tmk_core/tool/mbed/mbed-sdk/libraries/dsp/cmsis_dsp/FilteringFunctions/arm_conv_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_conv_opt_q7.c
   9 *
  10 * Description:  Convolution of Q7 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 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 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  * \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  * @details
  68  * <b>Scaling and Overflow Behavior:</b>
  69  *
  70  * \par
  71  * The function is implemented using a 32-bit internal accumulator.
  72  * Both the inputs are represented in 1.7 format and multiplications yield a 2.14 result.
  73  * The 2.14 intermediate results are accumulated in a 32-bit accumulator in 18.14 format.
  74  * This approach provides 17 guard bits and there is no risk of overflow as long as <code>max(srcALen, srcBLen)<131072</code>.
  75  * The 18.14 result is then truncated to 18.7 format by discarding the low 7 bits and then saturated to 1.7 format.
  76  *
  77  */
  78
  79 void arm_conv_opt_q7(
  80   q7_t * pSrcA,
  81   uint32_t srcALen,
  82   q7_t * pSrcB,
  83   uint32_t srcBLen,
  84   q7_t * pDst,
  85   q15_t * pScratch1,
  86   q15_t * pScratch2)
  87 {
  88
  89   q15_t *pScr2, *pScr1;                          /* Intermediate pointers for scratch pointers */
  90   q15_t x4;                                      /* Temporary input variable */
  91   q7_t *pIn1, *pIn2;                             /* inputA and inputB pointer */
  92   uint32_t j, k, blkCnt, tapCnt;                 /* loop counter */
  93   q7_t *px;                                      /* Temporary input1 pointer */
  94   q15_t *py;                                     /* Temporary input2 pointer */
  95   q31_t acc0, acc1, acc2, acc3;                  /* Accumulator */
  96   q31_t x1, x2, x3, y1;                          /* Temporary input variables */
  97   q7_t *pOut = pDst;                             /* output pointer */
  98   q7_t out0, out1, out2, out3;                   /* temporary variables */
  99
 100   /* The algorithm implementation is based on the lengths of the inputs. */
 101   /* srcB is always made to slide across srcA. */
 102   /* So srcBLen is always considered as shorter or equal to srcALen */
 103   if(srcALen >= srcBLen)
 104   {
 105     /* Initialization of inputA pointer */
 106     pIn1 = pSrcA;
 107
 108     /* Initialization of inputB pointer */
 109     pIn2 = pSrcB;
 110   }
 111   else
 112   {
 113     /* Initialization of inputA pointer */
 114     pIn1 = pSrcB;
 115
 116     /* Initialization of inputB pointer */
 117     pIn2 = pSrcA;
 118
 119     /* srcBLen is always considered as shorter or equal to srcALen */
 120     j = srcBLen;
 121     srcBLen = srcALen;
 122     srcALen = j;
 123   }
 124
 125   /* pointer to take end of scratch2 buffer */
 126   pScr2 = pScratch2;
 127
 128   /* points to smaller length sequence */
 129   px = pIn2 + srcBLen - 1;
 130
 131   /* Apply loop unrolling and do 4 Copies simultaneously. */
 132   k = srcBLen >> 2u;
 133
 134   /* First part of the processing with loop unrolling copies 4 data points at a time.
 135    ** a second loop below copies for the remaining 1 to 3 samples. */
 136   while(k > 0u)
 137   {
 138     /* copy second buffer in reversal manner */
 139     x4 = (q15_t) * px--;
 140     *pScr2++ = x4;
 141     x4 = (q15_t) * px--;
 142     *pScr2++ = x4;
 143     x4 = (q15_t) * px--;
 144     *pScr2++ = x4;
 145     x4 = (q15_t) * px--;
 146     *pScr2++ = x4;
 147
 148     /* Decrement the loop counter */
 149     k--;
 150   }
 151
 152   /* If the count is not a multiple of 4, copy remaining samples here.
 153    ** No loop unrolling is used. */
 154   k = srcBLen % 0x4u;
 155
 156   while(k > 0u)
 157   {
 158     /* copy second buffer in reversal manner for remaining samples */
 159     x4 = (q15_t) * px--;
 160     *pScr2++ = x4;
 161
 162     /* Decrement the loop counter */
 163     k--;
 164   }
 165
 166   /* Initialze temporary scratch pointer */
 167   pScr1 = pScratch1;
 168
 169   /* Fill (srcBLen - 1u) zeros in scratch buffer */
 170   arm_fill_q15(0, pScr1, (srcBLen - 1u));
 171
 172   /* Update temporary scratch pointer */
 173   pScr1 += (srcBLen - 1u);
 174
 175   /* Copy (srcALen) samples in scratch buffer */
 176   /* Apply loop unrolling and do 4 Copies simultaneously. */
 177   k = srcALen >> 2u;
 178
 179   /* First part of the processing with loop unrolling copies 4 data points at a time.
