CMSIS DSP Library from CMSIS 2.0. See http://www.onarm.com/cmsis/ for full details
Dependents: K22F_DSP_Matrix_least_square BNO055-ELEC3810 1BNO055 ECE4180Project--Slave2 ... more
src/Cortex-M4-M3/FilteringFunctions/arm_correlate_q15.c@0:1014af42efd9, 2011-03-10 (annotated)
- Committer:
- simon
- Date:
- Thu Mar 10 15:07:50 2011 +0000
- Revision:
- 0:1014af42efd9
Who changed what in which revision?
User | Revision | Line number | New contents of line |
---|---|---|---|
simon | 0:1014af42efd9 | 1 | /* ---------------------------------------------------------------------- |
simon | 0:1014af42efd9 | 2 | * Copyright (C) 2010 ARM Limited. All rights reserved. |
simon | 0:1014af42efd9 | 3 | * |
simon | 0:1014af42efd9 | 4 | * $Date: 29. November 2010 |
simon | 0:1014af42efd9 | 5 | * $Revision: V1.0.3 |
simon | 0:1014af42efd9 | 6 | * |
simon | 0:1014af42efd9 | 7 | * Project: CMSIS DSP Library |
simon | 0:1014af42efd9 | 8 | * Title: arm_correlate_q15.c |
simon | 0:1014af42efd9 | 9 | * |
simon | 0:1014af42efd9 | 10 | * Description: Q15 Correlation. |
simon | 0:1014af42efd9 | 11 | * |
simon | 0:1014af42efd9 | 12 | * Target Processor: Cortex-M4/Cortex-M3 |
simon | 0:1014af42efd9 | 13 | * |
simon | 0:1014af42efd9 | 14 | * Version 1.0.3 2010/11/29 |
simon | 0:1014af42efd9 | 15 | * Re-organized the CMSIS folders and updated documentation. |
simon | 0:1014af42efd9 | 16 | * |
simon | 0:1014af42efd9 | 17 | * Version 1.0.2 2010/11/11 |
simon | 0:1014af42efd9 | 18 | * Documentation updated. |
simon | 0:1014af42efd9 | 19 | * |
simon | 0:1014af42efd9 | 20 | * Version 1.0.1 2010/10/05 |
simon | 0:1014af42efd9 | 21 | * Production release and review comments incorporated. |
simon | 0:1014af42efd9 | 22 | * |
simon | 0:1014af42efd9 | 23 | * Version 1.0.0 2010/09/20 |
simon | 0:1014af42efd9 | 24 | * Production release and review comments incorporated |
simon | 0:1014af42efd9 | 25 | * |
simon | 0:1014af42efd9 | 26 | * Version 0.0.7 2010/06/10 |
simon | 0:1014af42efd9 | 27 | * Misra-C changes done |
simon | 0:1014af42efd9 | 28 | * |
simon | 0:1014af42efd9 | 29 | * -------------------------------------------------------------------- */ |
simon | 0:1014af42efd9 | 30 | |
simon | 0:1014af42efd9 | 31 | #include "arm_math.h" |
simon | 0:1014af42efd9 | 32 | |
simon | 0:1014af42efd9 | 33 | /** |
simon | 0:1014af42efd9 | 34 | * @ingroup groupFilters |
simon | 0:1014af42efd9 | 35 | */ |
simon | 0:1014af42efd9 | 36 | |
simon | 0:1014af42efd9 | 37 | /** |
simon | 0:1014af42efd9 | 38 | * @addtogroup Corr |
simon | 0:1014af42efd9 | 39 | * @{ |
simon | 0:1014af42efd9 | 40 | */ |
simon | 0:1014af42efd9 | 41 | |
simon | 0:1014af42efd9 | 42 | /** |
simon | 0:1014af42efd9 | 43 | * @brief Correlation of Q15 sequences |
simon | 0:1014af42efd9 | 44 | * @param[in] *pSrcA points to the first input sequence. |
simon | 0:1014af42efd9 | 45 | * @param[in] srcALen length of the first input sequence. |
simon | 0:1014af42efd9 | 46 | * @param[in] *pSrcB points to the second input sequence. |
simon | 0:1014af42efd9 | 47 | * @param[in] srcBLen length of the second input sequence. |
simon | 0:1014af42efd9 | 48 | * @param[out] *pDst points to the location where the output result is written. Length 2 * max(srcALen, srcBLen) - 1. |
simon | 0:1014af42efd9 | 49 | * @return none. |
simon | 0:1014af42efd9 | 50 | * |
simon | 0:1014af42efd9 | 51 | * @details |
simon | 0:1014af42efd9 | 52 | * <b>Scaling and Overflow Behavior:</b> |
simon | 0:1014af42efd9 | 53 | * |
simon | 0:1014af42efd9 | 54 | * \par |
simon | 0:1014af42efd9 | 55 | * The function is implemented using a 64-bit internal accumulator. |
simon | 0:1014af42efd9 | 56 | * Both inputs are in 1.15 format and multiplications yield a 2.30 result. |
simon | 0:1014af42efd9 | 57 | * The 2.30 intermediate results are accumulated in a 64-bit accumulator in 34.30 format. |
simon | 0:1014af42efd9 | 58 | * This approach provides 33 guard bits and there is no risk of overflow. |
simon | 0:1014af42efd9 | 59 | * The 34.30 result is then truncated to 34.15 format by discarding the low 15 bits and then saturated to 1.15 format. |
simon | 0:1014af42efd9 | 60 | * |
simon | 0:1014af42efd9 | 61 | * \par |
simon | 0:1014af42efd9 | 62 | * Refer to <code>arm_correlate_fast_q15()</code> for a faster but less precise version of this function. |
simon | 0:1014af42efd9 | 63 | */ |
simon | 0:1014af42efd9 | 64 | |
simon | 0:1014af42efd9 | 65 | void arm_correlate_q15( |
