apps/codecs/libspeex/ltp_bfin.h

   1 /* Copyright (C) 2005 Analog Devices */
   2 /**
   3    @file ltp_bfin.h
   4    @author Jean-Marc Valin
   5    @brief Long-Term Prediction functions (Blackfin version)
   6 */
   7 /*
   8    Redistribution and use in source and binary forms, with or without
   9    modification, are permitted provided that the following conditions
  10    are met:
  11
  12    - Redistributions of source code must retain the above copyright
  13    notice, this list of conditions and the following disclaimer.
  14
  15    - Redistributions in binary form must reproduce the above copyright
  16    notice, this list of conditions and the following disclaimer in the
  17    documentation and/or other materials provided with the distribution.
  18
  19    - Neither the name of the Xiph.org Foundation nor the names of its
  20    contributors may be used to endorse or promote products derived from
  21    this software without specific prior written permission.
  22
  23    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  24    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  25    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  26    A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
  27    CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
  28    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
  29    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
  30    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
  31    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
  32    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
  33    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  34 */
  35
  36 #define OVERRIDE_INNER_PROD
  37 spx_word32_t inner_prod(const spx_word16_t *x, const spx_word16_t *y, int len)
  38 {
  39    spx_word32_t sum=0;
  40    __asm__ __volatile__ (
  41       "P0 = %3;\n\t"
  42       "P1 = %1;\n\t"
  43       "P2 = %2;\n\t"
  44       "I0 = P1;\n\t"
  45       "I1 = P2;\n\t"
  46       "L0 = 0;\n\t"
  47       "L1 = 0;\n\t"
  48       "A0 = 0;\n\t"
  49       "R0.L = W[I0++] || R1.L = W[I1++];\n\t"
  50       "LOOP inner%= LC0 = P0;\n\t"
  51       "LOOP_BEGIN inner%=;\n\t"
  52          "A0 += R0.L*R1.L (IS) || R0.L = W[I0++] || R1.L = W[I1++];\n\t"
  53       "LOOP_END inner%=;\n\t"
  54       "A0 += R0.L*R1.L (IS);\n\t"
  55       "A0 = A0 >>> 6;\n\t"
  56       "R0 = A0;\n\t"
  57       "%0 = R0;\n\t"
  58    : "=m" (sum)
  59    : "m" (x), "m" (y), "d" (len-1)
  60    : "P0", "P1", "P2", "R0", "R1", "A0", "I0", "I1", "L0", "L1", "R3"
  61    );
  62    return sum;
  63 }
  64
  65 #define OVERRIDE_PITCH_XCORR
  66 void pitch_xcorr(const spx_word16_t *_x, const spx_word16_t *_y, spx_word32_t *corr, int len, int nb_pitch, char *stack)
  67 {
  68    corr += nb_pitch - 1;
  69    __asm__ __volatile__ (
  70       "P2 = %0;\n\t"
  71       "I0 = P2;\n\t" /* x in I0 */
  72       "B0 = P2;\n\t" /* x in B0 */
  73       "R0 = %3;\n\t" /* len in R0 */
  74       "P3 = %3;\n\t"
  75       "P3 += -2;\n\t" /* len in R0 */
  76       "P4 = %4;\n\t" /* nb_pitch in R0 */
  77       "R1 = R0 << 1;\n\t" /* number of bytes in x */
  78       "L0 = R1;\n\t"
  79       "P0 = %1;\n\t"
  80
  81       "P1 = %2;\n\t"
  82       "B1 = P1;\n\t"
  83       "L1 = 0;\n\t" /*Disable looping on I1*/
  84
  85       "r0 = [I0++];\n\t"
  86       "LOOP pitch%= LC0 = P4 >> 1;\n\t"
  87       "LOOP_BEGIN pitch%=;\n\t"
  88          "I1 = P0;\n\t"
  89          "A1 = A0 = 0;\n\t"
  90          "R1 = [I1++];\n\t"
  91          "LOOP inner_prod%= LC1 = P3 >> 1;\n\t"
  92          "LOOP_BEGIN inner_prod%=;\n\t"
  93             "A1 += R0.L*R1.H, A0 += R0.L*R1.L (IS) || R1.L = W[I1++];\n\t"
  94             "A1 += R0.H*R1.L, A0 += R0.H*R1.H (IS) || R1.H = W[I1++] || R0 = [I0++];\n\t"
  95          "LOOP_END inner_prod%=;\n\t"
  96          "A1 += R0.L*R1.H, A0 += R0.L*R1.L (IS) || R1.L = W[I1++];\n\t"
