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_f32.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_f32.c |
simon | 0:1014af42efd9 | 9 | * |
simon | 0:1014af42efd9 | 10 | * Description: Correlation for floating-point sequences. |
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 | * @defgroup Corr Correlation |
simon | 0:1014af42efd9 | 39 | * |
simon | 0:1014af42efd9 | 40 | * Correlation is a mathematical operation that is similar to convolution. |
simon | 0:1014af42efd9 | 41 | * As with convolution, correlation uses two signals to produce a third signal. |
simon | 0:1014af42efd9 | 42 | * The underlying algorithms in correlation and convolution are identical except that one of the inputs is flipped in convolution. |
simon | 0:1014af42efd9 | 43 | * Correlation is commonly used to measure the similarity between two signals. |
simon | 0:1014af42efd9 | 44 | * It has applications in pattern recognition, cryptanalysis, and searching. |
simon | 0:1014af42efd9 | 45 | * The CMSIS library provides correlation functions for Q7, Q15, Q31 and floating-point data types. |
simon | 0:1014af42efd9 | 46 | * Fast versions of the Q15 and Q31 functions are also provided. |
simon | 0:1014af42efd9 | 47 | * |
simon | 0:1014af42efd9 | 48 | * \par Algorithm |
simon | 0:1014af42efd9 | 49 | * Let <code>a[n]</code> and <code>b[n]</code> be sequences of length <code>srcALen</code> and <code>srcBLen</code> samples respectively. |
simon | 0:1014af42efd9 | 50 | * The convolution of the two signals is denoted by |
simon | 0:1014af42efd9 | 51 | * <pre> |
simon | 0:1014af42efd9 | 52 | * c[n] = a[n] * b[n] |
simon | 0:1014af42efd9 | 53 | * </pre> |
simon | 0:1014af42efd9 | 54 | * In correlation, one of the signals is flipped in time |
simon | 0:1014af42efd9 | 55 | * <pre> |
simon | 0:1014af42efd9 | 56 | * c[n] = a[n] * b[-n] |
simon | 0:1014af42efd9 | 57 | * </pre> |
simon | 0:1014af42efd9 | 58 | * |
simon | 0:1014af42efd9 | 59 | * \par |
simon | 0:1014af42efd9 | 60 | * and this is mathematically defined as |
simon | 0:1014af42efd9 | 61 | * \image html CorrelateEquation.gif |
simon | 0:1014af42efd9 | 62 | * \par |
simon | 0:1014af42efd9 | 63 | * The <code>pSrcA</code> points to the first input vector of length <code>srcALen</code> and <code>pSrcB</code> points to the second input vector of length <code>srcBLen</code>. |
simon | 0:1014af42efd9 | 64 | * The result <code>c[n]</code> is of length <code>2 * max(srcALen, srcBLen) - 1</code> and is defined over the interval <code>n=0, 1, 2, ..., (2 * max(srcALen, srcBLen) - 2)</code>. |
simon | 0:1014af42efd9 | 65 | * The output result is written to <code>pDst</code> and the calling function must allocate <code>2 * max(srcALen, srcBLen) - 1</code> words for the result. |
simon | 0:1014af42efd9 | 66 | * |
simon | 0:1014af42efd9 | 67 | * <b>Fixed-Point Behavior</b> |
simon | 0:1014af42efd9 | 68 | * \par |
simon | 0:1014af42efd9 | 69 | * Correlation requires summing up a large number of intermediate products. |
simon | 0:1014af42efd9 | 70 | * As such, the Q7, Q15, and Q31 functions run a risk of overflow and saturation. |
simon | 0:1014af42efd9 | 71 | * Refer to the function specific documentation below for further details of the particular algorithm used. |
simon | 0:1014af42efd9 | 72 | */ |
simon | 0:1014af42efd9 | 73 | |
simon | 0:1014af42efd9 | 74 | /** |
simon | 0:1014af42efd9 | 75 | * @addtogroup Corr |
simon | 0:1014af42efd9 | 76 | * @{ |
simon | 0:1014af42efd9 | 77 | */ |
simon | 0:1014af42efd9 | 78 | /** |
simon | 0:1014af42efd9 | 79 | * @brief Correlation of floating-point sequences |
simon | 0:1014af42efd9 | 80 | * @param[in] *pSrcA points to the first input sequence. |
simon | 0:1014af42efd9 | 81 | * @param[in] srcALen length of the first input sequence. |
simon | 0:1014af42efd9 | 82 | * @param[in] *pSrcB points to the second input sequence. |
simon | 0:1014af42efd9 | 83 | * @param[in] srcBLen length of the second input sequence. |
simon | 0:1014af42efd9 | 84 | * @param[out] *pDst points to the location where the output result is written. Length 2 * max(srcALen, srcBLen) - 1. |
simon | 0:1014af42efd9 | 85 | * @return none. |
simon | 0:1014af42efd9 | 86 | */ |
simon | 0:1014af42efd9 | 87 | |
