Statistics
| Branch: | Revision:

ffmpeg / libpostproc / postprocess_altivec_template.c @ 8008a043

History | View | Annotate | Download (50.7 KB)

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 call from
114
    // the (transposed) horizontable deblocks will be aligned, i naddition
115
    // to the naturraly 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
    {
165
      const vector unsigned char mmoP1 = (const vector unsigned char)
166
        AVV(0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f,
167
            0x00, 0x01, 0x12, 0x13, 0x08, 0x09, 0x1A, 0x1B);
168
      const vector unsigned char mmoP2 = (const vector unsigned char)
169
        AVV(0x04, 0x05, 0x16, 0x17, 0x0C, 0x0D, 0x1E, 0x1F,
170
            0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f);
171
      const vector unsigned char mmoP = (const vector unsigned char)
172
        vec_lvsl(8, (unsigned char*)0);
173

    
174
      vector signed short mmoL1 = vec_perm(v_srcAss0, v_srcAss2, mmoP1);
175
      vector signed short mmoL2 = vec_perm(v_srcAss4, v_srcAss6, mmoP2);
176
      vector signed short mmoL = vec_perm(mmoL1, mmoL2, mmoP);
177
      vector signed short mmoR1 = vec_perm(v_srcAss5, v_srcAss7, mmoP1);
178
      vector signed short mmoR2 = vec_perm(v_srcAss1, v_srcAss3, mmoP2);
179
      vector signed short mmoR = vec_perm(mmoR1, mmoR2, mmoP);
180
      vector signed short mmoDiff = vec_sub(mmoL, mmoR);
181
      vector unsigned short mmoSum = (vector unsigned short)vec_add(mmoDiff, v2QP);
182

    
183
      if (vec_any_gt(mmoSum, v4QP))
184
        return 0;
185
      else
186
        return 1;
187
    }
188
  else return 2;
189
}
190

    
191
static inline void doVertLowPass_altivec(uint8_t *src, int stride, PPContext *c) {
192
  /*
193
    this code makes no assumption on src or stride.
194
    One could remove the recomputation of the perm
195
    vector by assuming (stride % 16) == 0, unfortunately
196
    this is not always true. Quite a lot of load/stores
197
    can be removed by assuming proper alignement of
198
    src & stride :-(
199
  */
200
  uint8_t *src2 = src;
201
  const vector signed int zero = vec_splat_s32(0);
202
  const int properStride = (stride % 16);
203
  const int srcAlign = ((unsigned long)src2 % 16);
204
  DECLARE_ALIGNED(16, short, qp[8]);
205
  qp[0] = c->QP;
206
  vector signed short vqp = vec_ld(0, qp);
207
  vqp = vec_splat(vqp, 0);
208

    
209
  src2 += stride*3;
210

    
211
  vector signed short vb0, vb1, vb2, vb3, vb4, vb5, vb6, vb7, vb8, vb9;
212
  vector unsigned char vbA0, vbA1, vbA2, vbA3, vbA4, vbA5, vbA6, vbA7, vbA8, vbA9;
213
  vector unsigned char vbB0, vbB1, vbB2, vbB3, vbB4, vbB5, vbB6, vbB7, vbB8, vbB9;
214
  vector unsigned char vbT0, vbT1, vbT2, vbT3, vbT4, vbT5, vbT6, vbT7, vbT8, vbT9;
215

    
216
#define LOAD_LINE(i)                                                    \
217
  const vector unsigned char perml##i =                                 \
218
    vec_lvsl(i * stride, src2);                                         \
219
  vbA##i = vec_ld(i * stride, src2);                                    \
220
  vbB##i = vec_ld(i * stride + 16, src2);                               \
221
  vbT##i = vec_perm(vbA##i, vbB##i, perml##i);                          \
222
  vb##i =                                                               \
223
    (vector signed short)vec_mergeh((vector unsigned char)zero,         \
224
                                    (vector unsigned char)vbT##i)
225

    
226
#define LOAD_LINE_ALIGNED(i)                                            \
227
  register int j##i = i * stride;                                       \
228
  vbT##i = vec_ld(j##i, src2);                                          \
229
  vb##i =                                                               \
230
    (vector signed short)vec_mergeh((vector signed char)zero,           \
231
                                    (vector signed char)vbT##i)
232

    
233
    // special casing the aligned case is worthwhile, as all call from
234
    // the (transposed) horizontable deblocks will be aligned, in addition
235
    // to the naturraly aligned vertical deblocks.
236
    if (properStride && srcAlign) {
237
      LOAD_LINE_ALIGNED(0);
238
      LOAD_LINE_ALIGNED(1);
239
      LOAD_LINE_ALIGNED(2);
240
      LOAD_LINE_ALIGNED(3);
241
      LOAD_LINE_ALIGNED(4);
242
      LOAD_LINE_ALIGNED(5);
243
      LOAD_LINE_ALIGNED(6);
244
      LOAD_LINE_ALIGNED(7);
245
      LOAD_LINE_ALIGNED(8);
246
      LOAD_LINE_ALIGNED(9);
247
    } else {
248
      LOAD_LINE(0);
249
      LOAD_LINE(1);
250
      LOAD_LINE(2);
251
      LOAD_LINE(3);
252
      LOAD_LINE(4);
253
      LOAD_LINE(5);
254
      LOAD_LINE(6);
255
      LOAD_LINE(7);
256
      LOAD_LINE(8);
257
      LOAD_LINE(9);
258
    }
259
#undef LOAD_LINE
260
#undef LOAD_LINE_ALIGNED
261

    
262
  const vector unsigned short v_2 = vec_splat_u16(2);
263
  const vector unsigned short v_4 = vec_splat_u16(4);
264

    
265
  const vector signed short v_diff01 = vec_sub(vb0, vb1);
266
  const vector unsigned short v_cmp01 =
267
    (const vector unsigned short) vec_cmplt(vec_abs(v_diff01), vqp);
268
  const vector signed short v_first = vec_sel(vb1, vb0, v_cmp01);
269
  const vector signed short v_diff89 = vec_sub(vb8, vb9);
270
  const vector unsigned short v_cmp89 =
271
    (const vector unsigned short) vec_cmplt(vec_abs(v_diff89), vqp);
272
  const vector signed short v_last = vec_sel(vb8, vb9, v_cmp89);
273

    
274
  const vector signed short temp01 = vec_mladd(v_first, (vector signed short)v_4, vb1);
275
  const vector signed short temp02 = vec_add(vb2, vb3);
276
  const vector signed short temp03 = vec_add(temp01, (vector signed short)v_4);
277
  const vector signed short v_sumsB0 = vec_add(temp02, temp03);
278

    
279
  const vector signed short temp11 = vec_sub(v_sumsB0, v_first);
280
  const vector signed short v_sumsB1 = vec_add(temp11, vb4);
281

