libpostproc/postprocess_altivec_template.c

   1 /*
   2  * AltiVec optimizations (C) 2004 Romain Dolbeau <romain@dolbeau.org>
   3  *
   4  * based on code by Copyright (C) 2001-2003 Michael Niedermayer (michaelni@gmx.at)
   5  *
   6  * This file is part of FFmpeg.
   7  *
   8  * FFmpeg is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * FFmpeg is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * along with FFmpeg; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #include "avutil.h"
  24
  25 #define ALTIVEC_TRANSPOSE_8x8_SHORT(src_a,src_b,src_c,src_d,src_e,src_f,src_g,src_h) \
  26     do {                                                          \
  27         __typeof__(src_a) tempA1, tempB1, tempC1, tempD1;         \
  28         __typeof__(src_a) tempE1, tempF1, tempG1, tempH1;         \
  29         __typeof__(src_a) tempA2, tempB2, tempC2, tempD2;         \
  30         __typeof__(src_a) tempE2, tempF2, tempG2, tempH2;         \
  31         tempA1 = vec_mergeh (src_a, src_e);                       \
  32         tempB1 = vec_mergel (src_a, src_e);                       \
  33         tempC1 = vec_mergeh (src_b, src_f);                       \
  34         tempD1 = vec_mergel (src_b, src_f);                       \
  35         tempE1 = vec_mergeh (src_c, src_g);                       \
  36         tempF1 = vec_mergel (src_c, src_g);                       \
  37         tempG1 = vec_mergeh (src_d, src_h);                       \
  38         tempH1 = vec_mergel (src_d, src_h);                       \
  39         tempA2 = vec_mergeh (tempA1, tempE1);                     \
  40         tempB2 = vec_mergel (tempA1, tempE1);                     \
  41         tempC2 = vec_mergeh (tempB1, tempF1);                     \
  42         tempD2 = vec_mergel (tempB1, tempF1);                     \
  43         tempE2 = vec_mergeh (tempC1, tempG1);                     \
  44         tempF2 = vec_mergel (tempC1, tempG1);                     \
  45         tempG2 = vec_mergeh (tempD1, tempH1);                     \
  46         tempH2 = vec_mergel (tempD1, tempH1);                     \
  47         src_a = vec_mergeh (tempA2, tempE2);                      \
  48         src_b = vec_mergel (tempA2, tempE2);                      \
  49         src_c = vec_mergeh (tempB2, tempF2);                      \
  50         src_d = vec_mergel (tempB2, tempF2);                      \
  51         src_e = vec_mergeh (tempC2, tempG2);                      \
  52         src_f = vec_mergel (tempC2, tempG2);                      \
  53         src_g = vec_mergeh (tempD2, tempH2);                      \
  54         src_h = vec_mergel (tempD2, tempH2);                      \
  55     } while (0)
  56
  57
  58 static inline int vertClassify_altivec(uint8_t src[], int stride, PPContext *c) {
  59     /*
  60     this code makes no assumption on src or stride.
  61     One could remove the recomputation of the perm
  62     vector by assuming (stride % 16) == 0, unfortunately
  63     this is not always true.
  64     */
  65     DECLARE_ALIGNED(16, short, data[8]) =
  66                     {
  67                         ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1,
  68                         data[0] * 2 + 1,
  69                         c->QP * 2,
  70                         c->QP * 4
  71                     };
  72     int numEq;
  73     uint8_t *src2 = src;
  74     vector signed short v_dcOffset;
  75     vector signed short v2QP;
  76     vector unsigned short v4QP;
  77     vector unsigned short v_dcThreshold;
  78     const int properStride = (stride % 16);
  79     const int srcAlign = ((unsigned long)src2 % 16);
  80     const int two_vectors = ((srcAlign > 8) || properStride) ? 1 : 0;
  81     const vector signed int zero = vec_splat_s32(0);
  82     const vector signed short mask = vec_splat_s16(1);
  83     vector signed int v_numEq = vec_splat_s32(0);
  84     vector signed short v_data = vec_ld(0, data);
  85     vector signed short v_srcAss0, v_srcAss1, v_srcAss2, v_srcAss3,
  86                         v_srcAss4, v_srcAss5, v_srcAss6, v_srcAss7;
  87 //FIXME avoid this mess if possible
  88     register int j0 = 0,
  89                  j1 = stride,
  90                  j2 = 2 * stride,
  91                  j3 = 3 * stride,
  92                  j4 = 4 * stride,
  93                  j5 = 5 * stride,
  94                  j6 = 6 * stride,
  95                  j7 = 7 * stride;
  96     vector unsigned char v_srcA0, v_srcA1, v_srcA2, v_srcA3,
  97                          v_srcA4, v_srcA5, v_srcA6, v_srcA7;
  98
  99     v_dcOffset = vec_splat(v_data, 0);
 100     v_dcThreshold = (vector unsigned short)vec_splat(v_data, 1);
 101     v2QP = vec_splat(v_data, 2);
 102     v4QP = (vector unsigned short)vec_splat(v_data, 3);
 103
 104     src2 += stride * 4;
 105
 106
 107 #define LOAD_LINE(i)                                                    \
 108     {                                                                   \
 109     vector unsigned char perm##i = vec_lvsl(j##i, src2);                \
 110     vector unsigned char v_srcA2##i;                                    \
 111     vector unsigned char v_srcA1##i = vec_ld(j##i, src2);               \
 112     if (two_vectors)                                                    \
 113         v_srcA2##i = vec_ld(j##i + 16, src2);                           \
 114     v_srcA##i =                                                         \
 115         vec_perm(v_srcA1##i, v_srcA2##i, perm##i);                      \
 116     v_srcAss##i =                                                       \
 117         (vector signed short)vec_mergeh((vector signed char)zero,       \
 118                                         (vector signed char)v_srcA##i); }
 119
 120 #define LOAD_LINE_ALIGNED(i)                                            \
 121     v_srcA##i = vec_ld(j##i, src2);                                     \
 122     v_srcAss##i =                                                       \
 123         (vector signed short)vec_mergeh((vector signed char)zero,       \
 124                                         (vector signed char)v_srcA##i)
 125
 126     /* Special-casing the aligned case is worthwhile, as all calls from
 127      * the (transposed) horizontable deblocks will be aligned, in addition
 128      * to the naturally aligned vertical deblocks. */
 129     if (properStride && srcAlign) {
 130         LOAD_LINE_ALIGNED(0);
 131         LOAD_LINE_ALIGNED(1);
 132         LOAD_LINE_ALIGNED(2);
 133         LOAD_LINE_ALIGNED(3);
 134         LOAD_LINE_ALIGNED(4);
 135         LOAD_LINE_ALIGNED(5);
 136         LOAD_LINE_ALIGNED(6);
 137         LOAD_LINE_ALIGNED(7);
 138     } else {
 139         LOAD_LINE(0);
 140         LOAD_LINE(1);
 141         LOAD_LINE(2);
 142         LOAD_LINE(3);
 143         LOAD_LINE(4);
 144         LOAD_LINE(5);
 145         LOAD_LINE(6);
 146         LOAD_LINE(7);
 147     }
 148 #undef LOAD_LINE
 149 #undef LOAD_LINE_ALIGNED
 150
 151 #define ITER(i, j)                                                      \
 152     const vector signed short v_diff##i =                               \
 153         vec_sub(v_srcAss##i, v_srcAss##j);                              \
 154     const vector signed short v_sum##i =                                \