simon | 0:1014af42efd9 | 88 | void arm_correlate_f32( |
simon | 0:1014af42efd9 | 89 | float32_t * pSrcA, |
simon | 0:1014af42efd9 | 90 | uint32_t srcALen, |
simon | 0:1014af42efd9 | 91 | float32_t * pSrcB, |
simon | 0:1014af42efd9 | 92 | uint32_t srcBLen, |
simon | 0:1014af42efd9 | 93 | float32_t * pDst) |
simon | 0:1014af42efd9 | 94 | { |
simon | 0:1014af42efd9 | 95 | float32_t *pIn1; /* inputA pointer */ |
simon | 0:1014af42efd9 | 96 | float32_t *pIn2; /* inputB pointer */ |
simon | 0:1014af42efd9 | 97 | float32_t *pOut = pDst; /* output pointer */ |
simon | 0:1014af42efd9 | 98 | float32_t *px; /* Intermediate inputA pointer */ |
simon | 0:1014af42efd9 | 99 | float32_t *py; /* Intermediate inputB pointer */ |
simon | 0:1014af42efd9 | 100 | float32_t *pSrc1; /* Intermediate pointers */ |
simon | 0:1014af42efd9 | 101 | float32_t sum, acc0, acc1, acc2, acc3; /* Accumulators */ |
simon | 0:1014af42efd9 | 102 | float32_t x0, x1, x2, x3, c0; /* temporary variables for holding input and coefficient values */ |
simon | 0:1014af42efd9 | 103 | uint32_t j, k = 0u, count, blkCnt, outBlockSize, blockSize1, blockSize2, blockSize3; /* loop counters */ |
simon | 0:1014af42efd9 | 104 | int32_t inc = 1; /* Destination address modifier */ |
simon | 0:1014af42efd9 | 105 | |
simon | 0:1014af42efd9 | 106 | |
simon | 0:1014af42efd9 | 107 | /* The algorithm implementation is based on the lengths of the inputs. */ |
simon | 0:1014af42efd9 | 108 | /* srcB is always made to slide across srcA. */ |
simon | 0:1014af42efd9 | 109 | /* So srcBLen is always considered as shorter or equal to srcALen */ |
simon | 0:1014af42efd9 | 110 | /* But CORR(x, y) is reverse of CORR(y, x) */ |
simon | 0:1014af42efd9 | 111 | /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */ |
simon | 0:1014af42efd9 | 112 | /* and the destination pointer modifier, inc is set to -1 */ |
simon | 0:1014af42efd9 | 113 | /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */ |
simon | 0:1014af42efd9 | 114 | /* But to improve the performance, |
simon | 0:1014af42efd9 | 115 | * we include zeroes in the output instead of zero padding either of the the inputs*/ |
simon | 0:1014af42efd9 | 116 | /* If srcALen > srcBLen, |
simon | 0:1014af42efd9 | 117 | * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */ |
simon | 0:1014af42efd9 | 118 | /* If srcALen < srcBLen, |
simon | 0:1014af42efd9 | 119 | * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */ |
simon | 0:1014af42efd9 | 120 | if(srcALen >= srcBLen) |
simon | 0:1014af42efd9 | 121 | { |
simon | 0:1014af42efd9 | 122 | /* Initialization of inputA pointer */ |
simon | 0:1014af42efd9 | 123 | pIn1 = pSrcA; |
simon | 0:1014af42efd9 | 124 | |
simon | 0:1014af42efd9 | 125 | /* Initialization of inputB pointer */ |
simon | 0:1014af42efd9 | 126 | pIn2 = pSrcB; |
simon | 0:1014af42efd9 | 127 | |
simon | 0:1014af42efd9 | 128 | /* Number of output samples is calculated */ |
simon | 0:1014af42efd9 | 129 | outBlockSize = (2u * srcALen) - 1u; |
simon | 0:1014af42efd9 | 130 | |
simon | 0:1014af42efd9 | 131 | /* When srcALen > srcBLen, zero padding has to be done to srcB |
simon | 0:1014af42efd9 | 132 | * to make their lengths equal. |
simon | 0:1014af42efd9 | 133 | * Instead, (outBlockSize - (srcALen + srcBLen - 1)) |
simon | 0:1014af42efd9 | 134 | * number of output samples are made zero */ |
simon | 0:1014af42efd9 | 135 | j = outBlockSize - (srcALen + (srcBLen - 1u)); |
simon | 0:1014af42efd9 | 136 | |
simon | 0:1014af42efd9 | 137 | while(j > 0u) |
simon | 0:1014af42efd9 | 138 | { |
simon | 0:1014af42efd9 | 139 | /* Zero is stored in the destination buffer */ |
simon | 0:1014af42efd9 | 140 | *pOut++ = 0.0f; |
simon | 0:1014af42efd9 | 141 | |
simon | 0:1014af42efd9 | 142 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 143 | j--; |
simon | 0:1014af42efd9 | 144 | } |
simon | 0:1014af42efd9 | 145 | |
simon | 0:1014af42efd9 | 146 | } |
simon | 0:1014af42efd9 | 147 | else |
simon | 0:1014af42efd9 | 148 | { |
simon | 0:1014af42efd9 | 149 | /* Initialization of inputA pointer */ |
simon | 0:1014af42efd9 | 150 | pIn1 = pSrcB; |
simon | 0:1014af42efd9 | 151 | |
simon | 0:1014af42efd9 | 152 | /* Initialization of inputB pointer */ |