    
282
  const vector signed short temp21 = vec_sub(v_sumsB1, v_first);
283
  const vector signed short v_sumsB2 = vec_add(temp21, vb5);
284

    
285
  const vector signed short temp31 = vec_sub(v_sumsB2, v_first);
286
  const vector signed short v_sumsB3 = vec_add(temp31, vb6);
287

    
288
  const vector signed short temp41 = vec_sub(v_sumsB3, v_first);
289
  const vector signed short v_sumsB4 = vec_add(temp41, vb7);
290

    
291
  const vector signed short temp51 = vec_sub(v_sumsB4, vb1);
292
  const vector signed short v_sumsB5 = vec_add(temp51, vb8);
293

    
294
  const vector signed short temp61 = vec_sub(v_sumsB5, vb2);
295
  const vector signed short v_sumsB6 = vec_add(temp61, v_last);
296

    
297
  const vector signed short temp71 = vec_sub(v_sumsB6, vb3);
298
  const vector signed short v_sumsB7 = vec_add(temp71, v_last);
299

    
300
  const vector signed short temp81 = vec_sub(v_sumsB7, vb4);
301
  const vector signed short v_sumsB8 = vec_add(temp81, v_last);
302

    
303
  const vector signed short temp91 = vec_sub(v_sumsB8, vb5);
304
  const vector signed short v_sumsB9 = vec_add(temp91, v_last);
305

    
306
#define COMPUTE_VR(i, j, k)                                             \
307
  const vector signed short temps1##i =                                 \
308
    vec_add(v_sumsB##i, v_sumsB##k);                                    \
309
  const vector signed short temps2##i =                                 \
310
    vec_mladd(vb##j, (vector signed short)v_2, temps1##i);              \
311
  const vector signed short  vr##j = vec_sra(temps2##i, v_4)
312

    
313
  COMPUTE_VR(0, 1, 2);
314
  COMPUTE_VR(1, 2, 3);
315
  COMPUTE_VR(2, 3, 4);
316
  COMPUTE_VR(3, 4, 5);
317
  COMPUTE_VR(4, 5, 6);
318
  COMPUTE_VR(5, 6, 7);
319
  COMPUTE_VR(6, 7, 8);
320
  COMPUTE_VR(7, 8, 9);
321

    
322
  const vector signed char neg1 = vec_splat_s8(-1);
323
  const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
324
                                                                      0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
325

    
326
#define PACK_AND_STORE(i)                                       \
327
  const vector unsigned char perms##i =                         \
328
    vec_lvsr(i * stride, src2);                                 \
329
  const vector unsigned char vf##i =                            \
330
    vec_packsu(vr##i, (vector signed short)zero);               \
331
  const vector unsigned char vg##i =                            \
332
    vec_perm(vf##i, vbT##i, permHH);                            \
333
  const vector unsigned char mask##i =                          \
334
    vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \
335
  const vector unsigned char vg2##i =                           \
336
    vec_perm(vg##i, vg##i, perms##i);                           \
337
  const vector unsigned char svA##i =                           \
338
    vec_sel(vbA##i, vg2##i, mask##i);                           \
339
  const vector unsigned char svB##i =                           \
340
    vec_sel(vg2##i, vbB##i, mask##i);                           \
341
  vec_st(svA##i, i * stride, src2);                             \
342
  vec_st(svB##i, i * stride + 16, src2)
343

    
344
#define PACK_AND_STORE_ALIGNED(i)                               \
345
  const vector unsigned char vf##i =                            \
346
    vec_packsu(vr##i, (vector signed short)zero);               \
347
  const vector unsigned char vg##i =                            \
348
    vec_perm(vf##i, vbT##i, permHH);                            \
349
  vec_st(vg##i, i * stride, src2)
350

    
351
  // special casing the aligned case is worthwhile, as all call from
352
  // the (transposed) horizontable deblocks will be aligned, in addition
353
  // to the naturraly aligned vertical deblocks.
354
  if (properStride && srcAlign) {
355
    PACK_AND_STORE_ALIGNED(1);
356
    PACK_AND_STORE_ALIGNED(2);
357
    PACK_AND_STORE_ALIGNED(3);
358
    PACK_AND_STORE_ALIGNED(4);
359
    PACK_AND_STORE_ALIGNED(5);
360
    PACK_AND_STORE_ALIGNED(6);
361
    PACK_AND_STORE_ALIGNED(7);
362
    PACK_AND_STORE_ALIGNED(8);
363
  } else {
364
    PACK_AND_STORE(1);
365
    PACK_AND_STORE(2);
366
    PACK_AND_STORE(3);
367
    PACK_AND_STORE(4);
368
    PACK_AND_STORE(5);
369
    PACK_AND_STORE(6);
370
    PACK_AND_STORE(7);
371
    PACK_AND_STORE(8);
372
  }
373
#undef PACK_AND_STORE
374
#undef PACK_AND_STORE_ALIGNED
375
}
376

    
377

    
378

    
379
static inline void doVertDefFilter_altivec(uint8_t src[], int stride, PPContext *c) {
380
  /*
381
    this code makes no assumption on src or stride.
382
    One could remove the recomputation of the perm
383
    vector by assuming (stride % 16) == 0, unfortunately
384
    this is not always true. Quite a lot of load/stores
385
    can be removed by assuming proper alignement of
386
    src & stride :-(
387
  */
388
  uint8_t *src2 = src;
389
  const vector signed int zero = vec_splat_s32(0);
390
  DECLARE_ALIGNED(16, short, qp[8]);
391
  qp[0] = 8*c->QP;
392
  vector signed short vqp = vec_ld(0, qp);
393
  vqp = vec_splat(vqp, 0);
394

    
395
#define LOAD_LINE(i)                                                    \
396
  const vector unsigned char perm##i =                                  \
397
    vec_lvsl(i * stride, src2);                                         \
398
  const vector unsigned char vbA##i =                                   \
399
    vec_ld(i * stride, src2);                                           \
400
  const vector unsigned char vbB##i =                                   \
401
    vec_ld(i * stride + 16, src2);                                      \
402
  const vector unsigned char vbT##i =                                   \
403
    vec_perm(vbA##i, vbB##i, perm##i);                                  \
404
  const vector signed short vb##i =                                     \
405
    (vector signed short)vec_mergeh((vector unsigned char)zero,         \
406
                                    (vector unsigned char)vbT##i)
407

    
408
  src2 += stride*3;
409

    
410
  LOAD_LINE(1);
411
  LOAD_LINE(2);
412
  LOAD_LINE(3);
413
  LOAD_LINE(4);
414
  LOAD_LINE(5);
415
  LOAD_LINE(6);
416
  LOAD_LINE(7);
417
  LOAD_LINE(8);
418
#undef LOAD_LINE
419

