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