tmk_core/tool/mbed/mbed-sdk/libraries/dsp/cmsis_dsp/FilteringFunctions/arm_conv_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_opt_q15.c
   9 *
  10 * Description:  Convolution of Q15 sequences.
  11 *
  12 * Target Processor: Cortex-M4/Cortex-M3/Cortex-M0
  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.
  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  *
  68  * @details
  69  * <b>Scaling and Overflow Behavior:</b>
  70  *
  71  * \par
  72  * The function is implemented using a 64-bit internal accumulator.
  73  * Both inputs are in 1.15 format and multiplications yield a 2.30 result.
  74  * The 2.30 intermediate results are accumulated in a 64-bit accumulator in 34.30 format.
  75  * This approach provides 33 guard bits and there is no risk of overflow.
  76  * The 34.30 result is then truncated to 34.15 format by discarding the low 15 bits and then saturated to 1.15 format.
  77  *
  78  *
  79  * \par
  80  * Refer to <code>arm_conv_fast_q15()</code> for a faster but less precise version of this function for Cortex-M3 and Cortex-M4.
  81  *
  82  *
  83  */
  84
  85 void arm_conv_opt_q15(
  86   q15_t * pSrcA,
  87   uint32_t srcALen,
  88   q15_t * pSrcB,
  89   uint32_t srcBLen,
  90   q15_t * pDst,
  91   q15_t * pScratch1,
  92   q15_t * pScratch2)
  93 {
  94   q63_t acc0, acc1, acc2, acc3;                  /* Accumulator */
  95   q31_t x1, x2, x3;                              /* Temporary variables to hold state and coefficient values */
  96   q31_t y1, y2;                                  /* State variables */
  97   q15_t *pOut = pDst;                            /* output pointer */
  98   q15_t *pScr1 = pScratch1;                      /* Temporary pointer for scratch1 */
  99   q15_t *pScr2 = pScratch2;                      /* Temporary pointer for scratch1 */
 100   q15_t *pIn1;                                   /* inputA pointer */
 101   q15_t *pIn2;                                   /* inputB pointer */
 102   q15_t *px;                                     /* Intermediate inputA pointer  */
 103   q15_t *py;                                     /* Intermediate inputB pointer  */
 104   uint32_t j, k, blkCnt;                         /* loop counter */
 105   uint32_t tapCnt;                               /* loop count */
 106 #ifdef UNALIGNED_SUPPORT_DISABLE
 107
 108   q15_t a, b;
 109
 110 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 111
 112   /* The algorithm implementation is based on the lengths of the inputs. */
 113   /* srcB is always made to slide across srcA. */
 114   /* So srcBLen is always considered as shorter or equal to srcALen */
 115   if(srcALen >= srcBLen)
 116   {
 117     /* Initialization of inputA pointer */
 118     pIn1 = pSrcA;
 119
 120     /* Initialization of inputB pointer */
 121     pIn2 = pSrcB;
 122
 123   }
 124   else
 125   {
 126     /* Initialization of inputA pointer */
 127     pIn1 = pSrcB;
 128
 129     /* Initialization of inputB pointer */
 130     pIn2 = pSrcA;
 131
 132     /* srcBLen is always considered as shorter or equal to srcALen */
 133     j = srcBLen;
 134     srcBLen = srcALen;
 135     srcALen = j;
 136   }
 137
 138   /* pointer to take end of scratch2 buffer */
 139   pScr2 = pScratch2 + srcBLen - 1;
 140
 141   /* points to smaller length sequence */
 142   px = pIn2;
 143
 144   /* Apply loop unrolling and do 4 Copies simultaneously. */
 145   k = srcBLen >> 2u;
 146
 147   /* First part of the processing with loop unrolling copies 4 data points at a time.
 148    ** a second loop below copies for the remaining 1 to 3 samples. */
 149   /* Copy smaller length input sequence in reverse order into second scratch buffer */
 150   while(k > 0u)
 151   {
 152     /* copy second buffer in reversal manner */
 153     *pScr2-- = *px++;
 154     *pScr2-- = *px++;
 155     *pScr2-- = *px++;
 156     *pScr2-- = *px++;
 157
 158     /* Decrement the loop counter */
 159     k--;
 160   }
 161
 162   /* If the count is not a multiple of 4, copy remaining samples here.