simon | 0:1014af42efd9 | 66 | q15_t * pSrcA, |
simon | 0:1014af42efd9 | 67 | uint32_t srcALen, |
simon | 0:1014af42efd9 | 68 | q15_t * pSrcB, |
simon | 0:1014af42efd9 | 69 | uint32_t srcBLen, |
simon | 0:1014af42efd9 | 70 | q15_t * pDst) |
simon | 0:1014af42efd9 | 71 | { |
simon | 0:1014af42efd9 | 72 | q15_t *pIn1; /* inputA pointer */ |
simon | 0:1014af42efd9 | 73 | q15_t *pIn2; /* inputB pointer */ |
simon | 0:1014af42efd9 | 74 | q15_t *pOut = pDst; /* output pointer */ |
simon | 0:1014af42efd9 | 75 | q63_t sum, acc0, acc1, acc2, acc3; /* Accumulators */ |
simon | 0:1014af42efd9 | 76 | q15_t *px; /* Intermediate inputA pointer */ |
simon | 0:1014af42efd9 | 77 | q15_t *py; /* Intermediate inputB pointer */ |
simon | 0:1014af42efd9 | 78 | q15_t *pSrc1; /* Intermediate pointers */ |
simon | 0:1014af42efd9 | 79 | q31_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */ |
simon | 0:1014af42efd9 | 80 | uint32_t j, k = 0u, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counter */ |
simon | 0:1014af42efd9 | 81 | int32_t inc = 1; /* Destination address modifier */ |
simon | 0:1014af42efd9 | 82 | q31_t *pb; /* 32 bit pointer for inputB buffer */ |
simon | 0:1014af42efd9 | 83 | |
simon | 0:1014af42efd9 | 84 | |
simon | 0:1014af42efd9 | 85 | /* The algorithm implementation is based on the lengths of the inputs. */ |
simon | 0:1014af42efd9 | 86 | /* srcB is always made to slide across srcA. */ |
simon | 0:1014af42efd9 | 87 | /* So srcBLen is always considered as shorter or equal to srcALen */ |
simon | 0:1014af42efd9 | 88 | /* But CORR(x, y) is reverse of CORR(y, x) */ |
simon | 0:1014af42efd9 | 89 | /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */ |
simon | 0:1014af42efd9 | 90 | /* and the destination pointer modifier, inc is set to -1 */ |
simon | 0:1014af42efd9 | 91 | /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */ |
simon | 0:1014af42efd9 | 92 | /* But to improve the performance, |
simon | 0:1014af42efd9 | 93 | * we include zeroes in the output instead of zero padding either of the the inputs*/ |
simon | 0:1014af42efd9 | 94 | /* If srcALen > srcBLen, |
simon | 0:1014af42efd9 | 95 | * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */ |
simon | 0:1014af42efd9 | 96 | /* If srcALen < srcBLen, |
simon | 0:1014af42efd9 | 97 | * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */ |
simon | 0:1014af42efd9 | 98 | if(srcALen >= srcBLen) |
simon | 0:1014af42efd9 | 99 | { |
simon | 0:1014af42efd9 | 100 | /* Initialization of inputA pointer */ |
simon | 0:1014af42efd9 | 101 | pIn1 = (pSrcA); |
simon | 0:1014af42efd9 | 102 | |
simon | 0:1014af42efd9 | 103 | /* Initialization of inputB pointer */ |
simon | 0:1014af42efd9 | 104 | pIn2 = (pSrcB); |
simon | 0:1014af42efd9 | 105 | |
simon | 0:1014af42efd9 | 106 | /* Number of output samples is calculated */ |
simon | 0:1014af42efd9 | 107 | outBlockSize = (2u * srcALen) - 1u; |
simon | 0:1014af42efd9 | 108 | |
simon | 0:1014af42efd9 | 109 | /* When srcALen > srcBLen, zero padding is done to srcB |
simon | 0:1014af42efd9 | 110 | * to make their lengths equal. |
simon | 0:1014af42efd9 | 111 | * Instead, (outBlockSize - (srcALen + srcBLen - 1)) |
simon | 0:1014af42efd9 | 112 | * number of output samples are made zero */ |
simon | 0:1014af42efd9 | 113 | j = outBlockSize - (srcALen + (srcBLen - 1u)); |
simon | 0:1014af42efd9 | 114 | |
simon | 0:1014af42efd9 | 115 | while(j > 0u) |
simon | 0:1014af42efd9 | 116 | { |
simon | 0:1014af42efd9 | 117 | /* Zero is stored in the destination buffer */ |
simon | 0:1014af42efd9 | 118 | *pOut++ = 0; |
simon | 0:1014af42efd9 | 119 | |
simon | 0:1014af42efd9 | 120 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 121 | j--; |
simon | 0:1014af42efd9 | 122 | } |
simon | 0:1014af42efd9 | 123 | |
simon | 0:1014af42efd9 | 124 | } |
simon | 0:1014af42efd9 | 125 | else |
simon | 0:1014af42efd9 | 126 | { |
simon | 0:1014af42efd9 | 127 | /* Initialization of inputA pointer */ |
simon | 0:1014af42efd9 | 128 | pIn1 = (pSrcB); |
simon | 0:1014af42efd9 | 129 | |
simon | 0:1014af42efd9 | 130 | /* Initialization of inputB pointer */ |
simon | 0:1014af42efd9 | 131 | pIn2 = (pSrcA); |
simon | 0:1014af42efd9 | 132 | |
simon | 0:1014af42efd9 | 133 | /* srcBLen is always considered as shorter or equal to srcALen */ |
simon | 0:1014af42efd9 | 134 | j = srcBLen; |
simon | 0:1014af42efd9 | 135 | srcBLen = srcALen; |
