V4.0.1 of the ARM CMSIS DSP libraries. Note that arm_bitreversal2.s, arm_cfft_f32.c and arm_rfft_fast_f32.c had to be removed. arm_bitreversal2.s will not assemble with the online tools. So, the fast f32 FFT functions are not yet available. All the other FFT functions are available.
Dependents: MPU9150_Example fir_f32 fir_f32 MPU9150_nucleo_noni2cdev ... more
FilteringFunctions/arm_correlate_fast_q15.c@0:3d9c67d97d6f, 2014-07-28 (annotated)
- Committer:
- emh203
- Date:
- Mon Jul 28 15:03:15 2014 +0000
- Revision:
- 0:3d9c67d97d6f
1st working commit. Had to remove arm_bitreversal2.s arm_cfft_f32.c and arm_rfft_fast_f32.c. The .s will not assemble. For now I removed these functions so we could at least have a library for the other functions.
Who changed what in which revision?
User | Revision | Line number | New contents of line |
---|---|---|---|
emh203 | 0:3d9c67d97d6f | 1 | /* ---------------------------------------------------------------------- |
emh203 | 0:3d9c67d97d6f | 2 | * Copyright (C) 2010-2014 ARM Limited. All rights reserved. |
emh203 | 0:3d9c67d97d6f | 3 | * |
emh203 | 0:3d9c67d97d6f | 4 | * $Date: 12. March 2014 |
emh203 | 0:3d9c67d97d6f | 5 | * $Revision: V1.4.3 |
emh203 | 0:3d9c67d97d6f | 6 | * |
emh203 | 0:3d9c67d97d6f | 7 | * Project: CMSIS DSP Library |
emh203 | 0:3d9c67d97d6f | 8 | * Title: arm_correlate_fast_q15.c |
emh203 | 0:3d9c67d97d6f | 9 | * |
emh203 | 0:3d9c67d97d6f | 10 | * Description: Fast Q15 Correlation. |
emh203 | 0:3d9c67d97d6f | 11 | * |
emh203 | 0:3d9c67d97d6f | 12 | * Target Processor: Cortex-M4/Cortex-M3 |
emh203 | 0:3d9c67d97d6f | 13 | * |
emh203 | 0:3d9c67d97d6f | 14 | * Redistribution and use in source and binary forms, with or without |
emh203 | 0:3d9c67d97d6f | 15 | * modification, are permitted provided that the following conditions |
emh203 | 0:3d9c67d97d6f | 16 | * are met: |
emh203 | 0:3d9c67d97d6f | 17 | * - Redistributions of source code must retain the above copyright |
emh203 | 0:3d9c67d97d6f | 18 | * notice, this list of conditions and the following disclaimer. |
emh203 | 0:3d9c67d97d6f | 19 | * - Redistributions in binary form must reproduce the above copyright |
emh203 | 0:3d9c67d97d6f | 20 | * notice, this list of conditions and the following disclaimer in |
emh203 | 0:3d9c67d97d6f | 21 | * the documentation and/or other materials provided with the |
emh203 | 0:3d9c67d97d6f | 22 | * distribution. |
emh203 | 0:3d9c67d97d6f | 23 | * - Neither the name of ARM LIMITED nor the names of its contributors |
emh203 | 0:3d9c67d97d6f | 24 | * may be used to endorse or promote products derived from this |
emh203 | 0:3d9c67d97d6f | 25 | * software without specific prior written permission. |
emh203 | 0:3d9c67d97d6f | 26 | * |
emh203 | 0:3d9c67d97d6f | 27 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
emh203 | 0:3d9c67d97d6f | 28 | * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
emh203 | 0:3d9c67d97d6f | 29 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS |
emh203 | 0:3d9c67d97d6f | 30 | * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE |
emh203 | 0:3d9c67d97d6f | 31 | * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, |
emh203 | 0:3d9c67d97d6f | 32 | * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, |
emh203 | 0:3d9c67d97d6f | 33 | * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; |
emh203 | 0:3d9c67d97d6f | 34 | * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER |
emh203 | 0:3d9c67d97d6f | 35 | * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
emh203 | 0:3d9c67d97d6f | 36 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN |
emh203 | 0:3d9c67d97d6f | 37 | * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE |
emh203 | 0:3d9c67d97d6f | 38 | * POSSIBILITY OF SUCH DAMAGE. |
emh203 | 0:3d9c67d97d6f | 39 | * -------------------------------------------------------------------- */ |
emh203 | 0:3d9c67d97d6f | 40 | |
emh203 | 0:3d9c67d97d6f | 41 | #include "arm_math.h" |
emh203 | 0:3d9c67d97d6f | 42 | |
emh203 | 0:3d9c67d97d6f | 43 | /** |
emh203 | 0:3d9c67d97d6f | 44 | * @ingroup groupFilters |
emh203 | 0:3d9c67d97d6f | 45 | */ |
emh203 | 0:3d9c67d97d6f | 46 | |
emh203 | 0:3d9c67d97d6f | 47 | /** |
emh203 | 0:3d9c67d97d6f | 48 | * @addtogroup Corr |
emh203 | 0:3d9c67d97d6f | 49 | * @{ |
emh203 | 0:3d9c67d97d6f | 50 | */ |
emh203 | 0:3d9c67d97d6f | 51 | |
emh203 | 0:3d9c67d97d6f | 52 | /** |
emh203 | 0:3d9c67d97d6f | 53 | * @brief Correlation of Q15 sequences (fast version) for Cortex-M3 and Cortex-M4. |
emh203 | 0:3d9c67d97d6f | 54 | * @param[in] *pSrcA points to the first input sequence. |
emh203 | 0:3d9c67d97d6f | 55 | * @param[in] srcALen length of the first input sequence. |
emh203 | 0:3d9c67d97d6f | 56 | * @param[in] *pSrcB points to the second input sequence. |
emh203 | 0:3d9c67d97d6f | 57 | * @param[in] srcBLen length of the second input sequence. |
emh203 | 0:3d9c67d97d6f | 58 | * @param[out] *pDst points to the location where the output result is written. Length 2 * max(srcALen, srcBLen) - 1. |
emh203 | 0:3d9c67d97d6f | 59 | * @return none. |
emh203 | 0:3d9c67d97d6f | 60 | * |
emh203 | 0:3d9c67d97d6f | 61 | * <b>Scaling and Overflow Behavior:</b> |
emh203 | 0:3d9c67d97d6f | 62 | * |
emh203 | 0:3d9c67d97d6f | 63 | * \par |
emh203 | 0:3d9c67d97d6f | 64 | * This fast version uses a 32-bit accumulator with 2.30 format. |
emh203 | 0:3d9c67d97d6f | 65 | * The accumulator maintains full precision of the intermediate multiplication results but provides only a single guard bit. |
emh203 | 0:3d9c67d97d6f | 66 | * There is no saturation on intermediate additions. |
emh203 | 0:3d9c67d97d6f | 67 | * Thus, if the accumulator overflows it wraps around and distorts the result. |
emh203 | 0:3d9c67d97d6f | 68 | * The input signals should be scaled down to avoid intermediate overflows. |
emh203 | 0:3d9c67d97d6f | 69 | * Scale down one of the inputs by 1/min(srcALen, srcBLen) to avoid overflow since a |
emh203 | 0:3d9c67d97d6f | 70 | * maximum of min(srcALen, srcBLen) number of additions is carried internally. |
emh203 | 0:3d9c67d97d6f | 71 | * The 2.30 accumulator is right shifted by 15 bits and then saturated to 1.15 format to yield the final result. |
emh203 | 0:3d9c67d97d6f | 72 | * |
emh203 | 0:3d9c67d97d6f | 73 | * \par |
emh203 | 0:3d9c67d97d6f | 74 | * See <code>arm_correlate_q15()</code> for a slower implementation of this function which uses a 64-bit accumulator to avoid wrap around distortion. |
emh203 | 0:3d9c67d97d6f | 75 | */ |
emh203 | 0:3d9c67d97d6f | 76 | |
emh203 | 0:3d9c67d97d6f | 77 | void arm_correlate_fast_q15( |
emh203 | 0:3d9c67d97d6f | 78 | q15_t * pSrcA, |
emh203 | 0:3d9c67d97d6f | 79 | uint32_t srcALen, |
emh203 | 0:3d9c67d97d6f | 80 | q15_t * pSrcB, |
emh203 | 0:3d9c67d97d6f | 81 | uint32_t srcBLen, |
emh203 | 0:3d9c67d97d6f | 82 | q15_t * pDst) |
emh203 | 0:3d9c67d97d6f | 83 | { |
emh203 | 0:3d9c67d97d6f | 84 | #ifndef UNALIGNED_SUPPORT_DISABLE |
emh203 | 0:3d9c67d97d6f | 85 | |
emh203 | 0:3d9c67d97d6f | 86 | q15_t *pIn1; /* inputA pointer */ |
emh203 | 0:3d9c67d97d6f | 87 | q15_t *pIn2; /* inputB pointer */ |
emh203 | 0:3d9c67d97d6f | 88 | q15_t *pOut = pDst; /* output pointer */ |
emh203 | 0:3d9c67d97d6f | 89 | q31_t sum, acc0, acc1, acc2, acc3; /* Accumulators */ |
emh203 | 0:3d9c67d97d6f | 90 | q15_t *px; /* Intermediate inputA pointer */ |
emh203 | 0:3d9c67d97d6f | 91 | q15_t *py; /* Intermediate inputB pointer */ |
emh203 | 0:3d9c67d97d6f | 92 | q15_t *pSrc1; /* Intermediate pointers */ |
emh203 | 0:3d9c67d97d6f | 93 | q31_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */ |
emh203 | 0:3d9c67d97d6f | 94 | uint32_t j, k = 0u, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counter */ |
emh203 | 0:3d9c67d97d6f | 95 | int32_t inc = 1; /* Destination address modifier */ |
emh203 | 0:3d9c67d97d6f | 96 | |
emh203 | 0:3d9c67d97d6f | 97 | |
emh203 | 0:3d9c67d97d6f | 98 | /* The algorithm implementation is based on the lengths of the inputs. */ |
emh203 | 0:3d9c67d97d6f | 99 | /* srcB is always made to slide across srcA. */ |
emh203 | 0:3d9c67d97d6f | 100 | /* So srcBLen is always considered as shorter or equal to srcALen */ |
emh203 | 0:3d9c67d97d6f | 101 | /* But CORR(x, y) is reverse of CORR(y, x) */ |
emh203 | 0:3d9c67d97d6f | 102 | /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */ |
emh203 | 0:3d9c67d97d6f | 103 | /* and the destination pointer modifier, inc is set to -1 */ |
emh203 | 0:3d9c67d97d6f | 104 | /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */ |
emh203 | 0:3d9c67d97d6f | 105 | /* But to improve the performance, |
emh203 | 0:3d9c67d97d6f | 106 | * we include zeroes in the output instead of zero padding either of the the inputs*/ |
emh203 | 0:3d9c67d97d6f | 107 | /* If srcALen > srcBLen, |
emh203 | 0:3d9c67d97d6f | 108 | * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */ |
emh203 | 0:3d9c67d97d6f | 109 | /* If srcALen < srcBLen, |
emh203 | 0:3d9c67d97d6f | 110 | * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */ |
emh203 | 0:3d9c67d97d6f | 111 | if(srcALen >= srcBLen) |
emh203 | 0:3d9c67d97d6f | 112 | { |
emh203 | 0:3d9c67d97d6f | 113 | /* Initialization of inputA pointer */ |
emh203 | 0:3d9c67d97d6f | 114 | pIn1 = (pSrcA); |
emh203 | 0:3d9c67d97d6f | 115 | |
emh203 | 0:3d9c67d97d6f | 116 | /* Initialization of inputB pointer */ |
emh203 | 0:3d9c67d97d6f | 117 | pIn2 = (pSrcB); |
emh203 | 0:3d9c67d97d6f | 118 | |
emh203 | 0:3d9c67d97d6f | 119 | /* Number of output samples is calculated */ |
