tmk_core/tool/mbed/mbed-sdk/libraries/dsp/cmsis_dsp/FilteringFunctions/arm_correlate_opt_q7.c

   1 /* ----------------------------------------------------------------------
   2 * Copyright (C) 2010-2013 ARM Limited. All rights reserved.
   3 *
   4 * $Date:        17. January 2013
   5 * $Revision:    V1.4.1
   6 *
   7 * Project:          CMSIS DSP Library
   8 * Title:                arm_correlate_opt_q7.c
   9 *
  10 * Description:  Correlation of Q7 sequences.
  11 *
  12 * Target Processor: Cortex-M4/Cortex-M3
  13 *
  14 * Redistribution and use in source and binary forms, with or without
  15 * modification, are permitted provided that the following conditions
  16 * are met:
  17 *   - Redistributions of source code must retain the above copyright
  18 *     notice, this list of conditions and the following disclaimer.
  19 *   - Redistributions in binary form must reproduce the above copyright
  20 *     notice, this list of conditions and the following disclaimer in
  21 *     the documentation and/or other materials provided with the
  22 *     distribution.
  23 *   - Neither the name of ARM LIMITED nor the names of its contributors
  24 *     may be used to endorse or promote products derived from this
  25 *     software without specific prior written permission.
  26 *
  27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  30 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
  31 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  32 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
  33 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  34 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  35 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
  37 * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  38 * POSSIBILITY OF SUCH DAMAGE.
  39 * -------------------------------------------------------------------- */
  40
  41 #include "arm_math.h"
  42
  43 /**
  44  * @ingroup groupFilters
  45  */
  46
  47 /**
  48  * @addtogroup Corr
  49  * @{
  50  */
  51
  52 /**
  53  * @brief Correlation of Q7 sequences.
  54  * @param[in] *pSrcA points to the first input sequence.
  55  * @param[in] srcALen length of the first input sequence.
  56  * @param[in] *pSrcB points to the second input sequence.
  57  * @param[in] srcBLen length of the second input sequence.
  58  * @param[out] *pDst points to the location where the output result is written.  Length 2 * max(srcALen, srcBLen) - 1.
  59  * @param[in]  *pScratch1 points to scratch buffer(of type q15_t) of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2.
  60  * @param[in]  *pScratch2 points to scratch buffer (of type q15_t) of size min(srcALen, srcBLen).
  61  * @return none.
  62  *
  63  *
  64  * \par Restrictions
  65  *  If the silicon does not support unaligned memory access enable the macro UNALIGNED_SUPPORT_DISABLE
  66  *      In this case input, output, scratch1 and scratch2 buffers should be aligned by 32-bit
  67  *
  68  * @details
  69  * <b>Scaling and Overflow Behavior:</b>
  70  *
  71  * \par
  72  * The function is implemented using a 32-bit internal accumulator.
  73  * Both the inputs are represented in 1.7 format and multiplications yield a 2.14 result.
  74  * The 2.14 intermediate results are accumulated in a 32-bit accumulator in 18.14 format.
  75  * This approach provides 17 guard bits and there is no risk of overflow as long as <code>max(srcALen, srcBLen)<131072</code>.
  76  * The 18.14 result is then truncated to 18.7 format by discarding the low 7 bits and saturated to 1.7 format.