 163    ** No loop unrolling is used. */
 164   k = srcBLen % 0x4u;
 165
 166   while(k > 0u)
 167   {
 168     /* copy second buffer in reversal manner for remaining samples */
 169     *pScr2-- = *px++;
 170
 171     /* Decrement the loop counter */
 172     k--;
 173   }
 174
 175   /* Initialze temporary scratch pointer */
 176   pScr1 = pScratch1;
 177
 178   /* Assuming scratch1 buffer is aligned by 32-bit */
 179   /* Fill (srcBLen - 1u) zeros in scratch buffer */
 180   arm_fill_q15(0, pScr1, (srcBLen - 1u));
 181
 182   /* Update temporary scratch pointer */
 183   pScr1 += (srcBLen - 1u);
 184
 185   /* Copy bigger length sequence(srcALen) samples in scratch1 buffer */
 186
 187 #ifndef UNALIGNED_SUPPORT_DISABLE
 188
 189   /* Copy (srcALen) samples in scratch buffer */
 190   arm_copy_q15(pIn1, pScr1, srcALen);
 191
 192   /* Update pointers */
 193   pScr1 += srcALen;
 194
 195 #else
 196
 197   /* Apply loop unrolling and do 4 Copies simultaneously. */
 198   k = srcALen >> 2u;
 199
 200   /* First part of the processing with loop unrolling copies 4 data points at a time.
 201    ** a second loop below copies for the remaining 1 to 3 samples. */
 202   while(k > 0u)
 203   {
 204     /* copy second buffer in reversal manner */
 205     *pScr1++ = *pIn1++;
 206     *pScr1++ = *pIn1++;
 207     *pScr1++ = *pIn1++;
 208     *pScr1++ = *pIn1++;
 209
 210     /* Decrement the loop counter */
 211     k--;
 212   }
 213
 214   /* If the count is not a multiple of 4, copy remaining samples here.
 215    ** No loop unrolling is used. */
 216   k = srcALen % 0x4u;
 217
 218   while(k > 0u)
 219   {
 220     /* copy second buffer in reversal manner for remaining samples */
 221     *pScr1++ = *pIn1++;
 222
 223     /* Decrement the loop counter */
 224     k--;
 225   }
 226
 227 #endif
 228
 229
 230 #ifndef UNALIGNED_SUPPORT_DISABLE
 231
 232   /* Fill (srcBLen - 1u) zeros at end of scratch buffer */
 233   arm_fill_q15(0, pScr1, (srcBLen - 1u));
 234
 235   /* Update pointer */
 236   pScr1 += (srcBLen - 1u);
 237
 238 #else
 239
 240   /* Apply loop unrolling and do 4 Copies simultaneously. */
 241   k = (srcBLen - 1u) >> 2u;
 242
 243   /* First part of the processing with loop unrolling copies 4 data points at a time.
 244    ** a second loop below copies for the remaining 1 to 3 samples. */
 245   while(k > 0u)
 246   {
 247     /* copy second buffer in reversal manner */
 248     *pScr1++ = 0;
 249     *pScr1++ = 0;
 250     *pScr1++ = 0;
 251     *pScr1++ = 0;
 252
 253     /* Decrement the loop counter */
 254     k--;
 255   }
 256
 257   /* If the count is not a multiple of 4, copy remaining samples here.
 258    ** No loop unrolling is used. */
 259   k = (srcBLen - 1u) % 0x4u;
 260
 261   while(k > 0u)
 262   {
 263     /* copy second buffer in reversal manner for remaining samples */
 264     *pScr1++ = 0;
 265
 266     /* Decrement the loop counter */
 267     k--;
 268   }
 269
 270 #endif
 271
 272   /* Temporary pointer for scratch2 */
 273   py = pScratch2;
 274
 275
 276   /* Initialization of pIn2 pointer */
 277   pIn2 = py;
 278
 279   /* First part of the processing with loop unrolling process 4 data points at a time.