simon | 0:1014af42efd9 | 136 | srcALen = j; |
simon | 0:1014af42efd9 | 137 | |
simon | 0:1014af42efd9 | 138 | /* CORR(x, y) = Reverse order(CORR(y, x)) */ |
simon | 0:1014af42efd9 | 139 | /* Hence set the destination pointer to point to the last output sample */ |
simon | 0:1014af42efd9 | 140 | pOut = pDst + ((srcALen + srcBLen) - 2u); |
simon | 0:1014af42efd9 | 141 | |
simon | 0:1014af42efd9 | 142 | /* Destination address modifier is set to -1 */ |
simon | 0:1014af42efd9 | 143 | inc = -1; |
simon | 0:1014af42efd9 | 144 | |
simon | 0:1014af42efd9 | 145 | } |
simon | 0:1014af42efd9 | 146 | |
simon | 0:1014af42efd9 | 147 | /* The function is internally |
simon | 0:1014af42efd9 | 148 | * divided into three parts according to the number of multiplications that has to be |
simon | 0:1014af42efd9 | 149 | * taken place between inputA samples and inputB samples. In the first part of the |
simon | 0:1014af42efd9 | 150 | * algorithm, the multiplications increase by one for every iteration. |
simon | 0:1014af42efd9 | 151 | * In the second part of the algorithm, srcBLen number of multiplications are done. |
simon | 0:1014af42efd9 | 152 | * In the third part of the algorithm, the multiplications decrease by one |
simon | 0:1014af42efd9 | 153 | * for every iteration.*/ |
simon | 0:1014af42efd9 | 154 | /* The algorithm is implemented in three stages. |
simon | 0:1014af42efd9 | 155 | * The loop counters of each stage is initiated here. */ |
simon | 0:1014af42efd9 | 156 | blockSize1 = srcBLen - 1u; |
simon | 0:1014af42efd9 | 157 | blockSize2 = srcALen - (srcBLen - 1u); |
simon | 0:1014af42efd9 | 158 | blockSize3 = blockSize1; |
simon | 0:1014af42efd9 | 159 | |
simon | 0:1014af42efd9 | 160 | /* -------------------------- |
simon | 0:1014af42efd9 | 161 | * Initializations of stage1 |
simon | 0:1014af42efd9 | 162 | * -------------------------*/ |
simon | 0:1014af42efd9 | 163 | |
simon | 0:1014af42efd9 | 164 | /* sum = x[0] * y[srcBlen - 1] |
simon | 0:1014af42efd9 | 165 | * sum = x[0] * y[srcBlen - 2] + x[1] * y[srcBlen - 1] |
simon | 0:1014af42efd9 | 166 | * .... |
simon | 0:1014af42efd9 | 167 | * sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen - 1] * y[srcBLen - 1] |
simon | 0:1014af42efd9 | 168 | */ |
simon | 0:1014af42efd9 | 169 | |
simon | 0:1014af42efd9 | 170 | /* In this stage the MAC operations are increased by 1 for every iteration. |
simon | 0:1014af42efd9 | 171 | The count variable holds the number of MAC operations performed */ |
simon | 0:1014af42efd9 | 172 | count = 1u; |
simon | 0:1014af42efd9 | 173 | |
simon | 0:1014af42efd9 | 174 | /* Working pointer of inputA */ |
simon | 0:1014af42efd9 | 175 | px = pIn1; |
simon | 0:1014af42efd9 | 176 | |
simon | 0:1014af42efd9 | 177 | /* Working pointer of inputB */ |
simon | 0:1014af42efd9 | 178 | pSrc1 = pIn2 + (srcBLen - 1u); |
simon | 0:1014af42efd9 | 179 | py = pSrc1; |
simon | 0:1014af42efd9 | 180 | |
simon | 0:1014af42efd9 | 181 | /* ------------------------ |
simon | 0:1014af42efd9 | 182 | * Stage1 process |
simon | 0:1014af42efd9 | 183 | * ----------------------*/ |
simon | 0:1014af42efd9 | 184 | |
simon | 0:1014af42efd9 | 185 | /* The first loop starts here */ |
simon | 0:1014af42efd9 | 186 | while(blockSize1 > 0u) |
simon | 0:1014af42efd9 | 187 | { |
simon | 0:1014af42efd9 | 188 | /* Accumulator is made zero for every iteration */ |
simon | 0:1014af42efd9 | 189 | sum = 0; |
simon | 0:1014af42efd9 | 190 | |
simon | 0:1014af42efd9 | 191 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
simon | 0:1014af42efd9 | 192 | k = count >> 2; |
simon | 0:1014af42efd9 | 193 | |
simon | 0:1014af42efd9 | 194 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
simon | 0:1014af42efd9 | 195 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
simon | 0:1014af42efd9 | 196 | while(k > 0u) |
simon | 0:1014af42efd9 | 197 | { |
simon | 0:1014af42efd9 | 198 | /* x[0] * y[srcBLen - 4] , x[1] * y[srcBLen - 3] */ |
simon | 0:1014af42efd9 | 199 | sum = __SMLALD(*__SIMD32(px)++, *__SIMD32(py)++, sum); |
simon | 0:1014af42efd9 | 200 | /* x[3] * y[srcBLen - 1] , x[2] * y[srcBLen - 2] */ |
simon | 0:1014af42efd9 | 201 | sum = __SMLALD(*__SIMD32(px)++, *__SIMD32(py)++, sum); |
simon | 0:1014af42efd9 | 202 | |
simon | 0:1014af42efd9 | 203 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 204 | k--; |
simon | 0:1014af42efd9 | 205 | } |
simon | 0:1014af42efd9 | 206 | |
