arm_correlate_fast_q15.c

Go to the documentation of this file.

/* ----------------------------------------------------------------------      
* Copyright (C) 2011 ARM Limited. All rights reserved. 
*      
* $Date:        15. December 2011   
* $Revision:    V2.0.0  
*      
* Project:      Cortex-R DSP Library 
* Title:        arm_correlate_fast_q15.c      
*      
* Description:  Fast Q15 Correlation.      
*      
* Target Processor:          Cortex-R4/R5
*
* Version 1.0.0 2011/03/08
*     Alpha release.
*
* Version 1.0.1 2011/09/30
*     Beta release.
*
* Version 2.0.0 2011/12/15
*     Final release. 
* 
* -------------------------------------------------------------------- */     
#include "arm_math.h"     
     
void arm_correlate_fast_q15(     
  q15_t * pSrcA,     
  uint32_t srcALen,     
  q15_t * pSrcB,     
  uint32_t srcBLen,     
  q15_t * pDst,  
  q15_t * pScratch)     
{     
  q15_t *pIn1;                                   /* inputA pointer               */     
  q15_t *pIn2;                                   /* inputB pointer               */   
  q31_t acc0, acc1, acc2, acc3;                  /* Accumulators                  */     
  q15_t *py;                                     /* Intermediate inputB pointer  */     
  q31_t x1, x2, x3;                      /* temporary variables for holding input and coefficient values */     
  uint32_t j, blkCnt, outBlockSize;  /* loop counter                 */     
  int32_t inc = 1;                               /* Destination address modifier */     
  uint32_t tapCnt;  
  q31_t y1, y2;  
  q15_t *pScr;                                  /* Intermediate pointers        */   
  q15_t *pOut = pDst;                            /* output pointer               */    
  
  
  /* The algorithm implementation is based on the lengths of the inputs. */     
  /* srcB is always made to slide across srcA. */     
  /* So srcBLen is always considered as shorter or equal to srcALen */     
  /* But CORR(x, y) is reverse of CORR(y, x) */     
  /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */     
  /* and the destination pointer modifier, inc is set to -1 */     
  /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */     
  /* But to improve the performance,      
   * we include zeroes in the output instead of zero padding either of the the inputs*/     
  /* If srcALen > srcBLen,      
   * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */     
  /* If srcALen < srcBLen,      
   * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */     
  if(srcALen >= srcBLen)     
  {     
    /* Initialization of inputA pointer */     
    pIn1 = (pSrcA);     
     
    /* Initialization of inputB pointer */     
    pIn2 = (pSrcB);     
     
    /* Number of output samples is calculated */     
    outBlockSize = (2u * srcALen) - 1u;     
     
    /* When srcALen > srcBLen, zero padding is done to srcB      
     * to make their lengths equal.      
     * Instead, (outBlockSize - (srcALen + srcBLen - 1))      
     * number of output samples are made zero */     
    j = outBlockSize - (srcALen + (srcBLen - 1u));     
     
    /* Updating the pointer position to non zero value */    
    pOut += j;    
     
  }     
  else     
  {     
    /* Initialization of inputA pointer */     
    pIn1 = (pSrcB);     
     
    /* Initialization of inputB pointer */     
    pIn2 = (pSrcA);     
     
    /* srcBLen is always considered as shorter or equal to srcALen */     
    j = srcBLen;     
    srcBLen = srcALen;     
    srcALen = j;     
     
    /* CORR(x, y) = Reverse order(CORR(y, x)) */     
    /* Hence set the destination pointer to point to the last output sample */     
    pOut = pDst + ((srcALen + srcBLen) - 2u);     
     
    /* Destination address modifier is set to -1 */     
    inc = -1;     
     
  }  
    
  pScr = pScratch;  
  
    /* Fill (srcBLen - 1u) zeros in scratch buffer */  
  arm_fill_q15(0, pScr, (srcBLen - 1u));  
  
  /* Update temporary scratch pointer */  
  pScr += (srcBLen - 1u);  
  
#ifndef UNALIGNED_SUPPORT_DISABLE  
  
  /* Copy (srcALen) samples in scratch buffer */  
  arm_copy_q15(pIn1, pScr, srcALen);  
  
