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 "libavutil/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 #define LOAD_LINE(i)                                                    \
 107     {                                                                   \
 108     vector unsigned char perm##i = vec_lvsl(j##i, src2);                \
 109     vector unsigned char v_srcA2##i;                                    \
 110     vector unsigned char v_srcA1##i = vec_ld(j##i, src2);               \
 111     if (two_vectors)                                                    \
 112         v_srcA2##i = vec_ld(j##i + 16, src2);                           \
 113     v_srcA##i =                                                         \
 114         vec_perm(v_srcA1##i, v_srcA2##i, perm##i);                      \
 115     v_srcAss##i =                                                       \
 116         (vector signed short)vec_mergeh((vector signed char)zero,       \
 117                                         (vector signed char)v_srcA##i); }
 118
 119 #define LOAD_LINE_ALIGNED(i)                                            \
 120     v_srcA##i = vec_ld(j##i, src2);                                     \
 121     v_srcAss##i =                                                       \
 122         (vector signed short)vec_mergeh((vector signed char)zero,       \
 123                                         (vector signed char)v_srcA##i)
 124
 125     /* Special-casing the aligned case is worthwhile, as all calls from
 126      * the (transposed) horizontable deblocks will be aligned, in addition
 127      * to the naturally aligned vertical deblocks. */
 128     if (properStride && srcAlign) {
 129         LOAD_LINE_ALIGNED(0);
 130         LOAD_LINE_ALIGNED(1);
 131         LOAD_LINE_ALIGNED(2);
 132         LOAD_LINE_ALIGNED(3);
 133         LOAD_LINE_ALIGNED(4);
 134         LOAD_LINE_ALIGNED(5);
 135         LOAD_LINE_ALIGNED(6);
 136         LOAD_LINE_ALIGNED(7);
 137     } else {
 138         LOAD_LINE(0);
 139         LOAD_LINE(1);
 140         LOAD_LINE(2);
 141         LOAD_LINE(3);
 142         LOAD_LINE(4);
 143         LOAD_LINE(5);
 144         LOAD_LINE(6);
 145         LOAD_LINE(7);
 146     }
 147 #undef LOAD_LINE
 148 #undef LOAD_LINE_ALIGNED
 149
 150 #define ITER(i, j)                                                      \
 151     const vector signed short v_diff##i =                               \
 152         vec_sub(v_srcAss##i, v_srcAss##j);                              \
 153     const vector signed short v_sum##i =                                \
 154         vec_add(v_diff##i, v_dcOffset);                                 \
 155     const vector signed short v_comp##i =                               \
 156         (vector signed short)vec_cmplt((vector unsigned short)v_sum##i, \
 157                                        v_dcThreshold);                  \
 158     const vector signed short v_part##i = vec_and(mask, v_comp##i);
 159
 160     {
 161         ITER(0, 1)
 162         ITER(1, 2)
 163         ITER(2, 3)
 164         ITER(3, 4)
 165         ITER(4, 5)
 166         ITER(5, 6)
 167         ITER(6, 7)
 168
 169         v_numEq = vec_sum4s(v_part0, v_numEq);
 170         v_numEq = vec_sum4s(v_part1, v_numEq);
 171         v_numEq = vec_sum4s(v_part2, v_numEq);
 172         v_numEq = vec_sum4s(v_part3, v_numEq);
 173         v_numEq = vec_sum4s(v_part4, v_numEq);
 174         v_numEq = vec_sum4s(v_part5, v_numEq);
 175         v_numEq = vec_sum4s(v_part6, v_numEq);
 176     }
 177
 178 #undef ITER
 179
 180     v_numEq = vec_sums(v_numEq, zero);
 181
 182     v_numEq = vec_splat(v_numEq, 3);
 183     vec_ste(v_numEq, 0, &numEq);
 184
 185     if (numEq > c->ppMode.flatnessThreshold){
 186         const vector unsigned char mmoP1 = (const vector unsigned char)
 187             {0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f,
 188              0x00, 0x01, 0x12, 0x13, 0x08, 0x09, 0x1A, 0x1B};
 189         const vector unsigned char mmoP2 = (const vector unsigned char)
 190             {0x04, 0x05, 0x16, 0x17, 0x0C, 0x0D, 0x1E, 0x1F,
 191              0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f};
 192         const vector unsigned char mmoP = (const vector unsigned char)
 193             vec_lvsl(8, (unsigned char*)0);
 194
 195         vector signed short mmoL1 = vec_perm(v_srcAss0, v_srcAss2, mmoP1);
 196         vector signed short mmoL2 = vec_perm(v_srcAss4, v_srcAss6, mmoP2);
 197         vector signed short mmoL = vec_perm(mmoL1, mmoL2, mmoP);
 198         vector signed short mmoR1 = vec_perm(v_srcAss5, v_srcAss7, mmoP1);
 199         vector signed short mmoR2 = vec_perm(v_srcAss1, v_srcAss3, mmoP2);
 200         vector signed short mmoR = vec_perm(mmoR1, mmoR2, mmoP);
 201         vector signed short mmoDiff = vec_sub(mmoL, mmoR);
 202         vector unsigned short mmoSum = (vector unsigned short)vec_add(mmoDiff, v2QP);
 203
 204         if (vec_any_gt(mmoSum, v4QP))
 205             return 0;
 206         else
 207             return 1;
 208     }
 209     else return 2;
 210 }
 211
 212 static inline void doVertLowPass_altivec(uint8_t *src, int stride, PPContext *c) {
 213     /*
 214     this code makes no assumption on src or stride.