emh203 | 0:3d9c67d97d6f | 120 | outBlockSize = (2u * srcALen) - 1u; |
emh203 | 0:3d9c67d97d6f | 121 | |
emh203 | 0:3d9c67d97d6f | 122 | /* When srcALen > srcBLen, zero padding is done to srcB |
emh203 | 0:3d9c67d97d6f | 123 | * to make their lengths equal. |
emh203 | 0:3d9c67d97d6f | 124 | * Instead, (outBlockSize - (srcALen + srcBLen - 1)) |
emh203 | 0:3d9c67d97d6f | 125 | * number of output samples are made zero */ |
emh203 | 0:3d9c67d97d6f | 126 | j = outBlockSize - (srcALen + (srcBLen - 1u)); |
emh203 | 0:3d9c67d97d6f | 127 | |
emh203 | 0:3d9c67d97d6f | 128 | /* Updating the pointer position to non zero value */ |
emh203 | 0:3d9c67d97d6f | 129 | pOut += j; |
emh203 | 0:3d9c67d97d6f | 130 | |
emh203 | 0:3d9c67d97d6f | 131 | } |
emh203 | 0:3d9c67d97d6f | 132 | else |
emh203 | 0:3d9c67d97d6f | 133 | { |
emh203 | 0:3d9c67d97d6f | 134 | /* Initialization of inputA pointer */ |
emh203 | 0:3d9c67d97d6f | 135 | pIn1 = (pSrcB); |
emh203 | 0:3d9c67d97d6f | 136 | |
emh203 | 0:3d9c67d97d6f | 137 | /* Initialization of inputB pointer */ |
emh203 | 0:3d9c67d97d6f | 138 | pIn2 = (pSrcA); |
emh203 | 0:3d9c67d97d6f | 139 | |
emh203 | 0:3d9c67d97d6f | 140 | /* srcBLen is always considered as shorter or equal to srcALen */ |
emh203 | 0:3d9c67d97d6f | 141 | j = srcBLen; |
emh203 | 0:3d9c67d97d6f | 142 | srcBLen = srcALen; |
emh203 | 0:3d9c67d97d6f | 143 | srcALen = j; |
emh203 | 0:3d9c67d97d6f | 144 | |
emh203 | 0:3d9c67d97d6f | 145 | /* CORR(x, y) = Reverse order(CORR(y, x)) */ |
emh203 | 0:3d9c67d97d6f | 146 | /* Hence set the destination pointer to point to the last output sample */ |
emh203 | 0:3d9c67d97d6f | 147 | pOut = pDst + ((srcALen + srcBLen) - 2u); |
emh203 | 0:3d9c67d97d6f | 148 | |
emh203 | 0:3d9c67d97d6f | 149 | /* Destination address modifier is set to -1 */ |
emh203 | 0:3d9c67d97d6f | 150 | inc = -1; |
emh203 | 0:3d9c67d97d6f | 151 | |
emh203 | 0:3d9c67d97d6f | 152 | } |
emh203 | 0:3d9c67d97d6f | 153 | |
emh203 | 0:3d9c67d97d6f | 154 | /* The function is internally |
emh203 | 0:3d9c67d97d6f | 155 | * divided into three parts according to the number of multiplications that has to be |
emh203 | 0:3d9c67d97d6f | 156 | * taken place between inputA samples and inputB samples. In the first part of the |
emh203 | 0:3d9c67d97d6f | 157 | * algorithm, the multiplications increase by one for every iteration. |
emh203 | 0:3d9c67d97d6f | 158 | * In the second part of the algorithm, srcBLen number of multiplications are done. |
emh203 | 0:3d9c67d97d6f | 159 | * In the third part of the algorithm, the multiplications decrease by one |
emh203 | 0:3d9c67d97d6f | 160 | * for every iteration.*/ |
emh203 | 0:3d9c67d97d6f | 161 | /* The algorithm is implemented in three stages. |
emh203 | 0:3d9c67d97d6f | 162 | * The loop counters of each stage is initiated here. */ |
emh203 | 0:3d9c67d97d6f | 163 | blockSize1 = srcBLen - 1u; |
emh203 | 0:3d9c67d97d6f | 164 | blockSize2 = srcALen - (srcBLen - 1u); |
emh203 | 0:3d9c67d97d6f | 165 | blockSize3 = blockSize1; |
emh203 | 0:3d9c67d97d6f | 166 | |
emh203 | 0:3d9c67d97d6f | 167 | /* -------------------------- |
emh203 | 0:3d9c67d97d6f | 168 | * Initializations of stage1 |
emh203 | 0:3d9c67d97d6f | 169 | * -------------------------*/ |
emh203 | 0:3d9c67d97d6f | 170 | |
emh203 | 0:3d9c67d97d6f | 171 | /* sum = x[0] * y[srcBlen - 1] |
emh203 | 0:3d9c67d97d6f | 172 | * sum = x[0] * y[srcBlen - 2] + x[1] * y[srcBlen - 1] |
emh203 | 0:3d9c67d97d6f | 173 | * .... |
emh203 | 0:3d9c67d97d6f | 174 | * sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen - 1] * y[srcBLen - 1] |
emh203 | 0:3d9c67d97d6f | 175 | */ |
emh203 | 0:3d9c67d97d6f | 176 | |
emh203 | 0:3d9c67d97d6f | 177 | /* In this stage the MAC operations are increased by 1 for every iteration. |
emh203 | 0:3d9c67d97d6f | 178 | The count variable holds the number of MAC operations performed */ |
emh203 | 0:3d9c67d97d6f | 179 | count = 1u; |
emh203 | 0:3d9c67d97d6f | 180 | |
emh203 | 0:3d9c67d97d6f | 181 | /* Working pointer of inputA */ |
emh203 | 0:3d9c67d97d6f | 182 | px = pIn1; |
emh203 | 0:3d9c67d97d6f | 183 | |
emh203 | 0:3d9c67d97d6f | 184 | /* Working pointer of inputB */ |
emh203 | 0:3d9c67d97d6f | 185 | pSrc1 = pIn2 + (srcBLen - 1u); |
emh203 | 0:3d9c67d97d6f | 186 | py = pSrc1; |
emh203 | 0:3d9c67d97d6f | 187 | |
emh203 | 0:3d9c67d97d6f | 188 | /* ------------------------ |
emh203 | 0:3d9c67d97d6f | 189 | * Stage1 process |
emh203 | 0:3d9c67d97d6f | 190 | * ----------------------*/ |
emh203 | 0:3d9c67d97d6f | 191 | |
emh203 | 0:3d9c67d97d6f | 192 | /* The first loop starts here */ |
emh203 | 0:3d9c67d97d6f | 193 | while(blockSize1 > 0u) |
emh203 | 0:3d9c67d97d6f | 194 | { |
emh203 | 0:3d9c67d97d6f | 195 | /* Accumulator is made zero for every iteration */ |
emh203 | 0:3d9c67d97d6f | 196 | sum = 0; |
emh203 | 0:3d9c67d97d6f | 197 | |
emh203 | 0:3d9c67d97d6f | 198 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emh203 | 0:3d9c67d97d6f | 199 | k = count >> 2; |
emh203 | 0:3d9c67d97d6f | 200 | |
emh203 | 0:3d9c67d97d6f | 201 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
emh203 | 0:3d9c67d97d6f | 202 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
emh203 | 0:3d9c67d97d6f | 203 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 204 | { |
emh203 | 0:3d9c67d97d6f | 205 | /* x[0] * y[srcBLen - 4] , x[1] * y[srcBLen - 3] */ |
emh203 | 0:3d9c67d97d6f | 206 | sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); |
emh203 | 0:3d9c67d97d6f | 207 | /* x[3] * y[srcBLen - 1] , x[2] * y[srcBLen - 2] */ |
emh203 | 0:3d9c67d97d6f | 208 | sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); |
emh203 | 0:3d9c67d97d6f | 209 | |
emh203 | 0:3d9c67d97d6f | 210 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 211 | k--; |
emh203 | 0:3d9c67d97d6f | 212 | } |
emh203 | 0:3d9c67d97d6f | 213 | |
emh203 | 0:3d9c67d97d6f | 214 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
emh203 | 0:3d9c67d97d6f | 215 | ** No loop unrolling is used. */ |
emh203 | 0:3d9c67d97d6f | 216 | k = count % 0x4u; |
emh203 | 0:3d9c67d97d6f | 217 | |
emh203 | 0:3d9c67d97d6f | 218 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 219 | { |
emh203 | 0:3d9c67d97d6f | 220 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 221 | /* x[0] * y[srcBLen - 1] */ |
emh203 | 0:3d9c67d97d6f | 222 | sum = __SMLAD(*px++, *py++, sum); |
emh203 | 0:3d9c67d97d6f | 223 | |
emh203 | 0:3d9c67d97d6f | 224 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 225 | k--; |
emh203 | 0:3d9c67d97d6f | 226 | } |
emh203 | 0:3d9c67d97d6f | 227 | |
emh203 | 0:3d9c67d97d6f | 228 | /* Store the result in the accumulator in the destination buffer. */ |
emh203 | 0:3d9c67d97d6f | 229 | *pOut = (q15_t) (sum >> 15); |
emh203 | 0:3d9c67d97d6f | 230 | /* Destination pointer is updated according to the address modifier, inc */ |
emh203 | 0:3d9c67d97d6f | 231 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 232 | |
emh203 | 0:3d9c67d97d6f | 233 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emh203 | 0:3d9c67d97d6f | 234 | py = pSrc1 - count; |
emh203 | 0:3d9c67d97d6f | 235 | px = pIn1; |
emh203 | 0:3d9c67d97d6f | 236 | |
emh203 | 0:3d9c67d97d6f | 237 | /* Increment the MAC count */ |
emh203 | 0:3d9c67d97d6f | 238 | count++; |
emh203 | 0:3d9c67d97d6f | 239 | |
emh203 | 0:3d9c67d97d6f | 240 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 241 | blockSize1--; |
emh203 | 0:3d9c67d97d6f | 242 | } |
emh203 | 0:3d9c67d97d6f | 243 | |
emh203 | 0:3d9c67d97d6f | 244 | /* -------------------------- |
emh203 | 0:3d9c67d97d6f | 245 | * Initializations of stage2 |
emh203 | 0:3d9c67d97d6f | 246 | * ------------------------*/ |
emh203 | 0:3d9c67d97d6f | 247 | |
emh203 | 0:3d9c67d97d6f | 248 | /* sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen-1] * y[srcBLen-1] |
emh203 | 0:3d9c67d97d6f | 249 | * sum = x[1] * y[0] + x[2] * y[1] +...+ x[srcBLen] * y[srcBLen-1] |
emh203 | 0:3d9c67d97d6f | 250 | * .... |
emh203 | 0:3d9c67d97d6f | 251 | * sum = x[srcALen-srcBLen-2] * y[0] + x[srcALen-srcBLen-1] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
emh203 | 0:3d9c67d97d6f | 252 | */ |
emh203 | 0:3d9c67d97d6f | 253 | |
emh203 | 0:3d9c67d97d6f | 254 | /* Working pointer of inputA */ |
emh203 | 0:3d9c67d97d6f | 255 | px = pIn1; |
emh203 | 0:3d9c67d97d6f | 256 | |
emh203 | 0:3d9c67d97d6f | 257 | /* Working pointer of inputB */ |
emh203 | 0:3d9c67d97d6f | 258 | py = pIn2; |
emh203 | 0:3d9c67d97d6f | 259 | |
emh203 | 0:3d9c67d97d6f | 260 | /* count is index by which the pointer pIn1 to be incremented */ |
emh203 | 0:3d9c67d97d6f | 261 | count = 0u; |
emh203 | 0:3d9c67d97d6f | 262 | |
emh203 | 0:3d9c67d97d6f | 263 | /* ------------------- |
emh203 | 0:3d9c67d97d6f | 264 | * Stage2 process |
emh203 | 0:3d9c67d97d6f | 265 | * ------------------*/ |
emh203 | 0:3d9c67d97d6f | 266 | |
emh203 | 0:3d9c67d97d6f | 267 | /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed. |
emh203 | 0:3d9c67d97d6f | 268 | * So, to loop unroll over blockSize2, |
emh203 | 0:3d9c67d97d6f | 269 | * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */ |
emh203 | 0:3d9c67d97d6f | 270 | if(srcBLen >= 4u) |
emh203 | 0:3d9c67d97d6f | 271 | { |
emh203 | 0:3d9c67d97d6f | 272 | /* Loop unroll over blockSize2, by 4 */ |
emh203 | 0:3d9c67d97d6f | 273 | blkCnt = blockSize2 >> 2u; |
emh203 | 0:3d9c67d97d6f | 274 | |
emh203 | 0:3d9c67d97d6f | 275 | while(blkCnt > 0u) |
emh203 | 0:3d9c67d97d6f | 276 | { |
emh203 | 0:3d9c67d97d6f | 277 | /* Set all accumulators to zero */ |
emh203 | 0:3d9c67d97d6f | 278 | acc0 = 0; |
