arm_conv_fast_opt_q15.c

00001 /* ----------------------------------------------------------------------    
00002 * Copyright (C) 2010-2013 ARM Limited. All rights reserved.    
00003 *    
00004 * $Date:        17. January 2013
00005 * $Revision:    V1.4.1
00006 *    
00007 * Project:      CMSIS DSP Library    
00008 * Title:        arm_conv_fast_opt_q15.c    
00009 *    
00010 * Description:  Fast Q15 Convolution.    
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 Conv    
00049  * @{    
00050  */
00051 
00052 /**    
00053  * @brief Convolution of Q15 sequences (fast version) for Cortex-M3 and Cortex-M4.    
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 srcALen+srcBLen-1.    
00059  * @param[in]  *pScratch1 points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2.   
00060  * @param[in]  *pScratch2 points to scratch buffer of size min(srcALen, srcBLen).   
00061  * @return none.    
00062  *    
00063  * \par Restrictions    
00064  *  If the silicon does not support unaligned memory access enable the macro UNALIGNED_SUPPORT_DISABLE    
00065  *  In this case input, output, scratch1 and scratch2 buffers should be aligned by 32-bit    
00066  *     
00067  * <b>Scaling and Overflow Behavior:</b>    
00068  *    
00069  * \par    
00070  * This fast version uses a 32-bit accumulator with 2.30 format.    
00071  * The accumulator maintains full precision of the intermediate multiplication results    
00072  * but provides only a single guard bit. There is no saturation on intermediate additions.    
00073  * Thus, if the accumulator overflows it wraps around and distorts the result.    
00074  * The input signals should be scaled down to avoid intermediate overflows.    
00075  * Scale down the inputs by log2(min(srcALen, srcBLen)) (log2 is read as log to the base 2) times to avoid overflows,    
00076  * as maximum of min(srcALen, srcBLen) number of additions are carried internally.    
00077  * The 2.30 accumulator is right shifted by 15 bits and then saturated to 1.15 format to yield the final result.    
00078  *    
00079  * \par    
00080  * See <code>arm_conv_q15()</code> for a slower implementation of this function which uses 64-bit accumulation to avoid wrap around distortion.    
00081  */
00082 
00083 void arm_conv_fast_opt_q15(
00084   q15_t * pSrcA,
00085   uint32_t srcALen,
00086   q15_t * pSrcB,
00087   uint32_t srcBLen,
00088   q15_t * pDst,
00089   q15_t * pScratch1,
00090   q15_t * pScratch2)
00091 {
00092   q31_t acc0, acc1, acc2, acc3;                  /* Accumulators */
00093   q31_t x1, x2, x3;                              /* Temporary variables to hold state and coefficient values */
00094   q31_t y1, y2;                                  /* State variables */
00095   q15_t *pOut = pDst;                            /* output pointer */
00096   q15_t *pScr1 = pScratch1;                      /* Temporary pointer for scratch1 */
00097   q15_t *pScr2 = pScratch2;                      /* Temporary pointer for scratch1 */
00098   q15_t *pIn1;                                   /* inputA pointer */
00099   q15_t *pIn2;                                   /* inputB pointer */
00100   q15_t *px;                                     /* Intermediate inputA pointer  */
00101   q15_t *py;                                     /* Intermediate inputB pointer  */
00102   uint32_t j, k, blkCnt;                         /* loop counter */
00103   uint32_t tapCnt;                               /* loop count */
00104 #ifdef UNALIGNED_SUPPORT_DISABLE
00105 
00106   q15_t a, b;
00107 
00108 #endif  /*  #ifdef UNALIGNED_SUPPORT_DISABLE    */
00109 
00110   /* The algorithm implementation is based on the lengths of the inputs. */
00111   /* srcB is always made to slide across srcA. */
00112   /* So srcBLen is always considered as shorter or equal to srcALen */
00113   if(srcALen >= srcBLen)
00114   {
00115     /* Initialization of inputA pointer */
00116     pIn1 = pSrcA;
00117 
00118     /* Initialization of inputB pointer */
00119     pIn2 = pSrcB;
00120   }
00121   else
00122   {
00123     /* Initialization of inputA pointer */
00124     pIn1 = pSrcB;
00125 
00126     /* Initialization of inputB pointer */
00127     pIn2 = pSrcA;
00128 
00129     /* srcBLen is always considered as shorter or equal to srcALen */
00130     j = srcBLen;
00131     srcBLen = srcALen;
00132     srcALen = j;
00133   }
00134 
00135   /* Pointer to take end of scratch2 buffer */
00136   pScr2 = pScratch2 + srcBLen - 1;
00137 
00138   /* points to smaller length sequence */
00139   px = pIn2;
00140 
00141   /* Apply loop unrolling and do 4 Copies simultaneously. */
00142   k = srcBLen >> 2u;
00143 
00144   /* First part of the processing with loop unrolling copies 4 data points at a time.       