simon | 0:1014af42efd9 | 153 | pIn2 = pSrcA; |
simon | 0:1014af42efd9 | 154 | |
simon | 0:1014af42efd9 | 155 | /* srcBLen is always considered as shorter or equal to srcALen */ |
simon | 0:1014af42efd9 | 156 | j = srcBLen; |
simon | 0:1014af42efd9 | 157 | srcBLen = srcALen; |
simon | 0:1014af42efd9 | 158 | srcALen = j; |
simon | 0:1014af42efd9 | 159 | |
simon | 0:1014af42efd9 | 160 | /* CORR(x, y) = Reverse order(CORR(y, x)) */ |
simon | 0:1014af42efd9 | 161 | /* Hence set the destination pointer to point to the last output sample */ |
simon | 0:1014af42efd9 | 162 | pOut = pDst + ((srcALen + srcBLen) - 2u); |
simon | 0:1014af42efd9 | 163 | |
simon | 0:1014af42efd9 | 164 | /* Destination address modifier is set to -1 */ |
simon | 0:1014af42efd9 | 165 | inc = -1; |
simon | 0:1014af42efd9 | 166 | |
simon | 0:1014af42efd9 | 167 | } |
simon | 0:1014af42efd9 | 168 | |
simon | 0:1014af42efd9 | 169 | /* The function is internally |
simon | 0:1014af42efd9 | 170 | * divided into three parts according to the number of multiplications that has to be |
simon | 0:1014af42efd9 | 171 | * taken place between inputA samples and inputB samples. In the first part of the |
simon | 0:1014af42efd9 | 172 | * algorithm, the multiplications increase by one for every iteration. |
simon | 0:1014af42efd9 | 173 | * In the second part of the algorithm, srcBLen number of multiplications are done. |
simon | 0:1014af42efd9 | 174 | * In the third part of the algorithm, the multiplications decrease by one |
simon | 0:1014af42efd9 | 175 | * for every iteration.*/ |
simon | 0:1014af42efd9 | 176 | /* The algorithm is implemented in three stages. |
simon | 0:1014af42efd9 | 177 | * The loop counters of each stage is initiated here. */ |
simon | 0:1014af42efd9 | 178 | blockSize1 = srcBLen - 1u; |
simon | 0:1014af42efd9 | 179 | blockSize2 = srcALen - (srcBLen - 1u); |
simon | 0:1014af42efd9 | 180 | blockSize3 = blockSize1; |
simon | 0:1014af42efd9 | 181 | |
simon | 0:1014af42efd9 | 182 | /* -------------------------- |
simon | 0:1014af42efd9 | 183 | * Initializations of stage1 |
simon | 0:1014af42efd9 | 184 | * -------------------------*/ |
simon | 0:1014af42efd9 | 185 | |
simon | 0:1014af42efd9 | 186 | /* sum = x[0] * y[srcBlen - 1] |
simon | 0:1014af42efd9 | 187 | * sum = x[0] * y[srcBlen-2] + x[1] * y[srcBlen - 1] |
simon | 0:1014af42efd9 | 188 | * .... |
simon | 0:1014af42efd9 | 189 | * sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen - 1] * y[srcBLen - 1] |
simon | 0:1014af42efd9 | 190 | */ |
simon | 0:1014af42efd9 | 191 | |
simon | 0:1014af42efd9 | 192 | /* In this stage the MAC operations are increased by 1 for every iteration. |
simon | 0:1014af42efd9 | 193 | The count variable holds the number of MAC operations performed */ |
simon | 0:1014af42efd9 | 194 | count = 1u; |
simon | 0:1014af42efd9 | 195 | |
simon | 0:1014af42efd9 | 196 | /* Working pointer of inputA */ |
simon | 0:1014af42efd9 | 197 | px = pIn1; |
simon | 0:1014af42efd9 | 198 | |
simon | 0:1014af42efd9 | 199 | /* Working pointer of inputB */ |
simon | 0:1014af42efd9 | 200 | pSrc1 = pIn2 + (srcBLen - 1u); |
simon | 0:1014af42efd9 | 201 | py = pSrc1; |
simon | 0:1014af42efd9 | 202 | |
simon | 0:1014af42efd9 | 203 | /* ------------------------ |
simon | 0:1014af42efd9 | 204 | * Stage1 process |
simon | 0:1014af42efd9 | 205 | * ----------------------*/ |
simon | 0:1014af42efd9 | 206 | |
simon | 0:1014af42efd9 | 207 | /* The first stage starts here */ |
simon | 0:1014af42efd9 | 208 | while(blockSize1 > 0u) |
simon | 0:1014af42efd9 | 209 | { |
simon | 0:1014af42efd9 | 210 | /* Accumulator is made zero for every iteration */ |
simon | 0:1014af42efd9 | 211 | sum = 0.0f; |
simon | 0:1014af42efd9 | 212 | |
simon | 0:1014af42efd9 | 213 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
simon | 0:1014af42efd9 | 214 | k = count >> 2u; |
simon | 0:1014af42efd9 | 215 | |
simon | 0:1014af42efd9 | 216 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
simon | 0:1014af42efd9 | 217 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
simon | 0:1014af42efd9 | 218 | while(k > 0u) |
simon | 0:1014af42efd9 | 219 | { |
simon | 0:1014af42efd9 | 220 | /* x[0] * y[srcBLen - 4] */ |