  77  *
  78  *
  79  */
  80
  81
  82
  83 void arm_correlate_opt_q7(
  84   q7_t * pSrcA,
  85   uint32_t srcALen,
  86   q7_t * pSrcB,
  87   uint32_t srcBLen,
  88   q7_t * pDst,
  89   q15_t * pScratch1,
  90   q15_t * pScratch2)
  91 {
  92   q7_t *pOut = pDst;                             /* output pointer                */
  93   q15_t *pScr1 = pScratch1;                      /* Temporary pointer for scratch */
  94   q15_t *pScr2 = pScratch2;                      /* Temporary pointer for scratch */
  95   q7_t *pIn1;                                    /* inputA pointer                */
  96   q7_t *pIn2;                                    /* inputB pointer                */
  97   q15_t *py;                                     /* Intermediate inputB pointer   */
  98   q31_t acc0, acc1, acc2, acc3;                  /* Accumulators                  */
  99   uint32_t j, k = 0u, blkCnt;                    /* loop counter                  */
 100   int32_t inc = 1;                               /* output pointer increment          */
 101   uint32_t outBlockSize;                         /* loop counter                  */
 102   q15_t x4;                                      /* Temporary input variable      */
 103   uint32_t tapCnt;                               /* loop counter                  */
 104   q31_t x1, x2, x3, y1;                          /* Temporary input variables     */
 105
 106   /* The algorithm implementation is based on the lengths of the inputs. */
 107   /* srcB is always made to slide across srcA. */
 108   /* So srcBLen is always considered as shorter or equal to srcALen */
 109   /* But CORR(x, y) is reverse of CORR(y, x) */
 110   /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */
 111   /* and the destination pointer modifier, inc is set to -1 */
 112   /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */
 113   /* But to improve the performance,
 114    * we include zeroes in the output instead of zero padding either of the the inputs*/
 115   /* If srcALen > srcBLen,
 116    * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */
 117   /* If srcALen < srcBLen,
 118    * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */
 119   if(srcALen >= srcBLen)
 120   {
 121     /* Initialization of inputA pointer */
 122     pIn1 = (pSrcA);
 123
 124     /* Initialization of inputB pointer */
 125     pIn2 = (pSrcB);
 126
 127     /* Number of output samples is calculated */
 128     outBlockSize = (2u * srcALen) - 1u;
 129
 130     /* When srcALen > srcBLen, zero padding is done to srcB
 131      * to make their lengths equal.
 132      * Instead, (outBlockSize - (srcALen + srcBLen - 1))
 133      * number of output samples are made zero */
 134     j = outBlockSize - (srcALen + (srcBLen - 1u));
 135
 136     /* Updating the pointer position to non zero value */
 137     pOut += j;
 138
 139   }
 140   else
 141   {
 142     /* Initialization of inputA pointer */
 143     pIn1 = (pSrcB);
 144
 145     /* Initialization of inputB pointer */
 146     pIn2 = (pSrcA);
 147
 148     /* srcBLen is always considered as shorter or equal to srcALen */
 149     j = srcBLen;
 150     srcBLen = srcALen;
 151     srcALen = j;
 152
 153     /* CORR(x, y) = Reverse order(CORR(y, x)) */
 154     /* Hence set the destination pointer to point to the last output sample */
 155     pOut = pDst + ((srcALen + srcBLen) - 2u);
 156
 157     /* Destination address modifier is set to -1 */
 158     inc = -1;
 159
 160   }
 161
 162
 163   /* Copy (srcBLen) samples in scratch buffer */
 164   k = srcBLen >> 2u;
 165
 166   /* First part of the processing with loop unrolling copies 4 data points at a time.
 167    ** a second loop below copies for the remaining 1 to 3 samples. */
 168   while(k > 0u)
 169   {
 170     /* copy second buffer in reversal manner */
 171     x4 = (q15_t) * pIn2++;
 172     *pScr2++ = x4;
 173     x4 = (q15_t) * pIn2++;
 174     *pScr2++ = x4;
 175     x4 = (q15_t) * pIn2++;
 176     *pScr2++ = x4;
 177     x4 = (q15_t) * pIn2++;
 178     *pScr2++ = x4;
 179
 180     /* Decrement the loop counter */
 181     k--;
 182   }
 183
 184   /* If the count is not a multiple of 4, copy remaining samples here.
 185    ** No loop unrolling is used. */
 186   k = srcBLen % 0x4u;
 187
 188   while(k > 0u)
 189   {
 190     /* copy second buffer in reversal manner for remaining samples */
 191     x4 = (q15_t) * pIn2++;
 192     *pScr2++ = x4;
 193
 194     /* Decrement the loop counter */
 195     k--;
 196   }
 197
 198   /* Fill (srcBLen - 1u) zeros in scratch buffer */
 199   arm_fill_q15(0, pScr1, (srcBLen - 1u));
 200
 201   /* Update temporary scratch pointer */
 202   pScr1 += (srcBLen - 1u);
 203
 204   /* Copy (srcALen) samples in scratch buffer */
 205   k = srcALen >> 2u;
 206
 207   /* First part of the processing with loop unrolling copies 4 data points at a time.