00145    ** a second loop below copies for the remaining 1 to 3 samples. */
00146 
00147   /* Copy smaller length input sequence in reverse order into second scratch buffer */
00148   while(k > 0u)
00149   {
00150     /* copy second buffer in reversal manner */
00151     *pScr2-- = *px++;
00152     *pScr2-- = *px++;
00153     *pScr2-- = *px++;
00154     *pScr2-- = *px++;
00155 
00156     /* Decrement the loop counter */
00157     k--;
00158   }
00159 
00160   /* If the count is not a multiple of 4, copy remaining samples here.       
00161    ** No loop unrolling is used. */
00162   k = srcBLen % 0x4u;
00163 
00164   while(k > 0u)
00165   {
00166     /* copy second buffer in reversal manner for remaining samples */
00167     *pScr2-- = *px++;
00168 
00169     /* Decrement the loop counter */
00170     k--;
00171   }
00172 
00173   /* Initialze temporary scratch pointer */
00174   pScr1 = pScratch1;
00175 
00176   /* Assuming scratch1 buffer is aligned by 32-bit */
00177   /* Fill (srcBLen - 1u) zeros in scratch1 buffer */
00178   arm_fill_q15(0, pScr1, (srcBLen - 1u));
00179 
00180   /* Update temporary scratch pointer */
00181   pScr1 += (srcBLen - 1u);
00182 
00183   /* Copy bigger length sequence(srcALen) samples in scratch1 buffer */
00184 
00185 #ifndef UNALIGNED_SUPPORT_DISABLE
00186 
00187   /* Copy (srcALen) samples in scratch buffer */
00188   arm_copy_q15(pIn1, pScr1, srcALen);
00189 
00190   /* Update pointers */
00191   pScr1 += srcALen;
00192 
00193 #else
00194 
00195   /* Apply loop unrolling and do 4 Copies simultaneously. */
00196   k = srcALen >> 2u;
00197 
00198   /* First part of the processing with loop unrolling copies 4 data points at a time.       
00199    ** a second loop below copies for the remaining 1 to 3 samples. */
00200   while(k > 0u)
00201   {
00202     /* copy second buffer in reversal manner */
00203     *pScr1++ = *pIn1++;
00204     *pScr1++ = *pIn1++;
00205     *pScr1++ = *pIn1++;
00206     *pScr1++ = *pIn1++;
00207 
00208     /* Decrement the loop counter */
00209     k--;
00210   }
00211 
00212   /* If the count is not a multiple of 4, copy remaining samples here.       
00213    ** No loop unrolling is used. */
00214   k = srcALen % 0x4u;
00215 
00216   while(k > 0u)
00217   {
00218     /* copy second buffer in reversal manner for remaining samples */
00219     *pScr1++ = *pIn1++;
00220 
00221     /* Decrement the loop counter */
00222     k--;
00223   }
00224 
00225 #endif  /*  #ifndef UNALIGNED_SUPPORT_DISABLE   */
00226 
00227 
00228 #ifndef UNALIGNED_SUPPORT_DISABLE
00229 
00230   /* Fill (srcBLen - 1u) zeros at end of scratch buffer */
00231   arm_fill_q15(0, pScr1, (srcBLen - 1u));
00232 
00233   /* Update pointer */
00234   pScr1 += (srcBLen - 1u);
00235 
00236 #else
00237 
00238   /* Apply loop unrolling and do 4 Copies simultaneously. */
00239   k = (srcBLen - 1u) >> 2u;
00240 
00241   /* First part of the processing with loop unrolling copies 4 data points at a time.       
00242    ** a second loop below copies for the remaining 1 to 3 samples. */
00243   while(k > 0u)
00244   {
00245     /* copy second buffer in reversal manner */
00246     *pScr1++ = 0;
00247     *pScr1++ = 0;
00248     *pScr1++ = 0;
00249     *pScr1++ = 0;
00250 
00251     /* Decrement the loop counter */
00252     k--;
00253   }
00254 
00255   /* If the count is not a multiple of 4, copy remaining samples here.       
00256    ** No loop unrolling is used. */
00257   k = (srcBLen - 1u) % 0x4u;
00258 
00259   while(k > 0u)
00260   {
00261     /* copy second buffer in reversal manner for remaining samples */
00262     *pScr1++ = 0;
00263 
00264     /* Decrement the loop counter */
00265     k--;
00266   }
00267 
00268 #endif  /*  #ifndef UNALIGNED_SUPPORT_DISABLE   */
00269 
00270   /* Temporary pointer for scratch2 */
00271   py = pScratch2;
00272 
00273 
00274   /* Initialization of pIn2 pointer */
00275   pIn2 = py;
00276 
00277   /* First part of the processing with loop unrolling process 4 data points at a time.       