 215     One could remove the recomputation of the perm
 216     vector by assuming (stride % 16) == 0, unfortunately
 217     this is not always true. Quite a lot of load/stores
 218     can be removed by assuming proper alignment of
 219     src & stride :-(
 220     */
 221     uint8_t *src2 = src;
 222     const vector signed int zero = vec_splat_s32(0);
 223     const int properStride = (stride % 16);
 224     const int srcAlign = ((unsigned long)src2 % 16);
 225     DECLARE_ALIGNED(16, short, qp[8]) = {c->QP};
 226     vector signed short vqp = vec_ld(0, qp);
 227     vector signed short vb0, vb1, vb2, vb3, vb4, vb5, vb6, vb7, vb8, vb9;
 228     vector unsigned char vbA0, av_uninit(vbA1), av_uninit(vbA2), av_uninit(vbA3), av_uninit(vbA4), av_uninit(vbA5), av_uninit(vbA6), av_uninit(vbA7), av_uninit(vbA8), vbA9;
 229     vector unsigned char vbB0, av_uninit(vbB1), av_uninit(vbB2), av_uninit(vbB3), av_uninit(vbB4), av_uninit(vbB5), av_uninit(vbB6), av_uninit(vbB7), av_uninit(vbB8), vbB9;
 230     vector unsigned char vbT0, vbT1, vbT2, vbT3, vbT4, vbT5, vbT6, vbT7, vbT8, vbT9;
 231     vector unsigned char perml0, perml1, perml2, perml3, perml4,
 232                          perml5, perml6, perml7, perml8, perml9;
 233     register int j0 = 0,
 234                  j1 = stride,
 235                  j2 = 2 * stride,
 236                  j3 = 3 * stride,
 237                  j4 = 4 * stride,
 238                  j5 = 5 * stride,
 239                  j6 = 6 * stride,
 240                  j7 = 7 * stride,
 241                  j8 = 8 * stride,
 242                  j9 = 9 * stride;
 243
 244     vqp = vec_splat(vqp, 0);
 245
 246     src2 += stride*3;
 247
 248 #define LOAD_LINE(i)                                                    \
 249     perml##i = vec_lvsl(i * stride, src2);                              \
 250     vbA##i = vec_ld(i * stride, src2);                                  \
 251     vbB##i = vec_ld(i * stride + 16, src2);                             \
 252     vbT##i = vec_perm(vbA##i, vbB##i, perml##i);                        \
 253     vb##i =                                                             \
 254         (vector signed short)vec_mergeh((vector unsigned char)zero,     \
 255                                         (vector unsigned char)vbT##i)
 256
 257 #define LOAD_LINE_ALIGNED(i)                                            \
 258     vbT##i = vec_ld(j##i, src2);                                        \
 259     vb##i =                                                             \
 260         (vector signed short)vec_mergeh((vector signed char)zero,       \
 261                                         (vector signed char)vbT##i)
 262
 263       /* Special-casing the aligned case is worthwhile, as all calls from
 264        * the (transposed) horizontable deblocks will be aligned, in addition
 265        * to the naturally aligned vertical deblocks. */
 266     if (properStride && srcAlign) {
 267           LOAD_LINE_ALIGNED(0);
 268           LOAD_LINE_ALIGNED(1);
 269           LOAD_LINE_ALIGNED(2);
 270           LOAD_LINE_ALIGNED(3);
 271           LOAD_LINE_ALIGNED(4);
 272           LOAD_LINE_ALIGNED(5);
 273           LOAD_LINE_ALIGNED(6);
 274           LOAD_LINE_ALIGNED(7);
 275           LOAD_LINE_ALIGNED(8);
 276           LOAD_LINE_ALIGNED(9);
 277     } else {
 278           LOAD_LINE(0);
 279           LOAD_LINE(1);
 280           LOAD_LINE(2);
 281           LOAD_LINE(3);
 282           LOAD_LINE(4);
 283           LOAD_LINE(5);
 284           LOAD_LINE(6);
 285           LOAD_LINE(7);
 286           LOAD_LINE(8);
 287           LOAD_LINE(9);
 288     }
 289 #undef LOAD_LINE
 290 #undef LOAD_LINE_ALIGNED
 291     {
 292         const vector unsigned short v_2 = vec_splat_u16(2);
 293         const vector unsigned short v_4 = vec_splat_u16(4);
 294
 295         const vector signed short v_diff01 = vec_sub(vb0, vb1);
 296         const vector unsigned short v_cmp01 =
 297             (const vector unsigned short) vec_cmplt(vec_abs(v_diff01), vqp);
 298         const vector signed short v_first = vec_sel(vb1, vb0, v_cmp01);
 299         const vector signed short v_diff89 = vec_sub(vb8, vb9);
 300         const vector unsigned short v_cmp89 =
 301             (const vector unsigned short) vec_cmplt(vec_abs(v_diff89), vqp);
 302         const vector signed short v_last = vec_sel(vb8, vb9, v_cmp89);
 303
 304         const vector signed short temp01 = vec_mladd(v_first, (vector signed short)v_4, vb1);
 305         const vector signed short temp02 = vec_add(vb2, vb3);
 306         const vector signed short temp03 = vec_add(temp01, (vector signed short)v_4);
 307         const vector signed short v_sumsB0 = vec_add(temp02, temp03);
 308
 309         const vector signed short temp11 = vec_sub(v_sumsB0, v_first);
 310         const vector signed short v_sumsB1 = vec_add(temp11, vb4);
 311
 312         const vector signed short temp21 = vec_sub(v_sumsB1, v_first);
 313         const vector signed short v_sumsB2 = vec_add(temp21, vb5);
 314
 315         const vector signed short temp31 = vec_sub(v_sumsB2, v_first);
 316         const vector signed short v_sumsB3 = vec_add(temp31, vb6);
 317
 318         const vector signed short temp41 = vec_sub(v_sumsB3, v_first);
 319         const vector signed short v_sumsB4 = vec_add(temp41, vb7);
 320
 321         const vector signed short temp51 = vec_sub(v_sumsB4, vb1);
 322         const vector signed short v_sumsB5 = vec_add(temp51, vb8);
 323
 324         const vector signed short temp61 = vec_sub(v_sumsB5, vb2);
 325         const vector signed short v_sumsB6 = vec_add(temp61, v_last);
 326
 327         const vector signed short temp71 = vec_sub(v_sumsB6, vb3);
 328         const vector signed short v_sumsB7 = vec_add(temp71, v_last);
 329
 330         const vector signed short temp81 = vec_sub(v_sumsB7, vb4);
 331         const vector signed short v_sumsB8 = vec_add(temp81, v_last);
 332
 333         const vector signed short temp91 = vec_sub(v_sumsB8, vb5);
 334         const vector signed short v_sumsB9 = vec_add(temp91, v_last);
 335
 336     #define COMPUTE_VR(i, j, k)                                             \
 337         const vector signed short temps1##i =                               \
 338             vec_add(v_sumsB##i, v_sumsB##k);                                \
 339         const vector signed short temps2##i =                               \
 340             vec_mladd(vb##j, (vector signed short)v_2, temps1##i);          \
 341         const vector signed short  vr##j = vec_sra(temps2##i, v_4)
 342
 343         COMPUTE_VR(0, 1, 2);
 344         COMPUTE_VR(1, 2, 3);
 345         COMPUTE_VR(2, 3, 4);
 346         COMPUTE_VR(3, 4, 5);
 347         COMPUTE_VR(4, 5, 6);
 348         COMPUTE_VR(5, 6, 7);
 349         COMPUTE_VR(6, 7, 8);
 350         COMPUTE_VR(7, 8, 9);
 351
 352         const vector signed char neg1 = vec_splat_s8(-1);
 353         const vector unsigned char permHH = (const vector unsigned char){0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
 354                                                                          0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F};
 355
 356 #define PACK_AND_STORE(i)                                       \
 357 {   const vector unsigned char perms##i =                       \
 358         vec_lvsr(i * stride, src2);                             \
 359     const vector unsigned char vf##i =                          \
 360         vec_packsu(vr##i, (vector signed short)zero);           \
 361     const vector unsigned char vg##i =                          \
 362         vec_perm(vf##i, vbT##i, permHH);                        \
 363     const vector unsigned char mask##i =                        \
 364         vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \
 365     const vector unsigned char vg2##i =                         \
 366         vec_perm(vg##i, vg##i, perms##i);                       \
 367     const vector unsigned char svA##i =                         \
 368         vec_sel(vbA##i, vg2##i, mask##i);                       \
 369     const vector unsigned char svB##i =                         \
 370         vec_sel(vg2##i, vbB##i, mask##i);                       \
 371     vec_st(svA##i, i * stride, src2);                           \
 372     vec_st(svB##i, i * stride + 16, src2);}
 373
 374 #define PACK_AND_STORE_ALIGNED(i)                               \
 375 {   const vector unsigned char vf##i =                          \
 376         vec_packsu(vr##i, (vector signed short)zero);           \
 377     const vector unsigned char vg##i =                          \
 378         vec_perm(vf##i, vbT##i, permHH);                        \
 379     vec_st(vg##i, i * stride, src2);}
 380
 381         /* Special-casing the aligned case is worthwhile, as all calls from
 382          * the (transposed) horizontable deblocks will be aligned, in addition
 383          * to the naturally aligned vertical deblocks. */
 384         if (properStride && srcAlign) {
 385             PACK_AND_STORE_ALIGNED(1)
 386             PACK_AND_STORE_ALIGNED(2)
 387             PACK_AND_STORE_ALIGNED(3)
 388             PACK_AND_STORE_ALIGNED(4)
 389             PACK_AND_STORE_ALIGNED(5)
 390             PACK_AND_STORE_ALIGNED(6)
 391             PACK_AND_STORE_ALIGNED(7)
 392             PACK_AND_STORE_ALIGNED(8)
 393         } else {
 394             PACK_AND_STORE(1)
 395             PACK_AND_STORE(2)
 396             PACK_AND_STORE(3)
 397             PACK_AND_STORE(4)
 398             PACK_AND_STORE(5)
 399             PACK_AND_STORE(6)
 400             PACK_AND_STORE(7)
 401             PACK_AND_STORE(8)
 402         }
 403     #undef PACK_AND_STORE
 404     #undef PACK_AND_STORE_ALIGNED
 405     }
 406 }
 407
 408
 409
 410 static inline void doVertDefFilter_altivec(uint8_t src[], int stride, PPContext *c) {
 411     /*
 412     this code makes no assumption on src or stride.
