FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
postprocess_altivec_template.c
Go to the documentation of this file.
1 /*
2  * AltiVec optimizations (C) 2004 Romain Dolbeau <romain@dolbeau.org>
3  *
4  * based on code by Copyright (C) 2001-2003 Michael Niedermayer (michaelni@gmx.at)
5  *
6  * This file is part of FFmpeg.
7  *
8  * FFmpeg is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License as published by
10  * the Free Software Foundation; either version 2 of the License, or
11  * (at your option) any later version.
12  *
13  * FFmpeg is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16  * GNU General Public License for more details.
17  *
18  * You should have received a copy of the GNU General Public License
19  * along with FFmpeg; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22 
23 #include "libavutil/avutil.h"
24 
25 #define ALTIVEC_TRANSPOSE_8x8_SHORT(src_a,src_b,src_c,src_d,src_e,src_f,src_g,src_h) \
26  do { \
27  __typeof__(src_a) tempA1, tempB1, tempC1, tempD1; \
28  __typeof__(src_a) tempE1, tempF1, tempG1, tempH1; \
29  __typeof__(src_a) tempA2, tempB2, tempC2, tempD2; \
30  __typeof__(src_a) tempE2, tempF2, tempG2, tempH2; \
31  tempA1 = vec_mergeh (src_a, src_e); \
32  tempB1 = vec_mergel (src_a, src_e); \
33  tempC1 = vec_mergeh (src_b, src_f); \
34  tempD1 = vec_mergel (src_b, src_f); \
35  tempE1 = vec_mergeh (src_c, src_g); \
36  tempF1 = vec_mergel (src_c, src_g); \
37  tempG1 = vec_mergeh (src_d, src_h); \
38  tempH1 = vec_mergel (src_d, src_h); \
39  tempA2 = vec_mergeh (tempA1, tempE1); \
40  tempB2 = vec_mergel (tempA1, tempE1); \
41  tempC2 = vec_mergeh (tempB1, tempF1); \
42  tempD2 = vec_mergel (tempB1, tempF1); \
43  tempE2 = vec_mergeh (tempC1, tempG1); \
44  tempF2 = vec_mergel (tempC1, tempG1); \
45  tempG2 = vec_mergeh (tempD1, tempH1); \
46  tempH2 = vec_mergel (tempD1, tempH1); \
47  src_a = vec_mergeh (tempA2, tempE2); \
48  src_b = vec_mergel (tempA2, tempE2); \
49  src_c = vec_mergeh (tempB2, tempF2); \
50  src_d = vec_mergel (tempB2, tempF2); \
51  src_e = vec_mergeh (tempC2, tempG2); \
52  src_f = vec_mergel (tempC2, tempG2); \
53  src_g = vec_mergeh (tempD2, tempH2); \
54  src_h = vec_mergel (tempD2, tempH2); \
55  } while (0)
56 
57 
58 static inline int vertClassify_altivec(uint8_t src[], int stride, PPContext *c) {
59  /*
60  this code makes no assumption on src or stride.
61  One could remove the recomputation of the perm
62  vector by assuming (stride % 16) == 0, unfortunately
63  this is not always true.
64  */
65  short data_0 = ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
66  DECLARE_ALIGNED(16, short, data)[8] =
67  {
68  data_0,
69  data_0 * 2 + 1,
70  c->QP * 2,
71  c->QP * 4
72  };
73  int numEq;
74  uint8_t *src2 = src;
75  vector signed short v_dcOffset;
76  vector signed short v2QP;
77  vector unsigned short v4QP;
78  vector unsigned short v_dcThreshold;
79  const int properStride = (stride % 16);
80  const int srcAlign = ((unsigned long)src2 % 16);
81  const int two_vectors = ((srcAlign > 8) || properStride) ? 1 : 0;
82  const vector signed int zero = vec_splat_s32(0);
83  const vector signed short mask = vec_splat_s16(1);
84  vector signed int v_numEq = vec_splat_s32(0);
85  vector signed short v_data = vec_ld(0, data);
86  vector signed short v_srcAss0, v_srcAss1, v_srcAss2, v_srcAss3,
87  v_srcAss4, v_srcAss5, v_srcAss6, v_srcAss7;
88 //FIXME avoid this mess if possible
89  register int j0 = 0,
90  j1 = stride,
91  j2 = 2 * stride,
92  j3 = 3 * stride,
93  j4 = 4 * stride,
94  j5 = 5 * stride,
95  j6 = 6 * stride,
96  j7 = 7 * stride;
97  vector unsigned char v_srcA0, v_srcA1, v_srcA2, v_srcA3,
98  v_srcA4, v_srcA5, v_srcA6, v_srcA7;
99 
