arm_correlate_opt_q7.c

00001 /* ----------------------------------------------------------------------    
00002 * Copyright (C) 2010-2014 ARM Limited. All rights reserved.    
00003 *    
00004 * $Date:        12. March 2014
00005 * $Revision:    V1.4.3
00006 *    
00007 * Project:      CMSIS DSP Library    
00008 * Title:        arm_correlate_opt_q7.c    
00009 *    
00010 * Description:  Correlation of Q7 sequences.  
00011 *    
00012 * Target Processor: Cortex-M4/Cortex-M3
00013 *  
00014 * Redistribution and use in source and binary forms, with or without 
00015 * modification, are permitted provided that the following conditions
00016 * are met:
00017 *   - Redistributions of source code must retain the above copyright
00018 *     notice, this list of conditions and the following disclaimer.
00019 *   - Redistributions in binary form must reproduce the above copyright
00020 *     notice, this list of conditions and the following disclaimer in
00021 *     the documentation and/or other materials provided with the 
00022 *     distribution.
00023 *   - Neither the name of ARM LIMITED nor the names of its contributors
00024 *     may be used to endorse or promote products derived from this
00025 *     software without specific prior written permission.
00026 *
00027 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
00028 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
00029 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
00030 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 
00031 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
00032 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
00033 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
00034 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
00035 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
00036 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
00037 * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
00038 * POSSIBILITY OF SUCH DAMAGE.  
00039 * -------------------------------------------------------------------- */
00040 
00041 #include "arm_math.h"
00042 
00043 /**    
00044  * @ingroup groupFilters    
00045  */
00046 
00047 /**    
00048  * @addtogroup Corr    
00049  * @{    
00050  */
00051 
00052 /**    
00053  * @brief Correlation of Q7 sequences.    
00054  * @param[in] *pSrcA points to the first input sequence.    
00055  * @param[in] srcALen length of the first input sequence.    
00056  * @param[in] *pSrcB points to the second input sequence.    
00057  * @param[in] srcBLen length of the second input sequence.    
00058  * @param[out] *pDst points to the location where the output result is written.  Length 2 * max(srcALen, srcBLen) - 1.    
00059  * @param[in]  *pScratch1 points to scratch buffer(of type q15_t) of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2.    
00060  * @param[in]  *pScratch2 points to scratch buffer (of type q15_t) of size min(srcALen, srcBLen).    
00061  * @return none.    
00062  *    
00063  *    
00064  * \par Restrictions    
00065  *  If the silicon does not support unaligned memory access enable the macro UNALIGNED_SUPPORT_DISABLE    
00066  *  In this case input, output, scratch1 and scratch2 buffers should be aligned by 32-bit     
00067  *        
00068  * @details    
00069  * <b>Scaling and Overflow Behavior:</b>    
00070  *    
00071  * \par    
00072  * The function is implemented using a 32-bit internal accumulator.    
00073  * Both the inputs are represented in 1.7 format and multiplications yield a 2.14 result.    
00074  * The 2.14 intermediate results are accumulated in a 32-bit accumulator in 18.14 format.    
00075  * This approach provides 17 guard bits and there is no risk of overflow as long as <code>max(srcALen, srcBLen)<131072</code>.    
00076  * The 18.14 result is then truncated to 18.7 format by discarding the low 7 bits and saturated to 1.7 format.  