 180    ** a second loop below copies for the remaining 1 to 3 samples. */
 181   while(k > 0u)
 182   {
 183     /* copy second buffer in reversal manner */
 184     x4 = (q15_t) * pIn1++;
 185     *pScr1++ = x4;
 186     x4 = (q15_t) * pIn1++;
 187     *pScr1++ = x4;
 188     x4 = (q15_t) * pIn1++;
 189     *pScr1++ = x4;
 190     x4 = (q15_t) * pIn1++;
 191     *pScr1++ = x4;
 192
 193     /* Decrement the loop counter */
 194     k--;
 195   }
 196
 197   /* If the count is not a multiple of 4, copy remaining samples here.
 198    ** No loop unrolling is used. */
 199   k = srcALen % 0x4u;
 200
 201   while(k > 0u)
 202   {
 203     /* copy second buffer in reversal manner for remaining samples */
 204     x4 = (q15_t) * pIn1++;
 205     *pScr1++ = x4;
 206
 207     /* Decrement the loop counter */
 208     k--;
 209   }
 210
 211 #ifndef UNALIGNED_SUPPORT_DISABLE
 212
 213   /* Fill (srcBLen - 1u) zeros at end of scratch buffer */
 214   arm_fill_q15(0, pScr1, (srcBLen - 1u));
 215
 216   /* Update pointer */
 217   pScr1 += (srcBLen - 1u);
 218
 219 #else
 220
 221   /* Apply loop unrolling and do 4 Copies simultaneously. */
 222   k = (srcBLen - 1u) >> 2u;
 223
 224   /* First part of the processing with loop unrolling copies 4 data points at a time.
 225    ** a second loop below copies for the remaining 1 to 3 samples. */
 226   while(k > 0u)
 227   {
 228     /* copy second buffer in reversal manner */
 229     *pScr1++ = 0;
 230     *pScr1++ = 0;
 231     *pScr1++ = 0;
 232     *pScr1++ = 0;
 233
 234     /* Decrement the loop counter */
 235     k--;
 236   }
 237
 238   /* If the count is not a multiple of 4, copy remaining samples here.
 239    ** No loop unrolling is used. */
 240   k = (srcBLen - 1u) % 0x4u;
 241
 242   while(k > 0u)
 243   {
 244     /* copy second buffer in reversal manner for remaining samples */
 245     *pScr1++ = 0;
 246
 247     /* Decrement the loop counter */
 248     k--;
 249   }
 250
 251 #endif
 252
 253   /* Temporary pointer for scratch2 */
 254   py = pScratch2;
 255
 256   /* Initialization of pIn2 pointer */
 257   pIn2 = (q7_t *) py;
 258
 259   pScr2 = py;
 260
 261   /* Actual convolution process starts here */
 262   blkCnt = (srcALen + srcBLen - 1u) >> 2;
 263
 264   while(blkCnt > 0)
 265   {
 266     /* Initialze temporary scratch pointer as scratch1 */
 267     pScr1 = pScratch1;
 268
 269     /* Clear Accumlators */
 270     acc0 = 0;
 271     acc1 = 0;
 272     acc2 = 0;
 273     acc3 = 0;
 274
 275     /* Read two samples from scratch1 buffer */
 276     x1 = *__SIMD32(pScr1)++;
 277
 278     /* Read next two samples from scratch1 buffer */
 279     x2 = *__SIMD32(pScr1)++;
 280
 281     tapCnt = (srcBLen) >> 2u;
 282
 283     while(tapCnt > 0u)
 284     {
 285
 286       /* Read four samples from smaller buffer */
 287       y1 = _SIMD32_OFFSET(pScr2);
 288
 289       /* multiply and accumlate */
 290       acc0 = __SMLAD(x1, y1, acc0);
 291       acc2 = __SMLAD(x2, y1, acc2);
 292
 293       /* pack input data */
 294 #ifndef ARM_MATH_BIG_ENDIAN
 295       x3 = __PKHBT(x2, x1, 0);
 296 #else
 297       x3 = __PKHBT(x1, x2, 0);
 298 #endif
 299
 300       /* multiply and accumlate */
 301       acc1 = __SMLADX(x3, y1, acc1);