simon | 0:1014af42efd9 | 221 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 222 | /* x[1] * y[srcBLen - 3] */ |
simon | 0:1014af42efd9 | 223 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 224 | /* x[2] * y[srcBLen - 2] */ |
simon | 0:1014af42efd9 | 225 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 226 | /* x[3] * y[srcBLen - 1] */ |
simon | 0:1014af42efd9 | 227 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 228 | |
simon | 0:1014af42efd9 | 229 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 230 | k--; |
simon | 0:1014af42efd9 | 231 | } |
simon | 0:1014af42efd9 | 232 | |
simon | 0:1014af42efd9 | 233 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
simon | 0:1014af42efd9 | 234 | ** No loop unrolling is used. */ |
simon | 0:1014af42efd9 | 235 | k = count % 0x4u; |
simon | 0:1014af42efd9 | 236 | |
simon | 0:1014af42efd9 | 237 | while(k > 0u) |
simon | 0:1014af42efd9 | 238 | { |
simon | 0:1014af42efd9 | 239 | /* Perform the multiply-accumulate */ |
simon | 0:1014af42efd9 | 240 | /* x[0] * y[srcBLen - 1] */ |
simon | 0:1014af42efd9 | 241 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 242 | |
simon | 0:1014af42efd9 | 243 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 244 | k--; |
simon | 0:1014af42efd9 | 245 | } |
simon | 0:1014af42efd9 | 246 | |
simon | 0:1014af42efd9 | 247 | /* Store the result in the accumulator in the destination buffer. */ |
simon | 0:1014af42efd9 | 248 | *pOut = sum; |
simon | 0:1014af42efd9 | 249 | /* Destination pointer is updated according to the address modifier, inc */ |
simon | 0:1014af42efd9 | 250 | pOut += inc; |
simon | 0:1014af42efd9 | 251 | |
simon | 0:1014af42efd9 | 252 | /* Update the inputA and inputB pointers for next MAC calculation */ |
simon | 0:1014af42efd9 | 253 | py = pSrc1 - count; |
simon | 0:1014af42efd9 | 254 | px = pIn1; |
simon | 0:1014af42efd9 | 255 | |
simon | 0:1014af42efd9 | 256 | /* Increment the MAC count */ |
simon | 0:1014af42efd9 | 257 | count++; |
simon | 0:1014af42efd9 | 258 | |
simon | 0:1014af42efd9 | 259 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 260 | blockSize1--; |
simon | 0:1014af42efd9 | 261 | } |
simon | 0:1014af42efd9 | 262 | |
simon | 0:1014af42efd9 | 263 | /* -------------------------- |
simon | 0:1014af42efd9 | 264 | * Initializations of stage2 |
simon | 0:1014af42efd9 | 265 | * ------------------------*/ |
simon | 0:1014af42efd9 | 266 | |
simon | 0:1014af42efd9 | 267 | /* sum = x[0] * y[0] + x[1] * y[1] +...+ x[srcBLen-1] * y[srcBLen-1] |
simon | 0:1014af42efd9 | 268 | * sum = x[1] * y[0] + x[2] * y[1] +...+ x[srcBLen] * y[srcBLen-1] |
simon | 0:1014af42efd9 | 269 | * .... |
simon | 0:1014af42efd9 | 270 | * sum = x[srcALen-srcBLen-2] * y[0] + x[srcALen-srcBLen-1] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
simon | 0:1014af42efd9 | 271 | */ |
simon | 0:1014af42efd9 | 272 | |
simon | 0:1014af42efd9 | 273 | /* Working pointer of inputA */ |
simon | 0:1014af42efd9 | 274 | px = pIn1; |
simon | 0:1014af42efd9 | 275 | |
simon | 0:1014af42efd9 | 276 | /* Working pointer of inputB */ |
simon | 0:1014af42efd9 | 277 | py = pIn2; |
simon | 0:1014af42efd9 | 278 | |
simon | 0:1014af42efd9 | 279 | /* count is index by which the pointer pIn1 to be incremented */ |
simon | 0:1014af42efd9 | 280 | count = 1u; |
simon | 0:1014af42efd9 | 281 | |
simon | 0:1014af42efd9 | 282 | /* ------------------- |
simon | 0:1014af42efd9 | 283 | * Stage2 process |
simon | 0:1014af42efd9 | 284 | * ------------------*/ |
simon | 0:1014af42efd9 | 285 | |
simon | 0:1014af42efd9 | 286 | /* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed. |
simon | 0:1014af42efd9 | 287 | * So, to loop unroll over blockSize2, |
simon | 0:1014af42efd9 | 288 | * srcBLen should be greater than or equal to 4, to loop unroll the srcBLen loop */ |
simon | 0:1014af42efd9 | 289 | if(srcBLen >= 4u) |
simon | 0:1014af42efd9 | 290 | { |
simon | 0:1014af42efd9 | 291 | /* Loop unroll over blockSize2, by 4 */ |
simon | 0:1014af42efd9 | 292 | blkCnt = blockSize2 >> 2u; |
simon | 0:1014af42efd9 | 293 | |
simon | 0:1014af42efd9 | 294 | while(blkCnt > 0u) |
simon | 0:1014af42efd9 | 295 | { |
simon | 0:1014af42efd9 | 296 | /* Set all accumulators to zero */ |