    
420
  const vector signed short v_1 = vec_splat_s16(1);
421
  const vector signed short v_2 = vec_splat_s16(2);
422
  const vector signed short v_5 = vec_splat_s16(5);
423
  const vector signed short v_32 = vec_sl(v_1,
424
                                          (vector unsigned short)v_5);
425
  /* middle energy */
426
  const vector signed short l3minusl6 = vec_sub(vb3, vb6);
427
  const vector signed short l5minusl4 = vec_sub(vb5, vb4);
428
  const vector signed short twotimes_l3minusl6 = vec_mladd(v_2, l3minusl6, (vector signed short)zero);
429
  const vector signed short mE = vec_mladd(v_5, l5minusl4, twotimes_l3minusl6);
430
  const vector signed short absmE = vec_abs(mE);
431
  /* left & right energy */
432
  const vector signed short l1minusl4 = vec_sub(vb1, vb4);
433
  const vector signed short l3minusl2 = vec_sub(vb3, vb2);
434
  const vector signed short l5minusl8 = vec_sub(vb5, vb8);
435
  const vector signed short l7minusl6 = vec_sub(vb7, vb6);
436
  const vector signed short twotimes_l1minusl4 = vec_mladd(v_2, l1minusl4, (vector signed short)zero);
437
  const vector signed short twotimes_l5minusl8 = vec_mladd(v_2, l5minusl8, (vector signed short)zero);
438
  const vector signed short lE = vec_mladd(v_5, l3minusl2, twotimes_l1minusl4);
439
  const vector signed short rE = vec_mladd(v_5, l7minusl6, twotimes_l5minusl8);
440
  /* d */
441
  const vector signed short ddiff = vec_sub(absmE,
442
                                            vec_min(vec_abs(lE),
443
                                                    vec_abs(rE)));
444
  const vector signed short ddiffclamp = vec_max(ddiff, (vector signed short)zero);
445
  const vector signed short dtimes64 = vec_mladd(v_5, ddiffclamp, v_32);
446
  const vector signed short d = vec_sra(dtimes64, vec_splat_u16(6));
447
  const vector signed short minusd = vec_sub((vector signed short)zero, d);
448
  const vector signed short finald = vec_sel(minusd,
449
                                             d,
450
                                             vec_cmpgt(vec_sub((vector signed short)zero, mE),
451
                                                       (vector signed short)zero));
452
  /* q */
453
  const vector signed short qtimes2 = vec_sub(vb4, vb5);
454
  /* for a shift right to behave like /2, we need to add one
455
     to all negative integer */
456
  const vector signed short rounddown = vec_sel((vector signed short)zero,
457
                                                v_1,
458
                                                vec_cmplt(qtimes2, (vector signed short)zero));
459
  const vector signed short q = vec_sra(vec_add(qtimes2, rounddown), vec_splat_u16(1));
460
  /* clamp */
461
  const vector signed short dclamp_P1 = vec_max((vector signed short)zero, finald);
462
  const vector signed short dclamp_P = vec_min(dclamp_P1, q);
463
  const vector signed short dclamp_N1 = vec_min((vector signed short)zero, finald);
464
  const vector signed short dclamp_N = vec_max(dclamp_N1, q);
465

    
466
  const vector signed short dclampedfinal = vec_sel(dclamp_N,
467
                                                    dclamp_P,
468
                                                    vec_cmpgt(q, (vector signed short)zero));
469
  const vector signed short dornotd = vec_sel((vector signed short)zero,
470
                                              dclampedfinal,
471
                                              vec_cmplt(absmE, vqp));
472
  /* add/substract to l4 and l5 */
473
  const vector signed short vb4minusd = vec_sub(vb4, dornotd);
474
  const vector signed short vb5plusd = vec_add(vb5, dornotd);
475
  /* finally, stores */
476
  const vector unsigned char st4 = vec_packsu(vb4minusd, (vector signed short)zero);
477
  const vector unsigned char st5 = vec_packsu(vb5plusd, (vector signed short)zero);
478

    
479
  const vector signed char neg1 = vec_splat_s8(-1);
480
  const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
481
                                                                      0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
482

    
483
#define STORE(i)                                                \
484
  const vector unsigned char perms##i =                         \
485
    vec_lvsr(i * stride, src2);                                 \
486
  const vector unsigned char vg##i =                            \
487
    vec_perm(st##i, vbT##i, permHH);                            \
488
  const vector unsigned char mask##i =                          \
489
    vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \
490
  const vector unsigned char vg2##i =                           \
491
    vec_perm(vg##i, vg##i, perms##i);                           \
492
  const vector unsigned char svA##i =                           \
493
    vec_sel(vbA##i, vg2##i, mask##i);                           \
494
  const vector unsigned char svB##i =                           \
495
    vec_sel(vg2##i, vbB##i, mask##i);                           \
496
  vec_st(svA##i, i * stride, src2);                             \
497
  vec_st(svB##i, i * stride + 16, src2)
498

    
499
  STORE(4);
500
  STORE(5);
501
}
502

    
503
static inline void dering_altivec(uint8_t src[], int stride, PPContext *c) {
504
  /*
505
    this code makes no assumption on src or stride.
506
    One could remove the recomputation of the perm
507
    vector by assuming (stride % 16) == 0, unfortunately
508
    this is not always true. Quite a lot of load/stores
509
    can be removed by assuming proper alignement of
510
    src & stride :-(
511
  */
512
  uint8_t *srcCopy = src;
513
  DECLARE_ALIGNED(16, uint8_t, dt[16]);
514
  const vector signed int zero = vec_splat_s32(0);
515
  vector unsigned char v_dt;
516
  dt[0] = deringThreshold;
517
  v_dt = vec_splat(vec_ld(0, dt), 0);
518

    
519
#define LOAD_LINE(i)                                                    \
520
  const vector unsigned char perm##i =                                  \
521
    vec_lvsl(i * stride, srcCopy);                                      \
522
  vector unsigned char sA##i = vec_ld(i * stride, srcCopy);             \
523
  vector unsigned char sB##i = vec_ld(i * stride + 16, srcCopy);        \
524
  vector unsigned char src##i = vec_perm(sA##i, sB##i, perm##i)
525

    
526
  LOAD_LINE(0);
527
  LOAD_LINE(1);
528
  LOAD_LINE(2);
529
  LOAD_LINE(3);
530
  LOAD_LINE(4);
531
  LOAD_LINE(5);
532
  LOAD_LINE(6);
533
  LOAD_LINE(7);
534
  LOAD_LINE(8);
535
  LOAD_LINE(9);
536
#undef LOAD_LINE
537

    
538
  vector unsigned char v_avg;
539
  {
540
    const vector unsigned char trunc_perm = (vector unsigned char)
541
      AVV(0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08,
542
          0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18);
543
    const vector unsigned char trunc_src12 = vec_perm(src1, src2, trunc_perm);
544
    const vector unsigned char trunc_src34 = vec_perm(src3, src4, trunc_perm);
545
    const vector unsigned char trunc_src56 = vec_perm(src5, src6, trunc_perm);
546
    const vector unsigned char trunc_src78 = vec_perm(src7, src8, trunc_perm);
547