  97          "A1 += R0.H*R1.L, A0 += R0.H*R1.H (IS) || R0 = [I0++];\n\t"
  98          "A0 = A0 >>> 6;\n\t"
  99          "A1 = A1 >>> 6;\n\t"
 100          "R2 = A0, R3 = A1;\n\t"
 101          "[P1--] = r2;\n\t"
 102          "[P1--] = r3;\n\t"
 103          "P0 += 4;\n\t"
 104       "LOOP_END pitch%=;\n\t"
 105       "L0 = 0;\n\t"
 106    : : "m" (_x), "m" (_y), "m" (corr), "m" (len), "m" (nb_pitch)
 107    : "A0", "A1", "P0", "P1", "P2", "P3", "P4", "R0", "R1", "R2", "R3", "I0", "I1", "L0", "L1", "B0", "B1", "memory"
 108    );
 109 }
 110
 111 #define OVERRIDE_COMPUTE_PITCH_ERROR
 112 static inline spx_word32_t compute_pitch_error(spx_word16_t *C, spx_word16_t *g, spx_word16_t pitch_control)
 113 {
 114    spx_word32_t sum;
 115    __asm__ __volatile__
 116          (
 117          "A0 = 0;\n\t"
 118
 119          "R0 = W[%1++];\n\t"
 120          "R1.L = %2.L*%5.L (IS);\n\t"
 121          "A0 += R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
 122
 123          "R1.L = %3.L*%5.L (IS);\n\t"
 124          "A0 += R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
 125
 126          "R1.L = %4.L*%5.L (IS);\n\t"
 127          "A0 += R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
 128
 129          "R1.L = %2.L*%3.L (IS);\n\t"
 130          "A0 -= R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
 131
 132          "R1.L = %4.L*%3.L (IS);\n\t"
 133          "A0 -= R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
 134
 135          "R1.L = %4.L*%2.L (IS);\n\t"
 136          "A0 -= R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
 137
 138          "R1.L = %2.L*%2.L (IS);\n\t"
 139          "A0 -= R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
 140
 141          "R1.L = %3.L*%3.L (IS);\n\t"
 142          "A0 -= R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
 143
 144          "R1.L = %4.L*%4.L (IS);\n\t"
 145          "A0 -= R1.L*R0.L (IS);\n\t"
 146
 147          "%0 = A0;\n\t"
 148    : "=&D" (sum), "=a" (C)
 149    : "d" (g[0]), "d" (g[1]), "d" (g[2]), "d" (pitch_control), "1" (C)
 150    : "R0", "R1", "R2", "A0"
 151          );
 152    return sum;
 153 }
 154
 155 #define OVERRIDE_OPEN_LOOP_NBEST_PITCH
 156 #ifdef OVERRIDE_OPEN_LOOP_NBEST_PITCH
 157 void open_loop_nbest_pitch(spx_word16_t *sw, int start, int end, int len, int *pitch, spx_word16_t *gain, int N, char *stack)
 158 {
 159    int i,j,k;
 160    VARDECL(spx_word32_t *best_score);
 161    VARDECL(spx_word32_t *best_ener);
 162    spx_word32_t e0;
 163    VARDECL(spx_word32_t *corr);
 164    VARDECL(spx_word32_t *energy);
 165
 166    ALLOC(best_score, N, spx_word32_t);
 167    ALLOC(best_ener, N, spx_word32_t);
 168    ALLOC(corr, end-start+1, spx_word32_t);
 169    ALLOC(energy, end-start+2, spx_word32_t);
 170
 171    for (i=0;i<N;i++)
 172    {
 173         best_score[i]=-1;
 174         best_ener[i]=0;
 175         pitch[i]=start;
 176    }
 177
 178    energy[0]=inner_prod(sw-start, sw-start, len);
 179    e0=inner_prod(sw, sw, len);
 180
 181    /* energy update -------------------------------------*/
 182
 183       __asm__ __volatile__
 184       (
 185 "        P0 = %0;\n\t"
 186 "        I1 = %1;\n\t"
 187 "        L1 = 0;\n\t"
 188 "        I2 = %2;\n\t"
 189 "        L2 = 0;\n\t"
 190 "        R2 = [P0++];\n\t"
 191 "        R3 = 0;\n\t"
 192 "        LSETUP (eu1, eu2) LC1 = %3;\n\t"
 193 "eu1:      R1.L = W [I1--] || R0.L = W [I2--] ;\n\t"
 194 "          R1 = R1.L * R1.L (IS);\n\t"
 195 "          R0 = R0.L * R0.L (IS);\n\t"
 196 "          R1 >>>= 6;\n\t"
 197 "          R1 = R1 + R2;\n\t"
 198 "          R0 >>>= 6;\n\t"
 199 "          R1 = R1 - R0;\n\t"
 200 "          R2 = MAX(R1,R3);\n\t"
 201 "eu2:      [P0++] = R2;\n\t"
 202        : : "d" (energy), "d" (&sw[-start-1]), "d" (&sw[-start+len-1]),
 203            "a" (end-start)
 204        : "P0", "I1", "I2", "R0", "R1", "R2", "R3"
 205 #if (__GNUC__ == 4)
 206          , "LC1"
 207 #endif
 208        );
 209