simon | 0:1014af42efd9 | 297 | acc0 = 0.0f; |
simon | 0:1014af42efd9 | 298 | acc1 = 0.0f; |
simon | 0:1014af42efd9 | 299 | acc2 = 0.0f; |
simon | 0:1014af42efd9 | 300 | acc3 = 0.0f; |
simon | 0:1014af42efd9 | 301 | |
simon | 0:1014af42efd9 | 302 | /* read x[0], x[1], x[2] samples */ |
simon | 0:1014af42efd9 | 303 | x0 = *(px++); |
simon | 0:1014af42efd9 | 304 | x1 = *(px++); |
simon | 0:1014af42efd9 | 305 | x2 = *(px++); |
simon | 0:1014af42efd9 | 306 | |
simon | 0:1014af42efd9 | 307 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
simon | 0:1014af42efd9 | 308 | k = srcBLen >> 2u; |
simon | 0:1014af42efd9 | 309 | |
simon | 0:1014af42efd9 | 310 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
simon | 0:1014af42efd9 | 311 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
simon | 0:1014af42efd9 | 312 | do |
simon | 0:1014af42efd9 | 313 | { |
simon | 0:1014af42efd9 | 314 | /* Read y[0] sample */ |
simon | 0:1014af42efd9 | 315 | c0 = *(py++); |
simon | 0:1014af42efd9 | 316 | |
simon | 0:1014af42efd9 | 317 | /* Read x[3] sample */ |
simon | 0:1014af42efd9 | 318 | x3 = *(px++); |
simon | 0:1014af42efd9 | 319 | |
simon | 0:1014af42efd9 | 320 | /* Perform the multiply-accumulate */ |
simon | 0:1014af42efd9 | 321 | /* acc0 += x[0] * y[0] */ |
simon | 0:1014af42efd9 | 322 | acc0 += x0 * c0; |
simon | 0:1014af42efd9 | 323 | /* acc1 += x[1] * y[0] */ |
simon | 0:1014af42efd9 | 324 | acc1 += x1 * c0; |
simon | 0:1014af42efd9 | 325 | /* acc2 += x[2] * y[0] */ |
simon | 0:1014af42efd9 | 326 | acc2 += x2 * c0; |
simon | 0:1014af42efd9 | 327 | /* acc3 += x[3] * y[0] */ |
simon | 0:1014af42efd9 | 328 | acc3 += x3 * c0; |
simon | 0:1014af42efd9 | 329 | |
simon | 0:1014af42efd9 | 330 | /* Read y[1] sample */ |
simon | 0:1014af42efd9 | 331 | c0 = *(py++); |
simon | 0:1014af42efd9 | 332 | |
simon | 0:1014af42efd9 | 333 | /* Read x[4] sample */ |
simon | 0:1014af42efd9 | 334 | x0 = *(px++); |
simon | 0:1014af42efd9 | 335 | |
simon | 0:1014af42efd9 | 336 | /* Perform the multiply-accumulate */ |
simon | 0:1014af42efd9 | 337 | /* acc0 += x[1] * y[1] */ |
simon | 0:1014af42efd9 | 338 | acc0 += x1 * c0; |
simon | 0:1014af42efd9 | 339 | /* acc1 += x[2] * y[1] */ |
simon | 0:1014af42efd9 | 340 | acc1 += x2 * c0; |
simon | 0:1014af42efd9 | 341 | /* acc2 += x[3] * y[1] */ |
simon | 0:1014af42efd9 | 342 | acc2 += x3 * c0; |
simon | 0:1014af42efd9 | 343 | /* acc3 += x[4] * y[1] */ |
simon | 0:1014af42efd9 | 344 | acc3 += x0 * c0; |
simon | 0:1014af42efd9 | 345 | |
simon | 0:1014af42efd9 | 346 | /* Read y[2] sample */ |
simon | 0:1014af42efd9 | 347 | c0 = *(py++); |
simon | 0:1014af42efd9 | 348 | |
simon | 0:1014af42efd9 | 349 | /* Read x[5] sample */ |
simon | 0:1014af42efd9 | 350 | x1 = *(px++); |
simon | 0:1014af42efd9 | 351 | |
simon | 0:1014af42efd9 | 352 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 353 | /* acc0 += x[2] * y[2] */ |
simon | 0:1014af42efd9 | 354 | acc0 += x2 * c0; |
simon | 0:1014af42efd9 | 355 | /* acc1 += x[3] * y[2] */ |
simon | 0:1014af42efd9 | 356 | acc1 += x3 * c0; |
simon | 0:1014af42efd9 | 357 | /* acc2 += x[4] * y[2] */ |
simon | 0:1014af42efd9 | 358 | acc2 += x0 * c0; |
simon | 0:1014af42efd9 | 359 | /* acc3 += x[5] * y[2] */ |
simon | 0:1014af42efd9 | 360 | acc3 += x1 * c0; |
simon | 0:1014af42efd9 | 361 | |
simon | 0:1014af42efd9 | 362 | /* Read y[3] sample */ |
simon | 0:1014af42efd9 | 363 | c0 = *(py++); |
simon | 0:1014af42efd9 | 364 | |
simon | 0:1014af42efd9 | 365 | /* Read x[6] sample */ |
simon | 0:1014af42efd9 | 366 | x2 = *(px++); |
simon | 0:1014af42efd9 | 367 | |
simon | 0:1014af42efd9 | 368 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 369 | /* acc0 += x[3] * y[3] */ |
simon | 0:1014af42efd9 | 370 | acc0 += x3 * c0; |
simon | 0:1014af42efd9 | 371 | /* acc1 += x[4] * y[3] */ |
simon | 0:1014af42efd9 | 372 | acc1 += x0 * c0; |
simon | 0:1014af42efd9 | 373 | /* acc2 += x[5] * y[3] */ |
simon | 0:1014af42efd9 | 374 | acc2 += x1 * c0; |
simon | 0:1014af42efd9 | 375 | /* acc3 += x[6] * y[3] */ |
simon | 0:1014af42efd9 | 376 | acc3 += x2 * c0; |
simon | 0:1014af42efd9 | 377 | |
simon | 0:1014af42efd9 | 378 | |
simon | 0:1014af42efd9 | 379 | } while(--k); |
simon | 0:1014af42efd9 | 380 | |