    
548
#define EXTRACT(op) do {                                                \
549
      const vector unsigned char s##op##_1 = vec_##op(trunc_src12, trunc_src34); \
550
      const vector unsigned char s##op##_2 = vec_##op(trunc_src56, trunc_src78); \
551
      const vector unsigned char s##op##_6 = vec_##op(s##op##_1, s##op##_2); \
552
      const vector unsigned char s##op##_8h = vec_mergeh(s##op##_6, s##op##_6); \
553
      const vector unsigned char s##op##_8l = vec_mergel(s##op##_6, s##op##_6); \
554
      const vector unsigned char s##op##_9 = vec_##op(s##op##_8h, s##op##_8l); \
555
      const vector unsigned char s##op##_9h = vec_mergeh(s##op##_9, s##op##_9); \
556
      const vector unsigned char s##op##_9l = vec_mergel(s##op##_9, s##op##_9); \
557
      const vector unsigned char s##op##_10 = vec_##op(s##op##_9h, s##op##_9l); \
558
      const vector unsigned char s##op##_10h = vec_mergeh(s##op##_10, s##op##_10); \
559
      const vector unsigned char s##op##_10l = vec_mergel(s##op##_10, s##op##_10); \
560
      const vector unsigned char s##op##_11 = vec_##op(s##op##_10h, s##op##_10l); \
561
      const vector unsigned char s##op##_11h = vec_mergeh(s##op##_11, s##op##_11); \
562
      const vector unsigned char s##op##_11l = vec_mergel(s##op##_11, s##op##_11); \
563
      v_##op = vec_##op(s##op##_11h, s##op##_11l); } while (0)
564

    
565
    vector unsigned char v_min;
566
    vector unsigned char v_max;
567
    EXTRACT(min);
568
    EXTRACT(max);
569
#undef EXTRACT
570

    
571
    if (vec_all_lt(vec_sub(v_max, v_min), v_dt))
572
      return;
573

    
574
    v_avg = vec_avg(v_min, v_max);
575
  }
576

    
577
  DECLARE_ALIGNED(16, signed int, S[8]);
578
  {
579
    const vector unsigned short mask1 = (vector unsigned short)
580
      AVV(0x0001, 0x0002, 0x0004, 0x0008,
581
          0x0010, 0x0020, 0x0040, 0x0080);
582
    const vector unsigned short mask2 = (vector unsigned short)
583
      AVV(0x0100, 0x0200, 0x0000, 0x0000,
584
          0x0000, 0x0000, 0x0000, 0x0000);
585

    
586
    const vector unsigned int vuint32_16 = vec_sl(vec_splat_u32(1), vec_splat_u32(4));
587
    const vector unsigned int vuint32_1 = vec_splat_u32(1);
588

    
589
#define COMPARE(i)                                                      \
590
    vector signed int sum##i;                                           \
591
    do {                                                                \
592
      const vector unsigned char cmp##i =                               \
593
        (vector unsigned char)vec_cmpgt(src##i, v_avg);                 \
594
      const vector unsigned short cmpHi##i =                            \
595
        (vector unsigned short)vec_mergeh(cmp##i, cmp##i);              \
596
      const vector unsigned short cmpLi##i =                            \
597
        (vector unsigned short)vec_mergel(cmp##i, cmp##i);              \
598
      const vector signed short cmpHf##i =                              \
599
        (vector signed short)vec_and(cmpHi##i, mask1);                  \
600
      const vector signed short cmpLf##i =                              \
601
        (vector signed short)vec_and(cmpLi##i, mask2);                  \
602
      const vector signed int sump##i = vec_sum4s(cmpHf##i, zero);      \
603
      const vector signed int sumq##i = vec_sum4s(cmpLf##i, sump##i);   \
604
      sum##i  = vec_sums(sumq##i, zero); } while (0)
605

    
606
    COMPARE(0);
607
    COMPARE(1);
608
    COMPARE(2);
609
    COMPARE(3);
610
    COMPARE(4);
611
    COMPARE(5);
612
    COMPARE(6);
613
    COMPARE(7);
614
    COMPARE(8);
615
    COMPARE(9);
616
#undef COMPARE
617

    
618
    vector signed int sumA2;
619
    vector signed int sumB2;
620
    {
621
      const vector signed int sump02 = vec_mergel(sum0, sum2);
622
      const vector signed int sump13 = vec_mergel(sum1, sum3);
623
      const vector signed int sumA = vec_mergel(sump02, sump13);
624

    
625
      const vector signed int sump46 = vec_mergel(sum4, sum6);
626
      const vector signed int sump57 = vec_mergel(sum5, sum7);
627
      const vector signed int sumB = vec_mergel(sump46, sump57);
628

    
629
      const vector signed int sump8A = vec_mergel(sum8, zero);
630
      const vector signed int sump9B = vec_mergel(sum9, zero);
631
      const vector signed int sumC = vec_mergel(sump8A, sump9B);
632

    
633
      const vector signed int tA = vec_sl(vec_nor(zero, sumA), vuint32_16);
634
      const vector signed int tB = vec_sl(vec_nor(zero, sumB), vuint32_16);
635
      const vector signed int tC = vec_sl(vec_nor(zero, sumC), vuint32_16);
636
      const vector signed int t2A = vec_or(sumA, tA);
637
      const vector signed int t2B = vec_or(sumB, tB);
638
      const vector signed int t2C = vec_or(sumC, tC);
639
      const vector signed int t3A = vec_and(vec_sra(t2A, vuint32_1),
640
                                            vec_sl(t2A, vuint32_1));
641
      const vector signed int t3B = vec_and(vec_sra(t2B, vuint32_1),
642
                                            vec_sl(t2B, vuint32_1));
643
      const vector signed int t3C = vec_and(vec_sra(t2C, vuint32_1),
644
                                            vec_sl(t2C, vuint32_1));
645
      const vector signed int yA = vec_and(t2A, t3A);
646
      const vector signed int yB = vec_and(t2B, t3B);
647
      const vector signed int yC = vec_and(t2C, t3C);
648

    
649
      const vector unsigned char strangeperm1 = vec_lvsl(4, (unsigned char*)0);
650
      const vector unsigned char strangeperm2 = vec_lvsl(8, (unsigned char*)0);
651
      const vector signed int sumAd4 = vec_perm(yA, yB, strangeperm1);
652
      const vector signed int sumAd8 = vec_perm(yA, yB, strangeperm2);
653
      const vector signed int sumBd4 = vec_perm(yB, yC, strangeperm1);
654
      const vector signed int sumBd8 = vec_perm(yB, yC, strangeperm2);
655
      const vector signed int sumAp = vec_and(yA,
656
                                              vec_and(sumAd4,sumAd8));
657
      const vector signed int sumBp = vec_and(yB,
658
                                              vec_and(sumBd4,sumBd8));
659
      sumA2 = vec_or(sumAp,
660
                     vec_sra(sumAp,
661
                             vuint32_16));
662
      sumB2  = vec_or(sumBp,
663
                      vec_sra(sumBp,
664
                              vuint32_16));
665
    }
666
    vec_st(sumA2, 0, S);
667
    vec_st(sumB2, 16, S);
668
  }
669

