tmk_core/tool/mbed/mbed-sdk/libraries/dsp/cmsis_dsp/FilteringFunctions/arm_conv_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_conv_fast_opt_q15.c
   9 *
  10 * Description:  Fast Q15 Convolution.
  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 Conv
  49  * @{
  50  */
  51
  52 /**
  53  * @brief Convolution 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 srcALen+srcBLen-1.
  59  * @param[in]  *pScratch1 points to scratch buffer of size max(srcALen, srcBLen) + 2*min(srcALen, srcBLen) - 2.
  60  * @param[in]  *pScratch2 points to scratch buffer of size min(srcALen, srcBLen).
  61  * @return none.
  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, scratch1 and scratch2 buffers should be aligned by 32-bit
  66  *
  67  * <b>Scaling and Overflow Behavior:</b>
  68  *
  69  * \par
  70  * This fast version uses a 32-bit accumulator with 2.30 format.
  71  * The accumulator maintains full precision of the intermediate multiplication results
  72  * but provides only a single guard bit. There is no saturation on intermediate additions.
  73  * Thus, if the accumulator overflows it wraps around and distorts the result.
  74  * The input signals should be scaled down to avoid intermediate overflows.
  75  * Scale down the inputs by log2(min(srcALen, srcBLen)) (log2 is read as log to the base 2) times to avoid overflows,
  76  * as maximum of min(srcALen, srcBLen) number of additions are carried internally.
  77  * The 2.30 accumulator is right shifted by 15 bits and then saturated to 1.15 format to yield the final result.
  78  *
  79  * \par
  80  * See <code>arm_conv_q15()</code> for a slower implementation of this function which uses 64-bit accumulation to avoid wrap around distortion.
  81  */
  82
  83 void arm_conv_fast_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 * pScratch1,
  90   q15_t * pScratch2)
  91 {
  92   q31_t acc0, acc1, acc2, acc3;                  /* Accumulators */
  93   q31_t x1, x2, x3;                              /* Temporary variables to hold state and coefficient values */
  94   q31_t y1, y2;                                  /* State variables */
  95   q15_t *pOut = pDst;                            /* output pointer */
  96   q15_t *pScr1 = pScratch1;                      /* Temporary pointer for scratch1 */
  97   q15_t *pScr2 = pScratch2;                      /* Temporary pointer for scratch1 */
  98   q15_t *pIn1;                                   /* inputA pointer */
  99   q15_t *pIn2;                                   /* inputB pointer */
 100   q15_t *px;                                     /* Intermediate inputA pointer  */
 101   q15_t *py;                                     /* Intermediate inputB pointer  */
 102   uint32_t j, k, blkCnt;                         /* loop counter */
 103   uint32_t tapCnt;                               /* loop count */
 104 #ifdef UNALIGNED_SUPPORT_DISABLE
 105
 106   q15_t a, b;
 107
 108 #endif  /*      #ifdef UNALIGNED_SUPPORT_DISABLE        */
 109
 110   /* The algorithm implementation is based on the lengths of the inputs. */
 111   /* srcB is always made to slide across srcA. */
 112   /* So srcBLen is always considered as shorter or equal to srcALen */
 113   if(srcALen >= srcBLen)
 114   {
 115     /* Initialization of inputA pointer */
 116     pIn1 = pSrcA;
 117
 118     /* Initialization of inputB pointer */
 119     pIn2 = pSrcB;
 120   }
 121   else
 122   {
 123     /* Initialization of inputA pointer */
 124     pIn1 = pSrcB;
 125
 126     /* Initialization of inputB pointer */
 127     pIn2 = pSrcA;
 128
 129     /* srcBLen is always considered as shorter or equal to srcALen */
 130     j = srcBLen;
 131     srcBLen = srcALen;
 132     srcALen = j;
 133   }
 134
 135   /* Pointer to take end of scratch2 buffer */
 136   pScr2 = pScratch2 + srcBLen - 1;
 137
 138   /* points to smaller length sequence */
 139   px = pIn2;
 140
 141   /* Apply loop unrolling and do 4 Copies simultaneously. */
 142   k = srcBLen >> 2u;
 143
 144   /* First part of the processing with loop unrolling copies 4 data points at a time.