 413     One could remove the recomputation of the perm
 414     vector by assuming (stride % 16) == 0, unfortunately
 415     this is not always true. Quite a lot of load/stores
 416     can be removed by assuming proper alignment of
 417     src & stride :-(
 418     */
 419     uint8_t *src2 = src + stride*3;
 420     const vector signed int zero = vec_splat_s32(0);
 421     DECLARE_ALIGNED(16, short, qp[8]) = {8*c->QP};
 422     vector signed short vqp = vec_splat(
 423                                 (vector signed short)vec_ld(0, qp), 0);
 424
 425 #define LOAD_LINE(i)                                                    \
 426     const vector unsigned char perm##i =                                \
 427         vec_lvsl(i * stride, src2);                                     \
 428     const vector unsigned char vbA##i =                                 \
 429         vec_ld(i * stride, src2);                                       \
 430     const vector unsigned char vbB##i =                                 \
 431         vec_ld(i * stride + 16, src2);                                  \
 432     const vector unsigned char vbT##i =                                 \
 433         vec_perm(vbA##i, vbB##i, perm##i);                              \
 434     const vector signed short vb##i =                                   \
 435         (vector signed short)vec_mergeh((vector unsigned char)zero,     \
 436                                         (vector unsigned char)vbT##i)
 437
 438      LOAD_LINE(1);
 439      LOAD_LINE(2);
 440      LOAD_LINE(3);
 441      LOAD_LINE(4);
 442      LOAD_LINE(5);
 443      LOAD_LINE(6);
 444      LOAD_LINE(7);
 445      LOAD_LINE(8);
 446 #undef LOAD_LINE
 447
 448      const vector signed short v_1 = vec_splat_s16(1);
 449      const vector signed short v_2 = vec_splat_s16(2);
 450      const vector signed short v_5 = vec_splat_s16(5);
 451      const vector signed short v_32 = vec_sl(v_1,
 452                                              (vector unsigned short)v_5);
 453      /* middle energy */
 454      const vector signed short l3minusl6 = vec_sub(vb3, vb6);
 455      const vector signed short l5minusl4 = vec_sub(vb5, vb4);
 456      const vector signed short twotimes_l3minusl6 = vec_mladd(v_2, l3minusl6, (vector signed short)zero);
 457      const vector signed short mE = vec_mladd(v_5, l5minusl4, twotimes_l3minusl6);
 458      const vector signed short absmE = vec_abs(mE);
 459      /* left & right energy */
 460      const vector signed short l1minusl4 = vec_sub(vb1, vb4);
 461      const vector signed short l3minusl2 = vec_sub(vb3, vb2);
 462      const vector signed short l5minusl8 = vec_sub(vb5, vb8);
 463      const vector signed short l7minusl6 = vec_sub(vb7, vb6);
 464      const vector signed short twotimes_l1minusl4 = vec_mladd(v_2, l1minusl4, (vector signed short)zero);
 465      const vector signed short twotimes_l5minusl8 = vec_mladd(v_2, l5minusl8, (vector signed short)zero);
 466      const vector signed short lE = vec_mladd(v_5, l3minusl2, twotimes_l1minusl4);
 467      const vector signed short rE = vec_mladd(v_5, l7minusl6, twotimes_l5minusl8);
 468      /* d */
 469      const vector signed short ddiff = vec_sub(absmE,
 470                                                vec_min(vec_abs(lE),
 471                                                        vec_abs(rE)));
 472      const vector signed short ddiffclamp = vec_max(ddiff, (vector signed short)zero);
 473      const vector signed short dtimes64 = vec_mladd(v_5, ddiffclamp, v_32);
 474      const vector signed short d = vec_sra(dtimes64, vec_splat_u16(6));
 475      const vector signed short minusd = vec_sub((vector signed short)zero, d);
 476      const vector signed short finald = vec_sel(minusd,
 477                                                 d,
 478                                                 vec_cmpgt(vec_sub((vector signed short)zero, mE),
 479                                                           (vector signed short)zero));
 480      /* q */
 481      const vector signed short qtimes2 = vec_sub(vb4, vb5);
 482      /* for a shift right to behave like /2, we need to add one
 483         to all negative integer */
 484      const vector signed short rounddown = vec_sel((vector signed short)zero,
 485                                                    v_1,
 486                                                    vec_cmplt(qtimes2, (vector signed short)zero));
 487      const vector signed short q = vec_sra(vec_add(qtimes2, rounddown), vec_splat_u16(1));
 488      /* clamp */
 489      const vector signed short dclamp_P1 = vec_max((vector signed short)zero, finald);
 490      const vector signed short dclamp_P = vec_min(dclamp_P1, q);
 491      const vector signed short dclamp_N1 = vec_min((vector signed short)zero, finald);
 492      const vector signed short dclamp_N = vec_max(dclamp_N1, q);
 493
 494      const vector signed short dclampedfinal = vec_sel(dclamp_N,
 495                                                        dclamp_P,
 496                                                        vec_cmpgt(q, (vector signed short)zero));
 497      const vector signed short dornotd = vec_sel((vector signed short)zero,
 498                                                  dclampedfinal,
 499                                                  vec_cmplt(absmE, vqp));
 500      /* add/subtract to l4 and l5 */
 501      const vector signed short vb4minusd = vec_sub(vb4, dornotd);
 502      const vector signed short vb5plusd  = vec_add(vb5, dornotd);
 503      /* finally, stores */
 504      const vector unsigned char st4 = vec_packsu(vb4minusd, (vector signed short)zero);
 505      const vector unsigned char st5 = vec_packsu(vb5plusd,  (vector signed short)zero);
 506
 507      const vector signed char neg1 = vec_splat_s8(-1);
 508      const vector unsigned char permHH = (const vector unsigned char){0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
 509                                                                       0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F};
 510
 511 #define STORE(i)                                                \
 512 {    const vector unsigned char perms##i =                      \
 513          vec_lvsr(i * stride, src2);                            \
 514      const vector unsigned char vg##i =                         \
 515          vec_perm(st##i, vbT##i, permHH);                       \
 516      const vector unsigned char mask##i =                       \
 517          vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \
 518      const vector unsigned char vg2##i =                        \
 519          vec_perm(vg##i, vg##i, perms##i);                      \
 520      const vector unsigned char svA##i =                        \
 521          vec_sel(vbA##i, vg2##i, mask##i);                      \
 522      const vector unsigned char svB##i =                        \
 523          vec_sel(vg2##i, vbB##i, mask##i);                      \
 524      vec_st(svA##i, i * stride, src2);                          \
 525      vec_st(svB##i, i * stride + 16, src2);}
 526
 527      STORE(4)
 528      STORE(5)
 529 }
 530
 531 static inline void dering_altivec(uint8_t src[], int stride, PPContext *c) {
 532     /*
 533     this code makes no assumption on src or stride.
