tmk_core/tool/mbed/mbed-sdk/libraries/dsp/cmsis_dsp/FilteringFunctions/arm_correlate_fast_opt_q15.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_fast_opt_q15.c
   9 *
  10 * Description:  Fast Q15 Correlation.
  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 Q15 sequences (fast version) for Cortex-M3 and Cortex-M4.
  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]  *pScratch points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2.
  60  * @return none.
  61  *
  62  *
  63  * \par Restrictions
  64  *  If the silicon does not support unaligned memory access enable the macro UNALIGNED_SUPPORT_DISABLE
  65  *      In this case input, output, scratch buffers should be aligned by 32-bit
  66  *
  67  *
  68  * <b>Scaling and Overflow Behavior:</b>
  69  *
  70  * \par
  71  * This fast version uses a 32-bit accumulator with 2.30 format.
  72  * The accumulator maintains full precision of the intermediate multiplication results but provides only a single guard bit.
  73  * There is no saturation on intermediate additions.
  74  * Thus, if the accumulator overflows it wraps around and distorts the result.
  75  * The input signals should be scaled down to avoid intermediate overflows.
  76  * Scale down one of the inputs by 1/min(srcALen, srcBLen) to avoid overflow since a
  77  * maximum of min(srcALen, srcBLen) number of additions is carried internally.
  78  * The 2.30 accumulator is right shifted by 15 bits and then saturated to 1.15 format to yield the final result.
  79  *
  80  * \par
  81  * See <code>arm_correlate_q15()</code> for a slower implementation of this function which uses a 64-bit accumulator to avoid wrap around distortion.
  82  */
  83
  84 void arm_correlate_fast_opt_q15(
  85   q15_t * pSrcA,
  86   uint32_t srcALen,
  87   q15_t * pSrcB,
  88   uint32_t srcBLen,
  89   q15_t * pDst,
  90   q15_t * pScratch)
  91 {
  92   q15_t *pIn1;                                   /* inputA pointer               */
  93   q15_t *pIn2;                                   /* inputB pointer               */
  94   q31_t acc0, acc1, acc2, acc3;                  /* Accumulators                  */
  95   q15_t *py;                                     /* Intermediate inputB pointer  */
  96   q31_t x1, x2, x3;                              /* temporary variables for holding input and coefficient values */
  97   uint32_t j, blkCnt, outBlockSize;              /* loop counter                 */
  98   int32_t inc = 1;                               /* Destination address modifier */
  99   uint32_t tapCnt;
 100   q31_t y1, y2;
 101   q15_t *pScr;                                   /* Intermediate pointers        */
 102   q15_t *pOut = pDst;                            /* output pointer               */
 103 #ifdef UNALIGNED_SUPPORT_DISABLE
 104
 105   q15_t a, b;
 106
 107 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 108
 109   /* The algorithm implementation is based on the lengths of the inputs. */
 110   /* srcB is always made to slide across srcA. */
 111   /* So srcBLen is always considered as shorter or equal to srcALen */
 112   /* But CORR(x, y) is reverse of CORR(y, x) */
 113   /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */
 114   /* and the destination pointer modifier, inc is set to -1 */
 115   /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */
 116   /* But to improve the performance,
 117    * we include zeroes in the output instead of zero padding either of the the inputs*/
 118   /* If srcALen > srcBLen,
 119    * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */
 120   /* If srcALen < srcBLen,
 121    * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */
 122   if(srcALen >= srcBLen)
 123   {
 124     /* Initialization of inputA pointer */
 125     pIn1 = (pSrcA);
 126
 127     /* Initialization of inputB pointer */
 128     pIn2 = (pSrcB);
 129
 130     /* Number of output samples is calculated */
 131     outBlockSize = (2u * srcALen) - 1u;
 132
 133     /* When srcALen > srcBLen, zero padding is done to srcB
 134      * to make their lengths equal.