 145    ** a second loop below copies for the remaining 1 to 3 samples. */
 146
 147   /* Copy smaller length input sequence in reverse order into second scratch buffer */
 148   while(k > 0u)
 149   {
 150     /* copy second buffer in reversal manner */
 151     *pScr2-- = *px++;
 152     *pScr2-- = *px++;
 153     *pScr2-- = *px++;
 154     *pScr2-- = *px++;
 155
 156     /* Decrement the loop counter */
 157     k--;
 158   }
 159
 160   /* If the count is not a multiple of 4, copy remaining samples here.
 161    ** No loop unrolling is used. */
 162   k = srcBLen % 0x4u;
 163
 164   while(k > 0u)
 165   {
 166     /* copy second buffer in reversal manner for remaining samples */
 167     *pScr2-- = *px++;
 168
 169     /* Decrement the loop counter */
 170     k--;
 171   }
 172
 173   /* Initialze temporary scratch pointer */
 174   pScr1 = pScratch1;
 175
 176   /* Assuming scratch1 buffer is aligned by 32-bit */
 177   /* Fill (srcBLen - 1u) zeros in scratch1 buffer */
 178   arm_fill_q15(0, pScr1, (srcBLen - 1u));
 179
 180   /* Update temporary scratch pointer */
 181   pScr1 += (srcBLen - 1u);
 182
 183   /* Copy bigger length sequence(srcALen) samples in scratch1 buffer */
 184
 185 #ifndef UNALIGNED_SUPPORT_DISABLE
 186
 187   /* Copy (srcALen) samples in scratch buffer */
 188   arm_copy_q15(pIn1, pScr1, srcALen);
 189
 190   /* Update pointers */
 191   pScr1 += srcALen;
 192
 193 #else
 194
 195   /* Apply loop unrolling and do 4 Copies simultaneously. */
 196   k = srcALen >> 2u;
 197
 198   /* First part of the processing with loop unrolling copies 4 data points at a time.
 199    ** a second loop below copies for the remaining 1 to 3 samples. */
 200   while(k > 0u)
 201   {
 202     /* copy second buffer in reversal manner */
 203     *pScr1++ = *pIn1++;
 204     *pScr1++ = *pIn1++;
 205     *pScr1++ = *pIn1++;
 206     *pScr1++ = *pIn1++;
 207
 208     /* Decrement the loop counter */
 209     k--;
 210   }
 211
 212   /* If the count is not a multiple of 4, copy remaining samples here.
 213    ** No loop unrolling is used. */
 214   k = srcALen % 0x4u;
 215
 216   while(k > 0u)
 217   {
 218     /* copy second buffer in reversal manner for remaining samples */
 219     *pScr1++ = *pIn1++;
 220
 221     /* Decrement the loop counter */
 222     k--;
 223   }
 224
 225 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 226
 227
 228 #ifndef UNALIGNED_SUPPORT_DISABLE
 229
 230   /* Fill (srcBLen - 1u) zeros at end of scratch buffer */
 231   arm_fill_q15(0, pScr1, (srcBLen - 1u));
 232
 233   /* Update pointer */
 234   pScr1 += (srcBLen - 1u);
 235
 236 #else
 237
 238   /* Apply loop unrolling and do 4 Copies simultaneously. */
 239   k = (srcBLen - 1u) >> 2u;
 240
 241   /* First part of the processing with loop unrolling copies 4 data points at a time.
 242    ** a second loop below copies for the remaining 1 to 3 samples. */
 243   while(k > 0u)
 244   {
 245     /* copy second buffer in reversal manner */
 246     *pScr1++ = 0;
 247     *pScr1++ = 0;
 248     *pScr1++ = 0;
 249     *pScr1++ = 0;
 250
 251     /* Decrement the loop counter */
 252     k--;
 253   }
 254
 255   /* If the count is not a multiple of 4, copy remaining samples here.
 256    ** No loop unrolling is used. */
 257   k = (srcBLen - 1u) % 0x4u;
 258
 259   while(k > 0u)
 260   {
 261     /* copy second buffer in reversal manner for remaining samples */
 262     *pScr1++ = 0;
 263
 264     /* Decrement the loop counter */
 265     k--;
 266   }
 267
 268 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 269
 270   /* Temporary pointer for scratch2 */
 271   py = pScratch2;
 272
 273
 274   /* Initialization of pIn2 pointer */
 275   pIn2 = py;
 276
 277   /* First part of the processing with loop unrolling process 4 data points at a time.