emh203 | 0:3d9c67d97d6f | 279 | acc1 = 0; |
emh203 | 0:3d9c67d97d6f | 280 | acc2 = 0; |
emh203 | 0:3d9c67d97d6f | 281 | acc3 = 0; |
emh203 | 0:3d9c67d97d6f | 282 | |
emh203 | 0:3d9c67d97d6f | 283 | /* read x[0], x[1] samples */ |
emh203 | 0:3d9c67d97d6f | 284 | x0 = *__SIMD32(px); |
emh203 | 0:3d9c67d97d6f | 285 | /* read x[1], x[2] samples */ |
emh203 | 0:3d9c67d97d6f | 286 | x1 = _SIMD32_OFFSET(px + 1); |
emh203 | 0:3d9c67d97d6f | 287 | px += 2u; |
emh203 | 0:3d9c67d97d6f | 288 | |
emh203 | 0:3d9c67d97d6f | 289 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emh203 | 0:3d9c67d97d6f | 290 | k = srcBLen >> 2u; |
emh203 | 0:3d9c67d97d6f | 291 | |
emh203 | 0:3d9c67d97d6f | 292 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
emh203 | 0:3d9c67d97d6f | 293 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
emh203 | 0:3d9c67d97d6f | 294 | do |
emh203 | 0:3d9c67d97d6f | 295 | { |
emh203 | 0:3d9c67d97d6f | 296 | /* Read the first two inputB samples using SIMD: |
emh203 | 0:3d9c67d97d6f | 297 | * y[0] and y[1] */ |
emh203 | 0:3d9c67d97d6f | 298 | c0 = *__SIMD32(py)++; |
emh203 | 0:3d9c67d97d6f | 299 | |
emh203 | 0:3d9c67d97d6f | 300 | /* acc0 += x[0] * y[0] + x[1] * y[1] */ |
emh203 | 0:3d9c67d97d6f | 301 | acc0 = __SMLAD(x0, c0, acc0); |
emh203 | 0:3d9c67d97d6f | 302 | |
emh203 | 0:3d9c67d97d6f | 303 | /* acc1 += x[1] * y[0] + x[2] * y[1] */ |
emh203 | 0:3d9c67d97d6f | 304 | acc1 = __SMLAD(x1, c0, acc1); |
emh203 | 0:3d9c67d97d6f | 305 | |
emh203 | 0:3d9c67d97d6f | 306 | /* Read x[2], x[3] */ |
emh203 | 0:3d9c67d97d6f | 307 | x2 = *__SIMD32(px); |
emh203 | 0:3d9c67d97d6f | 308 | |
emh203 | 0:3d9c67d97d6f | 309 | /* Read x[3], x[4] */ |
emh203 | 0:3d9c67d97d6f | 310 | x3 = _SIMD32_OFFSET(px + 1); |
emh203 | 0:3d9c67d97d6f | 311 | |
emh203 | 0:3d9c67d97d6f | 312 | /* acc2 += x[2] * y[0] + x[3] * y[1] */ |
emh203 | 0:3d9c67d97d6f | 313 | acc2 = __SMLAD(x2, c0, acc2); |
emh203 | 0:3d9c67d97d6f | 314 | |
emh203 | 0:3d9c67d97d6f | 315 | /* acc3 += x[3] * y[0] + x[4] * y[1] */ |
emh203 | 0:3d9c67d97d6f | 316 | acc3 = __SMLAD(x3, c0, acc3); |
emh203 | 0:3d9c67d97d6f | 317 | |
emh203 | 0:3d9c67d97d6f | 318 | /* Read y[2] and y[3] */ |
emh203 | 0:3d9c67d97d6f | 319 | c0 = *__SIMD32(py)++; |
emh203 | 0:3d9c67d97d6f | 320 | |
emh203 | 0:3d9c67d97d6f | 321 | /* acc0 += x[2] * y[2] + x[3] * y[3] */ |
emh203 | 0:3d9c67d97d6f | 322 | acc0 = __SMLAD(x2, c0, acc0); |
emh203 | 0:3d9c67d97d6f | 323 | |
emh203 | 0:3d9c67d97d6f | 324 | /* acc1 += x[3] * y[2] + x[4] * y[3] */ |
emh203 | 0:3d9c67d97d6f | 325 | acc1 = __SMLAD(x3, c0, acc1); |
emh203 | 0:3d9c67d97d6f | 326 | |
emh203 | 0:3d9c67d97d6f | 327 | /* Read x[4], x[5] */ |
emh203 | 0:3d9c67d97d6f | 328 | x0 = _SIMD32_OFFSET(px + 2); |
emh203 | 0:3d9c67d97d6f | 329 | |
emh203 | 0:3d9c67d97d6f | 330 | /* Read x[5], x[6] */ |
emh203 | 0:3d9c67d97d6f | 331 | x1 = _SIMD32_OFFSET(px + 3); |
emh203 | 0:3d9c67d97d6f | 332 | px += 4u; |
emh203 | 0:3d9c67d97d6f | 333 | |
emh203 | 0:3d9c67d97d6f | 334 | /* acc2 += x[4] * y[2] + x[5] * y[3] */ |
emh203 | 0:3d9c67d97d6f | 335 | acc2 = __SMLAD(x0, c0, acc2); |
emh203 | 0:3d9c67d97d6f | 336 | |
emh203 | 0:3d9c67d97d6f | 337 | /* acc3 += x[5] * y[2] + x[6] * y[3] */ |
emh203 | 0:3d9c67d97d6f | 338 | acc3 = __SMLAD(x1, c0, acc3); |
emh203 | 0:3d9c67d97d6f | 339 | |
emh203 | 0:3d9c67d97d6f | 340 | } while(--k); |
emh203 | 0:3d9c67d97d6f | 341 | |
emh203 | 0:3d9c67d97d6f | 342 | /* For the next MAC operations, SIMD is not used |
emh203 | 0:3d9c67d97d6f | 343 | * So, the 16 bit pointer if inputB, py is updated */ |
emh203 | 0:3d9c67d97d6f | 344 | |
emh203 | 0:3d9c67d97d6f | 345 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
emh203 | 0:3d9c67d97d6f | 346 | ** No loop unrolling is used. */ |
emh203 | 0:3d9c67d97d6f | 347 | k = srcBLen % 0x4u; |
emh203 | 0:3d9c67d97d6f | 348 | |
emh203 | 0:3d9c67d97d6f | 349 | if(k == 1u) |
emh203 | 0:3d9c67d97d6f | 350 | { |
emh203 | 0:3d9c67d97d6f | 351 | /* Read y[4] */ |
emh203 | 0:3d9c67d97d6f | 352 | c0 = *py; |
emh203 | 0:3d9c67d97d6f | 353 | #ifdef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 354 | |
emh203 | 0:3d9c67d97d6f | 355 | c0 = c0 << 16u; |
emh203 | 0:3d9c67d97d6f | 356 | |
emh203 | 0:3d9c67d97d6f | 357 | #else |
emh203 | 0:3d9c67d97d6f | 358 | |
emh203 | 0:3d9c67d97d6f | 359 | c0 = c0 & 0x0000FFFF; |
emh203 | 0:3d9c67d97d6f | 360 | |
emh203 | 0:3d9c67d97d6f | 361 | #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 362 | |
emh203 | 0:3d9c67d97d6f | 363 | /* Read x[7] */ |
emh203 | 0:3d9c67d97d6f | 364 | x3 = *__SIMD32(px); |
emh203 | 0:3d9c67d97d6f | 365 | px++; |
emh203 | 0:3d9c67d97d6f | 366 | |
emh203 | 0:3d9c67d97d6f | 367 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 368 | acc0 = __SMLAD(x0, c0, acc0); |
emh203 | 0:3d9c67d97d6f | 369 | acc1 = __SMLAD(x1, c0, acc1); |
emh203 | 0:3d9c67d97d6f | 370 | acc2 = __SMLADX(x1, c0, acc2); |
emh203 | 0:3d9c67d97d6f | 371 | acc3 = __SMLADX(x3, c0, acc3); |
emh203 | 0:3d9c67d97d6f | 372 | } |
emh203 | 0:3d9c67d97d6f | 373 | |
emh203 | 0:3d9c67d97d6f | 374 | if(k == 2u) |
emh203 | 0:3d9c67d97d6f | 375 | { |
emh203 | 0:3d9c67d97d6f | 376 | /* Read y[4], y[5] */ |
emh203 | 0:3d9c67d97d6f | 377 | c0 = *__SIMD32(py); |
emh203 | 0:3d9c67d97d6f | 378 | |
emh203 | 0:3d9c67d97d6f | 379 | /* Read x[7], x[8] */ |
emh203 | 0:3d9c67d97d6f | 380 | x3 = *__SIMD32(px); |
emh203 | 0:3d9c67d97d6f | 381 | |
emh203 | 0:3d9c67d97d6f | 382 | /* Read x[9] */ |
emh203 | 0:3d9c67d97d6f | 383 | x2 = _SIMD32_OFFSET(px + 1); |
emh203 | 0:3d9c67d97d6f | 384 | px += 2u; |
emh203 | 0:3d9c67d97d6f | 385 | |
emh203 | 0:3d9c67d97d6f | 386 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 387 | acc0 = __SMLAD(x0, c0, acc0); |
emh203 | 0:3d9c67d97d6f | 388 | acc1 = __SMLAD(x1, c0, acc1); |
emh203 | 0:3d9c67d97d6f | 389 | acc2 = __SMLAD(x3, c0, acc2); |
emh203 | 0:3d9c67d97d6f | 390 | acc3 = __SMLAD(x2, c0, acc3); |
emh203 | 0:3d9c67d97d6f | 391 | } |
emh203 | 0:3d9c67d97d6f | 392 | |
emh203 | 0:3d9c67d97d6f | 393 | if(k == 3u) |
emh203 | 0:3d9c67d97d6f | 394 | { |
emh203 | 0:3d9c67d97d6f | 395 | /* Read y[4], y[5] */ |
emh203 | 0:3d9c67d97d6f | 396 | c0 = *__SIMD32(py)++; |
emh203 | 0:3d9c67d97d6f | 397 | |
emh203 | 0:3d9c67d97d6f | 398 | /* Read x[7], x[8] */ |
emh203 | 0:3d9c67d97d6f | 399 | x3 = *__SIMD32(px); |
emh203 | 0:3d9c67d97d6f | 400 | |
emh203 | 0:3d9c67d97d6f | 401 | /* Read x[9] */ |
emh203 | 0:3d9c67d97d6f | 402 | x2 = _SIMD32_OFFSET(px + 1); |
emh203 | 0:3d9c67d97d6f | 403 | |
emh203 | 0:3d9c67d97d6f | 404 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 405 | acc0 = __SMLAD(x0, c0, acc0); |
emh203 | 0:3d9c67d97d6f | 406 | acc1 = __SMLAD(x1, c0, acc1); |
emh203 | 0:3d9c67d97d6f | 407 | acc2 = __SMLAD(x3, c0, acc2); |
emh203 | 0:3d9c67d97d6f | 408 | acc3 = __SMLAD(x2, c0, acc3); |
emh203 | 0:3d9c67d97d6f | 409 | |
emh203 | 0:3d9c67d97d6f | 410 | c0 = (*py); |
emh203 | 0:3d9c67d97d6f | 411 | /* Read y[6] */ |
emh203 | 0:3d9c67d97d6f | 412 | #ifdef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 413 | |
emh203 | 0:3d9c67d97d6f | 414 | c0 = c0 << 16u; |
emh203 | 0:3d9c67d97d6f | 415 | #else |
emh203 | 0:3d9c67d97d6f | 416 | |
emh203 | 0:3d9c67d97d6f | 417 | c0 = c0 & 0x0000FFFF; |
emh203 | 0:3d9c67d97d6f | 418 | #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 419 | |
emh203 | 0:3d9c67d97d6f | 420 | /* Read x[10] */ |
emh203 | 0:3d9c67d97d6f | 421 | x3 = _SIMD32_OFFSET(px + 2); |
emh203 | 0:3d9c67d97d6f | 422 | px += 3u; |
emh203 | 0:3d9c67d97d6f | 423 | |
emh203 | 0:3d9c67d97d6f | 424 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 425 | acc0 = __SMLADX(x1, c0, acc0); |
emh203 | 0:3d9c67d97d6f | 426 | acc1 = __SMLAD(x2, c0, acc1); |
emh203 | 0:3d9c67d97d6f | 427 | acc2 = __SMLADX(x2, c0, acc2); |
emh203 | 0:3d9c67d97d6f | 428 | acc3 = __SMLADX(x3, c0, acc3); |
emh203 | 0:3d9c67d97d6f | 429 | } |
emh203 | 0:3d9c67d97d6f | 430 | |
emh203 | 0:3d9c67d97d6f | 431 | /* Store the result in the accumulator in the destination buffer. */ |
emh203 | 0:3d9c67d97d6f | 432 | *pOut = (q15_t) (acc0 >> 15); |
emh203 | 0:3d9c67d97d6f | 433 | /* Destination pointer is updated according to the address modifier, inc */ |
emh203 | 0:3d9c67d97d6f | 434 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 435 | |
emh203 | 0:3d9c67d97d6f | 436 | *pOut = (q15_t) (acc1 >> 15); |
emh203 | 0:3d9c67d97d6f | 437 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 438 | |
emh203 | 0:3d9c67d97d6f | 439 | *pOut = (q15_t) (acc2 >> 15); |
emh203 | 0:3d9c67d97d6f | 440 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 441 | |
emh203 | 0:3d9c67d97d6f | 442 | *pOut = (q15_t) (acc3 >> 15); |
emh203 | 0:3d9c67d97d6f | 443 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 444 | |
emh203 | 0:3d9c67d97d6f | 445 | /* Increment the pointer pIn1 index, count by 1 */ |
emh203 | 0:3d9c67d97d6f | 446 | count += 4u; |
emh203 | 0:3d9c67d97d6f | 447 | |
emh203 | 0:3d9c67d97d6f | 448 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emh203 | 0:3d9c67d97d6f | 449 | px = pIn1 + count; |
emh203 | 0:3d9c67d97d6f | 450 | py = pIn2; |
emh203 | 0:3d9c67d97d6f | 451 | |
emh203 | 0:3d9c67d97d6f | 452 | |
emh203 | 0:3d9c67d97d6f | 453 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 454 | blkCnt--; |
emh203 | 0:3d9c67d97d6f | 455 | } |
emh203 | 0:3d9c67d97d6f | 456 | |
emh203 | 0:3d9c67d97d6f | 457 | /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here. |
emh203 | 0:3d9c67d97d6f | 458 | ** No loop unrolling is used. */ |