  /* Update pointers */  
  pScr += srcALen;  
  
#else  
  
  /* Apply loop unrolling and do 4 Copies simultaneously. */  
  j = srcALen >> 2u;  
  
  /* First part of the processing with loop unrolling copies 4 data points at a time.     
   ** a second loop below copies for the remaining 1 to 3 samples. */  
  while(j > 0u)  
  {  
    /* copy second buffer in reversal manner */  
    *pScr++ = *pIn1++;  
    *pScr++ = *pIn1++;  
    *pScr++ = *pIn1++;  
    *pScr++ = *pIn1++;  
  
    /* Decrement the loop counter */  
    j--;  
  }  
  
  /* If the count is not a multiple of 4, copy remaining samples here.     
   ** No loop unrolling is used. */  
  j = srcALen % 0x4u;  
  
  while(j > 0u)  
  {  
    /* copy second buffer in reversal manner for remaining samples */  
    *pScr++ = *pIn1++;  
  
    /* Decrement the loop counter */  
    j--;  
  }  
  
#endif  
  
#ifndef UNALIGNED_SUPPORT_DISABLE  
  
  /* Fill (srcBLen - 1u) zeros at end of scratch buffer */  
  arm_fill_q15(0, pScr, (srcBLen - 1u));  
  
  /* Update pointer */  
  pScr += (srcBLen - 1u);  
  
#else  
  
/* Apply loop unrolling and do 4 Copies simultaneously. */  
  j = (srcBLen - 1u) >> 2u;  
  
  /* First part of the processing with loop unrolling copies 4 data points at a time.     
   ** a second loop below copies for the remaining 1 to 3 samples. */  
  while(j > 0u)  
  {  
    /* copy second buffer in reversal manner */  
    *pScr++ = 0;  
    *pScr++ = 0;  
    *pScr++ = 0;  
    *pScr++ = 0;  
  
    /* Decrement the loop counter */  
    j--;  
  }  
  
  /* If the count is not a multiple of 4, copy remaining samples here.     
   ** No loop unrolling is used. */  
  j = (srcBLen - 1u) % 0x4u;  
  
  while(j > 0u)  
  {  
    /* copy second buffer in reversal manner for remaining samples */  
    *pScr++ = 0;  
  
    /* Decrement the loop counter */  
    j--;  
  }  
  
#endif  
  
  /* Temporary pointer for scratch2 */  
  py = pIn2;  
  
    
  /* Actual convolution process starts here */  
  blkCnt = (srcALen + srcBLen - 1u) >> 2;  
  
  while(blkCnt > 0)  
  {  
    /* Initialze temporary scratch pointer as scratch1 */  
    pScr = pScratch;  
  
    /* Clear Accumlators */  
    acc0 = 0;  
    acc1 = 0;  
    acc2 = 0;  
    acc3 = 0;  
  
    /* Read four samples from scratch1 buffer */  
    x1 = *__SIMD32(pScr)++;  
  
     /* Read next four samples from scratch1 buffer */  
    x2 = *__SIMD32(pScr)++;  
      
    tapCnt = (srcBLen) >> 2u;  
  
    while(tapCnt > 0u)  
    {       
        
      /* Read four samples from smaller buffer */  
      y1 = _SIMD32_OFFSET(pIn2);  
      y2 = _SIMD32_OFFSET(pIn2 + 2u);  
        
      acc0 = __SMLAD(x1, y1, acc0);  
  
      acc2 = __SMLAD(x2, y1, acc2);  
  
#ifndef ARM_MATH_BIG_ENDIAN     
      x3 = __PKHBT(x2, x1, 0);  
#else  
      x3 = __PKHBT(x1, x2, 0);  
#endif  
  
      acc1 = __SMLADX(x3, y1, acc1);  
  
      x1 = _SIMD32_OFFSET(pScr);  
  
      acc0 = __SMLAD(x2, y2, acc0);  
  
      acc2 = __SMLAD(x1, y2, acc2);  
  