 135      * Instead, (outBlockSize - (srcALen + srcBLen - 1))
 136      * number of output samples are made zero */
 137     j = outBlockSize - (srcALen + (srcBLen - 1u));
 138
 139     /* Updating the pointer position to non zero value */
 140     pOut += j;
 141
 142   }
 143   else
 144   {
 145     /* Initialization of inputA pointer */
 146     pIn1 = (pSrcB);
 147
 148     /* Initialization of inputB pointer */
 149     pIn2 = (pSrcA);
 150
 151     /* srcBLen is always considered as shorter or equal to srcALen */
 152     j = srcBLen;
 153     srcBLen = srcALen;
 154     srcALen = j;
 155
 156     /* CORR(x, y) = Reverse order(CORR(y, x)) */
 157     /* Hence set the destination pointer to point to the last output sample */
 158     pOut = pDst + ((srcALen + srcBLen) - 2u);
 159
 160     /* Destination address modifier is set to -1 */
 161     inc = -1;
 162
 163   }
 164
 165   pScr = pScratch;
 166
 167   /* Fill (srcBLen - 1u) zeros in scratch buffer */
 168   arm_fill_q15(0, pScr, (srcBLen - 1u));
 169
 170   /* Update temporary scratch pointer */
 171   pScr += (srcBLen - 1u);
 172
 173 #ifndef UNALIGNED_SUPPORT_DISABLE
 174
 175   /* Copy (srcALen) samples in scratch buffer */
 176   arm_copy_q15(pIn1, pScr, srcALen);
 177
 178   /* Update pointers */
 179   pScr += srcALen;
 180
 181 #else
 182
 183   /* Apply loop unrolling and do 4 Copies simultaneously. */
 184   j = srcALen >> 2u;
 185
 186   /* First part of the processing with loop unrolling copies 4 data points at a time.
 187    ** a second loop below copies for the remaining 1 to 3 samples. */
 188   while(j > 0u)
 189   {
 190     /* copy second buffer in reversal manner */
 191     *pScr++ = *pIn1++;
 192     *pScr++ = *pIn1++;
 193     *pScr++ = *pIn1++;
 194     *pScr++ = *pIn1++;
 195
 196     /* Decrement the loop counter */
 197     j--;
 198   }
 199
 200   /* If the count is not a multiple of 4, copy remaining samples here.
 201    ** No loop unrolling is used. */
 202   j = srcALen % 0x4u;
 203
 204   while(j > 0u)
 205   {
 206     /* copy second buffer in reversal manner for remaining samples */
 207     *pScr++ = *pIn1++;
 208
 209     /* Decrement the loop counter */
 210     j--;
 211   }
 212
 213 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 214
 215 #ifndef UNALIGNED_SUPPORT_DISABLE
 216
 217   /* Fill (srcBLen - 1u) zeros at end of scratch buffer */
 218   arm_fill_q15(0, pScr, (srcBLen - 1u));
 219
 220   /* Update pointer */
 221   pScr += (srcBLen - 1u);
 222
 223 #else
 224
 225 /* Apply loop unrolling and do 4 Copies simultaneously. */
 226   j = (srcBLen - 1u) >> 2u;
 227
 228   /* First part of the processing with loop unrolling copies 4 data points at a time.
 229    ** a second loop below copies for the remaining 1 to 3 samples. */
 230   while(j > 0u)
 231   {
 232     /* copy second buffer in reversal manner */
 233     *pScr++ = 0;
 234     *pScr++ = 0;
 235     *pScr++ = 0;
 236     *pScr++ = 0;
 237
 238     /* Decrement the loop counter */
 239     j--;
 240   }
 241
 242   /* If the count is not a multiple of 4, copy remaining samples here.