 210    pitch_xcorr(sw, sw-end, corr, len, end-start+1, stack);
 211
 212    /* FIXME: Fixed-point and floating-point code should be merged */
 213    {
 214       VARDECL(spx_word16_t *corr16);
 215       VARDECL(spx_word16_t *ener16);
 216       ALLOC(corr16, end-start+1, spx_word16_t);
 217       ALLOC(ener16, end-start+1, spx_word16_t);
 218       /* Normalize to 180 so we can square it and it still fits in 16 bits */
 219       normalize16(corr, corr16, 180, end-start+1);
 220       normalize16(energy, ener16, 180, end-start+1);
 221
 222       if (N == 1) {
 223         /* optimised asm to handle N==1 case */
 224       __asm__ __volatile__
 225       (
 226 "        I0 = %1;\n\t"                     /* I0: corr16[]    */
 227 "        L0 = 0;\n\t"
 228 "        I1 = %2;\n\t"                     /* I1: energy      */
 229 "        L1 = 0;\n\t"
 230 "        R2 = -1;\n\t"                     /* R2: best score  */
 231 "        R3 = 0;\n\t"                      /* R3: best energy */
 232 "        P0 = %4;\n\t"                     /* P0: best pitch  */
 233 "        P1 = %4;\n\t"                     /* P1: counter     */
 234 "        LSETUP (sl1, sl2) LC1 = %3;\n\t"
 235 "sl1:      R0.L = W [I0++] || R1.L = W [I1++];\n\t"
 236 "          R0 = R0.L * R0.L (IS);\n\t"
 237 "          R1   += 1;\n\t"
 238 "          R4   = R0.L * R3.L;\n\t"
 239 "          R5   = R2.L * R1.L;\n\t"
 240 "          cc   = R5 < R4;\n\t"
 241 "          if cc R2 = R0;\n\t"
 242 "          if cc R3 = R1;\n\t"
 243 "          if cc P0 = P1;\n\t"
 244 "sl2:      P1 += 1;\n\t"
 245 "        %0 = P0;\n\t"
 246        : "=&d" (pitch[0])
 247        : "a" (corr16), "a" (ener16), "a" (end+1-start), "d" (start)
 248        : "P0", "P1", "I0", "I1", "R0", "R1", "R2", "R3", "R4", "R5"
 249 #if (__GNUC__ == 4)
 250          , "LC1"
 251 #endif
 252        );
 253
 254       }
 255       else {
 256         for (i=start;i<=end;i++)
 257           {
 258             spx_word16_t tmp = MULT16_16_16(corr16[i-start],corr16[i-start]);
 259             /* Instead of dividing the tmp by the energy, we multiply on the other side */
 260             if (MULT16_16(tmp,best_ener[N-1])>MULT16_16(best_score[N-1],ADD16(1,ener16[i-start])))
 261               {
 262                 /* We can safely put it last and then check */
 263                 best_score[N-1]=tmp;
 264                 best_ener[N-1]=ener16[i-start]+1;
 265                 pitch[N-1]=i;
 266                 /* Check if it comes in front of others */
 267                 for (j=0;j<N-1;j++)
 268                   {
 269                     if (MULT16_16(tmp,best_ener[j])>MULT16_16(best_score[j],ADD16(1,ener16[i-start])))
 270                       {
 271                         for (k=N-1;k>j;k--)
 272                           {
 273                             best_score[k]=best_score[k-1];
 274                             best_ener[k]=best_ener[k-1];
 275                             pitch[k]=pitch[k-1];
 276                           }
 277                         best_score[j]=tmp;
 278                         best_ener[j]=ener16[i-start]+1;
 279                         pitch[j]=i;
 280                         break;
 281                       }
 282                   }
 283               }
 284           }
 285       }
 286    }
 287
 288    /* Compute open-loop gain */
 289    if (gain)
 290    {
 291        for (j=0;j<N;j++)
 292        {
 293           spx_word16_t g;
 294           i=pitch[j];
 295           g = DIV32(corr[i-start], 10+SHR32(MULT16_16(spx_sqrt(e0),spx_sqrt(energy[i-start])),6));
 296           /* FIXME: g = max(g,corr/energy) */
 297                    if (g<0)
 298                    g = 0;
 299              gain[j]=g;
 300        }
 301    }
 302 }
 303 #endif
 304
 305 #define OVERRIDE_PITCH_GAIN_SEARCH_3TAP_VQ
 306 #ifdef OVERRIDE_PITCH_GAIN_SEARCH_3TAP_VQ
 307 static int pitch_gain_search_3tap_vq(
 308   const signed char *gain_cdbk,