 155         vec_add(v_diff##i, v_dcOffset);                                 \
 156     const vector signed short v_comp##i =                               \
 157         (vector signed short)vec_cmplt((vector unsigned short)v_sum##i, \
 158                                        v_dcThreshold);                  \
 159     const vector signed short v_part##i = vec_and(mask, v_comp##i);
 160
 161 {
 162     ITER(0, 1)
 163     ITER(1, 2)
 164     ITER(2, 3)
 165     ITER(3, 4)
 166     ITER(4, 5)
 167     ITER(5, 6)
 168     ITER(6, 7)
 169
 170     v_numEq = vec_sum4s(v_part0, v_numEq);
 171     v_numEq = vec_sum4s(v_part1, v_numEq);
 172     v_numEq = vec_sum4s(v_part2, v_numEq);
 173     v_numEq = vec_sum4s(v_part3, v_numEq);
 174     v_numEq = vec_sum4s(v_part4, v_numEq);
 175     v_numEq = vec_sum4s(v_part5, v_numEq);
 176     v_numEq = vec_sum4s(v_part6, v_numEq);
 177 }
 178
 179 #undef ITER
 180
 181     v_numEq = vec_sums(v_numEq, zero);
 182
 183     v_numEq = vec_splat(v_numEq, 3);
 184     vec_ste(v_numEq, 0, &numEq);
 185
 186     if (numEq > c->ppMode.flatnessThreshold){
 187         const vector unsigned char mmoP1 = (const vector unsigned char)
 188             AVV(0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f,
 189                 0x00, 0x01, 0x12, 0x13, 0x08, 0x09, 0x1A, 0x1B);
 190         const vector unsigned char mmoP2 = (const vector unsigned char)
 191             AVV(0x04, 0x05, 0x16, 0x17, 0x0C, 0x0D, 0x1E, 0x1F,
 192                 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f);
 193         const vector unsigned char mmoP = (const vector unsigned char)
 194             vec_lvsl(8, (unsigned char*)0);
 195
 196         vector signed short mmoL1 = vec_perm(v_srcAss0, v_srcAss2, mmoP1);
 197         vector signed short mmoL2 = vec_perm(v_srcAss4, v_srcAss6, mmoP2);
 198         vector signed short mmoL = vec_perm(mmoL1, mmoL2, mmoP);
 199         vector signed short mmoR1 = vec_perm(v_srcAss5, v_srcAss7, mmoP1);
 200         vector signed short mmoR2 = vec_perm(v_srcAss1, v_srcAss3, mmoP2);
 201         vector signed short mmoR = vec_perm(mmoR1, mmoR2, mmoP);
 202         vector signed short mmoDiff = vec_sub(mmoL, mmoR);
 203         vector unsigned short mmoSum = (vector unsigned short)vec_add(mmoDiff, v2QP);
 204
 205         if (vec_any_gt(mmoSum, v4QP))
 206             return 0;
 207         else
 208             return 1;
 209     }
 210     else return 2;
 211 }
 212
 213 static inline void doVertLowPass_altivec(uint8_t *src, int stride, PPContext *c) {
 214     /*
 215     this code makes no assumption on src or stride.
 216     One could remove the recomputation of the perm
 217     vector by assuming (stride % 16) == 0, unfortunately
 218     this is not always true. Quite a lot of load/stores
 219     can be removed by assuming proper alignment of
 220     src & stride :-(
 221     */
 222     uint8_t *src2 = src;
 223     const vector signed int zero = vec_splat_s32(0);
 224     const int properStride = (stride % 16);
 225     const int srcAlign = ((unsigned long)src2 % 16);
 226     DECLARE_ALIGNED(16, short, qp[8]);
 227     qp[0] = c->QP;
 228     vector signed short vqp = vec_ld(0, qp);
 229     vqp = vec_splat(vqp, 0);
 230
 231     src2 += stride*3;
 232
 233     vector signed short vb0, vb1, vb2, vb3, vb4, vb5, vb6, vb7, vb8, vb9;
 234     vector unsigned char vbA0, vbA1, vbA2, vbA3, vbA4, vbA5, vbA6, vbA7, vbA8, vbA9;
 235     vector unsigned char vbB0, vbB1, vbB2, vbB3, vbB4, vbB5, vbB6, vbB7, vbB8, vbB9;
 236     vector unsigned char vbT0, vbT1, vbT2, vbT3, vbT4, vbT5, vbT6, vbT7, vbT8, vbT9;
 237
 238 #define LOAD_LINE(i)                                                    \
 239     const vector unsigned char perml##i =                               \
 240         vec_lvsl(i * stride, src2);                                     \
 241     vbA##i = vec_ld(i * stride, src2);                                  \
 242     vbB##i = vec_ld(i * stride + 16, src2);                             \
 243     vbT##i = vec_perm(vbA##i, vbB##i, perml##i);                        \
 244     vb##i =                                                             \
 245         (vector signed short)vec_mergeh((vector unsigned char)zero,     \
 246                                         (vector unsigned char)vbT##i)
 247
 248 #define LOAD_LINE_ALIGNED(i)                                            \
 249     register int j##i = i * stride;                                     \
 250     vbT##i = vec_ld(j##i, src2);                                        \
 251     vb##i =                                                             \
 252         (vector signed short)vec_mergeh((vector signed char)zero,       \
 253                                         (vector signed char)vbT##i)
 254
 255       /* Special-casing the aligned case is worthwhile, as all calls from
 256        * the (transposed) horizontable deblocks will be aligned, in addition
 257        * to the naturally aligned vertical deblocks. */
 258       if (properStride && srcAlign) {
 259           LOAD_LINE_ALIGNED(0);
 260           LOAD_LINE_ALIGNED(1);
 261           LOAD_LINE_ALIGNED(2);
 262           LOAD_LINE_ALIGNED(3);
 263           LOAD_LINE_ALIGNED(4);
 264           LOAD_LINE_ALIGNED(5);
 265           LOAD_LINE_ALIGNED(6);
 266           LOAD_LINE_ALIGNED(7);
 267           LOAD_LINE_ALIGNED(8);
 268           LOAD_LINE_ALIGNED(9);
 269       } else {
 270           LOAD_LINE(0);
 271           LOAD_LINE(1);
 272           LOAD_LINE(2);
 273           LOAD_LINE(3);
 274           LOAD_LINE(4);
 275           LOAD_LINE(5);
 276           LOAD_LINE(6);
 277           LOAD_LINE(7);
 278           LOAD_LINE(8);
 279           LOAD_LINE(9);
 280       }
 281 #undef LOAD_LINE
 282 #undef LOAD_LINE_ALIGNED
 283
 284     const vector unsigned short v_2 = vec_splat_u16(2);
 285     const vector unsigned short v_4 = vec_splat_u16(4);
 286
 287     const vector signed short v_diff01 = vec_sub(vb0, vb1);
 288     const vector unsigned short v_cmp01 =
 289         (const vector unsigned short) vec_cmplt(vec_abs(v_diff01), vqp);
 290     const vector signed short v_first = vec_sel(vb1, vb0, v_cmp01);
 291     const vector signed short v_diff89 = vec_sub(vb8, vb9);
 292     const vector unsigned short v_cmp89 =
 293         (const vector unsigned short) vec_cmplt(vec_abs(v_diff89), vqp);
 294     const vector signed short v_last = vec_sel(vb8, vb9, v_cmp89);
 295
 296     const vector signed short temp01 = vec_mladd(v_first, (vector signed short)v_4, vb1);
 297     const vector signed short temp02 = vec_add(vb2, vb3);
 298     const vector signed short temp03 = vec_add(temp01, (vector signed short)v_4);
 299     const vector signed short v_sumsB0 = vec_add(temp02, temp03);
 300
 301     const vector signed short temp11 = vec_sub(v_sumsB0, v_first);
 302     const vector signed short v_sumsB1 = vec_add(temp11, vb4);
 303
 304     const vector signed short temp21 = vec_sub(v_sumsB1, v_first);
 305     const vector signed short v_sumsB2 = vec_add(temp21, vb5);
 306
 307     const vector signed short temp31 = vec_sub(v_sumsB2, v_first);
 308     const vector signed short v_sumsB3 = vec_add(temp31, vb6);
 309
 310     const vector signed short temp41 = vec_sub(v_sumsB3, v_first);
 311     const vector signed short v_sumsB4 = vec_add(temp41, vb7);
 312
 313     const vector signed short temp51 = vec_sub(v_sumsB4, vb1);
