tmk_core/tool/mbed/mbed-sdk/libraries/dsp/cmsis_dsp/FilteringFunctions/arm_correlate_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_opt_q15.c
   9 *
  10 * Description:  Correlation of Q15 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 Q15 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]  *pScratch points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2.
  60  * @return none.
  61  *
  62  * \par Restrictions
  63  *  If the silicon does not support unaligned memory access enable the macro UNALIGNED_SUPPORT_DISABLE
  64  *      In this case input, output, scratch buffers should be aligned by 32-bit
  65  *
  66  * @details
  67  * <b>Scaling and Overflow Behavior:</b>
  68  *
  69  * \par
  70  * The function is implemented using a 64-bit internal accumulator.
  71  * Both inputs are in 1.15 format and multiplications yield a 2.30 result.
  72  * The 2.30 intermediate results are accumulated in a 64-bit accumulator in 34.30 format.
  73  * This approach provides 33 guard bits and there is no risk of overflow.
  74  * The 34.30 result is then truncated to 34.15 format by discarding the low 15 bits and then saturated to 1.15 format.
  75  *
  76  * \par
  77  * Refer to <code>arm_correlate_fast_q15()</code> for a faster but less precise version of this function for Cortex-M3 and Cortex-M4.
  78  *
  79  *
  80  */
  81
  82
  83 void arm_correlate_opt_q15(
  84   q15_t * pSrcA,
  85   uint32_t srcALen,
  86   q15_t * pSrcB,
  87   uint32_t srcBLen,
  88   q15_t * pDst,
  89   q15_t * pScratch)
  90 {
  91   q15_t *pIn1;                                   /* inputA pointer               */
  92   q15_t *pIn2;                                   /* inputB pointer               */
  93   q63_t acc0, acc1, acc2, acc3;                  /* Accumulators                  */
  94   q15_t *py;                                     /* Intermediate inputB pointer  */
  95   q31_t x1, x2, x3;                              /* temporary variables for holding input1 and input2 values */
  96   uint32_t j, blkCnt, outBlockSize;              /* loop counter                 */
  97   int32_t inc = 1;                               /* output pointer increment     */
  98   uint32_t tapCnt;
  99   q31_t y1, y2;
 100   q15_t *pScr;                                   /* Intermediate pointers        */
 101   q15_t *pOut = pDst;                            /* output pointer               */
 102 #ifdef UNALIGNED_SUPPORT_DISABLE
 103
 104   q15_t a, b;
 105
 106 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 107
 108   /* The algorithm implementation is based on the lengths of the inputs. */
 109   /* srcB is always made to slide across srcA. */
 110   /* So srcBLen is always considered as shorter or equal to srcALen */
 111   /* But CORR(x, y) is reverse of CORR(y, x) */
 112   /* So, when srcBLen > srcALen, output pointer is made to point to the end of the output buffer */
 113   /* and the destination pointer modifier, inc is set to -1 */
 114   /* If srcALen > srcBLen, zero pad has to be done to srcB to make the two inputs of same length */
 115   /* But to improve the performance,
 116    * we include zeroes in the output instead of zero padding either of the the inputs*/
 117   /* If srcALen > srcBLen,
 118    * (srcALen - srcBLen) zeroes has to included in the starting of the output buffer */
 119   /* If srcALen < srcBLen,
 120    * (srcALen - srcBLen) zeroes has to included in the ending of the output buffer */
 121   if(srcALen >= srcBLen)
 122   {
 123     /* Initialization of inputA pointer */
 124     pIn1 = (pSrcA);
 125
 126     /* Initialization of inputB pointer */
 127     pIn2 = (pSrcB);
 128
 129     /* Number of output samples is calculated */
 130     outBlockSize = (2u * srcALen) - 1u;
 131
 132     /* When srcALen > srcBLen, zero padding is done to srcB
 133      * to make their lengths equal.