simon | 0:1014af42efd9 | 207 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
simon | 0:1014af42efd9 | 208 | ** No loop unrolling is used. */ |
simon | 0:1014af42efd9 | 209 | k = count % 0x4u; |
simon | 0:1014af42efd9 | 210 | |
simon | 0:1014af42efd9 | 211 | while(k > 0u) |
simon | 0:1014af42efd9 | 212 | { |
simon | 0:1014af42efd9 | 213 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 214 | /* x[0] * y[srcBLen - 1] */ |
simon | 0:1014af42efd9 | 215 | sum = __SMLALD(*px++, *py++, sum); |
simon | 0:1014af42efd9 | 216 | |
simon | 0:1014af42efd9 | 217 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 218 | k--; |
simon | 0:1014af42efd9 | 219 | } |
simon | 0:1014af42efd9 | 220 | |
simon | 0:1014af42efd9 | 221 | /* Store the result in the accumulator in the destination buffer. */ |
simon | 0:1014af42efd9 | 222 | *pOut = (q15_t) (__SSAT((sum >> 15), 16)); |
simon | 0:1014af42efd9 | 223 | /* Destination pointer is updated according to the address modifier, inc */ |
simon | 0:1014af42efd9 | 224 | pOut += inc; |
simon | 0:1014af42efd9 | 225 | |
simon | 0:1014af42efd9 | 226 | /* Update the inputA and inputB pointers for next MAC calculation */ |
simon | 0:1014af42efd9 | 227 | py = pSrc1 - count; |
simon | 0:1014af42efd9 | 228 | px = pIn1; |
simon | 0:1014af42efd9 | 229 | |
simon | 0:1014af42efd9 | 230 | /* Increment the MAC count */ |
simon | 0:1014af42efd9 | 231 | count++; |
simon | 0:1014af42efd9 | 232 | |
simon | 0:1014af42efd9 | 233 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 234 | blockSize1--; |
simon | 0:1014af42efd9 | 235 | } |
simon | 0:1014af42efd9 | 236 | |
simon | 0:1014af42efd9 | 237 | /* -------------------------- |
simon | 0:1014af42efd9 | 238 | * Initializations of stage2 |
simon | 0:1014af42efd9 | 239 | * ------------------------*/ |
simon | 0:1014af42efd9 | 240 | |
simon | 0:1014af42efd9 | 241 | /* sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen-1] * y[srcBLen-1] |
simon | 0:1014af42efd9 | 242 | * sum = x[1] * y[0] + x[2] * y[1] +...+ x[srcBLen] * y[srcBLen-1] |
simon | 0:1014af42efd9 | 243 | * .... |
simon | 0:1014af42efd9 | 244 | * sum = x[srcALen-srcBLen-2] * y[0] + x[srcALen-srcBLen-1] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
simon | 0:1014af42efd9 | 245 | */ |
simon | 0:1014af42efd9 | 246 | |
simon | 0:1014af42efd9 | 247 | /* Working pointer of inputA */ |
simon | 0:1014af42efd9 | 248 | px = pIn1; |
simon | 0:1014af42efd9 | 249 | |
simon | 0:1014af42efd9 | 250 | /* Working pointer of inputB */ |
simon | 0:1014af42efd9 | 251 | py = pIn2; |
simon | 0:1014af42efd9 | 252 | |
simon | 0:1014af42efd9 | 253 | /* Initialize inputB pointer of type q31 */ |
simon | 0:1014af42efd9 | 254 | pb = (q31_t *) (py); |
simon | 0:1014af42efd9 | 255 | |
simon | 0:1014af42efd9 | 256 | /* count is index by which the pointer pIn1 to be incremented */ |
simon | 0:1014af42efd9 | 257 | count = 0u; |
simon | 0:1014af42efd9 | 258 | |
simon | 0:1014af42efd9 | 259 | /* ------------------- |
simon | 0:1014af42efd9 | 260 | * Stage2 process |
simon | 0:1014af42efd9 | 261 | * ------------------*/ |
simon | 0:1014af42efd9 | 262 | |
simon | 0:1014af42efd9 | 263 | /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed. |
simon | 0:1014af42efd9 | 264 | * So, to loop unroll over blockSize2, |
simon | 0:1014af42efd9 | 265 | * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */ |
simon | 0:1014af42efd9 | 266 | if(srcBLen >= 4u) |
simon | 0:1014af42efd9 | 267 | { |
simon | 0:1014af42efd9 | 268 | /* Loop unroll over blockSize2, by 4 */ |
simon | 0:1014af42efd9 | 269 | blkCnt = blockSize2 >> 2u; |
simon | 0:1014af42efd9 | 270 | |
simon | 0:1014af42efd9 | 271 | while(blkCnt > 0u) |
simon | 0:1014af42efd9 | 272 | { |
simon | 0:1014af42efd9 | 273 | /* Set all accumulators to zero */ |
simon | 0:1014af42efd9 | 274 | acc0 = 0; |
simon | 0:1014af42efd9 | 275 | acc1 = 0; |
simon | 0:1014af42efd9 | 276 | acc2 = 0; |
simon | 0:1014af42efd9 | 277 | acc3 = 0; |
simon | 0:1014af42efd9 | 278 | |
simon | 0:1014af42efd9 | 279 | /* read x[0], x[1] samples */ |
simon | 0:1014af42efd9 | 280 | x0 = *(q31_t *) (px++); |
simon | 0:1014af42efd9 | 281 | /* read x[1], x[2] samples */ |
simon | 0:1014af42efd9 | 282 | x1 = *(q31_t *) (px++); |
simon | 0:1014af42efd9 | 283 | |
simon | 0:1014af42efd9 | 284 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
simon | 0:1014af42efd9 | 285 | k = srcBLen >> 2u; |
simon | 0:1014af42efd9 | 286 | |