 280    ** a second loop below process for the remaining 1 to 3 samples. */
 281
 282   /* Actual convolution process starts here */
 283   blkCnt = (srcALen + srcBLen - 1u) >> 2;
 284
 285   while(blkCnt > 0)
 286   {
 287     /* Initialze temporary scratch pointer as scratch1 */
 288     pScr1 = pScratch1;
 289
 290     /* Clear Accumlators */
 291     acc0 = 0;
 292     acc1 = 0;
 293     acc2 = 0;
 294     acc3 = 0;
 295
 296     /* Read two samples from scratch1 buffer */
 297     x1 = *__SIMD32(pScr1)++;
 298
 299     /* Read next two samples from scratch1 buffer */
 300     x2 = *__SIMD32(pScr1)++;
 301
 302     tapCnt = (srcBLen) >> 2u;
 303
 304     while(tapCnt > 0u)
 305     {
 306
 307 #ifndef UNALIGNED_SUPPORT_DISABLE
 308
 309       /* Read four samples from smaller buffer */
 310       y1 = _SIMD32_OFFSET(pIn2);
 311       y2 = _SIMD32_OFFSET(pIn2 + 2u);
 312
 313       /* multiply and accumlate */
 314       acc0 = __SMLALD(x1, y1, acc0);
 315       acc2 = __SMLALD(x2, y1, acc2);
 316
 317       /* pack input data */
 318 #ifndef ARM_MATH_BIG_ENDIAN
 319       x3 = __PKHBT(x2, x1, 0);
 320 #else
 321       x3 = __PKHBT(x1, x2, 0);
 322 #endif
 323
 324       /* multiply and accumlate */
 325       acc1 = __SMLALDX(x3, y1, acc1);
 326
 327       /* Read next two samples from scratch1 buffer */
 328       x1 = _SIMD32_OFFSET(pScr1);
 329
 330       /* multiply and accumlate */
 331       acc0 = __SMLALD(x2, y2, acc0);
 332       acc2 = __SMLALD(x1, y2, acc2);
 333
 334       /* pack input data */
 335 #ifndef ARM_MATH_BIG_ENDIAN
 336       x3 = __PKHBT(x1, x2, 0);
 337 #else
 338       x3 = __PKHBT(x2, x1, 0);
 339 #endif
 340
 341       acc3 = __SMLALDX(x3, y1, acc3);
 342       acc1 = __SMLALDX(x3, y2, acc1);
 343
 344       x2 = _SIMD32_OFFSET(pScr1 + 2u);
 345
 346 #ifndef ARM_MATH_BIG_ENDIAN
 347       x3 = __PKHBT(x2, x1, 0);
 348 #else
 349       x3 = __PKHBT(x1, x2, 0);
 350 #endif
 351
 352       acc3 = __SMLALDX(x3, y2, acc3);
 353
 354 #else
 355
 356       /* Read four samples from smaller buffer */
 357           a = *pIn2;
 358           b = *(pIn2 + 1);
 359
 360 #ifndef ARM_MATH_BIG_ENDIAN
 361       y1 = __PKHBT(a, b, 16);
 362 #else
 363       y1 = __PKHBT(b, a, 16);
 364 #endif
 365
 366           a = *(pIn2 + 2);
 367           b = *(pIn2 + 3);
 368 #ifndef ARM_MATH_BIG_ENDIAN
 369       y2 = __PKHBT(a, b, 16);
 370 #else
 371       y2 = __PKHBT(b, a, 16);
 372 #endif
 373
 374       acc0 = __SMLALD(x1, y1, acc0);
 375
 376       acc2 = __SMLALD(x2, y1, acc2);
 377
 378 #ifndef ARM_MATH_BIG_ENDIAN
 379       x3 = __PKHBT(x2, x1, 0);
 380 #else
 381       x3 = __PKHBT(x1, x2, 0);
 382 #endif
 383
 384       acc1 = __SMLALDX(x3, y1, acc1);
 385
 386           a = *pScr1;
 387           b = *(pScr1 + 1);
 388
 389 #ifndef ARM_MATH_BIG_ENDIAN
 390       x1 = __PKHBT(a, b, 16);
 391 #else