simon | 0:1014af42efd9 | 381 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
simon | 0:1014af42efd9 | 382 | ** No loop unrolling is used. */ |
simon | 0:1014af42efd9 | 383 | k = srcBLen % 0x4u; |
simon | 0:1014af42efd9 | 384 | |
simon | 0:1014af42efd9 | 385 | while(k > 0u) |
simon | 0:1014af42efd9 | 386 | { |
simon | 0:1014af42efd9 | 387 | /* Read y[4] sample */ |
simon | 0:1014af42efd9 | 388 | c0 = *(py++); |
simon | 0:1014af42efd9 | 389 | |
simon | 0:1014af42efd9 | 390 | /* Read x[7] sample */ |
simon | 0:1014af42efd9 | 391 | x3 = *(px++); |
simon | 0:1014af42efd9 | 392 | |
simon | 0:1014af42efd9 | 393 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 394 | /* acc0 += x[4] * y[4] */ |
simon | 0:1014af42efd9 | 395 | acc0 += x0 * c0; |
simon | 0:1014af42efd9 | 396 | /* acc1 += x[5] * y[4] */ |
simon | 0:1014af42efd9 | 397 | acc1 += x1 * c0; |
simon | 0:1014af42efd9 | 398 | /* acc2 += x[6] * y[4] */ |
simon | 0:1014af42efd9 | 399 | acc2 += x2 * c0; |
simon | 0:1014af42efd9 | 400 | /* acc3 += x[7] * y[4] */ |
simon | 0:1014af42efd9 | 401 | acc3 += x3 * c0; |
simon | 0:1014af42efd9 | 402 | |
simon | 0:1014af42efd9 | 403 | /* Reuse the present samples for the next MAC */ |
simon | 0:1014af42efd9 | 404 | x0 = x1; |
simon | 0:1014af42efd9 | 405 | x1 = x2; |
simon | 0:1014af42efd9 | 406 | x2 = x3; |
simon | 0:1014af42efd9 | 407 | |
simon | 0:1014af42efd9 | 408 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 409 | k--; |
simon | 0:1014af42efd9 | 410 | } |
simon | 0:1014af42efd9 | 411 | |
simon | 0:1014af42efd9 | 412 | /* Store the result in the accumulator in the destination buffer. */ |
simon | 0:1014af42efd9 | 413 | *pOut = acc0; |
simon | 0:1014af42efd9 | 414 | /* Destination pointer is updated according to the address modifier, inc */ |
simon | 0:1014af42efd9 | 415 | pOut += inc; |
simon | 0:1014af42efd9 | 416 | |
simon | 0:1014af42efd9 | 417 | *pOut = acc1; |
simon | 0:1014af42efd9 | 418 | pOut += inc; |
simon | 0:1014af42efd9 | 419 | |
simon | 0:1014af42efd9 | 420 | *pOut = acc2; |
simon | 0:1014af42efd9 | 421 | pOut += inc; |
simon | 0:1014af42efd9 | 422 | |
simon | 0:1014af42efd9 | 423 | *pOut = acc3; |
simon | 0:1014af42efd9 | 424 | pOut += inc; |
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 * 4u); |
simon | 0:1014af42efd9 | 428 | py = pIn2; |
simon | 0:1014af42efd9 | 429 | |
simon | 0:1014af42efd9 | 430 | /* Increment the pointer pIn1 index, count by 1 */ |
simon | 0:1014af42efd9 | 431 | count++; |
simon | 0:1014af42efd9 | 432 | |
simon | 0:1014af42efd9 | 433 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 434 | blkCnt--; |
simon | 0:1014af42efd9 | 435 | } |
simon | 0:1014af42efd9 | 436 | |
simon | 0:1014af42efd9 | 437 | /* If the blockSize2 is not a multiple of 4, compute any remaining output samples here. |
simon | 0:1014af42efd9 | 438 | ** No loop unrolling is used. */ |
simon | 0:1014af42efd9 | 439 | blkCnt = blockSize2 % 0x4u; |
simon | 0:1014af42efd9 | 440 | |
simon | 0:1014af42efd9 | 441 | while(blkCnt > 0u) |
simon | 0:1014af42efd9 | 442 | { |
simon | 0:1014af42efd9 | 443 | /* Accumulator is made zero for every iteration */ |
simon | 0:1014af42efd9 | 444 | sum = 0.0f; |
simon | 0:1014af42efd9 | 445 | |
simon | 0:1014af42efd9 | 446 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
simon | 0:1014af42efd9 | 447 | k = srcBLen >> 2u; |
simon | 0:1014af42efd9 | 448 | |
simon | 0:1014af42efd9 | 449 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
simon | 0:1014af42efd9 | 450 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
simon | 0:1014af42efd9 | 451 | while(k > 0u) |
simon | 0:1014af42efd9 | 452 | { |
simon | 0:1014af42efd9 | 453 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 454 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 455 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 456 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 457 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 458 | |
simon | 0:1014af42efd9 | 459 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 460 | k--; |
simon | 0:1014af42efd9 | 461 | } |
simon | 0:1014af42efd9 | 462 | |
simon | 0:1014af42efd9 | 463 | /* If the srcBLen is not a multiple of 4, compute any remaining MACs here. |