 208    ** a second loop below copies for the remaining 1 to 3 samples. */
 209   while(k > 0u)
 210   {
 211     /* copy second buffer in reversal manner */
 212     x4 = (q15_t) * pIn1++;
 213     *pScr1++ = x4;
 214     x4 = (q15_t) * pIn1++;
 215     *pScr1++ = x4;
 216     x4 = (q15_t) * pIn1++;
 217     *pScr1++ = x4;
 218     x4 = (q15_t) * pIn1++;
 219     *pScr1++ = x4;
 220
 221     /* Decrement the loop counter */
 222     k--;
 223   }
 224
 225   /* If the count is not a multiple of 4, copy remaining samples here.
 226    ** No loop unrolling is used. */
 227   k = srcALen % 0x4u;
 228
 229   while(k > 0u)
 230   {
 231     /* copy second buffer in reversal manner for remaining samples */
 232     x4 = (q15_t) * pIn1++;
 233     *pScr1++ = x4;
 234
 235     /* Decrement the loop counter */
 236     k--;
 237   }
 238
 239 #ifndef UNALIGNED_SUPPORT_DISABLE
 240
 241   /* Fill (srcBLen - 1u) zeros at end of scratch buffer */
 242   arm_fill_q15(0, pScr1, (srcBLen - 1u));
 243
 244   /* Update pointer */
 245   pScr1 += (srcBLen - 1u);
 246
 247 #else
 248
 249 /* Apply loop unrolling and do 4 Copies simultaneously. */
 250   k = (srcBLen - 1u) >> 2u;
 251
 252   /* First part of the processing with loop unrolling copies 4 data points at a time.
 253    ** a second loop below copies for the remaining 1 to 3 samples. */
 254   while(k > 0u)
 255   {
 256     /* copy second buffer in reversal manner */
 257     *pScr1++ = 0;
 258     *pScr1++ = 0;
 259     *pScr1++ = 0;
 260     *pScr1++ = 0;
 261
 262     /* Decrement the loop counter */
 263     k--;
 264   }
 265
 266   /* If the count is not a multiple of 4, copy remaining samples here.
 267    ** No loop unrolling is used. */
 268   k = (srcBLen - 1u) % 0x4u;
 269
 270   while(k > 0u)
 271   {
 272     /* copy second buffer in reversal manner for remaining samples */
 273     *pScr1++ = 0;
 274
 275     /* Decrement the loop counter */
 276     k--;
 277   }
 278
 279 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 280
 281   /* Temporary pointer for second sequence */
 282   py = pScratch2;
 283
 284   /* Initialization of pScr2 pointer */
 285   pScr2 = pScratch2;
 286
 287   /* Actual correlation process starts here */
 288   blkCnt = (srcALen + srcBLen - 1u) >> 2;
 289
 290   while(blkCnt > 0)
 291   {
 292     /* Initialze temporary scratch pointer as scratch1 */
 293     pScr1 = pScratch1;
 294
 295     /* Clear Accumlators */
 296     acc0 = 0;
 297     acc1 = 0;
 298     acc2 = 0;
 299     acc3 = 0;
 300
 301     /* Read two samples from scratch1 buffer */
 302     x1 = *__SIMD32(pScr1)++;
 303
 304     /* Read next two samples from scratch1 buffer */
 305     x2 = *__SIMD32(pScr1)++;
 306
 307     tapCnt = (srcBLen) >> 2u;
 308
 309     while(tapCnt > 0u)
 310     {
 311
 312       /* Read four samples from smaller buffer */
 313       y1 = _SIMD32_OFFSET(pScr2);
 314
 315       /* multiply and accumlate */
 316       acc0 = __SMLAD(x1, y1, acc0);
 317       acc2 = __SMLAD(x2, y1, acc2);
 318
 319       /* pack input data */
 320 #ifndef ARM_MATH_BIG_ENDIAN
 321       x3 = __PKHBT(x2, x1, 0);
 322 #else
 323       x3 = __PKHBT(x1, x2, 0);
 324 #endif
 325
 326       /* multiply and accumlate */
 327       acc1 = __SMLADX(x3, y1, acc1);
 328