 243    ** No loop unrolling is used. */
 244   j = (srcBLen - 1u) % 0x4u;
 245
 246   while(j > 0u)
 247   {
 248     /* copy second buffer in reversal manner for remaining samples */
 249     *pScr++ = 0;
 250
 251     /* Decrement the loop counter */
 252     j--;
 253   }
 254
 255 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 256
 257   /* Temporary pointer for scratch2 */
 258   py = pIn2;
 259
 260
 261   /* Actual correlation process starts here */
 262   blkCnt = (srcALen + srcBLen - 1u) >> 2;
 263
 264   while(blkCnt > 0)
 265   {
 266     /* Initialze temporary scratch pointer as scratch1 */
 267     pScr = pScratch;
 268
 269     /* Clear Accumlators */
 270     acc0 = 0;
 271     acc1 = 0;
 272     acc2 = 0;
 273     acc3 = 0;
 274
 275     /* Read four samples from scratch1 buffer */
 276     x1 = *__SIMD32(pScr)++;
 277
 278     /* Read next four samples from scratch1 buffer */
 279     x2 = *__SIMD32(pScr)++;
 280
 281     tapCnt = (srcBLen) >> 2u;
 282
 283     while(tapCnt > 0u)
 284     {
 285
 286 #ifndef UNALIGNED_SUPPORT_DISABLE
 287
 288       /* Read four samples from smaller buffer */
 289       y1 = _SIMD32_OFFSET(pIn2);
 290       y2 = _SIMD32_OFFSET(pIn2 + 2u);
 291
 292       acc0 = __SMLAD(x1, y1, acc0);
 293
 294       acc2 = __SMLAD(x2, y1, acc2);
 295
 296 #ifndef ARM_MATH_BIG_ENDIAN
 297       x3 = __PKHBT(x2, x1, 0);
 298 #else
 299       x3 = __PKHBT(x1, x2, 0);
 300 #endif
 301
 302       acc1 = __SMLADX(x3, y1, acc1);
 303
 304       x1 = _SIMD32_OFFSET(pScr);
 305
 306       acc0 = __SMLAD(x2, y2, acc0);
 307
 308       acc2 = __SMLAD(x1, y2, acc2);
 309
 310 #ifndef ARM_MATH_BIG_ENDIAN
 311       x3 = __PKHBT(x1, x2, 0);
 312 #else
 313       x3 = __PKHBT(x2, x1, 0);
 314 #endif
 315
 316       acc3 = __SMLADX(x3, y1, acc3);
 317
 318       acc1 = __SMLADX(x3, y2, acc1);
 319
 320       x2 = _SIMD32_OFFSET(pScr + 2u);
 321
 322 #ifndef ARM_MATH_BIG_ENDIAN
 323       x3 = __PKHBT(x2, x1, 0);
 324 #else
 325       x3 = __PKHBT(x1, x2, 0);
 326 #endif
 327
 328       acc3 = __SMLADX(x3, y2, acc3);
 329 #else
 330
 331       /* Read four samples from smaller buffer */
 332           a = *pIn2;
 333           b = *(pIn2 + 1);
 334
 335 #ifndef ARM_MATH_BIG_ENDIAN
 336       y1 = __PKHBT(a, b, 16);
 337 #else
 338       y1 = __PKHBT(b, a, 16);
 339 #endif
 340
 341           a = *(pIn2 + 2);
 342           b = *(pIn2 + 3);
 343 #ifndef ARM_MATH_BIG_ENDIAN
 344       y2 = __PKHBT(a, b, 16);
 345 #else
 346       y2 = __PKHBT(b, a, 16);
 347 #endif
 348
 349       acc0 = __SMLAD(x1, y1, acc0);
 350
 351       acc2 = __SMLAD(x2, y1, acc2);
 352
 353 #ifndef ARM_MATH_BIG_ENDIAN
 354       x3 = __PKHBT(x2, x1, 0);
 355 #else
 356       x3 = __PKHBT(x1, x2, 0);
 357 #endif
 358
 359       acc1 = __SMLADX(x3, y1, acc1);
 360
 361           a = *pScr;
 362           b = *(pScr + 1);
 363
 364 #ifndef ARM_MATH_BIG_ENDIAN
 365       x1 = __PKHBT(a, b, 16);
 366 #else
 367       x1 = __PKHBT(b, a, 16);
 368 #endif
 369
 370       acc0 = __SMLAD(x2, y2, acc0);
 371
 372       acc2 = __SMLAD(x1, y2, acc2);
 373
 374 #ifndef ARM_MATH_BIG_ENDIAN