emh203 | 0:3d9c67d97d6f | 459 | blkCnt = blockSize2 % 0x4u; |
emh203 | 0:3d9c67d97d6f | 460 | |
emh203 | 0:3d9c67d97d6f | 461 | while(blkCnt > 0u) |
emh203 | 0:3d9c67d97d6f | 462 | { |
emh203 | 0:3d9c67d97d6f | 463 | /* Accumulator is made zero for every iteration */ |
emh203 | 0:3d9c67d97d6f | 464 | sum = 0; |
emh203 | 0:3d9c67d97d6f | 465 | |
emh203 | 0:3d9c67d97d6f | 466 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emh203 | 0:3d9c67d97d6f | 467 | k = srcBLen >> 2u; |
emh203 | 0:3d9c67d97d6f | 468 | |
emh203 | 0:3d9c67d97d6f | 469 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
emh203 | 0:3d9c67d97d6f | 470 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
emh203 | 0:3d9c67d97d6f | 471 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 472 | { |
emh203 | 0:3d9c67d97d6f | 473 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 474 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 475 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 476 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 477 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 478 | |
emh203 | 0:3d9c67d97d6f | 479 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 480 | k--; |
emh203 | 0:3d9c67d97d6f | 481 | } |
emh203 | 0:3d9c67d97d6f | 482 | |
emh203 | 0:3d9c67d97d6f | 483 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
emh203 | 0:3d9c67d97d6f | 484 | ** No loop unrolling is used. */ |
emh203 | 0:3d9c67d97d6f | 485 | k = srcBLen % 0x4u; |
emh203 | 0:3d9c67d97d6f | 486 | |
emh203 | 0:3d9c67d97d6f | 487 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 488 | { |
emh203 | 0:3d9c67d97d6f | 489 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 490 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 491 | |
emh203 | 0:3d9c67d97d6f | 492 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 493 | k--; |
emh203 | 0:3d9c67d97d6f | 494 | } |
emh203 | 0:3d9c67d97d6f | 495 | |
emh203 | 0:3d9c67d97d6f | 496 | /* Store the result in the accumulator in the destination buffer. */ |
emh203 | 0:3d9c67d97d6f | 497 | *pOut = (q15_t) (sum >> 15); |
emh203 | 0:3d9c67d97d6f | 498 | /* Destination pointer is updated according to the address modifier, inc */ |
emh203 | 0:3d9c67d97d6f | 499 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 500 | |
emh203 | 0:3d9c67d97d6f | 501 | /* Increment the pointer pIn1 index, count by 1 */ |
emh203 | 0:3d9c67d97d6f | 502 | count++; |
emh203 | 0:3d9c67d97d6f | 503 | |
emh203 | 0:3d9c67d97d6f | 504 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emh203 | 0:3d9c67d97d6f | 505 | px = pIn1 + count; |
emh203 | 0:3d9c67d97d6f | 506 | py = pIn2; |
emh203 | 0:3d9c67d97d6f | 507 | |
emh203 | 0:3d9c67d97d6f | 508 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 509 | blkCnt--; |
emh203 | 0:3d9c67d97d6f | 510 | } |
emh203 | 0:3d9c67d97d6f | 511 | } |
emh203 | 0:3d9c67d97d6f | 512 | else |
emh203 | 0:3d9c67d97d6f | 513 | { |
emh203 | 0:3d9c67d97d6f | 514 | /* If the srcBLen is not a multiple of 4, |
emh203 | 0:3d9c67d97d6f | 515 | * the blockSize2 loop cannot be unrolled by 4 */ |
emh203 | 0:3d9c67d97d6f | 516 | blkCnt = blockSize2; |
emh203 | 0:3d9c67d97d6f | 517 | |
emh203 | 0:3d9c67d97d6f | 518 | while(blkCnt > 0u) |
emh203 | 0:3d9c67d97d6f | 519 | { |
emh203 | 0:3d9c67d97d6f | 520 | /* Accumulator is made zero for every iteration */ |
emh203 | 0:3d9c67d97d6f | 521 | sum = 0; |
emh203 | 0:3d9c67d97d6f | 522 | |
emh203 | 0:3d9c67d97d6f | 523 | /* Loop over srcBLen */ |
emh203 | 0:3d9c67d97d6f | 524 | k = srcBLen; |
emh203 | 0:3d9c67d97d6f | 525 | |
emh203 | 0:3d9c67d97d6f | 526 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 527 | { |
emh203 | 0:3d9c67d97d6f | 528 | /* Perform the multiply-accumulate */ |
emh203 | 0:3d9c67d97d6f | 529 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 530 | |
emh203 | 0:3d9c67d97d6f | 531 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 532 | k--; |
emh203 | 0:3d9c67d97d6f | 533 | } |
emh203 | 0:3d9c67d97d6f | 534 | |
emh203 | 0:3d9c67d97d6f | 535 | /* Store the result in the accumulator in the destination buffer. */ |
emh203 | 0:3d9c67d97d6f | 536 | *pOut = (q15_t) (sum >> 15); |
emh203 | 0:3d9c67d97d6f | 537 | /* Destination pointer is updated according to the address modifier, inc */ |
emh203 | 0:3d9c67d97d6f | 538 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 539 | |
emh203 | 0:3d9c67d97d6f | 540 | /* Increment the MAC count */ |
emh203 | 0:3d9c67d97d6f | 541 | count++; |
emh203 | 0:3d9c67d97d6f | 542 | |
emh203 | 0:3d9c67d97d6f | 543 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emh203 | 0:3d9c67d97d6f | 544 | px = pIn1 + count; |
emh203 | 0:3d9c67d97d6f | 545 | py = pIn2; |
emh203 | 0:3d9c67d97d6f | 546 | |
emh203 | 0:3d9c67d97d6f | 547 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 548 | blkCnt--; |
emh203 | 0:3d9c67d97d6f | 549 | } |
emh203 | 0:3d9c67d97d6f | 550 | } |
emh203 | 0:3d9c67d97d6f | 551 | |
emh203 | 0:3d9c67d97d6f | 552 | /* -------------------------- |
emh203 | 0:3d9c67d97d6f | 553 | * Initializations of stage3 |
emh203 | 0:3d9c67d97d6f | 554 | * -------------------------*/ |
emh203 | 0:3d9c67d97d6f | 555 | |
emh203 | 0:3d9c67d97d6f | 556 | /* sum += x[srcALen-srcBLen+1] * y[0] + x[srcALen-srcBLen+2] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
emh203 | 0:3d9c67d97d6f | 557 | * sum += x[srcALen-srcBLen+2] * y[0] + x[srcALen-srcBLen+3] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
emh203 | 0:3d9c67d97d6f | 558 | * .... |
emh203 | 0:3d9c67d97d6f | 559 | * sum += x[srcALen-2] * y[0] + x[srcALen-1] * y[1] |
emh203 | 0:3d9c67d97d6f | 560 | * sum += x[srcALen-1] * y[0] |
emh203 | 0:3d9c67d97d6f | 561 | */ |
emh203 | 0:3d9c67d97d6f | 562 | |
emh203 | 0:3d9c67d97d6f | 563 | /* In this stage the MAC operations are decreased by 1 for every iteration. |
emh203 | 0:3d9c67d97d6f | 564 | The count variable holds the number of MAC operations performed */ |
emh203 | 0:3d9c67d97d6f | 565 | count = srcBLen - 1u; |
emh203 | 0:3d9c67d97d6f | 566 | |
emh203 | 0:3d9c67d97d6f | 567 | /* Working pointer of inputA */ |
emh203 | 0:3d9c67d97d6f | 568 | pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u); |
emh203 | 0:3d9c67d97d6f | 569 | px = pSrc1; |
emh203 | 0:3d9c67d97d6f | 570 | |
emh203 | 0:3d9c67d97d6f | 571 | /* Working pointer of inputB */ |
emh203 | 0:3d9c67d97d6f | 572 | py = pIn2; |
emh203 | 0:3d9c67d97d6f | 573 | |
emh203 | 0:3d9c67d97d6f | 574 | /* ------------------- |
emh203 | 0:3d9c67d97d6f | 575 | * Stage3 process |
emh203 | 0:3d9c67d97d6f | 576 | * ------------------*/ |
emh203 | 0:3d9c67d97d6f | 577 | |
emh203 | 0:3d9c67d97d6f | 578 | while(blockSize3 > 0u) |
emh203 | 0:3d9c67d97d6f | 579 | { |
emh203 | 0:3d9c67d97d6f | 580 | /* Accumulator is made zero for every iteration */ |
emh203 | 0:3d9c67d97d6f | 581 | sum = 0; |
emh203 | 0:3d9c67d97d6f | 582 | |
emh203 | 0:3d9c67d97d6f | 583 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emh203 | 0:3d9c67d97d6f | 584 | k = count >> 2u; |
emh203 | 0:3d9c67d97d6f | 585 | |
emh203 | 0:3d9c67d97d6f | 586 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
emh203 | 0:3d9c67d97d6f | 587 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
emh203 | 0:3d9c67d97d6f | 588 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 589 | { |
emh203 | 0:3d9c67d97d6f | 590 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 591 | /* sum += x[srcALen - srcBLen + 4] * y[3] , sum += x[srcALen - srcBLen + 3] * y[2] */ |
emh203 | 0:3d9c67d97d6f | 592 | sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); |
emh203 | 0:3d9c67d97d6f | 593 | /* sum += x[srcALen - srcBLen + 2] * y[1] , sum += x[srcALen - srcBLen + 1] * y[0] */ |
emh203 | 0:3d9c67d97d6f | 594 | sum = __SMLAD(*__SIMD32(px)++, *__SIMD32(py)++, sum); |
emh203 | 0:3d9c67d97d6f | 595 | |
emh203 | 0:3d9c67d97d6f | 596 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 597 | k--; |
emh203 | 0:3d9c67d97d6f | 598 | } |
emh203 | 0:3d9c67d97d6f | 599 | |
emh203 | 0:3d9c67d97d6f | 600 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
emh203 | 0:3d9c67d97d6f | 601 | ** No loop unrolling is used. */ |
emh203 | 0:3d9c67d97d6f | 602 | k = count % 0x4u; |
emh203 | 0:3d9c67d97d6f | 603 | |
emh203 | 0:3d9c67d97d6f | 604 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 605 | { |
emh203 | 0:3d9c67d97d6f | 606 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 607 | sum = __SMLAD(*px++, *py++, sum); |
emh203 | 0:3d9c67d97d6f | 608 | |
emh203 | 0:3d9c67d97d6f | 609 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 610 | k--; |
emh203 | 0:3d9c67d97d6f | 611 | } |
emh203 | 0:3d9c67d97d6f | 612 | |
emh203 | 0:3d9c67d97d6f | 613 | /* Store the result in the accumulator in the destination buffer. */ |
emh203 | 0:3d9c67d97d6f | 614 | *pOut = (q15_t) (sum >> 15); |
emh203 | 0:3d9c67d97d6f | 615 | /* Destination pointer is updated according to the address modifier, inc */ |
emh203 | 0:3d9c67d97d6f | 616 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 617 | |
emh203 | 0:3d9c67d97d6f | 618 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emh203 | 0:3d9c67d97d6f | 619 | px = ++pSrc1; |
emh203 | 0:3d9c67d97d6f | 620 | py = pIn2; |
emh203 | 0:3d9c67d97d6f | 621 | |
emh203 | 0:3d9c67d97d6f | 622 | /* Decrement the MAC count */ |
emh203 | 0:3d9c67d97d6f | 623 | count--; |
emh203 | 0:3d9c67d97d6f | 624 | |
emh203 | 0:3d9c67d97d6f | 625 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 626 | blockSize3--; |
emh203 | 0:3d9c67d97d6f | 627 | } |