00278    ** a second loop below process for the remaining 1 to 3 samples. */
00279 
00280   /* Actual convolution process starts here */
00281   blkCnt = (srcALen + srcBLen - 1u) >> 2;
00282 
00283   while(blkCnt > 0)
00284   {
00285     /* Initialze temporary scratch pointer as scratch1 */
00286     pScr1 = pScratch1;
00287 
00288     /* Clear Accumlators */
00289     acc0 = 0;
00290     acc1 = 0;
00291     acc2 = 0;
00292     acc3 = 0;
00293 
00294     /* Read two samples from scratch1 buffer */
00295     x1 = *__SIMD32(pScr1)++;
00296 
00297     /* Read next two samples from scratch1 buffer */
00298     x2 = *__SIMD32(pScr1)++;
00299 
00300     tapCnt = (srcBLen) >> 2u;
00301 
00302     while(tapCnt > 0u)
00303     {
00304 
00305 #ifndef UNALIGNED_SUPPORT_DISABLE
00306 
00307       /* Read four samples from smaller buffer */
00308       y1 = _SIMD32_OFFSET(pIn2);
00309       y2 = _SIMD32_OFFSET(pIn2 + 2u);
00310 
00311       /* multiply and accumlate */
00312       acc0 = __SMLAD(x1, y1, acc0);
00313       acc2 = __SMLAD(x2, y1, acc2);
00314 
00315       /* pack input data */
00316 #ifndef ARM_MATH_BIG_ENDIAN
00317       x3 = __PKHBT(x2, x1, 0);
00318 #else
00319       x3 = __PKHBT(x1, x2, 0);
00320 #endif
00321 
00322       /* multiply and accumlate */
00323       acc1 = __SMLADX(x3, y1, acc1);
00324 
00325       /* Read next two samples from scratch1 buffer */
00326       x1 = _SIMD32_OFFSET(pScr1);
00327 
00328       /* multiply and accumlate */
00329       acc0 = __SMLAD(x2, y2, acc0);
00330       acc2 = __SMLAD(x1, y2, acc2);
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       acc1 = __SMLADX(x3, y2, acc1);
00341 
00342       x2 = _SIMD32_OFFSET(pScr1 + 2u);
00343 
00344 #ifndef ARM_MATH_BIG_ENDIAN
00345       x3 = __PKHBT(x2, x1, 0);
00346 #else
00347       x3 = __PKHBT(x1, x2, 0);
00348 #endif
00349 
00350       acc3 = __SMLADX(x3, y2, acc3);
00351 
00352 #else    
00353 
00354       /* Read four samples from smaller buffer */
00355       a = *pIn2;
00356       b = *(pIn2 + 1);
00357 
00358 #ifndef ARM_MATH_BIG_ENDIAN
00359       y1 = __PKHBT(a, b, 16);
00360 #else
00361       y1 = __PKHBT(b, a, 16);
00362 #endif
00363       
00364       a = *(pIn2 + 2);
00365       b = *(pIn2 + 3);
00366 #ifndef ARM_MATH_BIG_ENDIAN
00367       y2 = __PKHBT(a, b, 16);
00368 #else
00369       y2 = __PKHBT(b, a, 16);
00370 #endif              
00371 
00372       acc0 = __SMLAD(x1, y1, acc0);
00373 
00374       acc2 = __SMLAD(x2, y1, acc2);
00375 
00376 #ifndef ARM_MATH_BIG_ENDIAN
00377       x3 = __PKHBT(x2, x1, 0);
00378 #else
00379       x3 = __PKHBT(x1, x2, 0);
00380 #endif
00381 
00382       acc1 = __SMLADX(x3, y1, acc1);
00383 
00384       a = *pScr1;
00385       b = *(pScr1 + 1);
00386 
00387 #ifndef ARM_MATH_BIG_ENDIAN
00388       x1 = __PKHBT(a, b, 16);
00389 #else
00390       x1 = __PKHBT(b, a, 16);
00391 #endif
00392 
00393       acc0 = __SMLAD(x2, y2, acc0);
00394 
00395       acc2 = __SMLAD(x1, y2, acc2);
00396 
00397 #ifndef ARM_MATH_BIG_ENDIAN
00398       x3 = __PKHBT(x1, x2, 0);
00399 #else
00400       x3 = __PKHBT(x2, x1, 0);
00401 #endif
00402 
00403       acc3 = __SMLADX(x3, y1, acc3);
00404 
00405       acc1 = __SMLADX(x3, y2, acc1);
00406 
00407       a = *(pScr1 + 2);
00408       b = *(pScr1 + 3);
00409 
00410 #ifndef ARM_MATH_BIG_ENDIAN
00411       x2 = __PKHBT(a, b, 16);
00412 #else