 534     One could remove the recomputation of the perm
 535     vector by assuming (stride % 16) == 0, unfortunately
 536     this is not always true. Quite a lot of load/stores
 537     can be removed by assuming proper alignment of
 538     src & stride :-(
 539     */
 540     uint8_t *srcCopy = src;
 541     DECLARE_ALIGNED(16, uint8_t, dt[16]);
 542     const vector signed int zero = vec_splat_s32(0);
 543     vector unsigned char v_dt;
 544     dt[0] = deringThreshold;
 545     v_dt = vec_splat(vec_ld(0, dt), 0);
 546
 547 #define LOAD_LINE(i)                                                  \
 548     const vector unsigned char perm##i =                              \
 549         vec_lvsl(i * stride, srcCopy);                                \
 550     vector unsigned char sA##i = vec_ld(i * stride, srcCopy);         \
 551     vector unsigned char sB##i = vec_ld(i * stride + 16, srcCopy);    \
 552     vector unsigned char src##i = vec_perm(sA##i, sB##i, perm##i)
 553
 554     LOAD_LINE(0);
 555     LOAD_LINE(1);
 556     LOAD_LINE(2);
 557     LOAD_LINE(3);
 558     LOAD_LINE(4);
 559     LOAD_LINE(5);
 560     LOAD_LINE(6);
 561     LOAD_LINE(7);
 562     LOAD_LINE(8);
 563     LOAD_LINE(9);
 564 #undef LOAD_LINE
 565
 566     vector unsigned char v_avg;
 567     {
 568     const vector unsigned char trunc_perm = (vector unsigned char)
 569         {0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08,
 570          0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18};
 571     const vector unsigned char trunc_src12 = vec_perm(src1, src2, trunc_perm);
 572     const vector unsigned char trunc_src34 = vec_perm(src3, src4, trunc_perm);
 573     const vector unsigned char trunc_src56 = vec_perm(src5, src6, trunc_perm);
 574     const vector unsigned char trunc_src78 = vec_perm(src7, src8, trunc_perm);
 575
 576 #define EXTRACT(op) do {                                                \
 577     const vector unsigned char s##op##_1   = vec_##op(trunc_src12, trunc_src34); \
 578     const vector unsigned char s##op##_2   = vec_##op(trunc_src56, trunc_src78); \
 579     const vector unsigned char s##op##_6   = vec_##op(s##op##_1, s##op##_2);     \
 580     const vector unsigned char s##op##_8h  = vec_mergeh(s##op##_6, s##op##_6);   \
 581     const vector unsigned char s##op##_8l  = vec_mergel(s##op##_6, s##op##_6);   \
 582     const vector unsigned char s##op##_9   = vec_##op(s##op##_8h, s##op##_8l);   \
 583     const vector unsigned char s##op##_9h  = vec_mergeh(s##op##_9, s##op##_9);   \
 584     const vector unsigned char s##op##_9l  = vec_mergel(s##op##_9, s##op##_9);   \
 585     const vector unsigned char s##op##_10  = vec_##op(s##op##_9h, s##op##_9l);   \
 586     const vector unsigned char s##op##_10h = vec_mergeh(s##op##_10, s##op##_10); \
 587     const vector unsigned char s##op##_10l = vec_mergel(s##op##_10, s##op##_10); \
 588     const vector unsigned char s##op##_11  = vec_##op(s##op##_10h, s##op##_10l); \
 589     const vector unsigned char s##op##_11h = vec_mergeh(s##op##_11, s##op##_11); \
 590     const vector unsigned char s##op##_11l = vec_mergel(s##op##_11, s##op##_11); \
 591     v_##op = vec_##op(s##op##_11h, s##op##_11l); } while (0)
 592
 593     vector unsigned char v_min;
 594     vector unsigned char v_max;
 595     EXTRACT(min);
 596     EXTRACT(max);
 597 #undef EXTRACT
 598
 599     if (vec_all_lt(vec_sub(v_max, v_min), v_dt))
 600         return;
 601
 602     v_avg = vec_avg(v_min, v_max);
 603     }
 604
 605     DECLARE_ALIGNED(16, signed int, S[8]);
 606     {
 607     const vector unsigned short mask1 = (vector unsigned short)
 608                                         {0x0001, 0x0002, 0x0004, 0x0008,
 609                                          0x0010, 0x0020, 0x0040, 0x0080};
 610     const vector unsigned short mask2 = (vector unsigned short)
 611                                         {0x0100, 0x0200, 0x0000, 0x0000,
 612                                          0x0000, 0x0000, 0x0000, 0x0000};
 613
 614     const vector unsigned int vuint32_16 = vec_sl(vec_splat_u32(1), vec_splat_u32(4));
 615     const vector unsigned int vuint32_1 = vec_splat_u32(1);
 616
 617 #define COMPARE(i)                                                      \
 618     vector signed int sum##i;                                           \
 619     do {                                                                \
 620         const vector unsigned char cmp##i =                             \
 621             (vector unsigned char)vec_cmpgt(src##i, v_avg);             \
 622         const vector unsigned short cmpHi##i =                          \
 623             (vector unsigned short)vec_mergeh(cmp##i, cmp##i);          \
 624         const vector unsigned short cmpLi##i =                          \
 625             (vector unsigned short)vec_mergel(cmp##i, cmp##i);          \
 626         const vector signed short cmpHf##i =                            \
 627             (vector signed short)vec_and(cmpHi##i, mask1);              \
 628         const vector signed short cmpLf##i =                            \
 629             (vector signed short)vec_and(cmpLi##i, mask2);              \
 630         const vector signed int sump##i = vec_sum4s(cmpHf##i, zero);    \
 631         const vector signed int sumq##i = vec_sum4s(cmpLf##i, sump##i); \
 632         sum##i  = vec_sums(sumq##i, zero); } while (0)
 633
 634     COMPARE(0);
 635     COMPARE(1);
 636     COMPARE(2);
 637     COMPARE(3);
 638     COMPARE(4);
 639     COMPARE(5);
 640     COMPARE(6);
 641     COMPARE(7);
 642     COMPARE(8);
 643     COMPARE(9);
 644 #undef COMPARE
 645
 646     vector signed int sumA2;
 647     vector signed int sumB2;
 648     {
 649     const vector signed int sump02 = vec_mergel(sum0, sum2);
 650     const vector signed int sump13 = vec_mergel(sum1, sum3);
 651     const vector signed int sumA = vec_mergel(sump02, sump13);
 652
 653     const vector signed int sump46 = vec_mergel(sum4, sum6);
 654     const vector signed int sump57 = vec_mergel(sum5, sum7);
 655     const vector signed int sumB = vec_mergel(sump46, sump57);
 656
 657     const vector signed int sump8A = vec_mergel(sum8, zero);
 658     const vector signed int sump9B = vec_mergel(sum9, zero);
 659     const vector signed int sumC = vec_mergel(sump8A, sump9B);
 660
 661     const vector signed int tA = vec_sl(vec_nor(zero, sumA), vuint32_16);
 662     const vector signed int tB = vec_sl(vec_nor(zero, sumB), vuint32_16);
 663     const vector signed int tC = vec_sl(vec_nor(zero, sumC), vuint32_16);