emh203 | 0:3d9c67d97d6f | 628 | |
emh203 | 0:3d9c67d97d6f | 629 | #else |
emh203 | 0:3d9c67d97d6f | 630 | |
emh203 | 0:3d9c67d97d6f | 631 | q15_t *pIn1; /* inputA pointer */ |
emh203 | 0:3d9c67d97d6f | 632 | q15_t *pIn2; /* inputB pointer */ |
emh203 | 0:3d9c67d97d6f | 633 | q15_t *pOut = pDst; /* output pointer */ |
emh203 | 0:3d9c67d97d6f | 634 | q31_t sum, acc0, acc1, acc2, acc3; /* Accumulators */ |
emh203 | 0:3d9c67d97d6f | 635 | q15_t *px; /* Intermediate inputA pointer */ |
emh203 | 0:3d9c67d97d6f | 636 | q15_t *py; /* Intermediate inputB pointer */ |
emh203 | 0:3d9c67d97d6f | 637 | q15_t *pSrc1; /* Intermediate pointers */ |
emh203 | 0:3d9c67d97d6f | 638 | q31_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */ |
emh203 | 0:3d9c67d97d6f | 639 | uint32_t j, k = 0u, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counter */ |
emh203 | 0:3d9c67d97d6f | 640 | int32_t inc = 1; /* Destination address modifier */ |
emh203 | 0:3d9c67d97d6f | 641 | q15_t a, b; |
emh203 | 0:3d9c67d97d6f | 642 | |
emh203 | 0:3d9c67d97d6f | 643 | |
emh203 | 0:3d9c67d97d6f | 644 | /* The algorithm implementation is based on the lengths of the inputs. */ |
emh203 | 0:3d9c67d97d6f | 645 | /* srcB is always made to slide across srcA. */ |
emh203 | 0:3d9c67d97d6f | 646 | /* So srcBLen is always considered as shorter or equal to srcALen */ |
emh203 | 0:3d9c67d97d6f | 647 | /* But CORR(x, y) is reverse of CORR(y, x) */ |
emh203 | 0:3d9c67d97d6f | 648 | /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */ |
emh203 | 0:3d9c67d97d6f | 649 | /* and the destination pointer modifier, inc is set to -1 */ |
emh203 | 0:3d9c67d97d6f | 650 | /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */ |
emh203 | 0:3d9c67d97d6f | 651 | /* But to improve the performance, |
emh203 | 0:3d9c67d97d6f | 652 | * we include zeroes in the output instead of zero padding either of the the inputs*/ |
emh203 | 0:3d9c67d97d6f | 653 | /* If srcALen > srcBLen, |
emh203 | 0:3d9c67d97d6f | 654 | * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */ |
emh203 | 0:3d9c67d97d6f | 655 | /* If srcALen < srcBLen, |
emh203 | 0:3d9c67d97d6f | 656 | * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */ |
emh203 | 0:3d9c67d97d6f | 657 | if(srcALen >= srcBLen) |
emh203 | 0:3d9c67d97d6f | 658 | { |
emh203 | 0:3d9c67d97d6f | 659 | /* Initialization of inputA pointer */ |
emh203 | 0:3d9c67d97d6f | 660 | pIn1 = (pSrcA); |
emh203 | 0:3d9c67d97d6f | 661 | |
emh203 | 0:3d9c67d97d6f | 662 | /* Initialization of inputB pointer */ |
emh203 | 0:3d9c67d97d6f | 663 | pIn2 = (pSrcB); |
emh203 | 0:3d9c67d97d6f | 664 | |
emh203 | 0:3d9c67d97d6f | 665 | /* Number of output samples is calculated */ |
emh203 | 0:3d9c67d97d6f | 666 | outBlockSize = (2u * srcALen) - 1u; |
emh203 | 0:3d9c67d97d6f | 667 | |
emh203 | 0:3d9c67d97d6f | 668 | /* When srcALen > srcBLen, zero padding is done to srcB |
emh203 | 0:3d9c67d97d6f | 669 | * to make their lengths equal. |
emh203 | 0:3d9c67d97d6f | 670 | * Instead, (outBlockSize - (srcALen + srcBLen - 1)) |
emh203 | 0:3d9c67d97d6f | 671 | * number of output samples are made zero */ |
emh203 | 0:3d9c67d97d6f | 672 | j = outBlockSize - (srcALen + (srcBLen - 1u)); |
emh203 | 0:3d9c67d97d6f | 673 | |
emh203 | 0:3d9c67d97d6f | 674 | /* Updating the pointer position to non zero value */ |
emh203 | 0:3d9c67d97d6f | 675 | pOut += j; |
emh203 | 0:3d9c67d97d6f | 676 | |
emh203 | 0:3d9c67d97d6f | 677 | } |
emh203 | 0:3d9c67d97d6f | 678 | else |
emh203 | 0:3d9c67d97d6f | 679 | { |
emh203 | 0:3d9c67d97d6f | 680 | /* Initialization of inputA pointer */ |
emh203 | 0:3d9c67d97d6f | 681 | pIn1 = (pSrcB); |
emh203 | 0:3d9c67d97d6f | 682 | |
emh203 | 0:3d9c67d97d6f | 683 | /* Initialization of inputB pointer */ |
emh203 | 0:3d9c67d97d6f | 684 | pIn2 = (pSrcA); |
emh203 | 0:3d9c67d97d6f | 685 | |
emh203 | 0:3d9c67d97d6f | 686 | /* srcBLen is always considered as shorter or equal to srcALen */ |
emh203 | 0:3d9c67d97d6f | 687 | j = srcBLen; |
emh203 | 0:3d9c67d97d6f | 688 | srcBLen = srcALen; |
emh203 | 0:3d9c67d97d6f | 689 | srcALen = j; |
emh203 | 0:3d9c67d97d6f | 690 | |
emh203 | 0:3d9c67d97d6f | 691 | /* CORR(x, y) = Reverse order(CORR(y, x)) */ |
emh203 | 0:3d9c67d97d6f | 692 | /* Hence set the destination pointer to point to the last output sample */ |
emh203 | 0:3d9c67d97d6f | 693 | pOut = pDst + ((srcALen + srcBLen) - 2u); |
emh203 | 0:3d9c67d97d6f | 694 | |
emh203 | 0:3d9c67d97d6f | 695 | /* Destination address modifier is set to -1 */ |
emh203 | 0:3d9c67d97d6f | 696 | inc = -1; |
emh203 | 0:3d9c67d97d6f | 697 | |
emh203 | 0:3d9c67d97d6f | 698 | } |
emh203 | 0:3d9c67d97d6f | 699 | |
emh203 | 0:3d9c67d97d6f | 700 | /* The function is internally |
emh203 | 0:3d9c67d97d6f | 701 | * divided into three parts according to the number of multiplications that has to be |
emh203 | 0:3d9c67d97d6f | 702 | * taken place between inputA samples and inputB samples. In the first part of the |
emh203 | 0:3d9c67d97d6f | 703 | * algorithm, the multiplications increase by one for every iteration. |
emh203 | 0:3d9c67d97d6f | 704 | * In the second part of the algorithm, srcBLen number of multiplications are done. |
emh203 | 0:3d9c67d97d6f | 705 | * In the third part of the algorithm, the multiplications decrease by one |
emh203 | 0:3d9c67d97d6f | 706 | * for every iteration.*/ |
emh203 | 0:3d9c67d97d6f | 707 | /* The algorithm is implemented in three stages. |
emh203 | 0:3d9c67d97d6f | 708 | * The loop counters of each stage is initiated here. */ |
emh203 | 0:3d9c67d97d6f | 709 | blockSize1 = srcBLen - 1u; |
emh203 | 0:3d9c67d97d6f | 710 | blockSize2 = srcALen - (srcBLen - 1u); |
emh203 | 0:3d9c67d97d6f | 711 | blockSize3 = blockSize1; |
emh203 | 0:3d9c67d97d6f | 712 | |
emh203 | 0:3d9c67d97d6f | 713 | /* -------------------------- |
emh203 | 0:3d9c67d97d6f | 714 | * Initializations of stage1 |
emh203 | 0:3d9c67d97d6f | 715 | * -------------------------*/ |
emh203 | 0:3d9c67d97d6f | 716 | |
emh203 | 0:3d9c67d97d6f | 717 | /* sum = x[0] * y[srcBlen - 1] |
emh203 | 0:3d9c67d97d6f | 718 | * sum = x[0] * y[srcBlen - 2] + x[1] * y[srcBlen - 1] |
emh203 | 0:3d9c67d97d6f | 719 | * .... |
emh203 | 0:3d9c67d97d6f | 720 | * sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen - 1] * y[srcBLen - 1] |
emh203 | 0:3d9c67d97d6f | 721 | */ |
emh203 | 0:3d9c67d97d6f | 722 | |
emh203 | 0:3d9c67d97d6f | 723 | /* In this stage the MAC operations are increased by 1 for every iteration. |
emh203 | 0:3d9c67d97d6f | 724 | The count variable holds the number of MAC operations performed */ |
emh203 | 0:3d9c67d97d6f | 725 | count = 1u; |
emh203 | 0:3d9c67d97d6f | 726 | |
emh203 | 0:3d9c67d97d6f | 727 | /* Working pointer of inputA */ |
emh203 | 0:3d9c67d97d6f | 728 | px = pIn1; |
emh203 | 0:3d9c67d97d6f | 729 | |
emh203 | 0:3d9c67d97d6f | 730 | /* Working pointer of inputB */ |
emh203 | 0:3d9c67d97d6f | 731 | pSrc1 = pIn2 + (srcBLen - 1u); |
emh203 | 0:3d9c67d97d6f | 732 | py = pSrc1; |
emh203 | 0:3d9c67d97d6f | 733 | |
emh203 | 0:3d9c67d97d6f | 734 | /* ------------------------ |
emh203 | 0:3d9c67d97d6f | 735 | * Stage1 process |
emh203 | 0:3d9c67d97d6f | 736 | * ----------------------*/ |
emh203 | 0:3d9c67d97d6f | 737 | |
emh203 | 0:3d9c67d97d6f | 738 | /* The first loop starts here */ |
emh203 | 0:3d9c67d97d6f | 739 | while(blockSize1 > 0u) |
emh203 | 0:3d9c67d97d6f | 740 | { |
emh203 | 0:3d9c67d97d6f | 741 | /* Accumulator is made zero for every iteration */ |
emh203 | 0:3d9c67d97d6f | 742 | sum = 0; |
emh203 | 0:3d9c67d97d6f | 743 | |
emh203 | 0:3d9c67d97d6f | 744 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emh203 | 0:3d9c67d97d6f | 745 | k = count >> 2; |
emh203 | 0:3d9c67d97d6f | 746 | |
emh203 | 0:3d9c67d97d6f | 747 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
emh203 | 0:3d9c67d97d6f | 748 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
emh203 | 0:3d9c67d97d6f | 749 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 750 | { |
emh203 | 0:3d9c67d97d6f | 751 | /* x[0] * y[srcBLen - 4] , x[1] * y[srcBLen - 3] */ |
emh203 | 0:3d9c67d97d6f | 752 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 753 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 754 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 755 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 756 | |
emh203 | 0:3d9c67d97d6f | 757 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 758 | k--; |
emh203 | 0:3d9c67d97d6f | 759 | } |
emh203 | 0:3d9c67d97d6f | 760 | |
emh203 | 0:3d9c67d97d6f | 761 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
emh203 | 0:3d9c67d97d6f | 762 | ** No loop unrolling is used. */ |
emh203 | 0:3d9c67d97d6f | 763 | k = count % 0x4u; |
emh203 | 0:3d9c67d97d6f | 764 | |
emh203 | 0:3d9c67d97d6f | 765 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 766 | { |
emh203 | 0:3d9c67d97d6f | 767 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 768 | /* x[0] * y[srcBLen - 1] */ |
emh203 | 0:3d9c67d97d6f | 769 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 770 | |
emh203 | 0:3d9c67d97d6f | 771 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 772 | k--; |
emh203 | 0:3d9c67d97d6f | 773 | } |
emh203 | 0:3d9c67d97d6f | 774 | |
emh203 | 0:3d9c67d97d6f | 775 | /* Store the result in the accumulator in the destination buffer. */ |
emh203 | 0:3d9c67d97d6f | 776 | *pOut = (q15_t) (sum >> 15); |
emh203 | 0:3d9c67d97d6f | 777 | /* Destination pointer is updated according to the address modifier, inc */ |