    
670
  /* I'm not sure the following is actually faster
671
     than straight, unvectorized C code :-( */
672

    
673
  DECLARE_ALIGNED(16, int, tQP2[4]);
674
  tQP2[0]= c->QP/2 + 1;
675
  vector signed int vQP2 = vec_ld(0, tQP2);
676
  vQP2 = vec_splat(vQP2, 0);
677
  const vector signed int vsint32_8 = vec_splat_s32(8);
678
  const vector unsigned int vuint32_4 = vec_splat_u32(4);
679

    
680
  const vector unsigned char permA1 = (vector unsigned char)
681
    AVV(0x00, 0x01, 0x02, 0x10, 0x11, 0x12, 0x1F, 0x1F,
682
        0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F);
683
  const vector unsigned char permA2 = (vector unsigned char)
684
    AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x10, 0x11,
685
        0x12, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F);
686
  const vector unsigned char permA1inc = (vector unsigned char)
687
    AVV(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00,
688
        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
689
  const vector unsigned char permA2inc = (vector unsigned char)
690
    AVV(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01,
691
        0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
692
  const vector unsigned char magic = (vector unsigned char)
693
    AVV(0x01, 0x02, 0x01, 0x02, 0x04, 0x02, 0x01, 0x02,
694
        0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
695
  const vector unsigned char extractPerm = (vector unsigned char)
696
    AVV(0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01,
697
        0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01);
698
  const vector unsigned char extractPermInc = (vector unsigned char)
699
    AVV(0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01,
700
        0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01);
701
  const vector unsigned char identity = vec_lvsl(0,(unsigned char *)0);
702
  const vector unsigned char tenRight = (vector unsigned char)
703
    AVV(0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
704
        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
705
  const vector unsigned char eightLeft = (vector unsigned char)
706
    AVV(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
707
        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08);
708

    
709

    
710
#define F_INIT(i)                                       \
711
  vector unsigned char tenRightM##i = tenRight;         \
712
  vector unsigned char permA1M##i = permA1;             \
713
  vector unsigned char permA2M##i = permA2;             \
714
  vector unsigned char extractPermM##i = extractPerm
715

    
716
#define F2(i, j, k, l)                                                  \
717
  if (S[i] & (1 << (l+1))) {                                            \
718
    const vector unsigned char a_##j##_A##l =                           \
719
      vec_perm(src##i, src##j, permA1M##i);                             \
720
    const vector unsigned char a_##j##_B##l =                           \
721
      vec_perm(a_##j##_A##l, src##k, permA2M##i);                       \
722
    const vector signed int a_##j##_sump##l =                           \
723
      (vector signed int)vec_msum(a_##j##_B##l, magic,                  \
724
                                  (vector unsigned int)zero);           \
725
    vector signed int F_##j##_##l =                                     \
726
      vec_sr(vec_sums(a_##j##_sump##l, vsint32_8), vuint32_4);          \
727
    F_##j##_##l = vec_splat(F_##j##_##l, 3);                            \
728
    const vector signed int p_##j##_##l =                               \
729
      (vector signed int)vec_perm(src##j,                               \
730
                                  (vector unsigned char)zero,           \
731
                                  extractPermM##i);                     \
732
    const vector signed int sum_##j##_##l = vec_add( p_##j##_##l, vQP2);\
733
    const vector signed int diff_##j##_##l = vec_sub( p_##j##_##l, vQP2);\
734
    vector signed int newpm_##j##_##l;                                  \
735
    if (vec_all_lt(sum_##j##_##l, F_##j##_##l))                         \
736
      newpm_##j##_##l = sum_##j##_##l;                                  \
737
    else if (vec_all_gt(diff_##j##_##l, F_##j##_##l))                   \
738
      newpm_##j##_##l = diff_##j##_##l;                                 \
739
    else newpm_##j##_##l = F_##j##_##l;                                 \
740
    const vector unsigned char newpm2_##j##_##l =                       \
741
      vec_splat((vector unsigned char)newpm_##j##_##l, 15);             \
742
    const vector unsigned char mask##j##l = vec_add(identity,           \
743
                                                    tenRightM##i);      \
744
    src##j = vec_perm(src##j, newpm2_##j##_##l, mask##j##l);            \
745
  }                                                                     \
746
  permA1M##i = vec_add(permA1M##i, permA1inc);                          \
747
  permA2M##i = vec_add(permA2M##i, permA2inc);                          \
748
  tenRightM##i = vec_sro(tenRightM##i, eightLeft);                      \
749
  extractPermM##i = vec_add(extractPermM##i, extractPermInc)
750

    
751
#define ITER(i, j, k)                           \
752
  F_INIT(i);                                    \
753
  F2(i, j, k, 0);                               \
754
  F2(i, j, k, 1);                               \
755
  F2(i, j, k, 2);                               \
756
  F2(i, j, k, 3);                               \
757
  F2(i, j, k, 4);                               \
758
  F2(i, j, k, 5);                               \
759
  F2(i, j, k, 6);                               \
760
  F2(i, j, k, 7)
761

    
762
  ITER(0, 1, 2);
763
  ITER(1, 2, 3);
764
  ITER(2, 3, 4);
765
  ITER(3, 4, 5);
766
  ITER(4, 5, 6);
767
  ITER(5, 6, 7);
768
  ITER(6, 7, 8);
769
  ITER(7, 8, 9);
770

    
771
  const vector signed char neg1 = vec_splat_s8(-1);
772

    
773
#define STORE_LINE(i)                                   \
774
  const vector unsigned char permST##i =                \
775
    vec_lvsr(i * stride, srcCopy);                      \
776
  const vector unsigned char maskST##i =                \
777
    vec_perm((vector unsigned char)zero,                \
778
             (vector unsigned char)neg1, permST##i);    \
779
  src##i = vec_perm(src##i ,src##i, permST##i);         \
780
  sA##i= vec_sel(sA##i, src##i, maskST##i);             \
781
  sB##i= vec_sel(src##i, sB##i, maskST##i);             \
782
  vec_st(sA##i, i * stride, srcCopy);                   \
783
  vec_st(sB##i, i * stride + 16, srcCopy)
784

    
785
  STORE_LINE(1);
786
  STORE_LINE(2);
787
  STORE_LINE(3);
788
  STORE_LINE(4);
789
  STORE_LINE(5);
790
  STORE_LINE(6);
791
  STORE_LINE(7);
792
  STORE_LINE(8);
793