 302
 303       /* Read next two samples from scratch1 buffer */
 304       x1 = *__SIMD32(pScr1)++;
 305
 306       /* pack input data */
 307 #ifndef ARM_MATH_BIG_ENDIAN
 308       x3 = __PKHBT(x1, x2, 0);
 309 #else
 310       x3 = __PKHBT(x2, x1, 0);
 311 #endif
 312
 313       acc3 = __SMLADX(x3, y1, acc3);
 314
 315       /* Read four samples from smaller buffer */
 316       y1 = _SIMD32_OFFSET(pScr2 + 2u);
 317
 318       acc0 = __SMLAD(x2, y1, acc0);
 319
 320       acc2 = __SMLAD(x1, y1, acc2);
 321
 322       acc1 = __SMLADX(x3, y1, acc1);
 323
 324       x2 = *__SIMD32(pScr1)++;
 325
 326 #ifndef ARM_MATH_BIG_ENDIAN
 327       x3 = __PKHBT(x2, x1, 0);
 328 #else
 329       x3 = __PKHBT(x1, x2, 0);
 330 #endif
 331
 332       acc3 = __SMLADX(x3, y1, acc3);
 333
 334       pScr2 += 4u;
 335
 336
 337       /* Decrement the loop counter */
 338       tapCnt--;
 339     }
 340
 341
 342
 343     /* Update scratch pointer for remaining samples of smaller length sequence */
 344     pScr1 -= 4u;
 345
 346
 347     /* apply same above for remaining samples of smaller length sequence */
 348     tapCnt = (srcBLen) & 3u;
 349
 350     while(tapCnt > 0u)
 351     {
 352
 353       /* accumlate the results */
 354       acc0 += (*pScr1++ * *pScr2);
 355       acc1 += (*pScr1++ * *pScr2);
 356       acc2 += (*pScr1++ * *pScr2);
 357       acc3 += (*pScr1++ * *pScr2++);
 358
 359       pScr1 -= 3u;
 360
 361       /* Decrement the loop counter */
 362       tapCnt--;
 363     }
 364
 365     blkCnt--;
 366
 367     /* Store the result in the accumulator in the destination buffer. */
 368     out0 = (q7_t) (__SSAT(acc0 >> 7u, 8));
 369     out1 = (q7_t) (__SSAT(acc1 >> 7u, 8));
 370     out2 = (q7_t) (__SSAT(acc2 >> 7u, 8));
 371     out3 = (q7_t) (__SSAT(acc3 >> 7u, 8));
 372
 373     *__SIMD32(pOut)++ = __PACKq7(out0, out1, out2, out3);
 374
 375     /* Initialization of inputB pointer */
 376     pScr2 = py;
 377
 378     pScratch1 += 4u;
 379
 380   }
 381
 382
 383   blkCnt = (srcALen + srcBLen - 1u) & 0x3;
 384
 385   /* Calculate convolution for remaining samples of Bigger length sequence */
 386   while(blkCnt > 0)
 387   {
 388     /* Initialze temporary scratch pointer as scratch1 */
 389     pScr1 = pScratch1;
 390
 391     /* Clear Accumlators */
 392     acc0 = 0;
 393
 394     tapCnt = (srcBLen) >> 1u;
 395
 396     while(tapCnt > 0u)
 397     {
 398       acc0 += (*pScr1++ * *pScr2++);
 399       acc0 += (*pScr1++ * *pScr2++);
 400
 401       /* Decrement the loop counter */
 402       tapCnt--;
 403     }
 404
 405     tapCnt = (srcBLen) & 1u;
 406
 407     /* apply same above for remaining samples of smaller length sequence */
 408     while(tapCnt > 0u)
 409     {
 410
 411       /* accumlate the results */
 412       acc0 += (*pScr1++ * *pScr2++);
 413
 414       /* Decrement the loop counter */
 415       tapCnt--;
 416     }
 417
 418     blkCnt--;
 419
 420     /* Store the result in the accumulator in the destination buffer. */
 421     *pOut++ = (q7_t) (__SSAT(acc0 >> 7u, 8));
 422
 423     /* Initialization of inputB pointer */
 424     pScr2 = py;
 425
 426     pScratch1 += 1u;
 427
 428   }
 429
 430 }
 431
 432
 433 /**
 434  * @} end of Conv group
 435  */