 375       x3 = __PKHBT(x1, x2, 0);
 376 #else
 377       x3 = __PKHBT(x2, x1, 0);
 378 #endif
 379
 380       acc3 = __SMLADX(x3, y1, acc3);
 381
 382       acc1 = __SMLADX(x3, y2, acc1);
 383
 384           a = *(pScr + 2);
 385           b = *(pScr + 3);
 386
 387 #ifndef ARM_MATH_BIG_ENDIAN
 388       x2 = __PKHBT(a, b, 16);
 389 #else
 390       x2 = __PKHBT(b, a, 16);
 391 #endif
 392
 393 #ifndef ARM_MATH_BIG_ENDIAN
 394       x3 = __PKHBT(x2, x1, 0);
 395 #else
 396       x3 = __PKHBT(x1, x2, 0);
 397 #endif
 398
 399       acc3 = __SMLADX(x3, y2, acc3);
 400
 401 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 402
 403       pIn2 += 4u;
 404
 405       pScr += 4u;
 406
 407
 408       /* Decrement the loop counter */
 409       tapCnt--;
 410     }
 411
 412
 413
 414     /* Update scratch pointer for remaining samples of smaller length sequence */
 415     pScr -= 4u;
 416
 417
 418     /* apply same above for remaining samples of smaller length sequence */
 419     tapCnt = (srcBLen) & 3u;
 420
 421     while(tapCnt > 0u)
 422     {
 423
 424       /* accumlate the results */
 425       acc0 += (*pScr++ * *pIn2);
 426       acc1 += (*pScr++ * *pIn2);
 427       acc2 += (*pScr++ * *pIn2);
 428       acc3 += (*pScr++ * *pIn2++);
 429
 430       pScr -= 3u;
 431
 432       /* Decrement the loop counter */
 433       tapCnt--;
 434     }
 435
 436     blkCnt--;
 437
 438
 439     /* Store the results in the accumulators in the destination buffer. */
 440     *pOut = (__SSAT(acc0 >> 15u, 16));
 441     pOut += inc;
 442     *pOut = (__SSAT(acc1 >> 15u, 16));
 443     pOut += inc;
 444     *pOut = (__SSAT(acc2 >> 15u, 16));
 445     pOut += inc;
 446     *pOut = (__SSAT(acc3 >> 15u, 16));
 447     pOut += inc;
 448
 449
 450     /* Initialization of inputB pointer */
 451     pIn2 = py;
 452
 453     pScratch += 4u;
 454
 455   }
 456
 457
 458   blkCnt = (srcALen + srcBLen - 1u) & 0x3;
 459
 460   /* Calculate correlation for remaining samples of Bigger length sequence */
 461   while(blkCnt > 0)
 462   {
 463     /* Initialze temporary scratch pointer as scratch1 */
 464     pScr = pScratch;
 465
 466     /* Clear Accumlators */
 467     acc0 = 0;
 468
 469     tapCnt = (srcBLen) >> 1u;
 470
 471     while(tapCnt > 0u)
 472     {
 473
 474       acc0 += (*pScr++ * *pIn2++);
 475       acc0 += (*pScr++ * *pIn2++);
 476
 477       /* Decrement the loop counter */
 478       tapCnt--;
 479     }
 480
 481     tapCnt = (srcBLen) & 1u;
 482
 483     /* apply same above for remaining samples of smaller length sequence */
 484     while(tapCnt > 0u)
 485     {
 486
 487       /* accumlate the results */
 488       acc0 += (*pScr++ * *pIn2++);
 489
 490       /* Decrement the loop counter */
 491       tapCnt--;
 492     }
 493
 494     blkCnt--;
 495
 496     /* Store the result in the accumulator in the destination buffer. */
 497
 498     *pOut = (q15_t) (__SSAT((acc0 >> 15), 16));
 499
 500     pOut += inc;
 501
 502     /* Initialization of inputB pointer */
 503     pIn2 = py;
 504
 505     pScratch += 1u;
 506
 507   }
 508 }
 509
 510 /**
 511  * @} end of Corr group
 512  */