    
794
#undef STORE_LINE
795
#undef ITER
796
#undef F2
797
}
798

    
799
#define doHorizLowPass_altivec(a...) doHorizLowPass_C(a)
800
#define doHorizDefFilter_altivec(a...) doHorizDefFilter_C(a)
801
#define do_a_deblock_altivec(a...) do_a_deblock_C(a)
802

    
803
static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,
804
                                    uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)
805
{
806
  const vector signed int zero = vec_splat_s32(0);
807
  const vector signed short vsint16_1 = vec_splat_s16(1);
808
  vector signed int v_dp = zero;
809
  vector signed int v_sysdp = zero;
810
  int d, sysd, i;
811

    
812
  tempBluredPast[127]= maxNoise[0];
813
  tempBluredPast[128]= maxNoise[1];
814
  tempBluredPast[129]= maxNoise[2];
815

    
816
#define LOAD_LINE(src, i)                                               \
817
  register int j##src##i = i * stride;                                  \
818
  vector unsigned char perm##src##i = vec_lvsl(j##src##i, src);         \
819
  const vector unsigned char v_##src##A1##i = vec_ld(j##src##i, src);   \
820
  const vector unsigned char v_##src##A2##i = vec_ld(j##src##i + 16, src); \
821
  const vector unsigned char v_##src##A##i =                            \
822
    vec_perm(v_##src##A1##i, v_##src##A2##i, perm##src##i);             \
823
  vector signed short v_##src##Ass##i =                                 \
824
    (vector signed short)vec_mergeh((vector signed char)zero,           \
825
                                    (vector signed char)v_##src##A##i)
826

    
827
  LOAD_LINE(src, 0);
828
  LOAD_LINE(src, 1);
829
  LOAD_LINE(src, 2);
830
  LOAD_LINE(src, 3);
831
  LOAD_LINE(src, 4);
832
  LOAD_LINE(src, 5);
833
  LOAD_LINE(src, 6);
834
  LOAD_LINE(src, 7);
835

    
836
  LOAD_LINE(tempBlured, 0);
837
  LOAD_LINE(tempBlured, 1);
838
  LOAD_LINE(tempBlured, 2);
839
  LOAD_LINE(tempBlured, 3);
840
  LOAD_LINE(tempBlured, 4);
841
  LOAD_LINE(tempBlured, 5);
842
  LOAD_LINE(tempBlured, 6);
843
  LOAD_LINE(tempBlured, 7);
844
#undef LOAD_LINE
845

    
846
#define ACCUMULATE_DIFFS(i)                                     \
847
  vector signed short v_d##i = vec_sub(v_tempBluredAss##i,      \
848
                                       v_srcAss##i);            \
849
  v_dp = vec_msums(v_d##i, v_d##i, v_dp);                       \
850
  v_sysdp = vec_msums(v_d##i, vsint16_1, v_sysdp)
851

    
852
  ACCUMULATE_DIFFS(0);
853
  ACCUMULATE_DIFFS(1);
854
  ACCUMULATE_DIFFS(2);
855
  ACCUMULATE_DIFFS(3);
856
  ACCUMULATE_DIFFS(4);
857
  ACCUMULATE_DIFFS(5);
858
  ACCUMULATE_DIFFS(6);
859
  ACCUMULATE_DIFFS(7);
860
#undef ACCUMULATE_DIFFS
861

    
862
  v_dp = vec_sums(v_dp, zero);
863
  v_sysdp = vec_sums(v_sysdp, zero);
864

    
865
  v_dp = vec_splat(v_dp, 3);
866
  v_sysdp = vec_splat(v_sysdp, 3);
867

    
868
  vec_ste(v_dp, 0, &d);
869
  vec_ste(v_sysdp, 0, &sysd);
870

    
871
  i = d;
872
  d = (4*d
873
       +(*(tempBluredPast-256))
874
       +(*(tempBluredPast-1))+ (*(tempBluredPast+1))
875
       +(*(tempBluredPast+256))
876
       +4)>>3;
877

    
878
  *tempBluredPast=i;
879

    
880
  if (d > maxNoise[1]) {
881
    if (d < maxNoise[2]) {
882
#define OP(i) v_tempBluredAss##i = vec_avg(v_tempBluredAss##i, v_srcAss##i);
883

    
884
      OP(0);
885
      OP(1);
886
      OP(2);
887
      OP(3);
888
      OP(4);
889
      OP(5);
890
      OP(6);
891
      OP(7);
892
#undef OP
893
    } else {
894
#define OP(i) v_tempBluredAss##i = v_srcAss##i;
895

    
896
      OP(0);
897
      OP(1);
898
      OP(2);
899
      OP(3);
900
      OP(4);
901
      OP(5);
902
      OP(6);
903
      OP(7);
904
#undef OP
905
    }
906
  } else {
907
    if (d < maxNoise[0]) {
908
      const vector signed short vsint16_7 = vec_splat_s16(7);
909
      const vector signed short vsint16_4 = vec_splat_s16(4);
910
      const vector unsigned short vuint16_3 = vec_splat_u16(3);
911

    
912
#define OP(i)                                                   \
913
      const vector signed short v_temp##i =                     \
914
        vec_mladd(v_tempBluredAss##i,                           \
915
                  vsint16_7, v_srcAss##i);                      \
916
      const vector signed short v_temp2##i =                    \
917
        vec_add(v_temp##i, vsint16_4);                          \
918
      v_tempBluredAss##i = vec_sr(v_temp2##i, vuint16_3)
919

    
920
      OP(0);
921
      OP(1);
922
      OP(2);
923
      OP(3);
924
      OP(4);
925
      OP(5);
926
      OP(6);
927
      OP(7);
928
#undef OP
929
    } else {
930
      const vector signed short vsint16_3 = vec_splat_s16(3);
931
      const vector signed short vsint16_2 = vec_splat_s16(2);
932

    
933
#define OP(i)                                                   \
934
      const vector signed short v_temp##i =                     \
935
        vec_mladd(v_tempBluredAss##i,                           \
936
                  vsint16_3, v_srcAss##i);                      \
937
      const vector signed short v_temp2##i =                    \
938
        vec_add(v_temp##i, vsint16_2);                          \
939
      v_tempBluredAss##i = vec_sr(v_temp2##i, (vector unsigned short)vsint16_2)
940

    
941
      OP(0);
942
      OP(1);
943
      OP(2);
944
      OP(3);
945
      OP(4);
946
      OP(5);
947
      OP(6);
948
      OP(7);
949
#undef OP
950
    }
951
  }
952

    
953
  const vector signed char neg1 = vec_splat_s8(-1);
954
  const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
955
                                                                      0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
956