simon | 0:1014af42efd9 | 464 | ** No loop unrolling is used. */ |
simon | 0:1014af42efd9 | 465 | k = srcBLen % 0x4u; |
simon | 0:1014af42efd9 | 466 | |
simon | 0:1014af42efd9 | 467 | while(k > 0u) |
simon | 0:1014af42efd9 | 468 | { |
simon | 0:1014af42efd9 | 469 | /* Perform the multiply-accumulate */ |
simon | 0:1014af42efd9 | 470 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 471 | |
simon | 0:1014af42efd9 | 472 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 473 | k--; |
simon | 0:1014af42efd9 | 474 | } |
simon | 0:1014af42efd9 | 475 | |
simon | 0:1014af42efd9 | 476 | /* Store the result in the accumulator in the destination buffer. */ |
simon | 0:1014af42efd9 | 477 | *pOut = sum; |
simon | 0:1014af42efd9 | 478 | /* Destination pointer is updated according to the address modifier, inc */ |
simon | 0:1014af42efd9 | 479 | pOut += inc; |
simon | 0:1014af42efd9 | 480 | |
simon | 0:1014af42efd9 | 481 | /* Update the inputA and inputB pointers for next MAC calculation */ |
simon | 0:1014af42efd9 | 482 | px = pIn1 + count; |
simon | 0:1014af42efd9 | 483 | py = pIn2; |
simon | 0:1014af42efd9 | 484 | |
simon | 0:1014af42efd9 | 485 | /* Increment the pointer pIn1 index, count by 1 */ |
simon | 0:1014af42efd9 | 486 | count++; |
simon | 0:1014af42efd9 | 487 | |
simon | 0:1014af42efd9 | 488 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 489 | blkCnt--; |
simon | 0:1014af42efd9 | 490 | } |
simon | 0:1014af42efd9 | 491 | } |
simon | 0:1014af42efd9 | 492 | else |
simon | 0:1014af42efd9 | 493 | { |
simon | 0:1014af42efd9 | 494 | /* If the srcBLen is not a multiple of 4, |
simon | 0:1014af42efd9 | 495 | * the blockSize2 loop cannot be unrolled by 4 */ |
simon | 0:1014af42efd9 | 496 | blkCnt = blockSize2; |
simon | 0:1014af42efd9 | 497 | |
simon | 0:1014af42efd9 | 498 | while(blkCnt > 0u) |
simon | 0:1014af42efd9 | 499 | { |
simon | 0:1014af42efd9 | 500 | /* Accumulator is made zero for every iteration */ |
simon | 0:1014af42efd9 | 501 | sum = 0.0f; |
simon | 0:1014af42efd9 | 502 | |
simon | 0:1014af42efd9 | 503 | /* Loop over srcBLen */ |
simon | 0:1014af42efd9 | 504 | k = srcBLen; |
simon | 0:1014af42efd9 | 505 | |
simon | 0:1014af42efd9 | 506 | while(k > 0u) |
simon | 0:1014af42efd9 | 507 | { |
simon | 0:1014af42efd9 | 508 | /* Perform the multiply-accumulate */ |
simon | 0:1014af42efd9 | 509 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 510 | |
simon | 0:1014af42efd9 | 511 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 512 | k--; |
simon | 0:1014af42efd9 | 513 | } |
simon | 0:1014af42efd9 | 514 | |
simon | 0:1014af42efd9 | 515 | /* Store the result in the accumulator in the destination buffer. */ |
simon | 0:1014af42efd9 | 516 | *pOut = sum; |
simon | 0:1014af42efd9 | 517 | /* Destination pointer is updated according to the address modifier, inc */ |
simon | 0:1014af42efd9 | 518 | pOut += inc; |
simon | 0:1014af42efd9 | 519 | |
simon | 0:1014af42efd9 | 520 | /* Update the inputA and inputB pointers for next MAC calculation */ |
simon | 0:1014af42efd9 | 521 | px = pIn1 + count; |
simon | 0:1014af42efd9 | 522 | py = pIn2; |
simon | 0:1014af42efd9 | 523 | |
simon | 0:1014af42efd9 | 524 | /* Increment the pointer pIn1 index, count by 1 */ |
simon | 0:1014af42efd9 | 525 | count++; |
simon | 0:1014af42efd9 | 526 | |
simon | 0:1014af42efd9 | 527 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 528 | blkCnt--; |
simon | 0:1014af42efd9 | 529 | } |
simon | 0:1014af42efd9 | 530 | } |
simon | 0:1014af42efd9 | 531 | |
simon | 0:1014af42efd9 | 532 | /* -------------------------- |
simon | 0:1014af42efd9 | 533 | * Initializations of stage3 |
simon | 0:1014af42efd9 | 534 | * -------------------------*/ |
simon | 0:1014af42efd9 | 535 | |
simon | 0:1014af42efd9 | 536 | /* sum += x[srcALen-srcBLen+1] * y[0] + x[srcALen-srcBLen+2] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
simon | 0:1014af42efd9 | 537 | * sum += x[srcALen-srcBLen+2] * y[0] + x[srcALen-srcBLen+3] * y[1] +...+ x[srcALen-1] * y[srcBLen-1] |
simon | 0:1014af42efd9 | 538 | * .... |
simon | 0:1014af42efd9 | 539 | * sum += x[srcALen-2] * y[0] + x[srcALen-1] * y[1] |
simon | 0:1014af42efd9 | 540 | * sum += x[srcALen-1] * y[0] |