100  v_dcOffset = vec_splat(v_data, 0);
101  v_dcThreshold = (vector unsigned short)vec_splat(v_data, 1);
102  v2QP = vec_splat(v_data, 2);
103  v4QP = (vector unsigned short)vec_splat(v_data, 3);
104 
105  src2 += stride * 4;
106 
107 #define LOAD_LINE(i) \
108  { \
109  vector unsigned char perm##i = vec_lvsl(j##i, src2); \
110  vector unsigned char v_srcA2##i; \
111  vector unsigned char v_srcA1##i = vec_ld(j##i, src2); \
112  if (two_vectors) \
113  v_srcA2##i = vec_ld(j##i + 16, src2); \
114  v_srcA##i = \
115  vec_perm(v_srcA1##i, v_srcA2##i, perm##i); \
116  v_srcAss##i = \
117  (vector signed short)vec_mergeh((vector signed char)zero, \
118  (vector signed char)v_srcA##i); }
119 
120 #define LOAD_LINE_ALIGNED(i) \
121  v_srcA##i = vec_ld(j##i, src2); \
122  v_srcAss##i = \
123  (vector signed short)vec_mergeh((vector signed char)zero, \
124  (vector signed char)v_srcA##i)
125 
126  /* Special-casing the aligned case is worthwhile, as all calls from
127  * the (transposed) horizontable deblocks will be aligned, in addition
128  * to the naturally aligned vertical deblocks. */
129  if (properStride && srcAlign) {
138  } else {
139  LOAD_LINE(0);
140  LOAD_LINE(1);
141  LOAD_LINE(2);
142  LOAD_LINE(3);
143  LOAD_LINE(4);
144  LOAD_LINE(5);
145  LOAD_LINE(6);
146  LOAD_LINE(7);
147  }
148 #undef LOAD_LINE
149 #undef LOAD_LINE_ALIGNED
150 
151 #define ITER(i, j) \
152  const vector signed short v_diff##i = \
153  vec_sub(v_srcAss##i, v_srcAss##j); \
154  const vector signed short v_sum##i = \
155  vec_add(v_diff##i, v_dcOffset); \
156  const vector signed short v_comp##i = \
157  (vector signed short)vec_cmplt((vector unsigned short)v_sum##i, \
158  v_dcThreshold); \
159  const vector signed short v_part##i = vec_and(mask, v_comp##i);
160 
161  {
162  ITER(0, 1)
163  ITER(1, 2)
164  ITER(2, 3)
165  ITER(3, 4)
166  ITER(4, 5)
167  ITER(5, 6)
168  ITER(6, 7)
169 
170  v_numEq = vec_sum4s(v_part0, v_numEq);
171  v_numEq = vec_sum4s(v_part1, v_numEq);
172  v_numEq = vec_sum4s(v_part2, v_numEq);
173  v_numEq = vec_sum4s(v_part3, v_numEq);
174  v_numEq = vec_sum4s(v_part4, v_numEq);
175  v_numEq = vec_sum4s(v_part5, v_numEq);
176  v_numEq = vec_sum4s(v_part6, v_numEq);
177  }
178 
179 #undef ITER
180 
181  v_numEq = vec_sums(v_numEq, zero);
182 
183  v_numEq = vec_splat(v_numEq, 3);
184  vec_ste(v_numEq, 0, &numEq);
185 
186  if (numEq > c->ppMode.flatnessThreshold){
187  const vector unsigned char mmoP1 = (const vector unsigned char)
188  {0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f,
189  0x00, 0x01, 0x12, 0x13, 0x08, 0x09, 0x1A, 0x1B};
190  const vector unsigned char mmoP2 = (const vector unsigned char)
191  {0x04, 0x05, 0x16, 0x17, 0x0C, 0x0D, 0x1E, 0x1F,
192  0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f};
193  const vector unsigned char mmoP = (const vector unsigned char)
194  vec_lvsl(8, (unsigned char*)0);
195 
196  vector signed short mmoL1 = vec_perm(v_srcAss0, v_srcAss2, mmoP1);
197  vector signed short mmoL2 = vec_perm(v_srcAss4, v_srcAss6, mmoP2);
198  vector signed short mmoL = vec_perm(mmoL1, mmoL2, mmoP);
199  vector signed short mmoR1 = vec_perm(v_srcAss5, v_srcAss7, mmoP1);
200  vector signed short mmoR2 = vec_perm(v_srcAss1, v_srcAss3, mmoP2);
201  vector signed short mmoR = vec_perm(mmoR1, mmoR2, mmoP);
202  vector signed short mmoDiff = vec_sub(mmoL, mmoR);
203  vector unsigned short mmoSum = (vector unsigned short)vec_add(mmoDiff, v2QP);
204 
205  if (vec_any_gt(mmoSum, v4QP))
206  return 0;
207  else
208  return 1;
209  }
210  else return 2;
211 }
212 
213 static inline void doVertLowPass_altivec(uint8_t *src, int stride, PPContext *c) {
214  /*
215  this code makes no assumption on src or stride.