 309   int                gain_cdbk_size,
 310   spx_word16_t      *C16,
 311   spx_word16_t       max_gain
 312 )
 313 {
 314   const signed char *ptr=gain_cdbk;
 315   int                best_cdbk=0;
 316   spx_word32_t       best_sum=-VERY_LARGE32;
 317   spx_word32_t       sum=0;
 318   spx_word16_t       g[3];
 319   spx_word16_t       pitch_control=64;
 320   spx_word16_t       gain_sum;
 321   int                i;
 322
 323       /* fast asm version of VQ codebook search */
 324
 325       __asm__ __volatile__
 326       (
 327
 328 "        P0 = %2;\n\t"                     /* P0: ptr to gain_cdbk */
 329 "        L1 = 0;\n\t"                      /* no circ addr for L1  */
 330 "        %0 = 0;\n\t"                      /* %0: best_sum         */
 331 "        %1 = 0;\n\t"                      /* %1: best_cbdk        */
 332 "        P1 = 0;\n\t"                      /* P1: loop counter     */
 333
 334 "        LSETUP (pgs1, pgs2) LC1 = %4;\n\t"
 335 "pgs1:     R2  = B [P0++] (X);\n\t"        /* R2: g[0]             */
 336 "          R3  = B [P0++] (X);\n\t"        /* R3: g[1]             */
 337 "          R4  = B [P0++] (X);\n\t"        /* R4: g[2]             */
 338 "          R2 += 32;\n\t"
 339 "          R3 += 32;\n\t"
 340 "          R4 += 32;\n\t"
 341 "          R4.H = 64;\n\t"                 /* R4.H: pitch_control    */
 342
 343 "          R0  = B [P0++] (X);\n\t"
 344 "          B0  = R0;\n\t"                  /* BO: gain_sum         */
 345
 346            /* compute_pitch_error() -------------------------------*/
 347
 348 "          I1 = %3;\n\t"                   /* I1: ptr to C         */
 349 "          A0 = 0;\n\t"
 350
 351 "          R0.L = W[I1++];\n\t"
 352 "          R1.L = R2.L*R4.H (IS);\n\t"
 353 "          A0 += R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
 354
 355 "          R1.L = R3.L*R4.H (IS);\n\t"
 356 "          A0 += R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
 357
 358 "          R1.L = R4.L*R4.H (IS);\n\t"
 359 "          A0 += R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
 360
 361 "          R1.L = R2.L*R3.L (IS);\n\t"
 362 "          A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
 363
 364 "          R1.L = R4.L*R3.L (IS);\n\t"
 365 "          A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
 366
 367 "          R1.L = R4.L*R2.L (IS);\n\t"
 368 "          A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
 369
 370 "          R1.L = R2.L*R2.L (IS);\n\t"
 371 "          A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
 372
 373 "          R1.L = R3.L*R3.L (IS);\n\t"
 374 "          A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
 375
 376 "          R1.L = R4.L*R4.L (IS);\n\t"
 377 "          R0 = (A0 -= R1.L*R0.L) (IS);\n\t"
 378
 379 /*
 380     Re-arrange the if-then to code efficiently on the Blackfin:
 381
 382       if (sum>best_sum && gain_sum<=max_gain)   ------ (1)
 383
 384       if (sum>best_sum && !(gain_sum>max_gain)) ------ (2)
 385
 386       if (max_gain<=gain_sum) {                 ------ (3)
 387       sum = -VERY_LARGE32;
 388       }
 389       if (best_sum<=sum)
 390
 391     The blackin cc instructions are all of the form:
 392
 393       cc = x < y (or cc = x <= y)
 394 */
 395 "          R1 = B0\n\t"
 396 "          R2 = %5\n\t"
 397 "          R3 = %6\n\t"
 398 "          cc = R2 <= R1;\n\t"
 399 "          if cc R0 = R3;\n\t"
 400 "          cc = %0 <= R0;\n\t"
 401 "          if cc %0 = R0;\n\t"
 402 "          if cc %1 = P1;\n\t"
 403
 404 "pgs2:     P1 += 1;\n\t"
 405
 406        : "=&d" (best_sum), "=&d" (best_cdbk)
 407        : "a" (gain_cdbk), "a" (C16), "a" (gain_cdbk_size), "a" (max_gain),
 408          "b" (-VERY_LARGE32)
 409        : "R0", "R1", "R2", "R3", "R4", "P0",
 410          "P1", "I1", "L1", "A0", "B0"
 411 #if (__GNUC__ == 4)
 412          , "LC1"
 413 #endif
 414        );
 415
 416   return best_cdbk;
 417 }
 418 #endif
 419