simon | 0:1014af42efd9 | 541 | */ |
simon | 0:1014af42efd9 | 542 | |
simon | 0:1014af42efd9 | 543 | /* In this stage the MAC operations are decreased by 1 for every iteration. |
simon | 0:1014af42efd9 | 544 | The count variable holds the number of MAC operations performed */ |
simon | 0:1014af42efd9 | 545 | count = srcBLen - 1u; |
simon | 0:1014af42efd9 | 546 | |
simon | 0:1014af42efd9 | 547 | /* Working pointer of inputA */ |
simon | 0:1014af42efd9 | 548 | pSrc1 = pIn1 + (srcALen - (srcBLen - 1u)); |
simon | 0:1014af42efd9 | 549 | px = pSrc1; |
simon | 0:1014af42efd9 | 550 | |
simon | 0:1014af42efd9 | 551 | /* Working pointer of inputB */ |
simon | 0:1014af42efd9 | 552 | py = pIn2; |
simon | 0:1014af42efd9 | 553 | |
simon | 0:1014af42efd9 | 554 | /* ------------------- |
simon | 0:1014af42efd9 | 555 | * Stage3 process |
simon | 0:1014af42efd9 | 556 | * ------------------*/ |
simon | 0:1014af42efd9 | 557 | |
simon | 0:1014af42efd9 | 558 | while(blockSize3 > 0u) |
simon | 0:1014af42efd9 | 559 | { |
simon | 0:1014af42efd9 | 560 | /* Accumulator is made zero for every iteration */ |
simon | 0:1014af42efd9 | 561 | sum = 0.0f; |
simon | 0:1014af42efd9 | 562 | |
simon | 0:1014af42efd9 | 563 | /* Apply loop unrolling and compute 4 MACs simultaneously. */ |
simon | 0:1014af42efd9 | 564 | k = count >> 2u; |
simon | 0:1014af42efd9 | 565 | |
simon | 0:1014af42efd9 | 566 | /* First part of the processing with loop unrolling. Compute 4 MACs at a time. |
simon | 0:1014af42efd9 | 567 | ** a second loop below computes MACs for the remaining 1 to 3 samples. */ |
simon | 0:1014af42efd9 | 568 | while(k > 0u) |
simon | 0:1014af42efd9 | 569 | { |
simon | 0:1014af42efd9 | 570 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 571 | /* sum += x[srcALen - srcBLen + 4] * y[3] */ |
simon | 0:1014af42efd9 | 572 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 573 | /* sum += x[srcALen - srcBLen + 3] * y[2] */ |
simon | 0:1014af42efd9 | 574 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 575 | /* sum += x[srcALen - srcBLen + 2] * y[1] */ |
simon | 0:1014af42efd9 | 576 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 577 | /* sum += x[srcALen - srcBLen + 1] * y[0] */ |
simon | 0:1014af42efd9 | 578 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 579 | |
simon | 0:1014af42efd9 | 580 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 581 | k--; |
simon | 0:1014af42efd9 | 582 | } |
simon | 0:1014af42efd9 | 583 | |
simon | 0:1014af42efd9 | 584 | /* If the count is not a multiple of 4, compute any remaining MACs here. |
simon | 0:1014af42efd9 | 585 | ** No loop unrolling is used. */ |
simon | 0:1014af42efd9 | 586 | k = count % 0x4u; |
simon | 0:1014af42efd9 | 587 | |
simon | 0:1014af42efd9 | 588 | while(k > 0u) |
simon | 0:1014af42efd9 | 589 | { |
simon | 0:1014af42efd9 | 590 | /* Perform the multiply-accumulates */ |
simon | 0:1014af42efd9 | 591 | sum += *px++ * *py++; |
simon | 0:1014af42efd9 | 592 | |
simon | 0:1014af42efd9 | 593 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 594 | k--; |
simon | 0:1014af42efd9 | 595 | } |
simon | 0:1014af42efd9 | 596 | |
simon | 0:1014af42efd9 | 597 | /* Store the result in the accumulator in the destination buffer. */ |
simon | 0:1014af42efd9 | 598 | *pOut = sum; |
simon | 0:1014af42efd9 | 599 | /* Destination pointer is updated according to the address modifier, inc */ |
simon | 0:1014af42efd9 | 600 | pOut += inc; |
simon | 0:1014af42efd9 | 601 | |
simon | 0:1014af42efd9 | 602 | /* Update the inputA and inputB pointers for next MAC calculation */ |
simon | 0:1014af42efd9 | 603 | px = ++pSrc1; |
simon | 0:1014af42efd9 | 604 | py = pIn2; |
simon | 0:1014af42efd9 | 605 | |
simon | 0:1014af42efd9 | 606 | /* Decrement the MAC count */ |
simon | 0:1014af42efd9 | 607 | count--; |
simon | 0:1014af42efd9 | 608 | |
simon | 0:1014af42efd9 | 609 | /* Decrement the loop counter */ |
simon | 0:1014af42efd9 | 610 | blockSize3--; |
simon | 0:1014af42efd9 | 611 | } |
simon | 0:1014af42efd9 | 612 | |
simon | 0:1014af42efd9 | 613 | } |
simon | 0:1014af42efd9 | 614 | |
simon | 0:1014af42efd9 | 615 | /** |
simon | 0:1014af42efd9 | 616 | * @} end of Corr group |
simon | 0:1014af42efd9 | 617 | */ |