emh203 | 0:3d9c67d97d6f | 778 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 779 | |
emh203 | 0:3d9c67d97d6f | 780 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emh203 | 0:3d9c67d97d6f | 781 | py = pSrc1 - count; |
emh203 | 0:3d9c67d97d6f | 782 | px = pIn1; |
emh203 | 0:3d9c67d97d6f | 783 | |
emh203 | 0:3d9c67d97d6f | 784 | /* Increment the MAC count */ |
emh203 | 0:3d9c67d97d6f | 785 | count++; |
emh203 | 0:3d9c67d97d6f | 786 | |
emh203 | 0:3d9c67d97d6f | 787 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 788 | blockSize1--; |
emh203 | 0:3d9c67d97d6f | 789 | } |
emh203 | 0:3d9c67d97d6f | 790 | |
emh203 | 0:3d9c67d97d6f | 791 | /* -------------------------- |
emh203 | 0:3d9c67d97d6f | 792 | * Initializations of stage2 |
emh203 | 0:3d9c67d97d6f | 793 | * ------------------------*/ |
emh203 | 0:3d9c67d97d6f | 794 | |
emh203 | 0:3d9c67d97d6f | 795 | /* sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen-1] * y[srcBLen-1] |
emh203 | 0:3d9c67d97d6f | 796 | * sum = x[1] * y[0] + x[2] * y[1] +...+ x[srcBLen] * y[srcBLen-1] |
emh203 | 0:3d9c67d97d6f | 797 | * .... |
emh203 | 0:3d9c67d97d6f | 798 | * sum = x[srcALen-srcBLen-2] * y[0] + x[srcALen-srcBLen-1] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
emh203 | 0:3d9c67d97d6f | 799 | */ |
emh203 | 0:3d9c67d97d6f | 800 | |
emh203 | 0:3d9c67d97d6f | 801 | /* Working pointer of inputA */ |
emh203 | 0:3d9c67d97d6f | 802 | px = pIn1; |
emh203 | 0:3d9c67d97d6f | 803 | |
emh203 | 0:3d9c67d97d6f | 804 | /* Working pointer of inputB */ |
emh203 | 0:3d9c67d97d6f | 805 | py = pIn2; |
emh203 | 0:3d9c67d97d6f | 806 | |
emh203 | 0:3d9c67d97d6f | 807 | /* count is index by which the pointer pIn1 to be incremented */ |
emh203 | 0:3d9c67d97d6f | 808 | count = 0u; |
emh203 | 0:3d9c67d97d6f | 809 | |
emh203 | 0:3d9c67d97d6f | 810 | /* ------------------- |
emh203 | 0:3d9c67d97d6f | 811 | * Stage2 process |
emh203 | 0:3d9c67d97d6f | 812 | * ------------------*/ |
emh203 | 0:3d9c67d97d6f | 813 | |
emh203 | 0:3d9c67d97d6f | 814 | /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed. |
emh203 | 0:3d9c67d97d6f | 815 | * So, to loop unroll over blockSize2, |
emh203 | 0:3d9c67d97d6f | 816 | * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */ |
emh203 | 0:3d9c67d97d6f | 817 | if(srcBLen >= 4u) |
emh203 | 0:3d9c67d97d6f | 818 | { |
emh203 | 0:3d9c67d97d6f | 819 | /* Loop unroll over blockSize2, by 4 */ |
emh203 | 0:3d9c67d97d6f | 820 | blkCnt = blockSize2 >> 2u; |
emh203 | 0:3d9c67d97d6f | 821 | |
emh203 | 0:3d9c67d97d6f | 822 | while(blkCnt > 0u) |
emh203 | 0:3d9c67d97d6f | 823 | { |
emh203 | 0:3d9c67d97d6f | 824 | /* Set all accumulators to zero */ |
emh203 | 0:3d9c67d97d6f | 825 | acc0 = 0; |
emh203 | 0:3d9c67d97d6f | 826 | acc1 = 0; |
emh203 | 0:3d9c67d97d6f | 827 | acc2 = 0; |
emh203 | 0:3d9c67d97d6f | 828 | acc3 = 0; |
emh203 | 0:3d9c67d97d6f | 829 | |
emh203 | 0:3d9c67d97d6f | 830 | /* read x[0], x[1], x[2] samples */ |
emh203 | 0:3d9c67d97d6f | 831 | a = *px; |
emh203 | 0:3d9c67d97d6f | 832 | b = *(px + 1); |
emh203 | 0:3d9c67d97d6f | 833 | |
emh203 | 0:3d9c67d97d6f | 834 | #ifndef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 835 | |
emh203 | 0:3d9c67d97d6f | 836 | x0 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 837 | a = *(px + 2); |
emh203 | 0:3d9c67d97d6f | 838 | x1 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 839 | |
emh203 | 0:3d9c67d97d6f | 840 | #else |
emh203 | 0:3d9c67d97d6f | 841 | |
emh203 | 0:3d9c67d97d6f | 842 | x0 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 843 | a = *(px + 2); |
emh203 | 0:3d9c67d97d6f | 844 | x1 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 845 | |
emh203 | 0:3d9c67d97d6f | 846 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 847 | |
emh203 | 0:3d9c67d97d6f | 848 | px += 2u; |
emh203 | 0:3d9c67d97d6f | 849 | |
emh203 | 0:3d9c67d97d6f | 850 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emh203 | 0:3d9c67d97d6f | 851 | k = srcBLen >> 2u; |
emh203 | 0:3d9c67d97d6f | 852 | |
emh203 | 0:3d9c67d97d6f | 853 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
emh203 | 0:3d9c67d97d6f | 854 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
emh203 | 0:3d9c67d97d6f | 855 | do |
emh203 | 0:3d9c67d97d6f | 856 | { |
emh203 | 0:3d9c67d97d6f | 857 | /* Read the first two inputB samples using SIMD: |
emh203 | 0:3d9c67d97d6f | 858 | * y[0] and y[1] */ |
emh203 | 0:3d9c67d97d6f | 859 | a = *py; |
emh203 | 0:3d9c67d97d6f | 860 | b = *(py + 1); |
emh203 | 0:3d9c67d97d6f | 861 | |
emh203 | 0:3d9c67d97d6f | 862 | #ifndef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 863 | |
emh203 | 0:3d9c67d97d6f | 864 | c0 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 865 | |
emh203 | 0:3d9c67d97d6f | 866 | #else |
emh203 | 0:3d9c67d97d6f | 867 | |
emh203 | 0:3d9c67d97d6f | 868 | c0 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 869 | |
emh203 | 0:3d9c67d97d6f | 870 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 871 | |
emh203 | 0:3d9c67d97d6f | 872 | /* acc0 += x[0] * y[0] + x[1] * y[1] */ |
emh203 | 0:3d9c67d97d6f | 873 | acc0 = __SMLAD(x0, c0, acc0); |
emh203 | 0:3d9c67d97d6f | 874 | |
emh203 | 0:3d9c67d97d6f | 875 | /* acc1 += x[1] * y[0] + x[2] * y[1] */ |
emh203 | 0:3d9c67d97d6f | 876 | acc1 = __SMLAD(x1, c0, acc1); |
emh203 | 0:3d9c67d97d6f | 877 | |
emh203 | 0:3d9c67d97d6f | 878 | /* Read x[2], x[3], x[4] */ |
emh203 | 0:3d9c67d97d6f | 879 | a = *px; |
emh203 | 0:3d9c67d97d6f | 880 | b = *(px + 1); |
emh203 | 0:3d9c67d97d6f | 881 | |
emh203 | 0:3d9c67d97d6f | 882 | #ifndef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 883 | |
emh203 | 0:3d9c67d97d6f | 884 | x2 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 885 | a = *(px + 2); |
emh203 | 0:3d9c67d97d6f | 886 | x3 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 887 | |
emh203 | 0:3d9c67d97d6f | 888 | #else |
emh203 | 0:3d9c67d97d6f | 889 | |
emh203 | 0:3d9c67d97d6f | 890 | x2 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 891 | a = *(px + 2); |
emh203 | 0:3d9c67d97d6f | 892 | x3 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 893 | |
emh203 | 0:3d9c67d97d6f | 894 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 895 | |
emh203 | 0:3d9c67d97d6f | 896 | /* acc2 += x[2] * y[0] + x[3] * y[1] */ |
emh203 | 0:3d9c67d97d6f | 897 | acc2 = __SMLAD(x2, c0, acc2); |
emh203 | 0:3d9c67d97d6f | 898 | |
emh203 | 0:3d9c67d97d6f | 899 | /* acc3 += x[3] * y[0] + x[4] * y[1] */ |
emh203 | 0:3d9c67d97d6f | 900 | acc3 = __SMLAD(x3, c0, acc3); |
emh203 | 0:3d9c67d97d6f | 901 | |
emh203 | 0:3d9c67d97d6f | 902 | /* Read y[2] and y[3] */ |
emh203 | 0:3d9c67d97d6f | 903 | a = *(py + 2); |
emh203 | 0:3d9c67d97d6f | 904 | b = *(py + 3); |
emh203 | 0:3d9c67d97d6f | 905 | |
emh203 | 0:3d9c67d97d6f | 906 | py += 4u; |
emh203 | 0:3d9c67d97d6f | 907 | |
emh203 | 0:3d9c67d97d6f | 908 | #ifndef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 909 | |
emh203 | 0:3d9c67d97d6f | 910 | c0 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 911 | |
emh203 | 0:3d9c67d97d6f | 912 | #else |
emh203 | 0:3d9c67d97d6f | 913 | |
emh203 | 0:3d9c67d97d6f | 914 | c0 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 915 | |
emh203 | 0:3d9c67d97d6f | 916 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 917 | |
emh203 | 0:3d9c67d97d6f | 918 | /* acc0 += x[2] * y[2] + x[3] * y[3] */ |
emh203 | 0:3d9c67d97d6f | 919 | acc0 = __SMLAD(x2, c0, acc0); |
emh203 | 0:3d9c67d97d6f | 920 | |
emh203 | 0:3d9c67d97d6f | 921 | /* acc1 += x[3] * y[2] + x[4] * y[3] */ |
emh203 | 0:3d9c67d97d6f | 922 | acc1 = __SMLAD(x3, c0, acc1); |
emh203 | 0:3d9c67d97d6f | 923 | |
emh203 | 0:3d9c67d97d6f | 924 | /* Read x[4], x[5], x[6] */ |
emh203 | 0:3d9c67d97d6f | 925 | a = *(px + 2); |
emh203 | 0:3d9c67d97d6f | 926 | b = *(px + 3); |
emh203 | 0:3d9c67d97d6f | 927 | |
emh203 | 0:3d9c67d97d6f | 928 | #ifndef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 929 | |
emh203 | 0:3d9c67d97d6f | 930 | x0 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 931 | a = *(px + 4); |
emh203 | 0:3d9c67d97d6f | 932 | x1 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 933 | |
emh203 | 0:3d9c67d97d6f | 934 | #else |
emh203 | 0:3d9c67d97d6f | 935 | |
emh203 | 0:3d9c67d97d6f | 936 | x0 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 937 | a = *(px + 4); |
emh203 | 0:3d9c67d97d6f | 938 | x1 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 939 | |
emh203 | 0:3d9c67d97d6f | 940 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 941 | |
emh203 | 0:3d9c67d97d6f | 942 | px += 4u; |
emh203 | 0:3d9c67d97d6f | 943 | |
emh203 | 0:3d9c67d97d6f | 944 | /* acc2 += x[4] * y[2] + x[5] * y[3] */ |
emh203 | 0:3d9c67d97d6f | 945 | acc2 = __SMLAD(x0, c0, acc2); |
emh203 | 0:3d9c67d97d6f | 946 | |
emh203 | 0:3d9c67d97d6f | 947 | /* acc3 += x[5] * y[2] + x[6] * y[3] */ |
emh203 | 0:3d9c67d97d6f | 948 | acc3 = __SMLAD(x1, c0, acc3); |
emh203 | 0:3d9c67d97d6f | 949 | |
emh203 | 0:3d9c67d97d6f | 950 | } while(--k); |
emh203 | 0:3d9c67d97d6f | 951 | |
emh203 | 0:3d9c67d97d6f | 952 | /* For the next MAC operations, SIMD is not used |
emh203 | 0:3d9c67d97d6f | 953 | * So, the 16 bit pointer if inputB, py is updated */ |
emh203 | 0:3d9c67d97d6f | 954 | |
emh203 | 0:3d9c67d97d6f | 955 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
emh203 | 0:3d9c67d97d6f | 956 | ** No loop unrolling is used. */ |
emh203 | 0:3d9c67d97d6f | 957 | k = srcBLen % 0x4u; |
emh203 | 0:3d9c67d97d6f | 958 | |
emh203 | 0:3d9c67d97d6f | 959 | if(k == 1u) |