#ifndef ARM_MATH_BIG_ENDIAN     
      x3 = __PKHBT(x1, x2, 0);  
#else  
      x3 = __PKHBT(x2, x1, 0);  
#endif  
  
      acc3 = __SMLADX(x3, y1, acc3);  
  
      acc1 = __SMLADX(x3, y2, acc1);  
  
      x2 = _SIMD32_OFFSET(pScr + 2u);  
  
#ifndef ARM_MATH_BIG_ENDIAN     
      x3 = __PKHBT(x2, x1, 0);  
#else  
      x3 = __PKHBT(x1, x2, 0);  
#endif  
  
      acc3 = __SMLADX(x3, y2, acc3);  
  
      pIn2 += 4u;  
  
      pScr += 4u;  
  
    
      /* Decrement the loop counter */  
      tapCnt--;  
    }  
  
      
  
    /* Update scratch pointer for remaining samples of smaller length sequence */  
    pScr -= 4u;  
  
  
    /* apply same above for remaining samples of smaller length sequence */  
    tapCnt = (srcBLen) & 3u;  
  
    while(tapCnt > 0u)  
    {  
  
      /* accumlate the results */  
      acc0 += (*pScr++ * *pIn2);  
      acc1 += (*pScr++ * *pIn2);  
      acc2 += (*pScr++ * *pIn2);  
      acc3 += (*pScr++ * *pIn2++);  
  
      pScr -= 3u;  
  
      /* Decrement the loop counter */  
      tapCnt--;  
    }  
  
    blkCnt--;  
  
  
    /* Store the results in the accumulators in the destination buffer. */  
#ifdef CCS  
    *pOut = (__SSATA(acc0, 15u, 16));  
    pOut += inc;  
    *pOut = (__SSATA(acc1, 15u, 16));  
    pOut += inc;  
    *pOut = (__SSATA(acc2, 15u, 16));  
    pOut += inc;  
    *pOut = (__SSATA(acc3, 15u, 16));  
    pOut += inc;  
#else  
    *pOut = (__SSAT(acc0 >> 15u, 16));  
    pOut += inc;  
    *pOut = (__SSAT(acc1 >> 15u, 16));  
    pOut += inc;  
    *pOut = (__SSAT(acc2 >> 15u, 16));  
    pOut += inc;  
    *pOut = (__SSAT(acc3 >> 15u, 16));  
    pOut += inc;  
  
#endif  
   
    /* Initialization of inputB pointer */  
    pIn2 = py;  
  
    pScratch += 4u;  
  
  }  
  
  
  blkCnt = (srcALen + srcBLen - 1u) & 0x3;  
  
  /* Calculate convolution for remaining samples of Bigger length sequence */  
  while(blkCnt > 0)  
  {  
    /* Initialze temporary scratch pointer as scratch1 */  
    pScr = pScratch;  
  
    /* Clear Accumlators */  
    acc0 = 0;  
      
    tapCnt = (srcBLen) >> 1u;  
  
    while(tapCnt > 0u)  
    {  
  
      /* Read next four samples from scratch1 buffer */  
      x1 = *__SIMD32(pScr)++;  
        
      /* Read four samples from smaller buffer */  
      y1 = *__SIMD32(pIn2)++;   
        
      acc0 = __SMLAD(x1, y1, acc0);  
  
      /* Decrement the loop counter */  
      tapCnt--;  
    }  
  
    tapCnt = (srcBLen) & 1u;  
  
    /* apply same above for remaining samples of smaller length sequence */  
    while(tapCnt > 0u)  
    {  
  
      /* accumlate the results */  
      acc0 += (*pScr++ * *pIn2++);  
  
      /* Decrement the loop counter */  
      tapCnt--;  
    }  
  
    blkCnt--;  
  
    /* Store the result in the accumulator in the destination buffer. */  
#ifdef CCS  
    *pOut = (q15_t) (__SSATA((acc0), 15, 16));  
#else  
    *pOut = (q15_t) (__SSAT((acc0 >> 15), 16));  
#endif  
    pOut += inc;  
  
    /* Initialization of inputB pointer */  
    pIn2 = py;  
  
    pScratch += 1u;  
  
  }  
}