 664     const vector signed int t2A = vec_or(sumA, tA);
 665     const vector signed int t2B = vec_or(sumB, tB);
 666     const vector signed int t2C = vec_or(sumC, tC);
 667     const vector signed int t3A = vec_and(vec_sra(t2A, vuint32_1),
 668                                           vec_sl(t2A, vuint32_1));
 669     const vector signed int t3B = vec_and(vec_sra(t2B, vuint32_1),
 670                                           vec_sl(t2B, vuint32_1));
 671     const vector signed int t3C = vec_and(vec_sra(t2C, vuint32_1),
 672                                           vec_sl(t2C, vuint32_1));
 673     const vector signed int yA = vec_and(t2A, t3A);
 674     const vector signed int yB = vec_and(t2B, t3B);
 675     const vector signed int yC = vec_and(t2C, t3C);
 676
 677     const vector unsigned char strangeperm1 = vec_lvsl(4, (unsigned char*)0);
 678     const vector unsigned char strangeperm2 = vec_lvsl(8, (unsigned char*)0);
 679     const vector signed int sumAd4 = vec_perm(yA, yB, strangeperm1);
 680     const vector signed int sumAd8 = vec_perm(yA, yB, strangeperm2);
 681     const vector signed int sumBd4 = vec_perm(yB, yC, strangeperm1);
 682     const vector signed int sumBd8 = vec_perm(yB, yC, strangeperm2);
 683     const vector signed int sumAp = vec_and(yA,
 684                                             vec_and(sumAd4,sumAd8));
 685     const vector signed int sumBp = vec_and(yB,
 686                                             vec_and(sumBd4,sumBd8));
 687     sumA2 = vec_or(sumAp,
 688                    vec_sra(sumAp,
 689                            vuint32_16));
 690     sumB2  = vec_or(sumBp,
 691                     vec_sra(sumBp,
 692                             vuint32_16));
 693     }
 694     vec_st(sumA2, 0, S);
 695     vec_st(sumB2, 16, S);
 696     }
 697
 698     /* I'm not sure the following is actually faster
 699        than straight, unvectorized C code :-( */
 700
 701     DECLARE_ALIGNED(16, int, tQP2[4]);
 702     tQP2[0]= c->QP/2 + 1;
 703     vector signed int vQP2 = vec_ld(0, tQP2);
 704     vQP2 = vec_splat(vQP2, 0);
 705     const vector signed int vsint32_8 = vec_splat_s32(8);
 706     const vector unsigned int vuint32_4 = vec_splat_u32(4);
 707
 708     const vector unsigned char permA1 = (vector unsigned char)
 709         {0x00, 0x01, 0x02, 0x10, 0x11, 0x12, 0x1F, 0x1F,
 710          0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F};
 711     const vector unsigned char permA2 = (vector unsigned char)
 712         {0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x10, 0x11,
 713          0x12, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F};
 714     const vector unsigned char permA1inc = (vector unsigned char)
 715         {0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00,
 716          0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00};
 717     const vector unsigned char permA2inc = (vector unsigned char)
 718         {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01,
 719          0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00};
 720     const vector unsigned char magic = (vector unsigned char)
 721         {0x01, 0x02, 0x01, 0x02, 0x04, 0x02, 0x01, 0x02,
 722          0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00};
 723     const vector unsigned char extractPerm = (vector unsigned char)
 724         {0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01,
 725          0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01};
 726     const vector unsigned char extractPermInc = (vector unsigned char)
 727         {0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01,
 728          0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01};
 729     const vector unsigned char identity = vec_lvsl(0,(unsigned char *)0);
 730     const vector unsigned char tenRight = (vector unsigned char)
 731         {0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
 732          0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00};
 733     const vector unsigned char eightLeft = (vector unsigned char)
 734         {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
 735          0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08};
 736
 737
 738 #define F_INIT(i)                                       \
 739     vector unsigned char tenRightM##i = tenRight;       \
 740     vector unsigned char permA1M##i = permA1;           \
 741     vector unsigned char permA2M##i = permA2;           \
 742     vector unsigned char extractPermM##i = extractPerm
 743
 744 #define F2(i, j, k, l)                                                  \
 745     if (S[i] & (1 << (l+1))) {                                          \
 746         const vector unsigned char a_##j##_A##l =                       \
 747             vec_perm(src##i, src##j, permA1M##i);                       \
 748         const vector unsigned char a_##j##_B##l =                       \
 749             vec_perm(a_##j##_A##l, src##k, permA2M##i);                 \
 750         const vector signed int a_##j##_sump##l =                       \
 751             (vector signed int)vec_msum(a_##j##_B##l, magic,            \
 752                                         (vector unsigned int)zero);     \
 753         vector signed int F_##j##_##l =                                 \
 754             vec_sr(vec_sums(a_##j##_sump##l, vsint32_8), vuint32_4);    \
 755         F_##j##_##l = vec_splat(F_##j##_##l, 3);                        \
 756         const vector signed int p_##j##_##l =                           \
 757             (vector signed int)vec_perm(src##j,                         \
 758                                         (vector unsigned char)zero,     \
 759                                         extractPermM##i);               \
 760         const vector signed int sum_##j##_##l  = vec_add( p_##j##_##l, vQP2);\
 761         const vector signed int diff_##j##_##l = vec_sub( p_##j##_##l, vQP2);\
 762         vector signed int newpm_##j##_##l;                              \
 763         if (vec_all_lt(sum_##j##_##l, F_##j##_##l))                     \
 764             newpm_##j##_##l = sum_##j##_##l;                            \
 765         else if (vec_all_gt(diff_##j##_##l, F_##j##_##l))               \
 766             newpm_##j##_##l = diff_##j##_##l;                           \
 767         else newpm_##j##_##l = F_##j##_##l;                             \
 768         const vector unsigned char newpm2_##j##_##l =                   \
 769             vec_splat((vector unsigned char)newpm_##j##_##l, 15);       \