 392       x1 = __PKHBT(b, a, 16);
 393 #endif
 394
 395       acc0 = __SMLALD(x2, y2, acc0);
 396
 397       acc2 = __SMLALD(x1, y2, acc2);
 398
 399 #ifndef ARM_MATH_BIG_ENDIAN
 400       x3 = __PKHBT(x1, x2, 0);
 401 #else
 402       x3 = __PKHBT(x2, x1, 0);
 403 #endif
 404
 405       acc3 = __SMLALDX(x3, y1, acc3);
 406
 407       acc1 = __SMLALDX(x3, y2, acc1);
 408
 409           a = *(pScr1 + 2);
 410           b = *(pScr1 + 3);
 411
 412 #ifndef ARM_MATH_BIG_ENDIAN
 413       x2 = __PKHBT(a, b, 16);
 414 #else
 415       x2 = __PKHBT(b, a, 16);
 416 #endif
 417
 418 #ifndef ARM_MATH_BIG_ENDIAN
 419       x3 = __PKHBT(x2, x1, 0);
 420 #else
 421       x3 = __PKHBT(x1, x2, 0);
 422 #endif
 423
 424       acc3 = __SMLALDX(x3, y2, acc3);
 425
 426 #endif  /*      #ifndef UNALIGNED_SUPPORT_DISABLE       */
 427
 428       pIn2 += 4u;
 429       pScr1 += 4u;
 430
 431
 432       /* Decrement the loop counter */
 433       tapCnt--;
 434     }
 435
 436     /* Update scratch pointer for remaining samples of smaller length sequence */
 437     pScr1 -= 4u;
 438
 439     /* apply same above for remaining samples of smaller length sequence */
 440     tapCnt = (srcBLen) & 3u;
 441
 442     while(tapCnt > 0u)
 443     {
 444
 445       /* accumlate the results */
 446       acc0 += (*pScr1++ * *pIn2);
 447       acc1 += (*pScr1++ * *pIn2);
 448       acc2 += (*pScr1++ * *pIn2);
 449       acc3 += (*pScr1++ * *pIn2++);
 450
 451       pScr1 -= 3u;
 452
 453       /* Decrement the loop counter */
 454       tapCnt--;
 455     }
 456
 457     blkCnt--;
 458
 459
 460     /* Store the results in the accumulators in the destination buffer. */
 461
 462 #ifndef ARM_MATH_BIG_ENDIAN
 463
 464     *__SIMD32(pOut)++ =
 465       __PKHBT(__SSAT((acc0 >> 15), 16), __SSAT((acc1 >> 15), 16), 16);
 466
 467     *__SIMD32(pOut)++ =
 468       __PKHBT(__SSAT((acc2 >> 15), 16), __SSAT((acc3 >> 15), 16), 16);
 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 #endif /*      #ifndef ARM_MATH_BIG_ENDIAN       */
 480
 481     /* Initialization of inputB pointer */
 482     pIn2 = py;
 483
 484     pScratch1 += 4u;
 485
 486   }
 487
 488
 489   blkCnt = (srcALen + srcBLen - 1u) & 0x3;
 490
 491   /* Calculate convolution for remaining samples of Bigger length sequence */
 492   while(blkCnt > 0)
 493   {
 494     /* Initialze temporary scratch pointer as scratch1 */
 495     pScr1 = pScratch1;
 496
 497     /* Clear Accumlators */
 498     acc0 = 0;
 499
 500     tapCnt = (srcBLen) >> 1u;
 501
 502     while(tapCnt > 0u)
 503     {
 504
 505       /* Read next two samples from scratch1 buffer */
 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
 533     /* Initialization of inputB pointer */
 534     pIn2 = py;
 535
 536     pScratch1 += 1u;
 537
 538   }
 539
 540 }
 541
 542
 543 /**
 544  * @} end of Conv group
 545  */