 314     const vector signed short v_sumsB5 = vec_add(temp51, vb8);
 315
 316     const vector signed short temp61 = vec_sub(v_sumsB5, vb2);
 317     const vector signed short v_sumsB6 = vec_add(temp61, v_last);
 318
 319     const vector signed short temp71 = vec_sub(v_sumsB6, vb3);
 320     const vector signed short v_sumsB7 = vec_add(temp71, v_last);
 321
 322     const vector signed short temp81 = vec_sub(v_sumsB7, vb4);
 323     const vector signed short v_sumsB8 = vec_add(temp81, v_last);
 324
 325     const vector signed short temp91 = vec_sub(v_sumsB8, vb5);
 326     const vector signed short v_sumsB9 = vec_add(temp91, v_last);
 327
 328 #define COMPUTE_VR(i, j, k)                                             \
 329     const vector signed short temps1##i =                               \
 330         vec_add(v_sumsB##i, v_sumsB##k);                                \
 331     const vector signed short temps2##i =                               \
 332         vec_mladd(vb##j, (vector signed short)v_2, temps1##i);          \
 333     const vector signed short  vr##j = vec_sra(temps2##i, v_4)
 334
 335     COMPUTE_VR(0, 1, 2);
 336     COMPUTE_VR(1, 2, 3);
 337     COMPUTE_VR(2, 3, 4);
 338     COMPUTE_VR(3, 4, 5);
 339     COMPUTE_VR(4, 5, 6);
 340     COMPUTE_VR(5, 6, 7);
 341     COMPUTE_VR(6, 7, 8);
 342     COMPUTE_VR(7, 8, 9);
 343
 344     const vector signed char neg1 = vec_splat_s8(-1);
 345     const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
 346                                                                         0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
 347
 348 #define PACK_AND_STORE(i)                                       \
 349     const vector unsigned char perms##i =                       \
 350         vec_lvsr(i * stride, src2);                             \
 351     const vector unsigned char vf##i =                          \
 352         vec_packsu(vr##i, (vector signed short)zero);           \
 353     const vector unsigned char vg##i =                          \
 354         vec_perm(vf##i, vbT##i, permHH);                        \
 355     const vector unsigned char mask##i =                        \
 356         vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \
 357     const vector unsigned char vg2##i =                         \
 358         vec_perm(vg##i, vg##i, perms##i);                       \
 359     const vector unsigned char svA##i =                         \
 360         vec_sel(vbA##i, vg2##i, mask##i);                       \
 361     const vector unsigned char svB##i =                         \
 362         vec_sel(vg2##i, vbB##i, mask##i);                       \
 363     vec_st(svA##i, i * stride, src2);                           \
 364     vec_st(svB##i, i * stride + 16, src2)
 365
 366 #define PACK_AND_STORE_ALIGNED(i)                               \
 367     const vector unsigned char vf##i =                          \
 368         vec_packsu(vr##i, (vector signed short)zero);           \
 369     const vector unsigned char vg##i =                          \
 370         vec_perm(vf##i, vbT##i, permHH);                        \
 371     vec_st(vg##i, i * stride, src2)
 372
 373     /* Special-casing the aligned case is worthwhile, as all calls from
 374      * the (transposed) horizontable deblocks will be aligned, in addition
 375      * to the naturally aligned vertical deblocks. */
 376     if (properStride && srcAlign) {
 377         PACK_AND_STORE_ALIGNED(1);
 378         PACK_AND_STORE_ALIGNED(2);
 379         PACK_AND_STORE_ALIGNED(3);
 380         PACK_AND_STORE_ALIGNED(4);
 381         PACK_AND_STORE_ALIGNED(5);
 382         PACK_AND_STORE_ALIGNED(6);
 383         PACK_AND_STORE_ALIGNED(7);
 384         PACK_AND_STORE_ALIGNED(8);
 385     } else {
 386         PACK_AND_STORE(1);
 387         PACK_AND_STORE(2);
 388         PACK_AND_STORE(3);
 389         PACK_AND_STORE(4);
 390         PACK_AND_STORE(5);
 391         PACK_AND_STORE(6);
 392         PACK_AND_STORE(7);
 393         PACK_AND_STORE(8);
 394     }
 395 #undef PACK_AND_STORE
 396 #undef PACK_AND_STORE_ALIGNED
 397 }
 398
 399
 400
 401 static inline void doVertDefFilter_altivec(uint8_t src[], int stride, PPContext *c) {
 402     /*
 403     this code makes no assumption on src or stride.
 404     One could remove the recomputation of the perm
 405     vector by assuming (stride % 16) == 0, unfortunately
 406     this is not always true. Quite a lot of load/stores
 407     can be removed by assuming proper alignment of
 408     src & stride :-(
 409     */
 410     uint8_t *src2 = src;
 411     const vector signed int zero = vec_splat_s32(0);
 412     DECLARE_ALIGNED(16, short, qp[8]);
 413     qp[0] = 8*c->QP;
 414     vector signed short vqp = vec_ld(0, qp);
 415     vqp = vec_splat(vqp, 0);
 416
 417 #define LOAD_LINE(i)                                                    \
 418     const vector unsigned char perm##i =                                \
 419         vec_lvsl(i * stride, src2);                                     \
 420     const vector unsigned char vbA##i =                                 \
 421         vec_ld(i * stride, src2);                                       \
 422     const vector unsigned char vbB##i =                                 \
 423         vec_ld(i * stride + 16, src2);                                  \
 424     const vector unsigned char vbT##i =                                 \
 425         vec_perm(vbA##i, vbB##i, perm##i);                              \
 426     const vector signed short vb##i =                                   \
 427         (vector signed short)vec_mergeh((vector unsigned char)zero,     \
 428                                         (vector unsigned char)vbT##i)
 429
 430     src2 += stride*3;
 431
 432      LOAD_LINE(1);
 433      LOAD_LINE(2);
 434      LOAD_LINE(3);
 435      LOAD_LINE(4);
 436      LOAD_LINE(5);
 437      LOAD_LINE(6);
 438      LOAD_LINE(7);
 439      LOAD_LINE(8);
 440 #undef LOAD_LINE
 441
 442      const vector signed short v_1 = vec_splat_s16(1);
 443      const vector signed short v_2 = vec_splat_s16(2);
 444      const vector signed short v_5 = vec_splat_s16(5);
 445      const vector signed short v_32 = vec_sl(v_1,
 446                                              (vector unsigned short)v_5);
 447      /* middle energy */
 448      const vector signed short l3minusl6 = vec_sub(vb3, vb6);
 449      const vector signed short l5minusl4 = vec_sub(vb5, vb4);
 450      const vector signed short twotimes_l3minusl6 = vec_mladd(v_2, l3minusl6, (vector signed short)zero);
 451      const vector signed short mE = vec_mladd(v_5, l5minusl4, twotimes_l3minusl6);
 452      const vector signed short absmE = vec_abs(mE);
 453      /* left & right energy */
 454      const vector signed short l1minusl4 = vec_sub(vb1, vb4);
 455      const vector signed short l3minusl2 = vec_sub(vb3, vb2);
 456      const vector signed short l5minusl8 = vec_sub(vb5, vb8);
 457      const vector signed short l7minusl6 = vec_sub(vb7, vb6);
 458      const vector signed short twotimes_l1minusl4 = vec_mladd(v_2, l1minusl4, (vector signed short)zero);
 459      const vector signed short twotimes_l5minusl8 = vec_mladd(v_2, l5minusl8, (vector signed short)zero);
 460      const vector signed short lE = vec_mladd(v_5, l3minusl2, twotimes_l1minusl4);