 278    ** a second loop below process for the remaining 1 to 3 samples. */
 279
 280   /* Actual convolution process starts here */
 281   blkCnt = (srcALen + srcBLen - 1u) >> 2;
 282
 283   while(blkCnt > 0)
 284   {
 285     /* Initialze temporary scratch pointer as scratch1 */
 286     pScr1 = pScratch1;
 287
 288     /* Clear Accumlators */
 289     acc0 = 0;
 290     acc1 = 0;
 291     acc2 = 0;
 292     acc3 = 0;
 293
 294     /* Read two samples from scratch1 buffer */
 295     x1 = *__SIMD32(pScr1)++;
 296
 297     /* Read next two samples from scratch1 buffer */
 298     x2 = *__SIMD32(pScr1)++;
 299
 300     tapCnt = (srcBLen) >> 2u;
 301
 302     while(tapCnt > 0u)
 303     {
 304
 305 #ifndef UNALIGNED_SUPPORT_DISABLE
 306
 307       /* Read four samples from smaller buffer */
 308       y1 = _SIMD32_OFFSET(pIn2);
 309       y2 = _SIMD32_OFFSET(pIn2 + 2u);
 310
 311       /* multiply and accumlate */
 312       acc0 = __SMLAD(x1, y1, acc0);
 313       acc2 = __SMLAD(x2, y1, acc2);
 314
 315       /* pack input data */
 316 #ifndef ARM_MATH_BIG_ENDIAN
 317       x3 = __PKHBT(x2, x1, 0);
 318 #else
 319       x3 = __PKHBT(x1, x2, 0);
 320 #endif
 321
 322       /* multiply and accumlate */
 323       acc1 = __SMLADX(x3, y1, acc1);
 324
 325       /* Read next two samples from scratch1 buffer */
 326       x1 = _SIMD32_OFFSET(pScr1);
 327
 328       /* multiply and accumlate */
 329       acc0 = __SMLAD(x2, y2, acc0);
 330       acc2 = __SMLAD(x1, y2, acc2);
 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       acc1 = __SMLADX(x3, y2, acc1);
 341
 342       x2 = _SIMD32_OFFSET(pScr1 + 2u);
 343
 344 #ifndef ARM_MATH_BIG_ENDIAN
 345       x3 = __PKHBT(x2, x1, 0);
 346 #else
 347       x3 = __PKHBT(x1, x2, 0);
 348 #endif
 349
 350       acc3 = __SMLADX(x3, y2, acc3);
 351
 352 #else
 353
 354       /* Read four samples from smaller buffer */
 355           a = *pIn2;
 356           b = *(pIn2 + 1);
 357
 358 #ifndef ARM_MATH_BIG_ENDIAN
 359       y1 = __PKHBT(a, b, 16);
 360 #else
 361       y1 = __PKHBT(b, a, 16);
 362 #endif
 363
 364           a = *(pIn2 + 2);
 365           b = *(pIn2 + 3);
 366 #ifndef ARM_MATH_BIG_ENDIAN
 367       y2 = __PKHBT(a, b, 16);
 368 #else
 369       y2 = __PKHBT(b, a, 16);
 370 #endif
 371
 372       acc0 = __SMLAD(x1, y1, acc0);
 373
 374       acc2 = __SMLAD(x2, y1, acc2);
 375
 376 #ifndef ARM_MATH_BIG_ENDIAN
 377       x3 = __PKHBT(x2, x1, 0);
 378 #else
 379       x3 = __PKHBT(x1, x2, 0);
 380 #endif
 381
 382       acc1 = __SMLADX(x3, y1, acc1);
 383
 384           a = *pScr1;
 385           b = *(pScr1 + 1);
 386
 387 #ifndef ARM_MATH_BIG_ENDIAN
 388       x1 = __PKHBT(a, b, 16);
 389 #else
 390       x1 = __PKHBT(b, a, 16);
 391 #endif
 392
 393       acc0 = __SMLAD(x2, y2, acc0);
 394
 395       acc2 = __SMLAD(x1, y2, acc2);
 396
 397 #ifndef ARM_MATH_BIG_ENDIAN
 398       x3 = __PKHBT(x1, x2, 0);
 399 #else
 400       x3 = __PKHBT(x2, x1, 0);
 401 #endif
 402
 403       acc3 = __SMLADX(x3, y1, acc3);
 404
 405       acc1 = __SMLADX(x3, y2, acc1);
 406
 407           a = *(pScr1 + 2);
 408           b = *(pScr1 + 3);
 409
 410 #ifndef ARM_MATH_BIG_ENDIAN
 411       x2 = __PKHBT(a, b, 16);