00077  *  
00078  * 
00079  */
00080 
00081 
00082 
00083 void arm_correlate_opt_q7(
00084   q7_t * pSrcA,
00085   uint32_t srcALen,
00086   q7_t * pSrcB,
00087   uint32_t srcBLen,
00088   q7_t * pDst,
00089   q15_t * pScratch1,
00090   q15_t * pScratch2)
00091 {
00092   q7_t *pOut = pDst;                             /* output pointer                */
00093   q15_t *pScr1 = pScratch1;                      /* Temporary pointer for scratch */
00094   q15_t *pScr2 = pScratch2;                      /* Temporary pointer for scratch */
00095   q7_t *pIn1;                                    /* inputA pointer                */
00096   q7_t *pIn2;                                    /* inputB pointer                */
00097   q15_t *py;                                     /* Intermediate inputB pointer   */
00098   q31_t acc0, acc1, acc2, acc3;                  /* Accumulators                  */
00099   uint32_t j, k = 0u, blkCnt;                    /* loop counter                  */
00100   int32_t inc = 1;                               /* output pointer increment          */
00101   uint32_t outBlockSize;                         /* loop counter                  */
00102   q15_t x4;                                      /* Temporary input variable      */
00103   uint32_t tapCnt;                               /* loop counter                  */
00104   q31_t x1, x2, x3, y1;                          /* Temporary input variables     */
00105 
00106   /* The algorithm implementation is based on the lengths of the inputs. */
00107   /* srcB is always made to slide across srcA. */
00108   /* So srcBLen is always considered as shorter or equal to srcALen */
00109   /* But CORR(x, y) is reverse of CORR(y, x) */
00110   /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */
00111   /* and the destination pointer modifier, inc is set to -1 */
00112   /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */
00113   /* But to improve the performance,        
00114    * we include zeroes in the output instead of zero padding either of the the inputs*/
00115   /* If srcALen > srcBLen,        
00116    * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */
00117   /* If srcALen < srcBLen,        
00118    * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */
00119   if(srcALen >= srcBLen)
00120   {
00121     /* Initialization of inputA pointer */
00122     pIn1 = (pSrcA);
00123 
00124     /* Initialization of inputB pointer */
00125     pIn2 = (pSrcB);
00126 
00127     /* Number of output samples is calculated */
00128     outBlockSize = (2u * srcALen) - 1u;
00129 
00130     /* When srcALen > srcBLen, zero padding is done to srcB        
00131      * to make their lengths equal.        
00132      * Instead, (outBlockSize - (srcALen + srcBLen - 1))        
00133      * number of output samples are made zero */
00134     j = outBlockSize - (srcALen + (srcBLen - 1u));
00135 
00136     /* Updating the pointer position to non zero value */
00137     pOut += j;
00138 
00139   }
00140   else
00141   {
00142     /* Initialization of inputA pointer */
00143     pIn1 = (pSrcB);
00144 
00145     /* Initialization of inputB pointer */
00146     pIn2 = (pSrcA);
00147 
00148     /* srcBLen is always considered as shorter or equal to srcALen */
00149     j = srcBLen;
00150     srcBLen = srcALen;
00151     srcALen = j;
00152 
00153     /* CORR(x, y) = Reverse order(CORR(y, x)) */
00154     /* Hence set the destination pointer to point to the last output sample */
00155     pOut = pDst + ((srcALen + srcBLen) - 2u);
00156 
00157     /* Destination address modifier is set to -1 */
00158     inc = -1;
00159 
00160   }
00161 
00162 
00163   /* Copy (srcBLen) samples in scratch buffer */
00164   k = srcBLen >> 2u;
00165 
00166   /* First part of the processing with loop unrolling copies 4 data points at a time.       
00167    ** a second loop below copies for the remaining 1 to 3 samples. */
00168   while(k > 0u)
00169   {
00170     /* copy second buffer in reversal manner */
00171     x4 = (q15_t) * pIn2++;
00172     *pScr2++ = x4;
00173     x4 = (q15_t) * pIn2++;
00174     *pScr2++ = x4;
00175     x4 = (q15_t) * pIn2++;
00176     *pScr2++ = x4;
00177     x4 = (q15_t) * pIn2++;
00178     *pScr2++ = x4;
00179 
00180     /* Decrement the loop counter */
00181     k--;
00182   }
00183 
00184   /* If the count is not a multiple of 4, copy remaining samples here.       
00185    ** No loop unrolling is used. */
00186   k = srcBLen % 0x4u;
00187 
00188   while(k > 0u)
00189   {
00190     /* copy second buffer in reversal manner for remaining samples */
00191     x4 = (q15_t) * pIn2++;
00192     *pScr2++ = x4;
00193 
00194     /* Decrement the loop counter */
00195     k--;
00196   }
00197 
00198   /* Fill (srcBLen - 1u) zeros in scratch buffer */
00199   arm_fill_q15(0, pScr1, (srcBLen - 1u));
00200 
00201   /* Update temporary scratch pointer */
00202   pScr1 += (srcBLen - 1u);
00203 
00204   /* Copy (srcALen) samples in scratch buffer */
00205   k = srcALen >> 2u;
00206 
00207   /* First part of the processing with loop unrolling copies 4 data points at a time.       