 461      const vector signed short rE = vec_mladd(v_5, l7minusl6, twotimes_l5minusl8);
 462      /* d */
 463      const vector signed short ddiff = vec_sub(absmE,
 464                                                vec_min(vec_abs(lE),
 465                                                        vec_abs(rE)));
 466      const vector signed short ddiffclamp = vec_max(ddiff, (vector signed short)zero);
 467      const vector signed short dtimes64 = vec_mladd(v_5, ddiffclamp, v_32);
 468      const vector signed short d = vec_sra(dtimes64, vec_splat_u16(6));
 469      const vector signed short minusd = vec_sub((vector signed short)zero, d);
 470      const vector signed short finald = vec_sel(minusd,
 471                                                 d,
 472                                                 vec_cmpgt(vec_sub((vector signed short)zero, mE),
 473                                                           (vector signed short)zero));
 474      /* q */
 475      const vector signed short qtimes2 = vec_sub(vb4, vb5);
 476      /* for a shift right to behave like /2, we need to add one
 477         to all negative integer */
 478      const vector signed short rounddown = vec_sel((vector signed short)zero,
 479                                                    v_1,
 480                                                    vec_cmplt(qtimes2, (vector signed short)zero));
 481      const vector signed short q = vec_sra(vec_add(qtimes2, rounddown), vec_splat_u16(1));
 482      /* clamp */
 483      const vector signed short dclamp_P1 = vec_max((vector signed short)zero, finald);
 484      const vector signed short dclamp_P = vec_min(dclamp_P1, q);
 485      const vector signed short dclamp_N1 = vec_min((vector signed short)zero, finald);
 486      const vector signed short dclamp_N = vec_max(dclamp_N1, q);
 487
 488      const vector signed short dclampedfinal = vec_sel(dclamp_N,
 489                                                        dclamp_P,
 490                                                        vec_cmpgt(q, (vector signed short)zero));
 491      const vector signed short dornotd = vec_sel((vector signed short)zero,
 492                                                  dclampedfinal,
 493                                                  vec_cmplt(absmE, vqp));
 494      /* add/subtract to l4 and l5 */
 495      const vector signed short vb4minusd = vec_sub(vb4, dornotd);
 496      const vector signed short vb5plusd  = vec_add(vb5, dornotd);
 497      /* finally, stores */
 498      const vector unsigned char st4 = vec_packsu(vb4minusd, (vector signed short)zero);
 499      const vector unsigned char st5 = vec_packsu(vb5plusd,  (vector signed short)zero);
 500
 501      const vector signed char neg1 = vec_splat_s8(-1);
 502      const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
 503                                                                          0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
 504
 505 #define STORE(i)                                                \
 506      const vector unsigned char perms##i =                      \
 507          vec_lvsr(i * stride, src2);                            \
 508      const vector unsigned char vg##i =                         \
 509          vec_perm(st##i, vbT##i, permHH);                       \
 510      const vector unsigned char mask##i =                       \
 511          vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \
 512      const vector unsigned char vg2##i =                        \
 513          vec_perm(vg##i, vg##i, perms##i);                      \
 514      const vector unsigned char svA##i =                        \
 515          vec_sel(vbA##i, vg2##i, mask##i);                      \
 516      const vector unsigned char svB##i =                        \
 517          vec_sel(vg2##i, vbB##i, mask##i);                      \
 518      vec_st(svA##i, i * stride, src2);                          \
 519      vec_st(svB##i, i * stride + 16, src2)
 520
 521      STORE(4);
 522      STORE(5);
 523 }
 524
 525 static inline void dering_altivec(uint8_t src[], int stride, PPContext *c) {
 526     /*
 527     this code makes no assumption on src or stride.
 528     One could remove the recomputation of the perm
 529     vector by assuming (stride % 16) == 0, unfortunately
 530     this is not always true. Quite a lot of load/stores
 531     can be removed by assuming proper alignment of
 532     src & stride :-(
 533     */
 534     uint8_t *srcCopy = src;
 535     DECLARE_ALIGNED(16, uint8_t, dt[16]);
 536     const vector signed int zero = vec_splat_s32(0);
 537     vector unsigned char v_dt;
 538     dt[0] = deringThreshold;
 539     v_dt = vec_splat(vec_ld(0, dt), 0);
 540
 541 #define LOAD_LINE(i)                                                    \
 542     const vector unsigned char perm##i =                               \
 543         vec_lvsl(i * stride, srcCopy);                                 \
 544     vector unsigned char sA##i = vec_ld(i * stride, srcCopy);          \
 545     vector unsigned char sB##i = vec_ld(i * stride + 16, srcCopy);     \
 546     vector unsigned char src##i = vec_perm(sA##i, sB##i, perm##i)
 547
 548     LOAD_LINE(0);
 549     LOAD_LINE(1);
 550     LOAD_LINE(2);
 551     LOAD_LINE(3);
 552     LOAD_LINE(4);
 553     LOAD_LINE(5);
 554     LOAD_LINE(6);
 555     LOAD_LINE(7);
 556     LOAD_LINE(8);
 557     LOAD_LINE(9);
 558 #undef LOAD_LINE
 559
 560     vector unsigned char v_avg;
 561     {
 562     const vector unsigned char trunc_perm = (vector unsigned char)
 563         AVV(0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08,
 564             0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18);
 565     const vector unsigned char trunc_src12 = vec_perm(src1, src2, trunc_perm);
 566     const vector unsigned char trunc_src34 = vec_perm(src3, src4, trunc_perm);
 567     const vector unsigned char trunc_src56 = vec_perm(src5, src6, trunc_perm);
 568     const vector unsigned char trunc_src78 = vec_perm(src7, src8, trunc_perm);
 569
 570 #define EXTRACT(op) do {                                                \
 571     const vector unsigned char s##op##_1   = vec_##op(trunc_src12, trunc_src34); \
 572     const vector unsigned char s##op##_2   = vec_##op(trunc_src56, trunc_src78); \
 573     const vector unsigned char s##op##_6   = vec_##op(s##op##_1, s##op##_2);     \
 574     const vector unsigned char s##op##_8h  = vec_mergeh(s##op##_6, s##op##_6);   \
 575     const vector unsigned char s##op##_8l  = vec_mergel(s##op##_6, s##op##_6);   \
 576     const vector unsigned char s##op##_9   = vec_##op(s##op##_8h, s##op##_8l);   \
 577     const vector unsigned char s##op##_9h  = vec_mergeh(s##op##_9, s##op##_9);   \
 578     const vector unsigned char s##op##_9l  = vec_mergel(s##op##_9, s##op##_9);   \
 579     const vector unsigned char s##op##_10  = vec_##op(s##op##_9h, s##op##_9l);   \
 580     const vector unsigned char s##op##_10h = vec_mergeh(s##op##_10, s##op##_10); \
 581     const vector unsigned char s##op##_10l = vec_mergel(s##op##_10, s##op##_10); \
 582     const vector unsigned char s##op##_11  = vec_##op(s##op##_10h, s##op##_10l); \
 583     const vector unsigned char s##op##_11h = vec_mergeh(s##op##_11, s##op##_11); \
 584     const vector unsigned char s##op##_11l = vec_mergel(s##op##_11, s##op##_11); \