00413       x2 = __PKHBT(b, a, 16);
00414 #endif
00415 
00416 #ifndef ARM_MATH_BIG_ENDIAN
00417       x3 = __PKHBT(x2, x1, 0);
00418 #else
00419       x3 = __PKHBT(x1, x2, 0);
00420 #endif
00421 
00422       acc3 = __SMLADX(x3, y2, acc3);
00423 
00424 #endif  /*  #ifndef UNALIGNED_SUPPORT_DISABLE   */
00425 
00426       /* update scratch pointers */
00427       pIn2 += 4u;
00428       pScr1 += 4u;
00429 
00430 
00431       /* Decrement the loop counter */
00432       tapCnt--;
00433     }
00434 
00435     /* Update scratch pointer for remaining samples of smaller length sequence */
00436     pScr1 -= 4u;
00437 
00438     /* apply same above for remaining samples of smaller length sequence */
00439     tapCnt = (srcBLen) & 3u;
00440 
00441     while(tapCnt > 0u)
00442     {
00443 
00444       /* accumlate the results */
00445       acc0 += (*pScr1++ * *pIn2);
00446       acc1 += (*pScr1++ * *pIn2);
00447       acc2 += (*pScr1++ * *pIn2);
00448       acc3 += (*pScr1++ * *pIn2++);
00449 
00450       pScr1 -= 3u;
00451 
00452       /* Decrement the loop counter */
00453       tapCnt--;
00454     }
00455 
00456     blkCnt--;
00457 
00458 
00459     /* Store the results in the accumulators in the destination buffer. */
00460 
00461 #ifndef ARM_MATH_BIG_ENDIAN
00462 
00463     *__SIMD32(pOut)++ =
00464       __PKHBT(__SSAT((acc0 >> 15), 16), __SSAT((acc1 >> 15), 16), 16);
00465 
00466     *__SIMD32(pOut)++ =
00467       __PKHBT(__SSAT((acc2 >> 15), 16), __SSAT((acc3 >> 15), 16), 16);
00468 
00469 
00470 #else
00471 
00472     *__SIMD32(pOut)++ =
00473       __PKHBT(__SSAT((acc1 >> 15), 16), __SSAT((acc0 >> 15), 16), 16);
00474 
00475     *__SIMD32(pOut)++ =
00476       __PKHBT(__SSAT((acc3 >> 15), 16), __SSAT((acc2 >> 15), 16), 16);
00477 
00478 
00479 
00480 #endif /*      #ifndef ARM_MATH_BIG_ENDIAN       */
00481 
00482     /* Initialization of inputB pointer */
00483     pIn2 = py;
00484 
00485     pScratch1 += 4u;
00486 
00487   }
00488 
00489 
00490   blkCnt = (srcALen + srcBLen - 1u) & 0x3;
00491 
00492   /* Calculate convolution for remaining samples of Bigger length sequence */
00493   while(blkCnt > 0)
00494   {
00495     /* Initialze temporary scratch pointer as scratch1 */
00496     pScr1 = pScratch1;
00497 
00498     /* Clear Accumlators */
00499     acc0 = 0;
00500 
00501     tapCnt = (srcBLen) >> 1u;
00502 
00503     while(tapCnt > 0u)
00504     {
00505 
00506       acc0 += (*pScr1++ * *pIn2++);
00507       acc0 += (*pScr1++ * *pIn2++);
00508 
00509       /* Decrement the loop counter */
00510       tapCnt--;
00511     }
00512 
00513     tapCnt = (srcBLen) & 1u;
00514 
00515     /* apply same above for remaining samples of smaller length sequence */
00516     while(tapCnt > 0u)
00517     {
00518 
00519       /* accumlate the results */
00520       acc0 += (*pScr1++ * *pIn2++);
00521 
00522       /* Decrement the loop counter */
00523       tapCnt--;
00524     }
00525 
00526     blkCnt--;
00527 
00528     /* The result is in 2.30 format.  Convert to 1.15 with saturation.       
00529      ** Then store the output in the destination buffer. */
00530     *pOut++ = (q15_t) (__SSAT((acc0 >> 15), 16));
00531 
00532     /* Initialization of inputB pointer */
00533     pIn2 = py;
00534 
00535     pScratch1 += 1u;
00536 
00537   }
00538 
00539 }
00540 
00541 /**    
00542  * @} end of Conv group    
00543  */