 134      * Instead, (outBlockSize - (srcALen + srcBLen - 1))
 135      * number of output samples are made zero */
 136     j = outBlockSize - (srcALen + (srcBLen - 1u));
 137
 138     /* Updating the pointer position to non zero value */
 139     pOut += j;
 140
 141   }
 142   else
 143   {
 144     /* Initialization of inputA pointer */
 145     pIn1 = (pSrcB);
 146
 147     /* Initialization of inputB pointer */
 148     pIn2 = (pSrcA);
 149
 150     /* srcBLen is always considered as shorter or equal to srcALen */
 151     j = srcBLen;
 152     srcBLen = srcALen;
 153     srcALen = j;
 154
 155     /* CORR(x, y) = Reverse order(CORR(y, x)) */
 156     /* Hence set the destination pointer to point to the last output sample */
 157     pOut = pDst + ((srcALen + srcBLen) - 2u);
 158
 159     /* Destination address modifier is set to -1 */
 160     inc = -1;
 161
 162   }
 163
 164   pScr = pScratch;
 165
 166   /* Fill (srcBLen - 1u) zeros in scratch buffer */
 167   arm_fill_q15(0, pScr, (srcBLen - 1u));
 168
 169   /* Update temporary scratch pointer */
 170   pScr += (srcBLen - 1u);
 171
 172 #ifndef UNALIGNED_SUPPORT_DISABLE
 173
 174   /* Copy (srcALen) samples in scratch buffer */
 175   arm_copy_q15(pIn1, pScr, srcALen);
 176
 177   /* Update pointers */
 178   //pIn1 += srcALen;
 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 = __SMLALD(x1, y1, acc0);
 293
 294       acc2 = __SMLALD(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 = __SMLALDX(x3, y1, acc1);
 303
 304       x1 = _SIMD32_OFFSET(pScr);
 305
 306       acc0 = __SMLALD(x2, y2, acc0);
 307
 308       acc2 = __SMLALD(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 = __SMLALDX(x3, y1, acc3);
 317
 318       acc1 = __SMLALDX(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 = __SMLALDX(x3, y2, acc3);
 329
 330 #else
 331
 332       /* Read four samples from smaller buffer */
 333           a = *pIn2;
 334           b = *(pIn2 + 1);
 335
 336 #ifndef ARM_MATH_BIG_ENDIAN
 337       y1 = __PKHBT(a, b, 16);
 338 #else
 339       y1 = __PKHBT(b, a, 16);
 340 #endif
 341
 342           a = *(pIn2 + 2);
 343           b = *(pIn2 + 3);
 344 #ifndef ARM_MATH_BIG_ENDIAN
 345       y2 = __PKHBT(a, b, 16);
 346 #else
 347       y2 = __PKHBT(b, a, 16);
 348 #endif
 349
 350       acc0 = __SMLALD(x1, y1, acc0);
 351
 352       acc2 = __SMLALD(x2, y1, acc2);
 353
 354 #ifndef ARM_MATH_BIG_ENDIAN
 355       x3 = __PKHBT(x2, x1, 0);
 356 #else
 357       x3 = __PKHBT(x1, x2, 0);
 358 #endif
 359
 360       acc1 = __SMLALDX(x3, y1, acc1);
 361
 362           a = *pScr;
 363           b = *(pScr + 1);
 364
 365 #ifndef ARM_MATH_BIG_ENDIAN
 366       x1 = __PKHBT(a, b, 16);
 367 #else
 368       x1 = __PKHBT(b, a, 16);
 369 #endif
 370
 371       acc0 = __SMLALD(x2, y2, acc0);
 372
 373       acc2 = __SMLALD(x1, y2, acc2);
 374
 375 #ifndef ARM_MATH_BIG_ENDIAN
 376       x3 = __PKHBT(x1, x2, 0);
 377 #else
 378       x3 = __PKHBT(x2, x1, 0);
 379 #endif
 380
 381       acc3 = __SMLALDX(x3, y1, acc3);
 382
 383       acc1 = __SMLALDX(x3, y2, acc1);
 384
 385           a = *(pScr + 2);
 386           b = *(pScr + 3);
 387
 388 #ifndef ARM_MATH_BIG_ENDIAN
 389       x2 = __PKHBT(a, b, 16);
 390 #else
 391       x2 = __PKHBT(b, a, 16);
 392 #endif
 393
 394 #ifndef ARM_MATH_BIG_ENDIAN
 395       x3 = __PKHBT(x2, x1, 0);
 396 #else
 397       x3 = __PKHBT(x1, x2, 0);
 398 #endif
 399
 400       acc3 = __SMLALDX(x3, y2, acc3);
 401
 402 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 403
 404       pIn2 += 4u;
 405
 406       pScr += 4u;
 407
 408
 409       /* Decrement the loop counter */
 410       tapCnt--;
 411     }
 412
 413
 414
 415     /* Update scratch pointer for remaining samples of smaller length sequence */
 416     pScr -= 4u;
 417
 418
 419     /* apply same above for remaining samples of smaller length sequence */
 420     tapCnt = (srcBLen) & 3u;
 421
 422     while(tapCnt > 0u)
 423     {
 424
 425       /* accumlate the results */
 426       acc0 += (*pScr++ * *pIn2);
 427       acc1 += (*pScr++ * *pIn2);
 428       acc2 += (*pScr++ * *pIn2);
 429       acc3 += (*pScr++ * *pIn2++);
 430
 431       pScr -= 3u;
 432
 433       /* Decrement the loop counter */
 434       tapCnt--;
 435     }
 436
 437     blkCnt--;
 438
 439
 440     /* Store the results in the accumulators in the destination buffer. */
 441     *pOut = (__SSAT(acc0 >> 15u, 16));
 442     pOut += inc;
 443     *pOut = (__SSAT(acc1 >> 15u, 16));
 444     pOut += inc;
 445     *pOut = (__SSAT(acc2 >> 15u, 16));
 446     pOut += inc;
 447     *pOut = (__SSAT(acc3 >> 15u, 16));
 448     pOut += inc;
 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     *pOut = (q15_t) (__SSAT((acc0 >> 15), 16));
 498
 499     pOut += inc;
 500
 501     /* Initialization of inputB pointer */
 502     pIn2 = py;
 503
 504     pScratch += 1u;
 505
 506   }
 507
 508
 509 }
 510
 511 /**
 512  * @} end of Corr group
 513  */