 585     v_##op = vec_##op(s##op##_11h, s##op##_11l); } while (0)
 586
 587     vector unsigned char v_min;
 588     vector unsigned char v_max;
 589     EXTRACT(min);
 590     EXTRACT(max);
 591 #undef EXTRACT
 592
 593     if (vec_all_lt(vec_sub(v_max, v_min), v_dt))
 594         return;
 595
 596     v_avg = vec_avg(v_min, v_max);
 597     }
 598
 599     DECLARE_ALIGNED(16, signed int, S[8]);
 600     {
 601     const vector unsigned short mask1 = (vector unsigned short)
 602         AVV(0x0001, 0x0002, 0x0004, 0x0008,
 603             0x0010, 0x0020, 0x0040, 0x0080);
 604     const vector unsigned short mask2 = (vector unsigned short)
 605         AVV(0x0100, 0x0200, 0x0000, 0x0000,
 606             0x0000, 0x0000, 0x0000, 0x0000);
 607
 608     const vector unsigned int vuint32_16 = vec_sl(vec_splat_u32(1), vec_splat_u32(4));
 609     const vector unsigned int vuint32_1 = vec_splat_u32(1);
 610
 611 #define COMPARE(i)                                                      \
 612     vector signed int sum##i;                                           \
 613     do {                                                                \
 614         const vector unsigned char cmp##i =                             \
 615             (vector unsigned char)vec_cmpgt(src##i, v_avg);             \
 616         const vector unsigned short cmpHi##i =                          \
 617             (vector unsigned short)vec_mergeh(cmp##i, cmp##i);          \
 618         const vector unsigned short cmpLi##i =                          \
 619             (vector unsigned short)vec_mergel(cmp##i, cmp##i);          \
 620         const vector signed short cmpHf##i =                            \
 621             (vector signed short)vec_and(cmpHi##i, mask1);              \
 622         const vector signed short cmpLf##i =                            \
 623             (vector signed short)vec_and(cmpLi##i, mask2);              \
 624         const vector signed int sump##i = vec_sum4s(cmpHf##i, zero);    \
 625         const vector signed int sumq##i = vec_sum4s(cmpLf##i, sump##i); \
 626         sum##i  = vec_sums(sumq##i, zero); } while (0)
 627
 628     COMPARE(0);
 629     COMPARE(1);
 630     COMPARE(2);
 631     COMPARE(3);
 632     COMPARE(4);
 633     COMPARE(5);
 634     COMPARE(6);
 635     COMPARE(7);
 636     COMPARE(8);
 637     COMPARE(9);
 638 #undef COMPARE
 639
 640     vector signed int sumA2;
 641     vector signed int sumB2;
 642     {
 643     const vector signed int sump02 = vec_mergel(sum0, sum2);
 644     const vector signed int sump13 = vec_mergel(sum1, sum3);
 645     const vector signed int sumA = vec_mergel(sump02, sump13);
 646
 647     const vector signed int sump46 = vec_mergel(sum4, sum6);
 648     const vector signed int sump57 = vec_mergel(sum5, sum7);
 649     const vector signed int sumB = vec_mergel(sump46, sump57);
 650
 651     const vector signed int sump8A = vec_mergel(sum8, zero);
 652     const vector signed int sump9B = vec_mergel(sum9, zero);
 653     const vector signed int sumC = vec_mergel(sump8A, sump9B);
 654
 655     const vector signed int tA = vec_sl(vec_nor(zero, sumA), vuint32_16);
 656     const vector signed int tB = vec_sl(vec_nor(zero, sumB), vuint32_16);
 657     const vector signed int tC = vec_sl(vec_nor(zero, sumC), vuint32_16);
 658     const vector signed int t2A = vec_or(sumA, tA);
 659     const vector signed int t2B = vec_or(sumB, tB);
 660     const vector signed int t2C = vec_or(sumC, tC);
 661     const vector signed int t3A = vec_and(vec_sra(t2A, vuint32_1),
 662                                           vec_sl(t2A, vuint32_1));
 663     const vector signed int t3B = vec_and(vec_sra(t2B, vuint32_1),
 664                                           vec_sl(t2B, vuint32_1));
 665     const vector signed int t3C = vec_and(vec_sra(t2C, vuint32_1),
 666                                           vec_sl(t2C, vuint32_1));
 667     const vector signed int yA = vec_and(t2A, t3A);
 668     const vector signed int yB = vec_and(t2B, t3B);
 669     const vector signed int yC = vec_and(t2C, t3C);
 670
 671     const vector unsigned char strangeperm1 = vec_lvsl(4, (unsigned char*)0);
 672     const vector unsigned char strangeperm2 = vec_lvsl(8, (unsigned char*)0);
 673     const vector signed int sumAd4 = vec_perm(yA, yB, strangeperm1);
 674     const vector signed int sumAd8 = vec_perm(yA, yB, strangeperm2);
 675     const vector signed int sumBd4 = vec_perm(yB, yC, strangeperm1);
 676     const vector signed int sumBd8 = vec_perm(yB, yC, strangeperm2);
 677     const vector signed int sumAp = vec_and(yA,
 678                                             vec_and(sumAd4,sumAd8));
 679     const vector signed int sumBp = vec_and(yB,
 680                                             vec_and(sumBd4,sumBd8));
 681     sumA2 = vec_or(sumAp,
 682                    vec_sra(sumAp,
 683                            vuint32_16));
 684     sumB2  = vec_or(sumBp,
 685                     vec_sra(sumBp,
 686                             vuint32_16));
 687     }
 688     vec_st(sumA2, 0, S);
 689     vec_st(sumB2, 16, S);
 690     }
 691
 692     /* I'm not sure the following is actually faster
 693        than straight, unvectorized C code :-( */
 694
 695     DECLARE_ALIGNED(16, int, tQP2[4]);
 696     tQP2[0]= c->QP/2 + 1;
 697     vector signed int vQP2 = vec_ld(0, tQP2);
 698     vQP2 = vec_splat(vQP2, 0);
 699     const vector signed int vsint32_8 = vec_splat_s32(8);
 700     const vector unsigned int vuint32_4 = vec_splat_u32(4);
 701
 702     const vector unsigned char permA1 = (vector unsigned char)
 703         AVV(0x00, 0x01, 0x02, 0x10, 0x11, 0x12, 0x1F, 0x1F,
 704             0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F);
 705     const vector unsigned char permA2 = (vector unsigned char)
 706         AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x10, 0x11,
 707             0x12, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F);
 708     const vector unsigned char permA1inc = (vector unsigned char)
 709         AVV(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00,
 710             0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
 711     const vector unsigned char permA2inc = (vector unsigned char)
 712         AVV(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01,
 713             0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
 714     const vector unsigned char magic = (vector unsigned char)
 715         AVV(0x01, 0x02, 0x01, 0x02, 0x04, 0x02, 0x01, 0x02,
 716             0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
 717     const vector unsigned char extractPerm = (vector unsigned char)
 718         AVV(0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01,
 719             0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01);
 720     const vector unsigned char extractPermInc = (vector unsigned char)
 721         AVV(0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01,
 722             0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01);
 723     const vector unsigned char identity = vec_lvsl(0,(unsigned char *)0);
 724     const vector unsigned char tenRight = (vector unsigned char)
 725         AVV(0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
 726             0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
 727     const vector unsigned char eightLeft = (vector unsigned char)