00208    ** a second loop below copies for the remaining 1 to 3 samples. */
00209   while(k > 0u)
00210   {
00211     /* copy second buffer in reversal manner */
00212     x4 = (q15_t) * pIn1++;
00213     *pScr1++ = x4;
00214     x4 = (q15_t) * pIn1++;
00215     *pScr1++ = x4;
00216     x4 = (q15_t) * pIn1++;
00217     *pScr1++ = x4;
00218     x4 = (q15_t) * pIn1++;
00219     *pScr1++ = x4;
00220 
00221     /* Decrement the loop counter */
00222     k--;
00223   }
00224 
00225   /* If the count is not a multiple of 4, copy remaining samples here.       
00226    ** No loop unrolling is used. */
00227   k = srcALen % 0x4u;
00228 
00229   while(k > 0u)
00230   {
00231     /* copy second buffer in reversal manner for remaining samples */
00232     x4 = (q15_t) * pIn1++;
00233     *pScr1++ = x4;
00234 
00235     /* Decrement the loop counter */
00236     k--;
00237   }
00238 
00239 #ifndef UNALIGNED_SUPPORT_DISABLE
00240 
00241   /* Fill (srcBLen - 1u) zeros at end of scratch buffer */
00242   arm_fill_q15(0, pScr1, (srcBLen - 1u));
00243 
00244   /* Update pointer */
00245   pScr1 += (srcBLen - 1u);
00246 
00247 #else
00248 
00249 /* Apply loop unrolling and do 4 Copies simultaneously. */
00250   k = (srcBLen - 1u) >> 2u;
00251 
00252   /* First part of the processing with loop unrolling copies 4 data points at a time.       
00253    ** a second loop below copies for the remaining 1 to 3 samples. */
00254   while(k > 0u)
00255   {
00256     /* copy second buffer in reversal manner */
00257     *pScr1++ = 0;
00258     *pScr1++ = 0;
00259     *pScr1++ = 0;
00260     *pScr1++ = 0;
00261 
00262     /* Decrement the loop counter */
00263     k--;
00264   }
00265 
00266   /* If the count is not a multiple of 4, copy remaining samples here.       
00267    ** No loop unrolling is used. */
00268   k = (srcBLen - 1u) % 0x4u;
00269 
00270   while(k > 0u)
00271   {
00272     /* copy second buffer in reversal manner for remaining samples */
00273     *pScr1++ = 0;
00274 
00275     /* Decrement the loop counter */
00276     k--;
00277   }
00278 
00279 #endif  /*  #ifndef UNALIGNED_SUPPORT_DISABLE   */
00280 
00281   /* Temporary pointer for second sequence */
00282   py = pScratch2;
00283 
00284   /* Initialization of pScr2 pointer */
00285   pScr2 = pScratch2;
00286 
00287   /* Actual correlation process starts here */
00288   blkCnt = (srcALen + srcBLen - 1u) >> 2;
00289 
00290   while(blkCnt > 0)
00291   {
00292     /* Initialze temporary scratch pointer as scratch1 */
00293     pScr1 = pScratch1;
00294 
00295     /* Clear Accumlators */
00296     acc0 = 0;
00297     acc1 = 0;
00298     acc2 = 0;
00299     acc3 = 0;
00300 
00301     /* Read two samples from scratch1 buffer */
00302     x1 = *__SIMD32(pScr1)++;
00303 
00304     /* Read next two samples from scratch1 buffer */
00305     x2 = *__SIMD32(pScr1)++;
00306 
00307     tapCnt = (srcBLen) >> 2u;
00308 
00309     while(tapCnt > 0u)
00310     {
00311 
00312       /* Read four samples from smaller buffer */
00313       y1 = _SIMD32_OFFSET(pScr2);
00314 
00315       /* multiply and accumlate */
00316       acc0 = __SMLAD(x1, y1, acc0);
00317       acc2 = __SMLAD(x2, y1, acc2);
00318 
00319       /* pack input data */
00320 #ifndef ARM_MATH_BIG_ENDIAN
00321       x3 = __PKHBT(x2, x1, 0);
00322 #else
00323       x3 = __PKHBT(x1, x2, 0);
00324 #endif
00325 
00326       /* multiply and accumlate */
00327       acc1 = __SMLADX(x3, y1, acc1);
00328 