 412 #else
 413       x2 = __PKHBT(b, a, 16);
 414 #endif
 415
 416 #ifndef ARM_MATH_BIG_ENDIAN
 417       x3 = __PKHBT(x2, x1, 0);
 418 #else
 419       x3 = __PKHBT(x1, x2, 0);
 420 #endif
 421
 422       acc3 = __SMLADX(x3, y2, acc3);
 423
 424 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 425
 426       /* update scratch pointers */
 427       pIn2 += 4u;
 428       pScr1 += 4u;
 429
 430
 431       /* Decrement the loop counter */
 432       tapCnt--;
 433     }
 434
 435     /* Update scratch pointer for remaining samples of smaller length sequence */
 436     pScr1 -= 4u;
 437
 438     /* apply same above for remaining samples of smaller length sequence */
 439     tapCnt = (srcBLen) & 3u;
 440
 441     while(tapCnt > 0u)
 442     {
 443
 444       /* accumlate the results */
 445       acc0 += (*pScr1++ * *pIn2);
 446       acc1 += (*pScr1++ * *pIn2);
 447       acc2 += (*pScr1++ * *pIn2);
 448       acc3 += (*pScr1++ * *pIn2++);
 449
 450       pScr1 -= 3u;
 451
 452       /* Decrement the loop counter */
 453       tapCnt--;
 454     }
 455
 456     blkCnt--;
 457
 458
 459     /* Store the results in the accumulators in the destination buffer. */
 460
 461 #ifndef ARM_MATH_BIG_ENDIAN
 462
 463     *__SIMD32(pOut)++ =
 464       __PKHBT(__SSAT((acc0 >> 15), 16), __SSAT((acc1 >> 15), 16), 16);
 465
 466     *__SIMD32(pOut)++ =
 467       __PKHBT(__SSAT((acc2 >> 15), 16), __SSAT((acc3 >> 15), 16), 16);
 468
 469
 470 #else
 471
 472     *__SIMD32(pOut)++ =
 473       __PKHBT(__SSAT((acc1 >> 15), 16), __SSAT((acc0 >> 15), 16), 16);
 474
 475     *__SIMD32(pOut)++ =
 476       __PKHBT(__SSAT((acc3 >> 15), 16), __SSAT((acc2 >> 15), 16), 16);
 477
 478
 479
 480 #endif /*      #ifndef ARM_MATH_BIG_ENDIAN       */
 481
 482     /* Initialization of inputB pointer */
 483     pIn2 = py;
 484
 485     pScratch1 += 4u;
 486
 487   }
 488
 489
 490   blkCnt = (srcALen + srcBLen - 1u) & 0x3;
 491
 492   /* Calculate convolution for remaining samples of Bigger length sequence */
 493   while(blkCnt > 0)
 494   {
 495     /* Initialze temporary scratch pointer as scratch1 */
 496     pScr1 = pScratch1;
 497
 498     /* Clear Accumlators */
 499     acc0 = 0;
 500
 501     tapCnt = (srcBLen) >> 1u;
 502
 503     while(tapCnt > 0u)
 504     {
 505
 506       acc0 += (*pScr1++ * *pIn2++);
 507       acc0 += (*pScr1++ * *pIn2++);
 508
 509       /* Decrement the loop counter */
 510       tapCnt--;
 511     }
 512
 513     tapCnt = (srcBLen) & 1u;
 514
 515     /* apply same above for remaining samples of smaller length sequence */
 516     while(tapCnt > 0u)
 517     {
 518
 519       /* accumlate the results */
 520       acc0 += (*pScr1++ * *pIn2++);
 521
 522       /* Decrement the loop counter */
 523       tapCnt--;
 524     }
 525
 526     blkCnt--;
 527
 528     /* The result is in 2.30 format.  Convert to 1.15 with saturation.
 529      ** Then store the output in the destination buffer. */
 530     *pOut++ = (q15_t) (__SSAT((acc0 >> 15), 16));
 531
 532     /* Initialization of inputB pointer */
 533     pIn2 = py;
 534
 535     pScratch1 += 1u;
 536
 537   }
 538
 539 }
 540
 541 /**
 542  * @} end of Conv group
 543  */