 728         AVV(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
 729             0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08);
 730
 731
 732 #define F_INIT(i)                                       \
 733     vector unsigned char tenRightM##i = tenRight;       \
 734     vector unsigned char permA1M##i = permA1;           \
 735     vector unsigned char permA2M##i = permA2;           \
 736     vector unsigned char extractPermM##i = extractPerm
 737
 738 #define F2(i, j, k, l)                                                  \
 739     if (S[i] & (1 << (l+1))) {                                          \
 740         const vector unsigned char a_##j##_A##l =                       \
 741             vec_perm(src##i, src##j, permA1M##i);                       \
 742         const vector unsigned char a_##j##_B##l =                       \
 743             vec_perm(a_##j##_A##l, src##k, permA2M##i);                 \
 744         const vector signed int a_##j##_sump##l =                       \
 745             (vector signed int)vec_msum(a_##j##_B##l, magic,            \
 746                                         (vector unsigned int)zero);     \
 747         vector signed int F_##j##_##l =                                 \
 748             vec_sr(vec_sums(a_##j##_sump##l, vsint32_8), vuint32_4);    \
 749         F_##j##_##l = vec_splat(F_##j##_##l, 3);                        \
 750         const vector signed int p_##j##_##l =                           \
 751             (vector signed int)vec_perm(src##j,                         \
 752                                         (vector unsigned char)zero,     \
 753                                         extractPermM##i);               \
 754         const vector signed int sum_##j##_##l  = vec_add( p_##j##_##l, vQP2);\
 755         const vector signed int diff_##j##_##l = vec_sub( p_##j##_##l, vQP2);\
 756         vector signed int newpm_##j##_##l;                              \
 757         if (vec_all_lt(sum_##j##_##l, F_##j##_##l))                     \
 758             newpm_##j##_##l = sum_##j##_##l;                            \
 759         else if (vec_all_gt(diff_##j##_##l, F_##j##_##l))               \
 760             newpm_##j##_##l = diff_##j##_##l;                           \
 761         else newpm_##j##_##l = F_##j##_##l;                             \
 762         const vector unsigned char newpm2_##j##_##l =                   \
 763             vec_splat((vector unsigned char)newpm_##j##_##l, 15);       \
 764         const vector unsigned char mask##j##l = vec_add(identity,       \
 765                                                         tenRightM##i);  \
 766         src##j = vec_perm(src##j, newpm2_##j##_##l, mask##j##l);        \
 767     }                                                                   \
 768     permA1M##i = vec_add(permA1M##i, permA1inc);                        \
 769     permA2M##i = vec_add(permA2M##i, permA2inc);                        \
 770     tenRightM##i = vec_sro(tenRightM##i, eightLeft);                    \
 771     extractPermM##i = vec_add(extractPermM##i, extractPermInc)
 772
 773 #define ITER(i, j, k)                           \
 774     F_INIT(i);                                  \
 775     F2(i, j, k, 0);                             \
 776     F2(i, j, k, 1);                             \
 777     F2(i, j, k, 2);                             \
 778     F2(i, j, k, 3);                             \
 779     F2(i, j, k, 4);                             \
 780     F2(i, j, k, 5);                             \
 781     F2(i, j, k, 6);                             \
 782     F2(i, j, k, 7)
 783
 784     ITER(0, 1, 2);
 785     ITER(1, 2, 3);
 786     ITER(2, 3, 4);
 787     ITER(3, 4, 5);
 788     ITER(4, 5, 6);
 789     ITER(5, 6, 7);
 790     ITER(6, 7, 8);
 791     ITER(7, 8, 9);
 792
 793     const vector signed char neg1 = vec_splat_s8(-1);
 794
 795 #define STORE_LINE(i)                                   \
 796     const vector unsigned char permST##i =              \
 797         vec_lvsr(i * stride, srcCopy);                  \
 798     const vector unsigned char maskST##i =              \
 799         vec_perm((vector unsigned char)zero,            \
 800                  (vector unsigned char)neg1, permST##i);\
 801     src##i = vec_perm(src##i ,src##i, permST##i);       \
 802     sA##i= vec_sel(sA##i, src##i, maskST##i);           \
 803     sB##i= vec_sel(src##i, sB##i, maskST##i);           \
 804     vec_st(sA##i, i * stride, srcCopy);                 \
 805     vec_st(sB##i, i * stride + 16, srcCopy)
 806
 807     STORE_LINE(1);
 808     STORE_LINE(2);
 809     STORE_LINE(3);
 810     STORE_LINE(4);
 811     STORE_LINE(5);
 812     STORE_LINE(6);
 813     STORE_LINE(7);
 814     STORE_LINE(8);
 815
 816 #undef STORE_LINE
 817 #undef ITER
 818 #undef F2
 819 }
 820
 821 #define doHorizLowPass_altivec(a...) doHorizLowPass_C(a)
 822 #define doHorizDefFilter_altivec(a...) doHorizDefFilter_C(a)
 823 #define do_a_deblock_altivec(a...) do_a_deblock_C(a)
 824
 825 static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,
 826                                             uint8_t *tempBlurred, uint32_t *tempBlurredPast, int *maxNoise)
 827 {
 828     const vector signed int zero = vec_splat_s32(0);
 829     const vector signed short vsint16_1 = vec_splat_s16(1);
 830     vector signed int v_dp = zero;
 831     vector signed int v_sysdp = zero;
 832     int d, sysd, i;
 833
 834     tempBlurredPast[127]= maxNoise[0];
 835     tempBlurredPast[128]= maxNoise[1];
 836     tempBlurredPast[129]= maxNoise[2];
 837
 838 #define LOAD_LINE(src, i)                                               \
 839     register int j##src##i = i * stride;                                \
 840     vector unsigned char perm##src##i = vec_lvsl(j##src##i, src);       \
 841     const vector unsigned char v_##src##A1##i = vec_ld(j##src##i, src); \
 842     const vector unsigned char v_##src##A2##i = vec_ld(j##src##i + 16, src); \
 843     const vector unsigned char v_##src##A##i =                          \
 844         vec_perm(v_##src##A1##i, v_##src##A2##i, perm##src##i);         \
 845     vector signed short v_##src##Ass##i =                               \
 846         (vector signed short)vec_mergeh((vector signed char)zero,       \
 847                                         (vector signed char)v_##src##A##i)
 848
 849     LOAD_LINE(src, 0);
 850     LOAD_LINE(src, 1);
 851     LOAD_LINE(src, 2);
 852     LOAD_LINE(src, 3);
 853     LOAD_LINE(src, 4);
 854     LOAD_LINE(src, 5);
 855     LOAD_LINE(src, 6);
 856     LOAD_LINE(src, 7);
 857
 858     LOAD_LINE(tempBlurred, 0);
 859     LOAD_LINE(tempBlurred, 1);
 860     LOAD_LINE(tempBlurred, 2);
 861     LOAD_LINE(tempBlurred, 3);
 862     LOAD_LINE(tempBlurred, 4);
 863     LOAD_LINE(tempBlurred, 5);
 864     LOAD_LINE(tempBlurred, 6);
 865     LOAD_LINE(tempBlurred, 7);
 866 #undef LOAD_LINE
 867
 868 #define ACCUMULATE_DIFFS(i)                                     \
 869     vector signed short v_d##i = vec_sub(v_tempBlurredAss##i,    \
 870                                          v_srcAss##i);          \
 871     v_dp = vec_msums(v_d##i, v_d##i, v_dp);                     \
 872     v_sysdp = vec_msums(v_d##i, vsint16_1, v_sysdp)
 873
 874     ACCUMULATE_DIFFS(0);
 875     ACCUMULATE_DIFFS(1);
