoep | 1:24714b45cd1b | 1193 | sum = 0; |
xorjoep | 1:24714b45cd1b | 1194 | |
xorjoep | 1:24714b45cd1b | 1195 | /* Loop over srcBLen */ |
xorjoep | 1:24714b45cd1b | 1196 | k = srcBLen; |
xorjoep | 1:24714b45cd1b | 1197 | |
xorjoep | 1:24714b45cd1b | 1198 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 1199 | { |
xorjoep | 1:24714b45cd1b | 1200 | /* Perform the multiply-accumulate */ |
xorjoep | 1:24714b45cd1b | 1201 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 1202 | |
xorjoep | 1:24714b45cd1b | 1203 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 1204 | k--; |
xorjoep | 1:24714b45cd1b | 1205 | } |
xorjoep | 1:24714b45cd1b | 1206 | |
xorjoep | 1:24714b45cd1b | 1207 | /* Store the result in the accumulator in the destination buffer. */ |
xorjoep | 1:24714b45cd1b | 1208 | *pOut = (q15_t) (sum >> 15); |
xorjoep | 1:24714b45cd1b | 1209 | /* Destination pointer is updated according to the address modifier, inc */ |
xorjoep | 1:24714b45cd1b | 1210 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 1211 | |
xorjoep | 1:24714b45cd1b | 1212 | /* Increment the MAC count */ |
xorjoep | 1:24714b45cd1b | 1213 | count++; |
xorjoep | 1:24714b45cd1b | 1214 | |
xorjoep | 1:24714b45cd1b | 1215 | /* Update the inputA and inputB pointers for next MAC calculation */ |
xorjoep | 1:24714b45cd1b | 1216 | px = pIn1 + count; |
xorjoep | 1:24714b45cd1b | 1217 | py = pIn2; |
xorjoep | 1:24714b45cd1b | 1218 | |
xorjoep | 1:24714b45cd1b | 1219 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 1220 | blkCnt--; |
xorjoep | 1:24714b45cd1b | 1221 | } |
xorjoep | 1:24714b45cd1b | 1222 | } |
xorjoep | 1:24714b45cd1b | 1223 | |
xorjoep | 1:24714b45cd1b | 1224 | /* -------------------------- |
xorjoep | 1:24714b45cd1b | 1225 | * Initializations of stage3 |
xorjoep | 1:24714b45cd1b | 1226 | * -------------------------*/ |
xorjoep | 1:24714b45cd1b | 1227 | |
xorjoep | 1:24714b45cd1b | 1228 | /* sum += x[srcALen-srcBLen+1] * y[0] + x[srcALen-srcBLen+2] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
xorjoep | 1:24714b45cd1b | 1229 | * sum += x[srcALen-srcBLen+2] * y[0] + x[srcALen-srcBLen+3] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
xorjoep | 1:24714b45cd1b | 1230 | * .... |
xorjoep | 1:24714b45cd1b | 1231 | * sum += x[srcALen-2] * y[0] + x[srcALen-1] * y[1] |
xorjoep | 1:24714b45cd1b | 1232 | * sum += x[srcALen-1] * y[0] |
xorjoep | 1:24714b45cd1b | 1233 | */ |
xorjoep | 1:24714b45cd1b | 1234 | |
xorjoep | 1:24714b45cd1b | 1235 | /* In this stage the MAC operations are decreased by 1 for every iteration. |
xorjoep | 1:24714b45cd1b | 1236 | The count variable holds the number of MAC operations performed */ |
xorjoep | 1:24714b45cd1b | 1237 | count = srcBLen - 1U; |
xorjoep | 1:24714b45cd1b | 1238 | |
xorjoep | 1:24714b45cd1b | 1239 | /* Working pointer of inputA */ |
xorjoep | 1:24714b45cd1b | 1240 | pSrc1 = (pIn1 + srcALen) - (srcBLen - 1U); |
xorjoep | 1:24714b45cd1b | 1241 | px = pSrc1; |
xorjoep | 1:24714b45cd1b | 1242 | |
xorjoep | 1:24714b45cd1b | 1243 | /* Working pointer of inputB */ |
xorjoep | 1:24714b45cd1b | 1244 | py = pIn2; |
xorjoep | 1:24714b45cd1b | 1245 | |
xorjoep | 1:24714b45cd1b | 1246 | /* ------------------- |
xorjoep | 1:24714b45cd1b | 1247 | * Stage3 process |