emh203 | 0:3d9c67d97d6f | 960 | { |
emh203 | 0:3d9c67d97d6f | 961 | /* Read y[4] */ |
emh203 | 0:3d9c67d97d6f | 962 | c0 = *py; |
emh203 | 0:3d9c67d97d6f | 963 | #ifdef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 964 | |
emh203 | 0:3d9c67d97d6f | 965 | c0 = c0 << 16u; |
emh203 | 0:3d9c67d97d6f | 966 | |
emh203 | 0:3d9c67d97d6f | 967 | #else |
emh203 | 0:3d9c67d97d6f | 968 | |
emh203 | 0:3d9c67d97d6f | 969 | c0 = c0 & 0x0000FFFF; |
emh203 | 0:3d9c67d97d6f | 970 | |
emh203 | 0:3d9c67d97d6f | 971 | #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 972 | |
emh203 | 0:3d9c67d97d6f | 973 | /* Read x[7] */ |
emh203 | 0:3d9c67d97d6f | 974 | a = *px; |
emh203 | 0:3d9c67d97d6f | 975 | b = *(px + 1); |
emh203 | 0:3d9c67d97d6f | 976 | |
emh203 | 0:3d9c67d97d6f | 977 | px++;; |
emh203 | 0:3d9c67d97d6f | 978 | |
emh203 | 0:3d9c67d97d6f | 979 | #ifndef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 980 | |
emh203 | 0:3d9c67d97d6f | 981 | x3 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 982 | |
emh203 | 0:3d9c67d97d6f | 983 | #else |
emh203 | 0:3d9c67d97d6f | 984 | |
emh203 | 0:3d9c67d97d6f | 985 | x3 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 986 | |
emh203 | 0:3d9c67d97d6f | 987 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 988 | |
emh203 | 0:3d9c67d97d6f | 989 | px++; |
emh203 | 0:3d9c67d97d6f | 990 | |
emh203 | 0:3d9c67d97d6f | 991 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 992 | acc0 = __SMLAD(x0, c0, acc0); |
emh203 | 0:3d9c67d97d6f | 993 | acc1 = __SMLAD(x1, c0, acc1); |
emh203 | 0:3d9c67d97d6f | 994 | acc2 = __SMLADX(x1, c0, acc2); |
emh203 | 0:3d9c67d97d6f | 995 | acc3 = __SMLADX(x3, c0, acc3); |
emh203 | 0:3d9c67d97d6f | 996 | } |
emh203 | 0:3d9c67d97d6f | 997 | |
emh203 | 0:3d9c67d97d6f | 998 | if(k == 2u) |
emh203 | 0:3d9c67d97d6f | 999 | { |
emh203 | 0:3d9c67d97d6f | 1000 | /* Read y[4], y[5] */ |
emh203 | 0:3d9c67d97d6f | 1001 | a = *py; |
emh203 | 0:3d9c67d97d6f | 1002 | b = *(py + 1); |
emh203 | 0:3d9c67d97d6f | 1003 | |
emh203 | 0:3d9c67d97d6f | 1004 | #ifndef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 1005 | |
emh203 | 0:3d9c67d97d6f | 1006 | c0 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 1007 | |
emh203 | 0:3d9c67d97d6f | 1008 | #else |
emh203 | 0:3d9c67d97d6f | 1009 | |
emh203 | 0:3d9c67d97d6f | 1010 | c0 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 1011 | |
emh203 | 0:3d9c67d97d6f | 1012 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 1013 | |
emh203 | 0:3d9c67d97d6f | 1014 | /* Read x[7], x[8], x[9] */ |
emh203 | 0:3d9c67d97d6f | 1015 | a = *px; |
emh203 | 0:3d9c67d97d6f | 1016 | b = *(px + 1); |
emh203 | 0:3d9c67d97d6f | 1017 | |
emh203 | 0:3d9c67d97d6f | 1018 | #ifndef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 1019 | |
emh203 | 0:3d9c67d97d6f | 1020 | x3 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 1021 | a = *(px + 2); |
emh203 | 0:3d9c67d97d6f | 1022 | x2 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 1023 | |
emh203 | 0:3d9c67d97d6f | 1024 | #else |
emh203 | 0:3d9c67d97d6f | 1025 | |
emh203 | 0:3d9c67d97d6f | 1026 | x3 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 1027 | a = *(px + 2); |
emh203 | 0:3d9c67d97d6f | 1028 | x2 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 1029 | |
emh203 | 0:3d9c67d97d6f | 1030 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 1031 | |
emh203 | 0:3d9c67d97d6f | 1032 | px += 2u; |
emh203 | 0:3d9c67d97d6f | 1033 | |
emh203 | 0:3d9c67d97d6f | 1034 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 1035 | acc0 = __SMLAD(x0, c0, acc0); |
emh203 | 0:3d9c67d97d6f | 1036 | acc1 = __SMLAD(x1, c0, acc1); |
emh203 | 0:3d9c67d97d6f | 1037 | acc2 = __SMLAD(x3, c0, acc2); |
emh203 | 0:3d9c67d97d6f | 1038 | acc3 = __SMLAD(x2, c0, acc3); |
emh203 | 0:3d9c67d97d6f | 1039 | } |
emh203 | 0:3d9c67d97d6f | 1040 | |
emh203 | 0:3d9c67d97d6f | 1041 | if(k == 3u) |
emh203 | 0:3d9c67d97d6f | 1042 | { |
emh203 | 0:3d9c67d97d6f | 1043 | /* Read y[4], y[5] */ |
emh203 | 0:3d9c67d97d6f | 1044 | a = *py; |
emh203 | 0:3d9c67d97d6f | 1045 | b = *(py + 1); |
emh203 | 0:3d9c67d97d6f | 1046 | |
emh203 | 0:3d9c67d97d6f | 1047 | #ifndef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 1048 | |
emh203 | 0:3d9c67d97d6f | 1049 | c0 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 1050 | |
emh203 | 0:3d9c67d97d6f | 1051 | #else |
emh203 | 0:3d9c67d97d6f | 1052 | |
emh203 | 0:3d9c67d97d6f | 1053 | c0 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 1054 | |
emh203 | 0:3d9c67d97d6f | 1055 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 1056 | |
emh203 | 0:3d9c67d97d6f | 1057 | py += 2u; |
emh203 | 0:3d9c67d97d6f | 1058 | |
emh203 | 0:3d9c67d97d6f | 1059 | /* Read x[7], x[8], x[9] */ |
emh203 | 0:3d9c67d97d6f | 1060 | a = *px; |
emh203 | 0:3d9c67d97d6f | 1061 | b = *(px + 1); |
emh203 | 0:3d9c67d97d6f | 1062 | |
emh203 | 0:3d9c67d97d6f | 1063 | #ifndef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 1064 | |
emh203 | 0:3d9c67d97d6f | 1065 | x3 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 1066 | a = *(px + 2); |
emh203 | 0:3d9c67d97d6f | 1067 | x2 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 1068 | |
emh203 | 0:3d9c67d97d6f | 1069 | #else |
emh203 | 0:3d9c67d97d6f | 1070 | |
emh203 | 0:3d9c67d97d6f | 1071 | x3 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 1072 | a = *(px + 2); |
emh203 | 0:3d9c67d97d6f | 1073 | x2 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 1074 | |
emh203 | 0:3d9c67d97d6f | 1075 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 1076 | |
emh203 | 0:3d9c67d97d6f | 1077 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 1078 | acc0 = __SMLAD(x0, c0, acc0); |
emh203 | 0:3d9c67d97d6f | 1079 | acc1 = __SMLAD(x1, c0, acc1); |
emh203 | 0:3d9c67d97d6f | 1080 | acc2 = __SMLAD(x3, c0, acc2); |
emh203 | 0:3d9c67d97d6f | 1081 | acc3 = __SMLAD(x2, c0, acc3); |
emh203 | 0:3d9c67d97d6f | 1082 | |
emh203 | 0:3d9c67d97d6f | 1083 | c0 = (*py); |
emh203 | 0:3d9c67d97d6f | 1084 | /* Read y[6] */ |
emh203 | 0:3d9c67d97d6f | 1085 | #ifdef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 1086 | |
emh203 | 0:3d9c67d97d6f | 1087 | c0 = c0 << 16u; |
emh203 | 0:3d9c67d97d6f | 1088 | #else |
emh203 | 0:3d9c67d97d6f | 1089 | |
emh203 | 0:3d9c67d97d6f | 1090 | c0 = c0 & 0x0000FFFF; |
emh203 | 0:3d9c67d97d6f | 1091 | #endif /* #ifdef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 1092 | |
emh203 | 0:3d9c67d97d6f | 1093 | /* Read x[10] */ |
emh203 | 0:3d9c67d97d6f | 1094 | b = *(px + 3); |
emh203 | 0:3d9c67d97d6f | 1095 | |
emh203 | 0:3d9c67d97d6f | 1096 | #ifndef ARM_MATH_BIG_ENDIAN |
emh203 | 0:3d9c67d97d6f | 1097 | |
emh203 | 0:3d9c67d97d6f | 1098 | x3 = __PKHBT(a, b, 16); |
emh203 | 0:3d9c67d97d6f | 1099 | |
emh203 | 0:3d9c67d97d6f | 1100 | #else |
emh203 | 0:3d9c67d97d6f | 1101 | |
emh203 | 0:3d9c67d97d6f | 1102 | x3 = __PKHBT(b, a, 16); |
emh203 | 0:3d9c67d97d6f | 1103 | |
emh203 | 0:3d9c67d97d6f | 1104 | #endif /* #ifndef ARM_MATH_BIG_ENDIAN */ |
emh203 | 0:3d9c67d97d6f | 1105 | |
emh203 | 0:3d9c67d97d6f | 1106 | px += 3u; |
emh203 | 0:3d9c67d97d6f | 1107 | |
emh203 | 0:3d9c67d97d6f | 1108 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 1109 | acc0 = __SMLADX(x1, c0, acc0); |
emh203 | 0:3d9c67d97d6f | 1110 | acc1 = __SMLAD(x2, c0, acc1); |
emh203 | 0:3d9c67d97d6f | 1111 | acc2 = __SMLADX(x2, c0, acc2); |
emh203 | 0:3d9c67d97d6f | 1112 | acc3 = __SMLADX(x3, c0, acc3); |
emh203 | 0:3d9c67d97d6f | 1113 | } |
emh203 | 0:3d9c67d97d6f | 1114 | |
emh203 | 0:3d9c67d97d6f | 1115 | /* Store the result in the accumulator in the destination buffer. */ |
emh203 | 0:3d9c67d97d6f | 1116 | *pOut = (q15_t) (acc0 >> 15); |
emh203 | 0:3d9c67d97d6f | 1117 | /* Destination pointer is updated according to the address modifier, inc */ |
emh203 | 0:3d9c67d97d6f | 1118 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 1119 | |
emh203 | 0:3d9c67d97d6f | 1120 | *pOut = (q15_t) (acc1 >> 15); |
emh203 | 0:3d9c67d97d6f | 1121 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 1122 | |
emh203 | 0:3d9c67d97d6f | 1123 | *pOut = (q15_t) (acc2 >> 15); |
emh203 | 0:3d9c67d97d6f | 1124 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 1125 | |
emh203 | 0:3d9c67d97d6f | 1126 | *pOut = (q15_t) (acc3 >> 15); |
emh203 | 0:3d9c67d97d6f | 1127 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 1128 | |
emh203 | 0:3d9c67d97d6f | 1129 | /* Increment the pointer pIn1 index, count by 1 */ |
emh203 | 0:3d9c67d97d6f | 1130 | count += 4u; |
emh203 | 0:3d9c67d97d6f | 1131 | |
emh203 | 0:3d9c67d97d6f | 1132 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emh203 | 0:3d9c67d97d6f | 1133 | px = pIn1 + count; |
emh203 | 0:3d9c67d97d6f | 1134 | py = pIn2; |
emh203 | 0:3d9c67d97d6f | 1135 | |
emh203 | 0:3d9c67d97d6f | 1136 | |
emh203 | 0:3d9c67d97d6f | 1137 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 1138 | blkCnt--; |
emh203 | 0:3d9c67d97d6f | 1139 | } |
emh203 | 0:3d9c67d97d6f | 1140 | |
emh203 | 0:3d9c67d97d6f | 1141 | /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here. |
emh203 | 0:3d9c67d97d6f | 1142 | ** No loop unrolling is used. */ |
emh203 | 0:3d9c67d97d6f | 1143 | blkCnt = blockSize2 % 0x4u; |
emh203 | 0:3d9c67d97d6f | 1144 | |
emh203 | 0:3d9c67d97d6f | 1145 | while(blkCnt > 0u) |
emh203 | 0:3d9c67d97d6f | 1146 | { |
emh203 | 0:3d9c67d97d6f | 1147 | /* Accumulator is made zero for every iteration */ |
emh203 | 0:3d9c67d97d6f | 1148 | sum = 0; |
emh203 | 0:3d9c67d97d6f | 1149 | |