    
957
#define PACK_AND_STORE(src, i)                                  \
958
  const vector unsigned char perms##src##i =                    \
959
    vec_lvsr(i * stride, src);                                  \
960
  const vector unsigned char vf##src##i =                       \
961
    vec_packsu(v_tempBluredAss##i, (vector signed short)zero);  \
962
  const vector unsigned char vg##src##i =                       \
963
    vec_perm(vf##src##i, v_##src##A##i, permHH);                \
964
  const vector unsigned char mask##src##i =                     \
965
    vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##src##i); \
966
  const vector unsigned char vg2##src##i =                      \
967
    vec_perm(vg##src##i, vg##src##i, perms##src##i);            \
968
  const vector unsigned char svA##src##i =                      \
969
    vec_sel(v_##src##A1##i, vg2##src##i, mask##src##i);         \
970
  const vector unsigned char svB##src##i =                      \
971
    vec_sel(vg2##src##i, v_##src##A2##i, mask##src##i);         \
972
  vec_st(svA##src##i, i * stride, src);                         \
973
  vec_st(svB##src##i, i * stride + 16, src)
974

    
975
  PACK_AND_STORE(src, 0);
976
  PACK_AND_STORE(src, 1);
977
  PACK_AND_STORE(src, 2);
978
  PACK_AND_STORE(src, 3);
979
  PACK_AND_STORE(src, 4);
980
  PACK_AND_STORE(src, 5);
981
  PACK_AND_STORE(src, 6);
982
  PACK_AND_STORE(src, 7);
983
  PACK_AND_STORE(tempBlured, 0);
984
  PACK_AND_STORE(tempBlured, 1);
985
  PACK_AND_STORE(tempBlured, 2);
986
  PACK_AND_STORE(tempBlured, 3);
987
  PACK_AND_STORE(tempBlured, 4);
988
  PACK_AND_STORE(tempBlured, 5);
989
  PACK_AND_STORE(tempBlured, 6);
990
  PACK_AND_STORE(tempBlured, 7);
991
#undef PACK_AND_STORE
992
}
993

    
994
static inline void transpose_16x8_char_toPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
995
  const vector unsigned char zero = vec_splat_u8(0);
996

    
997
#define LOAD_DOUBLE_LINE(i, j)                                          \
998
  vector unsigned char perm1##i = vec_lvsl(i * stride, src);            \
999
  vector unsigned char perm2##i = vec_lvsl(j * stride, src);            \
1000
  vector unsigned char srcA##i = vec_ld(i * stride, src);               \
1001
  vector unsigned char srcB##i = vec_ld(i * stride + 16, src);          \
1002
  vector unsigned char srcC##i = vec_ld(j * stride, src);               \
1003
  vector unsigned char srcD##i = vec_ld(j * stride+ 16, src);           \
1004
  vector unsigned char src##i = vec_perm(srcA##i, srcB##i, perm1##i);   \
1005
  vector unsigned char src##j = vec_perm(srcC##i, srcD##i, perm2##i)
1006

    
1007
  LOAD_DOUBLE_LINE(0, 1);
1008
  LOAD_DOUBLE_LINE(2, 3);
1009
  LOAD_DOUBLE_LINE(4, 5);
1010
  LOAD_DOUBLE_LINE(6, 7);
1011
#undef LOAD_DOUBLE_LINE
1012

    
1013
  vector unsigned char tempA = vec_mergeh(src0, zero);
1014
  vector unsigned char tempB = vec_mergel(src0, zero);
1015
  vector unsigned char tempC = vec_mergeh(src1, zero);
1016
  vector unsigned char tempD = vec_mergel(src1, zero);
1017
  vector unsigned char tempE = vec_mergeh(src2, zero);
1018
  vector unsigned char tempF = vec_mergel(src2, zero);
1019
  vector unsigned char tempG = vec_mergeh(src3, zero);
1020
  vector unsigned char tempH = vec_mergel(src3, zero);
1021
  vector unsigned char tempI = vec_mergeh(src4, zero);
1022
  vector unsigned char tempJ = vec_mergel(src4, zero);
1023
  vector unsigned char tempK = vec_mergeh(src5, zero);
1024
  vector unsigned char tempL = vec_mergel(src5, zero);
1025
  vector unsigned char tempM = vec_mergeh(src6, zero);
1026
  vector unsigned char tempN = vec_mergel(src6, zero);
1027
  vector unsigned char tempO = vec_mergeh(src7, zero);
1028
  vector unsigned char tempP = vec_mergel(src7, zero);
1029

    
1030
  vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1031
  vector unsigned char temp1 = vec_mergel(tempA, tempI);
1032
  vector unsigned char temp2 = vec_mergeh(tempB, tempJ);
1033
  vector unsigned char temp3 = vec_mergel(tempB, tempJ);
1034
  vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1035
  vector unsigned char temp5 = vec_mergel(tempC, tempK);
1036
  vector unsigned char temp6 = vec_mergeh(tempD, tempL);
1037
  vector unsigned char temp7 = vec_mergel(tempD, tempL);
1038
  vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1039
  vector unsigned char temp9 = vec_mergel(tempE, tempM);
1040
  vector unsigned char temp10 = vec_mergeh(tempF, tempN);
1041
  vector unsigned char temp11 = vec_mergel(tempF, tempN);
1042
  vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1043
  vector unsigned char temp13 = vec_mergel(tempG, tempO);
1044
  vector unsigned char temp14 = vec_mergeh(tempH, tempP);
1045
  vector unsigned char temp15 = vec_mergel(tempH, tempP);
1046

    
1047
  tempA = vec_mergeh(temp0, temp8);
1048
  tempB = vec_mergel(temp0, temp8);
1049
  tempC = vec_mergeh(temp1, temp9);
1050
  tempD = vec_mergel(temp1, temp9);
1051
  tempE = vec_mergeh(temp2, temp10);
1052
  tempF = vec_mergel(temp2, temp10);
1053
  tempG = vec_mergeh(temp3, temp11);
1054
  tempH = vec_mergel(temp3, temp11);
1055
  tempI = vec_mergeh(temp4, temp12);
1056
  tempJ = vec_mergel(temp4, temp12);
1057
  tempK = vec_mergeh(temp5, temp13);
1058
  tempL = vec_mergel(temp5, temp13);
1059
  tempM = vec_mergeh(temp6, temp14);
1060
  tempN = vec_mergel(temp6, temp14);
1061
  tempO = vec_mergeh(temp7, temp15);
1062
  tempP = vec_mergel(temp7, temp15);
1063