simon | 0:1014af42efd9 | 287 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
simon | 0:1014af42efd9 | 288 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
simon | 0:1014af42efd9 | 289 | do |
simon | 0:1014af42efd9 | 290 | { |
simon | 0:1014af42efd9 | 291 | /* Read the first two inputB samples using SIMD: |
simon | 0:1014af42efd9 | 292 | * y[0] and y[1] */ |
simon | 0:1014af42efd9 | 293 | c0 = *(pb++); |
simon | 0:1014af42efd9 | 294 | |
simon | 0:1014af42efd9 | 295 | /* acc0 += x[0] * y[0] + x[1] * y[1] */ |
simon | 0:1014af42efd9 | 296 | acc0 = __SMLALD(x0, c0, acc0); |
simon | 0:1014af42efd9 | 297 | |
simon | 0:1014af42efd9 | 298 | /* acc1 += x[1] * y[0] + x[2] * y[1] */ |
simon | 0:1014af42efd9 | 299 | acc1 = __SMLALD(x1, c0, acc1); |
simon | 0:1014af42efd9 | 300 | |
simon | 0:1014af42efd9 | 301 | /* Read x[2], x[3] */ |
simon | 0:1014af42efd9 | 302 | x2 = *(q31_t *) (px++); |
simon | 0:1014af42efd9 | 303 | |
simon | 0:1014af42efd9 | 304 | /* Read x[3], x[4] */ |
simon | 0:1014af42efd9 | 305 | x3 = *(q31_t *) (px++); |
simon | 0:1014af42efd9 | 306 | |
simon | 0:1014af42efd9 | 307 | /* acc2 += x[2] * y[0] + x[3] * y[1] */ |
simon | 0:1014af42efd9 | 308 | acc2 = __SMLALD(x2, c0, acc2); |
simon | 0:1014af42efd9 | 309 | |
simon | 0:1014af42efd9 | 310 | /* acc3 += x[3] * y[0] + x[4] * y[1] */ |
simon | 0:1014af42efd9 | 311 | acc3 = __SMLALD(x3, c0, acc3); |
simon | 0:1014af42efd9 | 312 | |
simon | 0:1014af42efd9 | 313 | /* Read y[2] and y[3] */ |
simon | 0:1014af42efd9 | 314 | c0 = *(pb++); |
simon | 0:1014af42efd9 | 315 | |
simon | 0:1014af42efd9 | 316 | /* acc0 += x[2] * y[2] + x[3] * y[3] */ |
simon | 0:1014af42efd9 | 317 | acc0 = __SMLALD(x2, c0, acc0); |
simon | 0:1014af42efd9 | 318 | |
simon | 0:1014af42efd9 | 319 | /* acc1 += x[3] * y[2] + x[4] * y[3] */ |
simon | 0:1014af42efd9 | 320 | acc1 = __SMLALD(x3, c0, acc1); |
simon | 0:1014af42efd9 | 321 | |
simon | 0:1014af42efd9 | 322 | /* Read x[4], x[5] */ |
simon | 0:1014af42efd9 | 323 | x0 = *(q31_t *) (px++); |
simon | 0:1014af42efd9 | 324 | |
simon | 0:1014af42efd9 | 325 | /* Read x[5], x[6] */ |
simon | 0:1014af42efd9 | 326 | x1 = *(q31_t *) (px++); |
simon | 0:1014af42efd9 | 327 | |
simon | 0:1014af42efd9 | 328 | /* acc2 += x[4] * y[2] + x[5] * y[3] */ |
simon | 0:1014af42efd9 | 329 | acc2 = __SMLALD(x0, c0, acc2); |
simon | 0:1014af42efd9 | 330 | |
simon | 0:1014af42efd9 | 331 | /* acc3 += x[5] * y[2] + x[6] * y[3] */ |
simon | 0:1014af42efd9 | 332 | acc3 = __SMLALD(x1, c0, acc3); |
simon | 0:1014af42efd9 | 333 | |
simon | 0:1014af42efd9 | 334 | } while(--k); |
simon | 0:1014af42efd9 | 335 | |
simon | 0:1014af42efd9 | 336 | /* For the next MAC operations, SIMD is not used |
simon | 0:1014af42efd9 | 337 | * So, the 16 bit pointer if inputB, py is updated */ |
simon | 0:1014af42efd9 | 338 | py = (q15_t *) (pb); |
simon | 0:1014af42efd9 | 339 | |
simon | 0:1014af42efd9 | 340 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
simon | 0:1014af42efd9 | 341 | ** No loop unrolling is used. */ |
simon | 0:1014af42efd9 | 342 | k = srcBLen % 0x4u; |
simon | 0:1014af42efd9 | 343 | |
simon | 0:1014af42efd9 | 344 | if(k == 1u) |
simon | 0:1014af42efd9 | 345 | { |
simon | 0:1014af42efd9 | 346 | /* Read y[4] */ |
simon | 0:1014af42efd9 | 347 | c0 = *py; |
simon | 0:1014af42efd9 | 348 | c0 = c0 & 0x0000FFFF; |
simon | 0:1014af42efd9 | 349 | |
simon | 0:1014af42efd9 | 350 | /* Read x[7] */ |
simon | 0:1014af42efd9 | 351 | x3 = *(q31_t *) px++; |
simon | 0:1014af42efd9 | 352 | |
simon | 0:1014af42efd9 | 353 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 354 | acc0 = __SMLALD(x0, c0, acc0); |
simon | 0:1014af42efd9 | 355 | acc1 = __SMLALD(x1, c0, acc1); |
simon | 0:1014af42efd9 | 356 | acc2 = __SMLALDX(x1, c0, acc2); |
simon | 0:1014af42efd9 | 357 | acc3 = __SMLALDX(x3, c0, acc3); |
simon | 0:1014af42efd9 | 358 | } |
simon | 0:1014af42efd9 | 359 | |
simon | 0:1014af42efd9 | 360 | if(k == 2u) |
simon | 0:1014af42efd9 | 361 | { |
simon | 0:1014af42efd9 | 362 | /* Read y[4], y[5] */ |
simon | 0:1014af42efd9 | 363 | c0 = *(pb); |
simon | 0:1014af42efd9 | 364 | |
simon | 0:1014af42efd9 | 365 | /* Read x[7], x[8] */ |
simon | 0:1014af42efd9 | 366 | x3 = *(q31_t *) px++; |
simon | 0:1014af42efd9 | 367 | |
simon | 0:1014af42efd9 | 368 | /* Read x[9] */ |
simon | 0:1014af42efd9 | 369 | x2 = *(q31_t *) px++; |
simon | 0:1014af42efd9 | 370 | |