emh203 | 0:3d9c67d97d6f | 1150 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emh203 | 0:3d9c67d97d6f | 1151 | k = srcBLen >> 2u; |
emh203 | 0:3d9c67d97d6f | 1152 | |
emh203 | 0:3d9c67d97d6f | 1153 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
emh203 | 0:3d9c67d97d6f | 1154 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
emh203 | 0:3d9c67d97d6f | 1155 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 1156 | { |
emh203 | 0:3d9c67d97d6f | 1157 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 1158 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 1159 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 1160 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 1161 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 1162 | |
emh203 | 0:3d9c67d97d6f | 1163 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 1164 | k--; |
emh203 | 0:3d9c67d97d6f | 1165 | } |
emh203 | 0:3d9c67d97d6f | 1166 | |
emh203 | 0:3d9c67d97d6f | 1167 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
emh203 | 0:3d9c67d97d6f | 1168 | ** No loop unrolling is used. */ |
emh203 | 0:3d9c67d97d6f | 1169 | k = srcBLen % 0x4u; |
emh203 | 0:3d9c67d97d6f | 1170 | |
emh203 | 0:3d9c67d97d6f | 1171 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 1172 | { |
emh203 | 0:3d9c67d97d6f | 1173 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 1174 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 1175 | |
emh203 | 0:3d9c67d97d6f | 1176 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 1177 | k--; |
emh203 | 0:3d9c67d97d6f | 1178 | } |
emh203 | 0:3d9c67d97d6f | 1179 | |
emh203 | 0:3d9c67d97d6f | 1180 | /* Store the result in the accumulator in the destination buffer. */ |
emh203 | 0:3d9c67d97d6f | 1181 | *pOut = (q15_t) (sum >> 15); |
emh203 | 0:3d9c67d97d6f | 1182 | /* Destination pointer is updated according to the address modifier, inc */ |
emh203 | 0:3d9c67d97d6f | 1183 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 1184 | |
emh203 | 0:3d9c67d97d6f | 1185 | /* Increment the pointer pIn1 index, count by 1 */ |
emh203 | 0:3d9c67d97d6f | 1186 | count++; |
emh203 | 0:3d9c67d97d6f | 1187 | |
emh203 | 0:3d9c67d97d6f | 1188 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emh203 | 0:3d9c67d97d6f | 1189 | px = pIn1 + count; |
emh203 | 0:3d9c67d97d6f | 1190 | py = pIn2; |
emh203 | 0:3d9c67d97d6f | 1191 | |
emh203 | 0:3d9c67d97d6f | 1192 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 1193 | blkCnt--; |
emh203 | 0:3d9c67d97d6f | 1194 | } |
emh203 | 0:3d9c67d97d6f | 1195 | } |
emh203 | 0:3d9c67d97d6f | 1196 | else |
emh203 | 0:3d9c67d97d6f | 1197 | { |
emh203 | 0:3d9c67d97d6f | 1198 | /* If the srcBLen is not a multiple of 4, |
emh203 | 0:3d9c67d97d6f | 1199 | * the blockSize2 loop cannot be unrolled by 4 */ |
emh203 | 0:3d9c67d97d6f | 1200 | blkCnt = blockSize2; |
emh203 | 0:3d9c67d97d6f | 1201 | |
emh203 | 0:3d9c67d97d6f | 1202 | while(blkCnt > 0u) |
emh203 | 0:3d9c67d97d6f | 1203 | { |
emh203 | 0:3d9c67d97d6f | 1204 | /* Accumulator is made zero for every iteration */ |
emh203 | 0:3d9c67d97d6f | 1205 | sum = 0; |
emh203 | 0:3d9c67d97d6f | 1206 | |
emh203 | 0:3d9c67d97d6f | 1207 | /* Loop over srcBLen */ |
emh203 | 0:3d9c67d97d6f | 1208 | k = srcBLen; |
emh203 | 0:3d9c67d97d6f | 1209 | |
emh203 | 0:3d9c67d97d6f | 1210 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 1211 | { |
emh203 | 0:3d9c67d97d6f | 1212 | /* Perform the multiply-accumulate */ |
emh203 | 0:3d9c67d97d6f | 1213 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 1214 | |
emh203 | 0:3d9c67d97d6f | 1215 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 1216 | k--; |
emh203 | 0:3d9c67d97d6f | 1217 | } |
emh203 | 0:3d9c67d97d6f | 1218 | |
emh203 | 0:3d9c67d97d6f | 1219 | /* Store the result in the accumulator in the destination buffer. */ |
emh203 | 0:3d9c67d97d6f | 1220 | *pOut = (q15_t) (sum >> 15); |
emh203 | 0:3d9c67d97d6f | 1221 | /* Destination pointer is updated according to the address modifier, inc */ |
emh203 | 0:3d9c67d97d6f | 1222 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 1223 | |
emh203 | 0:3d9c67d97d6f | 1224 | /* Increment the MAC count */ |
emh203 | 0:3d9c67d97d6f | 1225 | count++; |
emh203 | 0:3d9c67d97d6f | 1226 | |
emh203 | 0:3d9c67d97d6f | 1227 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emh203 | 0:3d9c67d97d6f | 1228 | px = pIn1 + count; |
emh203 | 0:3d9c67d97d6f | 1229 | py = pIn2; |
emh203 | 0:3d9c67d97d6f | 1230 | |
emh203 | 0:3d9c67d97d6f | 1231 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 1232 | blkCnt--; |
emh203 | 0:3d9c67d97d6f | 1233 | } |
emh203 | 0:3d9c67d97d6f | 1234 | } |
emh203 | 0:3d9c67d97d6f | 1235 | |
emh203 | 0:3d9c67d97d6f | 1236 | /* -------------------------- |
emh203 | 0:3d9c67d97d6f | 1237 | * Initializations of stage3 |
emh203 | 0:3d9c67d97d6f | 1238 | * -------------------------*/ |
emh203 | 0:3d9c67d97d6f | 1239 | |
emh203 | 0:3d9c67d97d6f | 1240 | /* sum += x[srcALen-srcBLen+1] * y[0] + x[srcALen-srcBLen+2] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
emh203 | 0:3d9c67d97d6f | 1241 | * sum += x[srcALen-srcBLen+2] * y[0] + x[srcALen-srcBLen+3] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
emh203 | 0:3d9c67d97d6f | 1242 | * .... |
emh203 | 0:3d9c67d97d6f | 1243 | * sum += x[srcALen-2] * y[0] + x[srcALen-1] * y[1] |
emh203 | 0:3d9c67d97d6f | 1244 | * sum += x[srcALen-1] * y[0] |
emh203 | 0:3d9c67d97d6f | 1245 | */ |
emh203 | 0:3d9c67d97d6f | 1246 | |
emh203 | 0:3d9c67d97d6f | 1247 | /* In this stage the MAC operations are decreased by 1 for every iteration. |
emh203 | 0:3d9c67d97d6f | 1248 | The count variable holds the number of MAC operations performed */ |
emh203 | 0:3d9c67d97d6f | 1249 | count = srcBLen - 1u; |
emh203 | 0:3d9c67d97d6f | 1250 | |
emh203 | 0:3d9c67d97d6f | 1251 | /* Working pointer of inputA */ |
emh203 | 0:3d9c67d97d6f | 1252 | pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u); |
emh203 | 0:3d9c67d97d6f | 1253 | px = pSrc1; |
emh203 | 0:3d9c67d97d6f | 1254 | |
emh203 | 0:3d9c67d97d6f | 1255 | /* Working pointer of inputB */ |
emh203 | 0:3d9c67d97d6f | 1256 | py = pIn2; |
emh203 | 0:3d9c67d97d6f | 1257 | |
emh203 | 0:3d9c67d97d6f | 1258 | /* ------------------- |
emh203 | 0:3d9c67d97d6f | 1259 | * Stage3 process |
emh203 | 0:3d9c67d97d6f | 1260 | * ------------------*/ |
emh203 | 0:3d9c67d97d6f | 1261 | |
emh203 | 0:3d9c67d97d6f | 1262 | while(blockSize3 > 0u) |
emh203 | 0:3d9c67d97d6f | 1263 | { |
emh203 | 0:3d9c67d97d6f | 1264 | /* Accumulator is made zero for every iteration */ |
emh203 | 0:3d9c67d97d6f | 1265 | sum = 0; |
emh203 | 0:3d9c67d97d6f | 1266 | |
emh203 | 0:3d9c67d97d6f | 1267 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
emh203 | 0:3d9c67d97d6f | 1268 | k = count >> 2u; |
emh203 | 0:3d9c67d97d6f | 1269 | |
emh203 | 0:3d9c67d97d6f | 1270 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
emh203 | 0:3d9c67d97d6f | 1271 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
emh203 | 0:3d9c67d97d6f | 1272 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 1273 | { |
emh203 | 0:3d9c67d97d6f | 1274 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 1275 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 1276 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 1277 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 1278 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 1279 | |
emh203 | 0:3d9c67d97d6f | 1280 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 1281 | k--; |
emh203 | 0:3d9c67d97d6f | 1282 | } |
emh203 | 0:3d9c67d97d6f | 1283 | |
emh203 | 0:3d9c67d97d6f | 1284 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
emh203 | 0:3d9c67d97d6f | 1285 | ** No loop unrolling is used. */ |
emh203 | 0:3d9c67d97d6f | 1286 | k = count % 0x4u; |
emh203 | 0:3d9c67d97d6f | 1287 | |
emh203 | 0:3d9c67d97d6f | 1288 | while(k > 0u) |
emh203 | 0:3d9c67d97d6f | 1289 | { |
emh203 | 0:3d9c67d97d6f | 1290 | /* Perform the multiply-accumulates */ |
emh203 | 0:3d9c67d97d6f | 1291 | sum += ((q31_t) * px++ * *py++); |
emh203 | 0:3d9c67d97d6f | 1292 | |
emh203 | 0:3d9c67d97d6f | 1293 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 1294 | k--; |
emh203 | 0:3d9c67d97d6f | 1295 | } |
emh203 | 0:3d9c67d97d6f | 1296 | |
emh203 | 0:3d9c67d97d6f | 1297 | /* Store the result in the accumulator in the destination buffer. */ |
emh203 | 0:3d9c67d97d6f | 1298 | *pOut = (q15_t) (sum >> 15); |
emh203 | 0:3d9c67d97d6f | 1299 | /* Destination pointer is updated according to the address modifier, inc */ |
emh203 | 0:3d9c67d97d6f | 1300 | pOut += inc; |
emh203 | 0:3d9c67d97d6f | 1301 | |
emh203 | 0:3d9c67d97d6f | 1302 | /* Update the inputA and inputB pointers for next MAC calculation */ |
emh203 | 0:3d9c67d97d6f | 1303 | px = ++pSrc1; |
emh203 | 0:3d9c67d97d6f | 1304 | py = pIn2; |
emh203 | 0:3d9c67d97d6f | 1305 | |
emh203 | 0:3d9c67d97d6f | 1306 | /* Decrement the MAC count */ |
emh203 | 0:3d9c67d97d6f | 1307 | count--; |
emh203 | 0:3d9c67d97d6f | 1308 | |
emh203 | 0:3d9c67d97d6f | 1309 | /* Decrement the loop counter */ |
emh203 | 0:3d9c67d97d6f | 1310 | blockSize3--; |
emh203 | 0:3d9c67d97d6f | 1311 | } |
emh203 | 0:3d9c67d97d6f | 1312 | |
emh203 | 0:3d9c67d97d6f | 1313 | #endif /* #ifndef UNALIGNED_SUPPORT_DISABLE */ |
emh203 | 0:3d9c67d97d6f | 1314 | |
emh203 | 0:3d9c67d97d6f | 1315 | } |
emh203 | 0:3d9c67d97d6f | 1316 | |
emh203 | 0:3d9c67d97d6f | 1317 | /** |
emh203 | 0:3d9c67d97d6f | 1318 | * @} end of Corr group |
emh203 | 0:3d9c67d97d6f | 1319 | */ |