00329       /* Read next two samples from scratch1 buffer */
00330       x1 = *__SIMD32(pScr1)++;
00331 
00332       /* pack input data */
00333 #ifndef ARM_MATH_BIG_ENDIAN
00334       x3 = __PKHBT(x1, x2, 0);
00335 #else
00336       x3 = __PKHBT(x2, x1, 0);
00337 #endif
00338 
00339       acc3 = __SMLADX(x3, y1, acc3);
00340 
00341       /* Read four samples from smaller buffer */
00342       y1 = _SIMD32_OFFSET(pScr2 + 2u);
00343 
00344       acc0 = __SMLAD(x2, y1, acc0);
00345 
00346       acc2 = __SMLAD(x1, y1, acc2);
00347 
00348       acc1 = __SMLADX(x3, y1, acc1);
00349 
00350       x2 = *__SIMD32(pScr1)++;
00351 
00352 #ifndef ARM_MATH_BIG_ENDIAN
00353       x3 = __PKHBT(x2, x1, 0);
00354 #else
00355       x3 = __PKHBT(x1, x2, 0);
00356 #endif
00357 
00358       acc3 = __SMLADX(x3, y1, acc3);
00359 
00360       pScr2 += 4u;
00361 
00362 
00363       /* Decrement the loop counter */
00364       tapCnt--;
00365     }
00366 
00367 
00368 
00369     /* Update scratch pointer for remaining samples of smaller length sequence */
00370     pScr1 -= 4u;
00371 
00372 
00373     /* apply same above for remaining samples of smaller length sequence */
00374     tapCnt = (srcBLen) & 3u;
00375 
00376     while(tapCnt > 0u)
00377     {
00378 
00379       /* accumlate the results */
00380       acc0 += (*pScr1++ * *pScr2);
00381       acc1 += (*pScr1++ * *pScr2);
00382       acc2 += (*pScr1++ * *pScr2);
00383       acc3 += (*pScr1++ * *pScr2++);
00384 
00385       pScr1 -= 3u;
00386 
00387       /* Decrement the loop counter */
00388       tapCnt--;
00389     }
00390 
00391     blkCnt--;
00392 
00393     /* Store the result in the accumulator in the destination buffer. */
00394     *pOut = (q7_t) (__SSAT(acc0 >> 7u, 8));
00395     pOut += inc;
00396     *pOut = (q7_t) (__SSAT(acc1 >> 7u, 8));
00397     pOut += inc;
00398     *pOut = (q7_t) (__SSAT(acc2 >> 7u, 8));
00399     pOut += inc;
00400     *pOut = (q7_t) (__SSAT(acc3 >> 7u, 8));
00401     pOut += inc;
00402 
00403     /* Initialization of inputB pointer */
00404     pScr2 = py;
00405 
00406     pScratch1 += 4u;
00407 
00408   }
00409 
00410 
00411   blkCnt = (srcALen + srcBLen - 1u) & 0x3;
00412 
00413   /* Calculate correlation for remaining samples of Bigger length sequence */
00414   while(blkCnt > 0)
00415   {
00416     /* Initialze temporary scratch pointer as scratch1 */
00417     pScr1 = pScratch1;
00418 
00419     /* Clear Accumlators */
00420     acc0 = 0;
00421 
00422     tapCnt = (srcBLen) >> 1u;
00423 
00424     while(tapCnt > 0u)
00425     {
00426       acc0 += (*pScr1++ * *pScr2++);
00427       acc0 += (*pScr1++ * *pScr2++);
00428 
00429       /* Decrement the loop counter */
00430       tapCnt--;
00431     }
00432 
00433     tapCnt = (srcBLen) & 1u;
00434 
00435     /* apply same above for remaining samples of smaller length sequence */
00436     while(tapCnt > 0u)
00437     {
00438 
00439       /* accumlate the results */
00440       acc0 += (*pScr1++ * *pScr2++);
00441 
00442       /* Decrement the loop counter */
00443       tapCnt--;
00444     }
00445 
00446     blkCnt--;
00447 
00448     /* Store the result in the accumulator in the destination buffer. */
00449     *pOut = (q7_t) (__SSAT(acc0 >> 7u, 8));
00450 
00451     pOut += inc;
00452 
00453     /* Initialization of inputB pointer */
00454     pScr2 = py;
00455 
00456     pScratch1 += 1u;
00457 
00458   }
00459 
00460 }
00461 
00462 /**    
00463  * @} end of Corr group    
00464  */