 876     ACCUMULATE_DIFFS(2);
 877     ACCUMULATE_DIFFS(3);
 878     ACCUMULATE_DIFFS(4);
 879     ACCUMULATE_DIFFS(5);
 880     ACCUMULATE_DIFFS(6);
 881     ACCUMULATE_DIFFS(7);
 882 #undef ACCUMULATE_DIFFS
 883
 884     v_dp = vec_sums(v_dp, zero);
 885     v_sysdp = vec_sums(v_sysdp, zero);
 886
 887     v_dp = vec_splat(v_dp, 3);
 888     v_sysdp = vec_splat(v_sysdp, 3);
 889
 890     vec_ste(v_dp, 0, &d);
 891     vec_ste(v_sysdp, 0, &sysd);
 892
 893     i = d;
 894     d = (4*d
 895          +(*(tempBlurredPast-256))
 896          +(*(tempBlurredPast-1))+ (*(tempBlurredPast+1))
 897          +(*(tempBlurredPast+256))
 898          +4)>>3;
 899
 900     *tempBlurredPast=i;
 901
 902     if (d > maxNoise[1]) {
 903         if (d < maxNoise[2]) {
 904 #define OP(i) v_tempBlurredAss##i = vec_avg(v_tempBlurredAss##i, v_srcAss##i);
 905
 906             OP(0);
 907             OP(1);
 908             OP(2);
 909             OP(3);
 910             OP(4);
 911             OP(5);
 912             OP(6);
 913             OP(7);
 914 #undef OP
 915         } else {
 916 #define OP(i) v_tempBlurredAss##i = v_srcAss##i;
 917
 918             OP(0);
 919             OP(1);
 920             OP(2);
 921             OP(3);
 922             OP(4);
 923             OP(5);
 924             OP(6);
 925             OP(7);
 926 #undef OP
 927         }
 928     } else {
 929         if (d < maxNoise[0]) {
 930             const vector signed short vsint16_7 = vec_splat_s16(7);
 931             const vector signed short vsint16_4 = vec_splat_s16(4);
 932             const vector unsigned short vuint16_3 = vec_splat_u16(3);
 933
 934 #define OP(i)                                                   \
 935             const vector signed short v_temp##i =               \
 936                 vec_mladd(v_tempBlurredAss##i,                   \
 937                           vsint16_7, v_srcAss##i);              \
 938             const vector signed short v_temp2##i =              \
 939                 vec_add(v_temp##i, vsint16_4);                  \
 940             v_tempBlurredAss##i = vec_sr(v_temp2##i, vuint16_3)
 941
 942             OP(0);
 943             OP(1);
 944             OP(2);
 945             OP(3);
 946             OP(4);
 947             OP(5);
 948             OP(6);
 949             OP(7);
 950 #undef OP
 951         } else {
 952             const vector signed short vsint16_3 = vec_splat_s16(3);
 953             const vector signed short vsint16_2 = vec_splat_s16(2);
 954
 955 #define OP(i)                                                   \
 956             const vector signed short v_temp##i =               \
 957                 vec_mladd(v_tempBlurredAss##i,                   \
 958                           vsint16_3, v_srcAss##i);              \
 959             const vector signed short v_temp2##i =              \
 960                 vec_add(v_temp##i, vsint16_2);                  \
 961             v_tempBlurredAss##i = vec_sr(v_temp2##i, (vector unsigned short)vsint16_2)
 962
 963             OP(0);
 964             OP(1);
 965             OP(2);
 966             OP(3);
 967             OP(4);
 968             OP(5);
 969             OP(6);
 970             OP(7);
 971 #undef OP
 972         }
 973     }
 974
 975     const vector signed char neg1 = vec_splat_s8(-1);
 976     const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
 977                                                                         0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
 978
 979 #define PACK_AND_STORE(src, i)                                  \
 980     const vector unsigned char perms##src##i =                  \
 981         vec_lvsr(i * stride, src);                              \
 982     const vector unsigned char vf##src##i =                     \
 983         vec_packsu(v_tempBlurredAss##i, (vector signed short)zero); \
 984     const vector unsigned char vg##src##i =                     \
 985         vec_perm(vf##src##i, v_##src##A##i, permHH);            \
 986     const vector unsigned char mask##src##i =                   \
 987         vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##src##i); \
 988     const vector unsigned char vg2##src##i =                    \
 989         vec_perm(vg##src##i, vg##src##i, perms##src##i);        \
 990     const vector unsigned char svA##src##i =                    \
 991         vec_sel(v_##src##A1##i, vg2##src##i, mask##src##i);     \
 992     const vector unsigned char svB##src##i =                    \
 993         vec_sel(vg2##src##i, v_##src##A2##i, mask##src##i);     \
 994     vec_st(svA##src##i, i * stride, src);                       \
 995     vec_st(svB##src##i, i * stride + 16, src)
 996
 997     PACK_AND_STORE(src, 0);
 998     PACK_AND_STORE(src, 1);
 999     PACK_AND_STORE(src, 2);
1000     PACK_AND_STORE(src, 3);
1001     PACK_AND_STORE(src, 4);
1002     PACK_AND_STORE(src, 5);
1003     PACK_AND_STORE(src, 6);
1004     PACK_AND_STORE(src, 7);
1005     PACK_AND_STORE(tempBlurred, 0);
1006     PACK_AND_STORE(tempBlurred, 1);
1007     PACK_AND_STORE(tempBlurred, 2);
1008     PACK_AND_STORE(tempBlurred, 3);
1009     PACK_AND_STORE(tempBlurred, 4);
1010     PACK_AND_STORE(tempBlurred, 5);
1011     PACK_AND_STORE(tempBlurred, 6);
1012     PACK_AND_STORE(tempBlurred, 7);
1013 #undef PACK_AND_STORE
1014 }
1015
1016 static inline void transpose_16x8_char_toPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1017     const vector unsigned char zero = vec_splat_u8(0);
1018
1019 #define LOAD_DOUBLE_LINE(i, j)                                          \
1020     vector unsigned char perm1##i = vec_lvsl(i * stride, src);          \
1021     vector unsigned char perm2##i = vec_lvsl(j * stride, src);          \
1022     vector unsigned char srcA##i = vec_ld(i * stride, src);             \
1023     vector unsigned char srcB##i = vec_ld(i * stride + 16, src);        \
1024     vector unsigned char srcC##i = vec_ld(j * stride, src);             \
1025     vector unsigned char srcD##i = vec_ld(j * stride+ 16, src);         \
1026     vector unsigned char src##i = vec_perm(srcA##i, srcB##i, perm1##i); \
1027     vector unsigned char src##j = vec_perm(srcC##i, srcD##i, perm2##i)
1028
1029     LOAD_DOUBLE_LINE(0, 1);
1030     LOAD_DOUBLE_LINE(2, 3);
1031     LOAD_DOUBLE_LINE(4, 5);
1032     LOAD_DOUBLE_LINE(6, 7);
1033 #undef LOAD_DOUBLE_LINE
1034
1035     vector unsigned char tempA = vec_mergeh(src0, zero);
1036     vector unsigned char tempB = vec_mergel(src0, zero);
1037     vector unsigned char tempC = vec_mergeh(src1, zero);
1038     vector unsigned char tempD = vec_mergel(src1, zero);
1039     vector unsigned char tempE = vec_mergeh(src2, zero);
1040     vector unsigned char tempF = vec_mergel(src2, zero);
1041     vector unsigned char tempG = vec_mergeh(src3, zero);
1042     vector unsigned char tempH = vec_mergel(src3, zero);
1043     vector unsigned char tempI = vec_mergeh(src4, zero);
1044     vector unsigned char tempJ = vec_mergel(src4, zero);
1045     vector unsigned char tempK = vec_mergeh(src5, zero);
1046     vector unsigned char tempL = vec_mergel(src5, zero);
1047     vector unsigned char tempM = vec_mergeh(src6, zero);
1048     vector unsigned char tempN = vec_mergel(src6, zero);
1049     vector unsigned char tempO = vec_mergeh(src7, zero);
1050     vector unsigned char tempP = vec_mergel(src7, zero);