    
1064
  temp0 = vec_mergeh(tempA, tempI);
1065
  temp1 = vec_mergel(tempA, tempI);
1066
  temp2 = vec_mergeh(tempB, tempJ);
1067
  temp3 = vec_mergel(tempB, tempJ);
1068
  temp4 = vec_mergeh(tempC, tempK);
1069
  temp5 = vec_mergel(tempC, tempK);
1070
  temp6 = vec_mergeh(tempD, tempL);
1071
  temp7 = vec_mergel(tempD, tempL);
1072
  temp8 = vec_mergeh(tempE, tempM);
1073
  temp9 = vec_mergel(tempE, tempM);
1074
  temp10 = vec_mergeh(tempF, tempN);
1075
  temp11 = vec_mergel(tempF, tempN);
1076
  temp12 = vec_mergeh(tempG, tempO);
1077
  temp13 = vec_mergel(tempG, tempO);
1078
  temp14 = vec_mergeh(tempH, tempP);
1079
  temp15 = vec_mergel(tempH, tempP);
1080

    
1081
  vec_st(temp0, 0, dst);
1082
  vec_st(temp1, 16, dst);
1083
  vec_st(temp2, 32, dst);
1084
  vec_st(temp3, 48, dst);
1085
  vec_st(temp4, 64, dst);
1086
  vec_st(temp5, 80, dst);
1087
  vec_st(temp6, 96, dst);
1088
  vec_st(temp7, 112, dst);
1089
  vec_st(temp8, 128, dst);
1090
  vec_st(temp9, 144, dst);
1091
  vec_st(temp10, 160, dst);
1092
  vec_st(temp11, 176, dst);
1093
  vec_st(temp12, 192, dst);
1094
  vec_st(temp13, 208, dst);
1095
  vec_st(temp14, 224, dst);
1096
  vec_st(temp15, 240, dst);
1097
}
1098

    
1099
static inline void transpose_8x16_char_fromPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1100
  const vector unsigned char zero = vec_splat_u8(0);
1101

    
1102
#define LOAD_DOUBLE_LINE(i, j)                                  \
1103
  vector unsigned char src##i = vec_ld(i * 16, src);            \
1104
  vector unsigned char src##j = vec_ld(j * 16, src)
1105

    
1106
  LOAD_DOUBLE_LINE(0, 1);
1107
  LOAD_DOUBLE_LINE(2, 3);
1108
  LOAD_DOUBLE_LINE(4, 5);
1109
  LOAD_DOUBLE_LINE(6, 7);
1110
  LOAD_DOUBLE_LINE(8, 9);
1111
  LOAD_DOUBLE_LINE(10, 11);
1112
  LOAD_DOUBLE_LINE(12, 13);
1113
  LOAD_DOUBLE_LINE(14, 15);
1114
#undef LOAD_DOUBLE_LINE
1115

    
1116
  vector unsigned char tempA = vec_mergeh(src0, src8);
1117
  vector unsigned char tempB;
1118
  vector unsigned char tempC = vec_mergeh(src1, src9);
1119
  vector unsigned char tempD;
1120
  vector unsigned char tempE = vec_mergeh(src2, src10);
1121
  vector unsigned char tempG = vec_mergeh(src3, src11);
1122
  vector unsigned char tempI = vec_mergeh(src4, src12);
1123
  vector unsigned char tempJ;
1124
  vector unsigned char tempK = vec_mergeh(src5, src13);
1125
  vector unsigned char tempL;
1126
  vector unsigned char tempM = vec_mergeh(src6, src14);
1127
  vector unsigned char tempO = vec_mergeh(src7, src15);
1128

    
1129
  vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1130
  vector unsigned char temp1 = vec_mergel(tempA, tempI);
1131
  vector unsigned char temp2;
1132
  vector unsigned char temp3;
1133
  vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1134
  vector unsigned char temp5 = vec_mergel(tempC, tempK);
1135
  vector unsigned char temp6;
1136
  vector unsigned char temp7;
1137
  vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1138
  vector unsigned char temp9 = vec_mergel(tempE, tempM);
1139
  vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1140
  vector unsigned char temp13 = vec_mergel(tempG, tempO);
1141

    
1142
  tempA = vec_mergeh(temp0, temp8);
1143
  tempB = vec_mergel(temp0, temp8);
1144
  tempC = vec_mergeh(temp1, temp9);
1145
  tempD = vec_mergel(temp1, temp9);
1146
  tempI = vec_mergeh(temp4, temp12);
1147
  tempJ = vec_mergel(temp4, temp12);
1148
  tempK = vec_mergeh(temp5, temp13);
1149
  tempL = vec_mergel(temp5, temp13);
1150

    
1151
  temp0 = vec_mergeh(tempA, tempI);
1152
  temp1 = vec_mergel(tempA, tempI);
1153
  temp2 = vec_mergeh(tempB, tempJ);
1154
  temp3 = vec_mergel(tempB, tempJ);
1155
  temp4 = vec_mergeh(tempC, tempK);
1156
  temp5 = vec_mergel(tempC, tempK);
1157
  temp6 = vec_mergeh(tempD, tempL);
1158
  temp7 = vec_mergel(tempD, tempL);
1159

    
1160

    
1161
  const vector signed char neg1 = vec_splat_s8(-1);
1162
#define STORE_DOUBLE_LINE(i, j)                                         \
1163
  vector unsigned char dstA##i = vec_ld(i * stride, dst);               \
1164
  vector unsigned char dstB##i = vec_ld(i * stride + 16, dst);          \
1165
  vector unsigned char dstA##j = vec_ld(j * stride, dst);               \
1166
  vector unsigned char dstB##j = vec_ld(j * stride+ 16, dst);           \
1167
  vector unsigned char align##i = vec_lvsr(i * stride, dst);            \
1168
  vector unsigned char align##j = vec_lvsr(j * stride, dst);            \
1169
  vector unsigned char mask##i = vec_perm(zero, (vector unsigned char)neg1, align##i); \
1170
  vector unsigned char mask##j = vec_perm(zero, (vector unsigned char)neg1, align##j); \
1171
  vector unsigned char dstR##i = vec_perm(temp##i, temp##i, align##i);  \
1172
  vector unsigned char dstR##j = vec_perm(temp##j, temp##j, align##j);  \
1173
  vector unsigned char dstAF##i = vec_sel(dstA##i, dstR##i, mask##i);   \
1174
  vector unsigned char dstBF##i = vec_sel(dstR##i, dstB##i, mask##i);   \
1175
  vector unsigned char dstAF##j = vec_sel(dstA##j, dstR##j, mask##j);   \
1176
  vector unsigned char dstBF##j = vec_sel(dstR##j, dstB##j, mask##j);   \
1177
  vec_st(dstAF##i, i * stride, dst);                                    \
1178
  vec_st(dstBF##i, i * stride + 16, dst);                               \
1179
  vec_st(dstAF##j, j * stride, dst);                                    \
1180
  vec_st(dstBF##j, j * stride + 16, dst)
1181

    
1182
  STORE_DOUBLE_LINE(0,1);
1183
  STORE_DOUBLE_LINE(2,3);
1184
  STORE_DOUBLE_LINE(4,5);
1185
  STORE_DOUBLE_LINE(6,7);
1186
}