 770         const vector unsigned char mask##j##l = vec_add(identity,       \
 771                                                         tenRightM##i);  \
 772         src##j = vec_perm(src##j, newpm2_##j##_##l, mask##j##l);        \
 773     }                                                                   \
 774     permA1M##i = vec_add(permA1M##i, permA1inc);                        \
 775     permA2M##i = vec_add(permA2M##i, permA2inc);                        \
 776     tenRightM##i = vec_sro(tenRightM##i, eightLeft);                    \
 777     extractPermM##i = vec_add(extractPermM##i, extractPermInc)
 778
 779 #define ITER(i, j, k)                           \
 780     F_INIT(i);                                  \
 781     F2(i, j, k, 0);                             \
 782     F2(i, j, k, 1);                             \
 783     F2(i, j, k, 2);                             \
 784     F2(i, j, k, 3);                             \
 785     F2(i, j, k, 4);                             \
 786     F2(i, j, k, 5);                             \
 787     F2(i, j, k, 6);                             \
 788     F2(i, j, k, 7)
 789
 790     ITER(0, 1, 2);
 791     ITER(1, 2, 3);
 792     ITER(2, 3, 4);
 793     ITER(3, 4, 5);
 794     ITER(4, 5, 6);
 795     ITER(5, 6, 7);
 796     ITER(6, 7, 8);
 797     ITER(7, 8, 9);
 798
 799     const vector signed char neg1 = vec_splat_s8(-1);
 800
 801 #define STORE_LINE(i)                                   \
 802     const vector unsigned char permST##i =              \
 803         vec_lvsr(i * stride, srcCopy);                  \
 804     const vector unsigned char maskST##i =              \
 805         vec_perm((vector unsigned char)zero,            \
 806                  (vector unsigned char)neg1, permST##i);\
 807     src##i = vec_perm(src##i ,src##i, permST##i);       \
 808     sA##i= vec_sel(sA##i, src##i, maskST##i);           \
 809     sB##i= vec_sel(src##i, sB##i, maskST##i);           \
 810     vec_st(sA##i, i * stride, srcCopy);                 \
 811     vec_st(sB##i, i * stride + 16, srcCopy)
 812
 813     STORE_LINE(1);
 814     STORE_LINE(2);
 815     STORE_LINE(3);
 816     STORE_LINE(4);
 817     STORE_LINE(5);
 818     STORE_LINE(6);
 819     STORE_LINE(7);
 820     STORE_LINE(8);
 821
 822 #undef STORE_LINE
 823 #undef ITER
 824 #undef F2
 825 }
 826
 827 #define doHorizLowPass_altivec(a...) doHorizLowPass_C(a)
 828 #define doHorizDefFilter_altivec(a...) doHorizDefFilter_C(a)
 829 #define do_a_deblock_altivec(a...) do_a_deblock_C(a)
 830
 831 static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,
 832                                             uint8_t *tempBlurred, uint32_t *tempBlurredPast, int *maxNoise)
 833 {
 834     const vector signed int zero = vec_splat_s32(0);
 835     const vector signed short vsint16_1 = vec_splat_s16(1);
 836     vector signed int v_dp = zero;
 837     vector signed int v_sysdp = zero;
 838     int d, sysd, i;
 839
 840     tempBlurredPast[127]= maxNoise[0];
 841     tempBlurredPast[128]= maxNoise[1];
 842     tempBlurredPast[129]= maxNoise[2];
 843
 844 #define LOAD_LINE(src, i)                                               \
 845     register int j##src##i = i * stride;                                \
 846     vector unsigned char perm##src##i = vec_lvsl(j##src##i, src);       \
 847     const vector unsigned char v_##src##A1##i = vec_ld(j##src##i, src); \
 848     const vector unsigned char v_##src##A2##i = vec_ld(j##src##i + 16, src); \
 849     const vector unsigned char v_##src##A##i =                          \
 850         vec_perm(v_##src##A1##i, v_##src##A2##i, perm##src##i);         \
 851     vector signed short v_##src##Ass##i =                               \
 852         (vector signed short)vec_mergeh((vector signed char)zero,       \
 853                                         (vector signed char)v_##src##A##i)
 854
 855     LOAD_LINE(src, 0);
 856     LOAD_LINE(src, 1);
 857     LOAD_LINE(src, 2);
 858     LOAD_LINE(src, 3);
 859     LOAD_LINE(src, 4);
 860     LOAD_LINE(src, 5);
 861     LOAD_LINE(src, 6);
 862     LOAD_LINE(src, 7);
 863
 864     LOAD_LINE(tempBlurred, 0);
 865     LOAD_LINE(tempBlurred, 1);
 866     LOAD_LINE(tempBlurred, 2);
 867     LOAD_LINE(tempBlurred, 3);
 868     LOAD_LINE(tempBlurred, 4);
 869     LOAD_LINE(tempBlurred, 5);
 870     LOAD_LINE(tempBlurred, 6);
 871     LOAD_LINE(tempBlurred, 7);
 872 #undef LOAD_LINE
 873
 874 #define ACCUMULATE_DIFFS(i)                                     \
 875     vector signed short v_d##i = vec_sub(v_tempBlurredAss##i,   \
 876                                          v_srcAss##i);          \
 877     v_dp = vec_msums(v_d##i, v_d##i, v_dp);                     \
 878     v_sysdp = vec_msums(v_d##i, vsint16_1, v_sysdp)
 879
 880     ACCUMULATE_DIFFS(0);
 881     ACCUMULATE_DIFFS(1);
 882     ACCUMULATE_DIFFS(2);
 883     ACCUMULATE_DIFFS(3);
 884     ACCUMULATE_DIFFS(4);
 885     ACCUMULATE_DIFFS(5);
 886     ACCUMULATE_DIFFS(6);
 887     ACCUMULATE_DIFFS(7);
 888 #undef ACCUMULATE_DIFFS
 889
 890     v_dp = vec_sums(v_dp, zero);
 891     v_sysdp = vec_sums(v_sysdp, zero);
 892
 893     v_dp = vec_splat(v_dp, 3);
 894     v_sysdp = vec_splat(v_sysdp, 3);
 895
 896     vec_ste(v_dp, 0, &d);
 897     vec_ste(v_sysdp, 0, &sysd);
 898
 899     i = d;
 900     d = (4*d
 901          +(*(tempBlurredPast-256))
 902          +(*(tempBlurredPast-1))+ (*(tempBlurredPast+1))
 903          +(*(tempBlurredPast+256))
 904          +4)>>3;
 905
 906     *tempBlurredPast=i;
 907
 908     if (d > maxNoise[1]) {
 909         if (d < maxNoise[2]) {
 910 #define OP(i) v_tempBlurredAss##i = vec_avg(v_tempBlurredAss##i, v_srcAss##i);
 911
 912             OP(0);
 913             OP(1);
 914             OP(2);
 915             OP(3);
 916             OP(4);
 917             OP(5);
 918             OP(6);
 919             OP(7);
 920 #undef OP
 921         } else {
 922 #define OP(i) v_tempBlurredAss##i = v_srcAss##i;
 923
 924             OP(0);
 925             OP(1);
 926             OP(2);
 927             OP(3);
 928             OP(4);
 929             OP(5);
 930             OP(6);
 931             OP(7);
 932 #undef OP
 933         }
 934     } else {
 935         if (d < maxNoise[0]) {
 936             const vector signed short vsint16_7 = vec_splat_s16(7);
 937             const vector signed short vsint16_4 = vec_splat_s16(4);
 938             const vector unsigned short vuint16_3 = vec_splat_u16(3);
 939
 940 #define OP(i)                                                   \