1051
1052     vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1053     vector unsigned char temp1 = vec_mergel(tempA, tempI);
1054     vector unsigned char temp2 = vec_mergeh(tempB, tempJ);
1055     vector unsigned char temp3 = vec_mergel(tempB, tempJ);
1056     vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1057     vector unsigned char temp5 = vec_mergel(tempC, tempK);
1058     vector unsigned char temp6 = vec_mergeh(tempD, tempL);
1059     vector unsigned char temp7 = vec_mergel(tempD, tempL);
1060     vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1061     vector unsigned char temp9 = vec_mergel(tempE, tempM);
1062     vector unsigned char temp10 = vec_mergeh(tempF, tempN);
1063     vector unsigned char temp11 = vec_mergel(tempF, tempN);
1064     vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1065     vector unsigned char temp13 = vec_mergel(tempG, tempO);
1066     vector unsigned char temp14 = vec_mergeh(tempH, tempP);
1067     vector unsigned char temp15 = vec_mergel(tempH, tempP);
1068
1069     tempA = vec_mergeh(temp0, temp8);
1070     tempB = vec_mergel(temp0, temp8);
1071     tempC = vec_mergeh(temp1, temp9);
1072     tempD = vec_mergel(temp1, temp9);
1073     tempE = vec_mergeh(temp2, temp10);
1074     tempF = vec_mergel(temp2, temp10);
1075     tempG = vec_mergeh(temp3, temp11);
1076     tempH = vec_mergel(temp3, temp11);
1077     tempI = vec_mergeh(temp4, temp12);
1078     tempJ = vec_mergel(temp4, temp12);
1079     tempK = vec_mergeh(temp5, temp13);
1080     tempL = vec_mergel(temp5, temp13);
1081     tempM = vec_mergeh(temp6, temp14);
1082     tempN = vec_mergel(temp6, temp14);
1083     tempO = vec_mergeh(temp7, temp15);
1084     tempP = vec_mergel(temp7, temp15);
1085
1086     temp0 = vec_mergeh(tempA, tempI);
1087     temp1 = vec_mergel(tempA, tempI);
1088     temp2 = vec_mergeh(tempB, tempJ);
1089     temp3 = vec_mergel(tempB, tempJ);
1090     temp4 = vec_mergeh(tempC, tempK);
1091     temp5 = vec_mergel(tempC, tempK);
1092     temp6 = vec_mergeh(tempD, tempL);
1093     temp7 = vec_mergel(tempD, tempL);
1094     temp8 = vec_mergeh(tempE, tempM);
1095     temp9 = vec_mergel(tempE, tempM);
1096     temp10 = vec_mergeh(tempF, tempN);
1097     temp11 = vec_mergel(tempF, tempN);
1098     temp12 = vec_mergeh(tempG, tempO);
1099     temp13 = vec_mergel(tempG, tempO);
1100     temp14 = vec_mergeh(tempH, tempP);
1101     temp15 = vec_mergel(tempH, tempP);
1102
1103     vec_st(temp0, 0, dst);
1104     vec_st(temp1, 16, dst);
1105     vec_st(temp2, 32, dst);
1106     vec_st(temp3, 48, dst);
1107     vec_st(temp4, 64, dst);
1108     vec_st(temp5, 80, dst);
1109     vec_st(temp6, 96, dst);
1110     vec_st(temp7, 112, dst);
1111     vec_st(temp8, 128, dst);
1112     vec_st(temp9, 144, dst);
1113     vec_st(temp10, 160, dst);
1114     vec_st(temp11, 176, dst);
1115     vec_st(temp12, 192, dst);
1116     vec_st(temp13, 208, dst);
1117     vec_st(temp14, 224, dst);
1118     vec_st(temp15, 240, dst);
1119 }
1120
1121 static inline void transpose_8x16_char_fromPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1122     const vector unsigned char zero = vec_splat_u8(0);
1123
1124 #define LOAD_DOUBLE_LINE(i, j)                                  \
1125     vector unsigned char src##i = vec_ld(i * 16, src);            \
1126     vector unsigned char src##j = vec_ld(j * 16, src)
1127
1128     LOAD_DOUBLE_LINE(0, 1);
1129     LOAD_DOUBLE_LINE(2, 3);
1130     LOAD_DOUBLE_LINE(4, 5);
1131     LOAD_DOUBLE_LINE(6, 7);
1132     LOAD_DOUBLE_LINE(8, 9);
1133     LOAD_DOUBLE_LINE(10, 11);
1134     LOAD_DOUBLE_LINE(12, 13);
1135     LOAD_DOUBLE_LINE(14, 15);
1136 #undef LOAD_DOUBLE_LINE
1137
1138     vector unsigned char tempA = vec_mergeh(src0, src8);
1139     vector unsigned char tempB;
1140     vector unsigned char tempC = vec_mergeh(src1, src9);
1141     vector unsigned char tempD;
1142     vector unsigned char tempE = vec_mergeh(src2, src10);
1143     vector unsigned char tempG = vec_mergeh(src3, src11);
1144     vector unsigned char tempI = vec_mergeh(src4, src12);
1145     vector unsigned char tempJ;
1146     vector unsigned char tempK = vec_mergeh(src5, src13);
1147     vector unsigned char tempL;
1148     vector unsigned char tempM = vec_mergeh(src6, src14);
1149     vector unsigned char tempO = vec_mergeh(src7, src15);
1150
1151     vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1152     vector unsigned char temp1 = vec_mergel(tempA, tempI);
1153     vector unsigned char temp2;
1154     vector unsigned char temp3;
1155     vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1156     vector unsigned char temp5 = vec_mergel(tempC, tempK);
1157     vector unsigned char temp6;
1158     vector unsigned char temp7;
1159     vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1160     vector unsigned char temp9 = vec_mergel(tempE, tempM);
1161     vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1162     vector unsigned char temp13 = vec_mergel(tempG, tempO);
1163
1164     tempA = vec_mergeh(temp0, temp8);
1165     tempB = vec_mergel(temp0, temp8);
1166     tempC = vec_mergeh(temp1, temp9);
1167     tempD = vec_mergel(temp1, temp9);
1168     tempI = vec_mergeh(temp4, temp12);
1169     tempJ = vec_mergel(temp4, temp12);
1170     tempK = vec_mergeh(temp5, temp13);
1171     tempL = vec_mergel(temp5, temp13);
1172
1173     temp0 = vec_mergeh(tempA, tempI);
1174     temp1 = vec_mergel(tempA, tempI);
1175     temp2 = vec_mergeh(tempB, tempJ);
1176     temp3 = vec_mergel(tempB, tempJ);
1177     temp4 = vec_mergeh(tempC, tempK);
1178     temp5 = vec_mergel(tempC, tempK);
1179     temp6 = vec_mergeh(tempD, tempL);
1180     temp7 = vec_mergel(tempD, tempL);
1181
1182
1183     const vector signed char neg1 = vec_splat_s8(-1);
1184 #define STORE_DOUBLE_LINE(i, j)                                         \
1185     vector unsigned char dstA##i = vec_ld(i * stride, dst);             \
1186     vector unsigned char dstB##i = vec_ld(i * stride + 16, dst);        \
1187     vector unsigned char dstA##j = vec_ld(j * stride, dst);             \
1188     vector unsigned char dstB##j = vec_ld(j * stride+ 16, dst);         \
1189     vector unsigned char align##i = vec_lvsr(i * stride, dst);          \
1190     vector unsigned char align##j = vec_lvsr(j * stride, dst);          \
1191     vector unsigned char mask##i = vec_perm(zero, (vector unsigned char)neg1, align##i); \
1192     vector unsigned char mask##j = vec_perm(zero, (vector unsigned char)neg1, align##j); \
1193     vector unsigned char dstR##i = vec_perm(temp##i, temp##i, align##i);\
1194     vector unsigned char dstR##j = vec_perm(temp##j, temp##j, align##j);\
1195     vector unsigned char dstAF##i = vec_sel(dstA##i, dstR##i, mask##i); \
1196     vector unsigned char dstBF##i = vec_sel(dstR##i, dstB##i, mask##i); \
1197     vector unsigned char dstAF##j = vec_sel(dstA##j, dstR##j, mask##j); \
1198     vector unsigned char dstBF##j = vec_sel(dstR##j, dstB##j, mask##j); \
1199     vec_st(dstAF##i, i * stride, dst);                                  \
1200     vec_st(dstBF##i, i * stride + 16, dst);                             \
1201     vec_st(dstAF##j, j * stride, dst);                                  \
1202     vec_st(dstBF##j, j * stride + 16, dst)
1203
1204     STORE_DOUBLE_LINE(0,1);
1205     STORE_DOUBLE_LINE(2,3);
1206     STORE_DOUBLE_LINE(4,5);
1207     STORE_DOUBLE_LINE(6,7);
1208 }