 329       /* Read next two samples from scratch1 buffer */
 330       x1 = *__SIMD32(pScr1)++;
 331
 332       /* pack input data */
 333 #ifndef ARM_MATH_BIG_ENDIAN
 334       x3 = __PKHBT(x1, x2, 0);
 335 #else
 336       x3 = __PKHBT(x2, x1, 0);
 337 #endif
 338
 339       acc3 = __SMLADX(x3, y1, acc3);
 340
 341       /* Read four samples from smaller buffer */
 342       y1 = _SIMD32_OFFSET(pScr2 + 2u);
 343
 344       acc0 = __SMLAD(x2, y1, acc0);
 345
 346       acc2 = __SMLAD(x1, y1, acc2);
 347
 348       acc1 = __SMLADX(x3, y1, acc1);
 349
 350       x2 = *__SIMD32(pScr1)++;
 351
 352 #ifndef ARM_MATH_BIG_ENDIAN
 353       x3 = __PKHBT(x2, x1, 0);
 354 #else
 355       x3 = __PKHBT(x1, x2, 0);
 356 #endif
 357
 358       acc3 = __SMLADX(x3, y1, acc3);
 359
 360       pScr2 += 4u;
 361
 362
 363       /* Decrement the loop counter */
 364       tapCnt--;
 365     }
 366
 367
 368
 369     /* Update scratch pointer for remaining samples of smaller length sequence */
 370     pScr1 -= 4u;
 371
 372
 373     /* apply same above for remaining samples of smaller length sequence */
 374     tapCnt = (srcBLen) & 3u;
 375
 376     while(tapCnt > 0u)
 377     {
 378
 379       /* accumlate the results */
 380       acc0 += (*pScr1++ * *pScr2);
 381       acc1 += (*pScr1++ * *pScr2);
 382       acc2 += (*pScr1++ * *pScr2);
 383       acc3 += (*pScr1++ * *pScr2++);
 384
 385       pScr1 -= 3u;
 386
 387       /* Decrement the loop counter */
 388       tapCnt--;
 389     }
 390
 391     blkCnt--;
 392
 393     /* Store the result in the accumulator in the destination buffer. */
 394     *pOut = (q7_t) (__SSAT(acc0 >> 7u, 8));
 395     pOut += inc;
 396     *pOut = (q7_t) (__SSAT(acc1 >> 7u, 8));
 397     pOut += inc;
 398     *pOut = (q7_t) (__SSAT(acc2 >> 7u, 8));
 399     pOut += inc;
 400     *pOut = (q7_t) (__SSAT(acc3 >> 7u, 8));
 401     pOut += inc;
 402
 403     /* Initialization of inputB pointer */
 404     pScr2 = py;
 405
 406     pScratch1 += 4u;
 407
 408   }
 409
 410
 411   blkCnt = (srcALen + srcBLen - 1u) & 0x3;
 412
 413   /* Calculate correlation for remaining samples of Bigger length sequence */
 414   while(blkCnt > 0)
 415   {
 416     /* Initialze temporary scratch pointer as scratch1 */
 417     pScr1 = pScratch1;
 418
 419     /* Clear Accumlators */
 420     acc0 = 0;
 421
 422     tapCnt = (srcBLen) >> 1u;
 423
 424     while(tapCnt > 0u)
 425     {
 426       acc0 += (*pScr1++ * *pScr2++);
 427       acc0 += (*pScr1++ * *pScr2++);
 428
 429       /* Decrement the loop counter */
 430       tapCnt--;
 431     }
 432
 433     tapCnt = (srcBLen) & 1u;
 434
 435     /* apply same above for remaining samples of smaller length sequence */
 436     while(tapCnt > 0u)
 437     {
 438
 439       /* accumlate the results */
 440       acc0 += (*pScr1++ * *pScr2++);
 441
 442       /* Decrement the loop counter */
 443       tapCnt--;
 444     }
 445
 446     blkCnt--;
 447
 448     /* Store the result in the accumulator in the destination buffer. */
 449     *pOut = (q7_t) (__SSAT(acc0 >> 7u, 8));
 450
 451     pOut += inc;
 452
 453     /* Initialization of inputB pointer */
 454     pScr2 = py;
 455
 456     pScratch1 += 1u;
 457
 458   }
 459
 460 }
 461
 462 /**
 463  * @} end of Corr group
 464  */