 941             const vector signed short v_temp##i =               \
 942                 vec_mladd(v_tempBlurredAss##i,                  \
 943                           vsint16_7, v_srcAss##i);              \
 944             const vector signed short v_temp2##i =              \
 945                 vec_add(v_temp##i, vsint16_4);                  \
 946             v_tempBlurredAss##i = vec_sr(v_temp2##i, vuint16_3)
 947
 948             OP(0);
 949             OP(1);
 950             OP(2);
 951             OP(3);
 952             OP(4);
 953             OP(5);
 954             OP(6);
 955             OP(7);
 956 #undef OP
 957         } else {
 958             const vector signed short vsint16_3 = vec_splat_s16(3);
 959             const vector signed short vsint16_2 = vec_splat_s16(2);
 960
 961 #define OP(i)                                                   \
 962             const vector signed short v_temp##i =               \
 963                 vec_mladd(v_tempBlurredAss##i,                  \
 964                           vsint16_3, v_srcAss##i);              \
 965             const vector signed short v_temp2##i =              \
 966                 vec_add(v_temp##i, vsint16_2);                  \
 967             v_tempBlurredAss##i = vec_sr(v_temp2##i, (vector unsigned short)vsint16_2)
 968
 969             OP(0);
 970             OP(1);
 971             OP(2);
 972             OP(3);
 973             OP(4);
 974             OP(5);
 975             OP(6);
 976             OP(7);
 977 #undef OP
 978         }
 979     }
 980
 981     const vector signed char neg1 = vec_splat_s8(-1);
 982     const vector unsigned char permHH = (const vector unsigned char){0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
 983                                                                      0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F};
 984
 985 #define PACK_AND_STORE(src, i)                                  \
 986     const vector unsigned char perms##src##i =                  \
 987         vec_lvsr(i * stride, src);                              \
 988     const vector unsigned char vf##src##i =                     \
 989         vec_packsu(v_tempBlurredAss##i, (vector signed short)zero); \
 990     const vector unsigned char vg##src##i =                     \
 991         vec_perm(vf##src##i, v_##src##A##i, permHH);            \
 992     const vector unsigned char mask##src##i =                   \
 993         vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##src##i); \
 994     const vector unsigned char vg2##src##i =                    \
 995         vec_perm(vg##src##i, vg##src##i, perms##src##i);        \
 996     const vector unsigned char svA##src##i =                    \
 997         vec_sel(v_##src##A1##i, vg2##src##i, mask##src##i);     \
 998     const vector unsigned char svB##src##i =                    \
 999         vec_sel(vg2##src##i, v_##src##A2##i, mask##src##i);     \
1000     vec_st(svA##src##i, i * stride, src);                       \
1001     vec_st(svB##src##i, i * stride + 16, src)
1002
1003     PACK_AND_STORE(src, 0);
1004     PACK_AND_STORE(src, 1);
1005     PACK_AND_STORE(src, 2);
1006     PACK_AND_STORE(src, 3);
1007     PACK_AND_STORE(src, 4);
1008     PACK_AND_STORE(src, 5);
1009     PACK_AND_STORE(src, 6);
1010     PACK_AND_STORE(src, 7);
1011     PACK_AND_STORE(tempBlurred, 0);
1012     PACK_AND_STORE(tempBlurred, 1);
1013     PACK_AND_STORE(tempBlurred, 2);
1014     PACK_AND_STORE(tempBlurred, 3);
1015     PACK_AND_STORE(tempBlurred, 4);
1016     PACK_AND_STORE(tempBlurred, 5);
1017     PACK_AND_STORE(tempBlurred, 6);
1018     PACK_AND_STORE(tempBlurred, 7);
1019 #undef PACK_AND_STORE
1020 }
1021
1022 static inline void transpose_16x8_char_toPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1023     const vector unsigned char zero = vec_splat_u8(0);
1024
1025 #define LOAD_DOUBLE_LINE(i, j)                                          \
1026     vector unsigned char perm1##i = vec_lvsl(i * stride, src);          \
1027     vector unsigned char perm2##i = vec_lvsl(j * stride, src);          \
1028     vector unsigned char srcA##i = vec_ld(i * stride, src);             \
1029     vector unsigned char srcB##i = vec_ld(i * stride + 16, src);        \
1030     vector unsigned char srcC##i = vec_ld(j * stride, src);             \
1031     vector unsigned char srcD##i = vec_ld(j * stride+ 16, src);         \
1032     vector unsigned char src##i = vec_perm(srcA##i, srcB##i, perm1##i); \
1033     vector unsigned char src##j = vec_perm(srcC##i, srcD##i, perm2##i)
1034
1035     LOAD_DOUBLE_LINE(0, 1);
1036     LOAD_DOUBLE_LINE(2, 3);
1037     LOAD_DOUBLE_LINE(4, 5);
1038     LOAD_DOUBLE_LINE(6, 7);
1039 #undef LOAD_DOUBLE_LINE
1040
1041     vector unsigned char tempA = vec_mergeh(src0, zero);
1042     vector unsigned char tempB = vec_mergel(src0, zero);
1043     vector unsigned char tempC = vec_mergeh(src1, zero);
1044     vector unsigned char tempD = vec_mergel(src1, zero);
1045     vector unsigned char tempE = vec_mergeh(src2, zero);
1046     vector unsigned char tempF = vec_mergel(src2, zero);
1047     vector unsigned char tempG = vec_mergeh(src3, zero);
1048     vector unsigned char tempH = vec_mergel(src3, zero);
1049     vector unsigned char tempI = vec_mergeh(src4, zero);
1050     vector unsigned char tempJ = vec_mergel(src4, zero);
1051     vector unsigned char tempK = vec_mergeh(src5, zero);
1052     vector unsigned char tempL = vec_mergel(src5, zero);
1053     vector unsigned char tempM = vec_mergeh(src6, zero);
1054     vector unsigned char tempN = vec_mergel(src6, zero);
1055     vector unsigned char tempO = vec_mergeh(src7, zero);
1056     vector unsigned char tempP = vec_mergel(src7, zero);
1057
1058     vector unsigned char temp0  = vec_mergeh(tempA, tempI);
1059     vector unsigned char temp1  = vec_mergel(tempA, tempI);
1060     vector unsigned char temp2  = vec_mergeh(tempB, tempJ);
1061     vector unsigned char temp3  = vec_mergel(tempB, tempJ);
1062     vector unsigned char temp4  = vec_mergeh(tempC, tempK);
1063     vector unsigned char temp5  = vec_mergel(tempC, tempK);
1064     vector unsigned char temp6  = vec_mergeh(tempD, tempL);
1065     vector unsigned char temp7  = vec_mergel(tempD, tempL);
1066     vector unsigned char temp8  = vec_mergeh(tempE, tempM);
1067     vector unsigned char temp9  = vec_mergel(tempE, tempM);
1068     vector unsigned char temp10 = vec_mergeh(tempF, tempN);
1069     vector unsigned char temp11 = vec_mergel(tempF, tempN);