simon | 0:1014af42efd9 | 371 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 372 | acc0 = __SMLALD(x0, c0, acc0); |
simon | 0:1014af42efd9 | 373 | acc1 = __SMLALD(x1, c0, acc1); |
simon | 0:1014af42efd9 | 374 | acc2 = __SMLALD(x3, c0, acc2); |
simon | 0:1014af42efd9 | 375 | acc3 = __SMLALD(x2, c0, acc3); |
simon | 0:1014af42efd9 | 376 | } |
simon | 0:1014af42efd9 | 377 | |
simon | 0:1014af42efd9 | 378 | if(k == 3u) |
simon | 0:1014af42efd9 | 379 | { |
simon | 0:1014af42efd9 | 380 | /* Read y[4], y[5] */ |
simon | 0:1014af42efd9 | 381 | c0 = *pb++; |
simon | 0:1014af42efd9 | 382 | |
simon | 0:1014af42efd9 | 383 | /* Read x[7], x[8] */ |
simon | 0:1014af42efd9 | 384 | x3 = *(q31_t *) px++; |
simon | 0:1014af42efd9 | 385 | |
simon | 0:1014af42efd9 | 386 | /* Read x[9] */ |
simon | 0:1014af42efd9 | 387 | x2 = *(q31_t *) px++; |
simon | 0:1014af42efd9 | 388 | |
simon | 0:1014af42efd9 | 389 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 390 | acc0 = __SMLALD(x0, c0, acc0); |
simon | 0:1014af42efd9 | 391 | acc1 = __SMLALD(x1, c0, acc1); |
simon | 0:1014af42efd9 | 392 | acc2 = __SMLALD(x3, c0, acc2); |
simon | 0:1014af42efd9 | 393 | acc3 = __SMLALD(x2, c0, acc3); |
simon | 0:1014af42efd9 | 394 | |
simon | 0:1014af42efd9 | 395 | /* Read y[6] */ |
simon | 0:1014af42efd9 | 396 | c0 = (q15_t) (*pb); |
simon | 0:1014af42efd9 | 397 | c0 = c0 & 0x0000FFFF; |
simon | 0:1014af42efd9 | 398 | |
simon | 0:1014af42efd9 | 399 | /* Read x[10] */ |
simon | 0:1014af42efd9 | 400 | x3 = *(q31_t *) px++; |
simon | 0:1014af42efd9 | 401 | |
simon | 0:1014af42efd9 | 402 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 403 | acc0 = __SMLALDX(x1, c0, acc0); |
simon | 0:1014af42efd9 | 404 | acc1 = __SMLALD(x2, c0, acc1); |
simon | 0:1014af42efd9 | 405 | acc2 = __SMLALDX(x2, c0, acc2); |
simon | 0:1014af42efd9 | 406 | acc3 = __SMLALDX(x3, c0, acc3); |
simon | 0:1014af42efd9 | 407 | } |
simon | 0:1014af42efd9 | 408 | |
simon | 0:1014af42efd9 | 409 | /* Store the result in the accumulator in the destination buffer. */ |
simon | 0:1014af42efd9 | 410 | *pOut = (q15_t) (__SSAT(acc0 >> 15, 16)); |
simon | 0:1014af42efd9 | 411 | /* Destination pointer is updated according to the address modifier, inc */ |
simon | 0:1014af42efd9 | 412 | pOut += inc; |
simon | 0:1014af42efd9 | 413 | |
simon | 0:1014af42efd9 | 414 | *pOut = (q15_t) (__SSAT(acc1 >> 15, 16)); |
simon | 0:1014af42efd9 | 415 | pOut += inc; |
simon | 0:1014af42efd9 | 416 | |
simon | 0:1014af42efd9 | 417 | *pOut = (q15_t) (__SSAT(acc2 >> 15, 16)); |
simon | 0:1014af42efd9 | 418 | pOut += inc; |
simon | 0:1014af42efd9 | 419 | |
simon | 0:1014af42efd9 | 420 | *pOut = (q15_t) (__SSAT(acc3 >> 15, 16)); |
simon | 0:1014af42efd9 | 421 | pOut += inc; |
simon | 0:1014af42efd9 | 422 | |
simon | 0:1014af42efd9 | 423 | /* Increment the count by 4 as 4 output values are computed */ |
simon | 0:1014af42efd9 | 424 | count += 4u; |
simon | 0:1014af42efd9 | 425 | |
simon | 0:1014af42efd9 | 426 | /* Update the inputA and inputB pointers for next MAC calculation */ |
simon | 0:1014af42efd9 | 427 | px = pIn1 + count; |
simon | 0:1014af42efd9 | 428 | py = pIn2; |
simon | 0:1014af42efd9 | 429 | pb = (q31_t *) (py); |
simon | 0:1014af42efd9 | 430 | |
simon | 0:1014af42efd9 | 431 | |
simon | 0:1014af42efd9 | 432 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 433 | blkCnt--; |
simon | 0:1014af42efd9 | 434 | } |
simon | 0:1014af42efd9 | 435 | |
simon | 0:1014af42efd9 | 436 | /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here. |
simon | 0:1014af42efd9 | 437 | ** No loop unrolling is used. */ |
simon | 0:1014af42efd9 | 438 | blkCnt = blockSize2 % 0x4u; |
simon | 0:1014af42efd9 | 439 | |
simon | 0:1014af42efd9 | 440 | while(blkCnt > 0u) |
simon | 0:1014af42efd9 | 441 | { |
simon | 0:1014af42efd9 | 442 | /* Accumulator is made zero for every iteration */ |
simon | 0:1014af42efd9 | 443 | sum = 0; |
simon | 0:1014af42efd9 | 444 | |
simon | 0:1014af42efd9 | 445 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
simon | 0:1014af42efd9 | 446 | k = srcBLen >> 2u; |
simon | 0:1014af42efd9 | 447 | |
simon | 0:1014af42efd9 | 448 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
simon | 0:1014af42efd9 | 449 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
simon | 0:1014af42efd9 | 450 | while(k > 0u) |
simon | 0:1014af42efd9 | 451 | { |