xorjoep | 1:24714b45cd1b | 1248 | * ------------------*/ |
xorjoep | 1:24714b45cd1b | 1249 | |
xorjoep | 1:24714b45cd1b | 1250 | while (blockSize3 > 0U) |
xorjoep | 1:24714b45cd1b | 1251 | { |
xorjoep | 1:24714b45cd1b | 1252 | /* Accumulator is made zero for every iteration */ |
xorjoep | 1:24714b45cd1b | 1253 | sum = 0; |
xorjoep | 1:24714b45cd1b | 1254 | |
xorjoep | 1:24714b45cd1b | 1255 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
xorjoep | 1:24714b45cd1b | 1256 | k = count >> 2U; |
xorjoep | 1:24714b45cd1b | 1257 | |
xorjoep | 1:24714b45cd1b | 1258 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
xorjoep | 1:24714b45cd1b | 1259 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
xorjoep | 1:24714b45cd1b | 1260 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 1261 | { |
xorjoep | 1:24714b45cd1b | 1262 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 1263 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 1264 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 1265 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 1266 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 1267 | |
xorjoep | 1:24714b45cd1b | 1268 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 1269 | k--; |
xorjoep | 1:24714b45cd1b | 1270 | } |
xorjoep | 1:24714b45cd1b | 1271 | |
xorjoep | 1:24714b45cd1b | 1272 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
xorjoep | 1:24714b45cd1b | 1273 | ** No loop unrolling is used. */ |
xorjoep | 1:24714b45cd1b | 1274 | k = count % 0x4U; |
xorjoep | 1:24714b45cd1b | 1275 | |
xorjoep | 1:24714b45cd1b | 1276 | while (k > 0U) |
xorjoep | 1:24714b45cd1b | 1277 | { |
xorjoep | 1:24714b45cd1b | 1278 | /* Perform the multiply-accumulates */ |
xorjoep | 1:24714b45cd1b | 1279 | sum += ((q31_t) * px++ * *py++); |
xorjoep | 1:24714b45cd1b | 1280 | |
xorjoep | 1:24714b45cd1b | 1281 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 1282 | k--; |
xorjoep | 1:24714b45cd1b | 1283 | } |
xorjoep | 1:24714b45cd1b | 1284 | |
xorjoep | 1:24714b45cd1b | 1285 | /* Store the result in the accumulator in the destination buffer. */ |
xorjoep | 1:24714b45cd1b | 1286 | *pOut = (q15_t) (sum >> 15); |
xorjoep | 1:24714b45cd1b | 1287 | /* Destination pointer is updated according to the address modifier, inc */ |
xorjoep | 1:24714b45cd1b | 1288 | pOut += inc; |
xorjoep | 1:24714b45cd1b | 1289 | |
xorjoep | 1:24714b45cd1b | 1290 | /* Update the inputA and inputB pointers for next MAC calculation */ |
xorjoep | 1:24714b45cd1b | 1291 | px = ++pSrc1; |
xorjoep | 1:24714b45cd1b | 1292 | py = pIn2; |
xorjoep | 1:24714b45cd1b | 1293 | |
xorjoep | 1:24714b45cd1b | 1294 | /* Decrement the MAC count */ |
xorjoep | 1:24714b45cd1b | 1295 | count--; |
xorjoep | 1:24714b45cd1b | 1296 | |
xorjoep | 1:24714b45cd1b | 1297 | /* Decrement the loop counter */ |
xorjoep | 1:24714b45cd1b | 1298 | blockSize3--; |
xorjoep | 1:24714b45cd1b | 1299 | } |
xorjoep | 1:24714b45cd1b | 1300 | |
xorjoep | 1:24714b45cd1b | 1301 | #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ |
xorjoep | 1:24714b45cd1b | 1302 | |
xorjoep | 1:24714b45cd1b | 1303 | } |
xorjoep | 1:24714b45cd1b | 1304 | |
xorjoep | 1:24714b45cd1b | 1305 | /** |
xorjoep | 1:24714b45cd1b | 1306 | * @} end of Corr group |
xorjoep | 1:24714b45cd1b | 1307 | */ |