1070     vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1071     vector unsigned char temp13 = vec_mergel(tempG, tempO);
1072     vector unsigned char temp14 = vec_mergeh(tempH, tempP);
1073     vector unsigned char temp15 = vec_mergel(tempH, tempP);
1074
1075     tempA = vec_mergeh(temp0, temp8);
1076     tempB = vec_mergel(temp0, temp8);
1077     tempC = vec_mergeh(temp1, temp9);
1078     tempD = vec_mergel(temp1, temp9);
1079     tempE = vec_mergeh(temp2, temp10);
1080     tempF = vec_mergel(temp2, temp10);
1081     tempG = vec_mergeh(temp3, temp11);
1082     tempH = vec_mergel(temp3, temp11);
1083     tempI = vec_mergeh(temp4, temp12);
1084     tempJ = vec_mergel(temp4, temp12);
1085     tempK = vec_mergeh(temp5, temp13);
1086     tempL = vec_mergel(temp5, temp13);
1087     tempM = vec_mergeh(temp6, temp14);
1088     tempN = vec_mergel(temp6, temp14);
1089     tempO = vec_mergeh(temp7, temp15);
1090     tempP = vec_mergel(temp7, temp15);
1091
1092     temp0  = vec_mergeh(tempA, tempI);
1093     temp1  = vec_mergel(tempA, tempI);
1094     temp2  = vec_mergeh(tempB, tempJ);
1095     temp3  = vec_mergel(tempB, tempJ);
1096     temp4  = vec_mergeh(tempC, tempK);
1097     temp5  = vec_mergel(tempC, tempK);
1098     temp6  = vec_mergeh(tempD, tempL);
1099     temp7  = vec_mergel(tempD, tempL);
1100     temp8  = vec_mergeh(tempE, tempM);
1101     temp9  = vec_mergel(tempE, tempM);
1102     temp10 = vec_mergeh(tempF, tempN);
1103     temp11 = vec_mergel(tempF, tempN);
1104     temp12 = vec_mergeh(tempG, tempO);
1105     temp13 = vec_mergel(tempG, tempO);
1106     temp14 = vec_mergeh(tempH, tempP);
1107     temp15 = vec_mergel(tempH, tempP);
1108
1109     vec_st(temp0,    0, dst);
1110     vec_st(temp1,   16, dst);
1111     vec_st(temp2,   32, dst);
1112     vec_st(temp3,   48, dst);
1113     vec_st(temp4,   64, dst);
1114     vec_st(temp5,   80, dst);
1115     vec_st(temp6,   96, dst);
1116     vec_st(temp7,  112, dst);
1117     vec_st(temp8,  128, dst);
1118     vec_st(temp9,  144, dst);
1119     vec_st(temp10, 160, dst);
1120     vec_st(temp11, 176, dst);
1121     vec_st(temp12, 192, dst);
1122     vec_st(temp13, 208, dst);
1123     vec_st(temp14, 224, dst);
1124     vec_st(temp15, 240, dst);
1125 }
1126
1127 static inline void transpose_8x16_char_fromPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1128     const vector unsigned char zero = vec_splat_u8(0);
1129
1130 #define LOAD_DOUBLE_LINE(i, j)                                  \
1131     vector unsigned char src##i = vec_ld(i * 16, src);            \
1132     vector unsigned char src##j = vec_ld(j * 16, src)
1133
1134     LOAD_DOUBLE_LINE(0, 1);
1135     LOAD_DOUBLE_LINE(2, 3);
1136     LOAD_DOUBLE_LINE(4, 5);
1137     LOAD_DOUBLE_LINE(6, 7);
1138     LOAD_DOUBLE_LINE(8, 9);
1139     LOAD_DOUBLE_LINE(10, 11);
1140     LOAD_DOUBLE_LINE(12, 13);
1141     LOAD_DOUBLE_LINE(14, 15);
1142 #undef LOAD_DOUBLE_LINE
1143
1144     vector unsigned char tempA = vec_mergeh(src0, src8);
1145     vector unsigned char tempB;
1146     vector unsigned char tempC = vec_mergeh(src1, src9);
1147     vector unsigned char tempD;
1148     vector unsigned char tempE = vec_mergeh(src2, src10);
1149     vector unsigned char tempG = vec_mergeh(src3, src11);
1150     vector unsigned char tempI = vec_mergeh(src4, src12);
1151     vector unsigned char tempJ;
1152     vector unsigned char tempK = vec_mergeh(src5, src13);
1153     vector unsigned char tempL;
1154     vector unsigned char tempM = vec_mergeh(src6, src14);
1155     vector unsigned char tempO = vec_mergeh(src7, src15);
1156
1157     vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1158     vector unsigned char temp1 = vec_mergel(tempA, tempI);
1159     vector unsigned char temp2;
1160     vector unsigned char temp3;
1161     vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1162     vector unsigned char temp5 = vec_mergel(tempC, tempK);
1163     vector unsigned char temp6;
1164     vector unsigned char temp7;
1165     vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1166     vector unsigned char temp9 = vec_mergel(tempE, tempM);
1167     vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1168     vector unsigned char temp13 = vec_mergel(tempG, tempO);
1169
1170     tempA = vec_mergeh(temp0, temp8);
1171     tempB = vec_mergel(temp0, temp8);
1172     tempC = vec_mergeh(temp1, temp9);
1173     tempD = vec_mergel(temp1, temp9);
1174     tempI = vec_mergeh(temp4, temp12);
1175     tempJ = vec_mergel(temp4, temp12);
1176     tempK = vec_mergeh(temp5, temp13);
1177     tempL = vec_mergel(temp5, temp13);
1178
1179     temp0 = vec_mergeh(tempA, tempI);
1180     temp1 = vec_mergel(tempA, tempI);
1181     temp2 = vec_mergeh(tempB, tempJ);
1182     temp3 = vec_mergel(tempB, tempJ);
1183     temp4 = vec_mergeh(tempC, tempK);
1184     temp5 = vec_mergel(tempC, tempK);
1185     temp6 = vec_mergeh(tempD, tempL);
1186     temp7 = vec_mergel(tempD, tempL);
1187
1188
1189     const vector signed char neg1 = vec_splat_s8(-1);
1190 #define STORE_DOUBLE_LINE(i, j)                                         \
1191     vector unsigned char dstA##i = vec_ld(i * stride, dst);             \
1192     vector unsigned char dstB##i = vec_ld(i * stride + 16, dst);        \
1193     vector unsigned char dstA##j = vec_ld(j * stride, dst);             \
1194     vector unsigned char dstB##j = vec_ld(j * stride+ 16, dst);         \
1195     vector unsigned char align##i = vec_lvsr(i * stride, dst);          \
1196     vector unsigned char align##j = vec_lvsr(j * stride, dst);          \
1197     vector unsigned char mask##i = vec_perm(zero, (vector unsigned char)neg1, align##i); \
1198     vector unsigned char mask##j = vec_perm(zero, (vector unsigned char)neg1, align##j); \
1199     vector unsigned char dstR##i = vec_perm(temp##i, temp##i, align##i);\
1200     vector unsigned char dstR##j = vec_perm(temp##j, temp##j, align##j);\
1201     vector unsigned char dstAF##i = vec_sel(dstA##i, dstR##i, mask##i); \
1202     vector unsigned char dstBF##i = vec_sel(dstR##i, dstB##i, mask##i); \
1203     vector unsigned char dstAF##j = vec_sel(dstA##j, dstR##j, mask##j); \
1204     vector unsigned char dstBF##j = vec_sel(dstR##j, dstB##j, mask##j); \
1205     vec_st(dstAF##i, i * stride, dst);                                  \
1206     vec_st(dstBF##i, i * stride + 16, dst);                             \
1207     vec_st(dstAF##j, j * stride, dst);                                  \
1208     vec_st(dstBF##j, j * stride + 16, dst)
1209
1210     STORE_DOUBLE_LINE(0,1);
1211     STORE_DOUBLE_LINE(2,3);
1212     STORE_DOUBLE_LINE(4,5);
1213     STORE_DOUBLE_LINE(6,7);
1214 }