simon | 0:1014af42efd9 | 452 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 453 | sum += ((q63_t) * px++ * *py++); |
simon | 0:1014af42efd9 | 454 | sum += ((q63_t) * px++ * *py++); |
simon | 0:1014af42efd9 | 455 | sum += ((q63_t) * px++ * *py++); |
simon | 0:1014af42efd9 | 456 | sum += ((q63_t) * px++ * *py++); |
simon | 0:1014af42efd9 | 457 | |
simon | 0:1014af42efd9 | 458 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 459 | k--; |
simon | 0:1014af42efd9 | 460 | } |
simon | 0:1014af42efd9 | 461 | |
simon | 0:1014af42efd9 | 462 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
simon | 0:1014af42efd9 | 463 | ** No loop unrolling is used. */ |
simon | 0:1014af42efd9 | 464 | k = srcBLen % 0x4u; |
simon | 0:1014af42efd9 | 465 | |
simon | 0:1014af42efd9 | 466 | while(k > 0u) |
simon | 0:1014af42efd9 | 467 | { |
simon | 0:1014af42efd9 | 468 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 469 | sum += ((q63_t) * px++ * *py++); |
simon | 0:1014af42efd9 | 470 | |
simon | 0:1014af42efd9 | 471 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 472 | k--; |
simon | 0:1014af42efd9 | 473 | } |
simon | 0:1014af42efd9 | 474 | |
simon | 0:1014af42efd9 | 475 | /* Store the result in the accumulator in the destination buffer. */ |
simon | 0:1014af42efd9 | 476 | *pOut = (q15_t) (__SSAT(sum >> 15, 16)); |
simon | 0:1014af42efd9 | 477 | /* Destination pointer is updated according to the address modifier, inc */ |
simon | 0:1014af42efd9 | 478 | pOut += inc; |
simon | 0:1014af42efd9 | 479 | |
simon | 0:1014af42efd9 | 480 | /* Increment count by 1, as one output value is computed */ |
simon | 0:1014af42efd9 | 481 | count++; |
simon | 0:1014af42efd9 | 482 | |
simon | 0:1014af42efd9 | 483 | /* Update the inputA and inputB pointers for next MAC calculation */ |
simon | 0:1014af42efd9 | 484 | px = pIn1 + count; |
simon | 0:1014af42efd9 | 485 | py = pIn2; |
simon | 0:1014af42efd9 | 486 | |
simon | 0:1014af42efd9 | 487 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 488 | blkCnt--; |
simon | 0:1014af42efd9 | 489 | } |
simon | 0:1014af42efd9 | 490 | } |
simon | 0:1014af42efd9 | 491 | else |
simon | 0:1014af42efd9 | 492 | { |
simon | 0:1014af42efd9 | 493 | /* If the srcBLen is not a multiple of 4, |
simon | 0:1014af42efd9 | 494 | * the blockSize2 loop cannot be unrolled by 4 */ |
simon | 0:1014af42efd9 | 495 | blkCnt = blockSize2; |
simon | 0:1014af42efd9 | 496 | |
simon | 0:1014af42efd9 | 497 | while(blkCnt > 0u) |
simon | 0:1014af42efd9 | 498 | { |
simon | 0:1014af42efd9 | 499 | /* Accumulator is made zero for every iteration */ |
simon | 0:1014af42efd9 | 500 | sum = 0; |
simon | 0:1014af42efd9 | 501 | |
simon | 0:1014af42efd9 | 502 | /* Loop over srcBLen */ |
simon | 0:1014af42efd9 | 503 | k = srcBLen; |
simon | 0:1014af42efd9 | 504 | |
simon | 0:1014af42efd9 | 505 | while(k > 0u) |
simon | 0:1014af42efd9 | 506 | { |
simon | 0:1014af42efd9 | 507 | /* Perform the multiply-accumulate */ |
simon | 0:1014af42efd9 | 508 | sum += ((q63_t) * px++ * *py++); |
simon | 0:1014af42efd9 | 509 | |
simon | 0:1014af42efd9 | 510 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 511 | k--; |
simon | 0:1014af42efd9 | 512 | } |
simon | 0:1014af42efd9 | 513 | |
simon | 0:1014af42efd9 | 514 | /* Store the result in the accumulator in the destination buffer. */ |
simon | 0:1014af42efd9 | 515 | *pOut = (q15_t) (__SSAT(sum >> 15, 16)); |
simon | 0:1014af42efd9 | 516 | /* Destination pointer is updated according to the address modifier, inc */ |
simon | 0:1014af42efd9 | 517 | pOut += inc; |
simon | 0:1014af42efd9 | 518 | |
simon | 0:1014af42efd9 | 519 | /* Increment the MAC count */ |
simon | 0:1014af42efd9 | 520 | count++; |
simon | 0:1014af42efd9 | 521 | |
simon | 0:1014af42efd9 | 522 | /* Update the inputA and inputB pointers for next MAC calculation */ |
simon | 0:1014af42efd9 | 523 | px = pIn1 + count; |
simon | 0:1014af42efd9 | 524 | py = pIn2; |
simon | 0:1014af42efd9 | 525 | |
simon | 0:1014af42efd9 | 526 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 527 | blkCnt--; |
simon | 0:1014af42efd9 | 528 | } |
simon | 0:1014af42efd9 | 529 | } |
simon | 0:1014af42efd9 | 530 | |
simon | 0:1014af42efd9 | 531 | /* -------------------------- |
simon | 0:1014af42efd9 | 532 | * Initializations of stage3 |
simon | 0:1014af42efd9 | 533 | * -------------------------*/ |
simon | 0:1014af42efd9 | 534 | |
simon | 0:1014af42efd9 | 535 | /* sum += x[srcALen-srcBLen+1] * y[0] + x[srcALen-srcBLen+2] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
simon | 0:1014af42efd9 | 536 | * sum += x[srcALen-srcBLen+2] * y[0] + x[srcALen-srcBLen+3] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
simon | 0:1014af42efd9 | 537 | * .... |
simon | 0:1014af42efd9 | 538 | * sum += x[srcALen-2] * y[0] + x[srcALen-1] * y[1] |
simon | 0:1014af42efd9 | 539 | * sum += x[srcALen-1] * y[0] |
simon | 0:1014af42efd9 | 540 | */ |
simon | 0:1014af42efd9 | 541 | |
simon | 0:1014af42efd9 | 542 | /* In this stage the MAC operations are decreased by 1 for every iteration. |
simon | 0:1014af42efd9 | 543 | The count variable holds the number of MAC operations performed */ |
simon | 0:1014af42efd9 | 544 | count = srcBLen - 1u; |
simon | 0:1014af42efd9 | 545 | |
simon | 0:1014af42efd9 | 546 | /* Working pointer of inputA */ |
simon | 0:1014af42efd9 | 547 | pSrc1 = (pIn1 + srcALen) - (srcBLen - 1u); |
simon | 0:1014af42efd9 | 548 | px = pSrc1; |
simon | 0:1014af42efd9 | 549 | |
simon | 0:1014af42efd9 | 550 | /* Working pointer of inputB */ |
simon | 0:1014af42efd9 | 551 | py = pIn2; |
simon | 0:1014af42efd9 | 552 | |
simon | 0:1014af42efd9 | 553 | /* ------------------- |
simon | 0:1014af42efd9 | 554 | * Stage3 process |
simon | 0:1014af42efd9 | 555 | * ------------------*/ |
simon | 0:1014af42efd9 | 556 | |
simon | 0:1014af42efd9 | 557 | while(blockSize3 > 0u) |
simon | 0:1014af42efd9 | 558 | { |
simon | 0:1014af42efd9 | 559 | /* Accumulator is made zero for every iteration */ |
simon | 0:1014af42efd9 | 560 | sum = 0; |
simon | 0:1014af42efd9 | 561 | |
simon | 0:1014af42efd9 | 562 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
simon | 0:1014af42efd9 | 563 | k = count >> 2u; |
simon | 0:1014af42efd9 | 564 | |
simon | 0:1014af42efd9 | 565 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
simon | 0:1014af42efd9 | 566 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
simon | 0:1014af42efd9 | 567 | while(k > 0u) |
simon | 0:1014af42efd9 | 568 | { |
simon | 0:1014af42efd9 | 569 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 570 | /* sum += x[srcALen - srcBLen + 4] * y[3] , sum += x[srcALen - srcBLen + 3] * y[2] */ |
simon | 0:1014af42efd9 | 571 | sum = __SMLALD(*__SIMD32(px)++, *__SIMD32(py)++, sum); |
simon | 0:1014af42efd9 | 572 | /* sum += x[srcALen - srcBLen + 2] * y[1] , sum += x[srcALen - srcBLen + 1] * y[0] */ |
simon | 0:1014af42efd9 | 573 | sum = __SMLALD(*__SIMD32(px)++, *__SIMD32(py)++, sum); |
simon | 0:1014af42efd9 | 574 | |
simon | 0:1014af42efd9 | 575 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 576 | k--; |
simon | 0:1014af42efd9 | 577 | } |
simon | 0:1014af42efd9 | 578 | |
simon | 0:1014af42efd9 | 579 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
simon | 0:1014af42efd9 | 580 | ** No loop unrolling is used. */ |
simon | 0:1014af42efd9 | 581 | k = count % 0x4u; |
simon | 0:1014af42efd9 | 582 | |
simon | 0:1014af42efd9 | 583 | while(k > 0u) |
simon | 0:1014af42efd9 | 584 | { |
simon | 0:1014af42efd9 | 585 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 586 | sum = __SMLALD(*px++, *py++, sum); |
simon | 0:1014af42efd9 | 587 | |
simon | 0:1014af42efd9 | 588 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 589 | k--; |
simon | 0:1014af42efd9 | 590 | } |
simon | 0:1014af42efd9 | 591 | |
simon | 0:1014af42efd9 | 592 | /* Store the result in the accumulator in the destination buffer. */ |
simon | 0:1014af42efd9 | 593 | *pOut = (q15_t) (__SSAT((sum >> 15), 16)); |
simon | 0:1014af42efd9 | 594 | /* Destination pointer is updated according to the address modifier, inc */ |
simon | 0:1014af42efd9 | 595 | pOut += inc; |
simon | 0:1014af42efd9 | 596 | |
simon | 0:1014af42efd9 | 597 | /* Update the inputA and inputB pointers for next MAC calculation */ |
simon | 0:1014af42efd9 | 598 | px = ++pSrc1; |
simon | 0:1014af42efd9 | 599 | py = pIn2; |
simon | 0:1014af42efd9 | 600 | |
simon | 0:1014af42efd9 | 601 | /* Decrement the MAC count */ |
simon | 0:1014af42efd9 | 602 | count--; |
simon | 0:1014af42efd9 | 603 | |
simon | 0:1014af42efd9 | 604 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 605 | blockSize3--; |
simon | 0:1014af42efd9 | 606 | } |
simon | 0:1014af42efd9 | 607 | |
simon | 0:1014af42efd9 | 608 | } |
simon | 0:1014af42efd9 | 609 | |
simon | 0:1014af42efd9 | 610 | /** |
simon | 0:1014af42efd9 | 611 | * @} end of Corr group |
simon | 0:1014af42efd9 | 612 | */ |