216  One could remove the recomputation of the perm
217  vector by assuming (stride % 16) == 0, unfortunately
218  this is not always true. Quite a lot of load/stores
219  can be removed by assuming proper alignment of
220  src & stride :-(
221  */
222  uint8_t *src2 = src;
223  const vector signed int zero = vec_splat_s32(0);
224  const int properStride = (stride % 16);
225  const int srcAlign = ((unsigned long)src2 % 16);
226  DECLARE_ALIGNED(16, short, qp)[8] = {c->QP};
227  vector signed short vqp = vec_ld(0, qp);
228  vector signed short vb0, vb1, vb2, vb3, vb4, vb5, vb6, vb7, vb8, vb9;
229  vector unsigned char vbA0, av_uninit(vbA1), av_uninit(vbA2), av_uninit(vbA3), av_uninit(vbA4), av_uninit(vbA5), av_uninit(vbA6), av_uninit(vbA7), av_uninit(vbA8), vbA9;
230  vector unsigned char vbB0, av_uninit(vbB1), av_uninit(vbB2), av_uninit(vbB3), av_uninit(vbB4), av_uninit(vbB5), av_uninit(vbB6), av_uninit(vbB7), av_uninit(vbB8), vbB9;
231  vector unsigned char vbT0, vbT1, vbT2, vbT3, vbT4, vbT5, vbT6, vbT7, vbT8, vbT9;
232  vector unsigned char perml0, perml1, perml2, perml3, perml4,
233  perml5, perml6, perml7, perml8, perml9;
234  register int j0 = 0,
235  j1 = stride,
236  j2 = 2 * stride,
237  j3 = 3 * stride,
238  j4 = 4 * stride,
239  j5 = 5 * stride,
240  j6 = 6 * stride,
241  j7 = 7 * stride,
242  j8 = 8 * stride,
243  j9 = 9 * stride;
244 
245  vqp = vec_splat(vqp, 0);
246 
247  src2 += stride*3;
248 
249 #define LOAD_LINE(i) \
250  perml##i = vec_lvsl(i * stride, src2); \
251  vbA##i = vec_ld(i * stride, src2); \
252  vbB##i = vec_ld(i * stride + 16, src2); \
253  vbT##i = vec_perm(vbA##i, vbB##i, perml##i); \
254  vb##i = \
255  (vector signed short)vec_mergeh((vector unsigned char)zero, \
256  (vector unsigned char)vbT##i)
257 
258 #define LOAD_LINE_ALIGNED(i) \
259  vbT##i = vec_ld(j##i, src2); \
260  vb##i = \
261  (vector signed short)vec_mergeh((vector signed char)zero, \
262  (vector signed char)vbT##i)
263 
264  /* Special-casing the aligned case is worthwhile, as all calls from
265  * the (transposed) horizontable deblocks will be aligned, in addition
266  * to the naturally aligned vertical deblocks. */
267  if (properStride && srcAlign) {
278  } else {
279  LOAD_LINE(0);
280  LOAD_LINE(1);
281  LOAD_LINE(2);
282  LOAD_LINE(3);
283  LOAD_LINE(4);
284  LOAD_LINE(5);
285  LOAD_LINE(6);
286  LOAD_LINE(7);
287  LOAD_LINE(8);
288  LOAD_LINE(9);
289  }
290 #undef LOAD_LINE
291 #undef LOAD_LINE_ALIGNED
292  {
293  const vector unsigned short v_2 = vec_splat_u16(2);
294  const vector unsigned short v_4 = vec_splat_u16(4);
295 
296  const vector signed short v_diff01 = vec_sub(vb0, vb1);
297  const vector unsigned short v_cmp01 =
298  (const vector unsigned short) vec_cmplt(vec_abs(v_diff01), vqp);
299  const vector signed short v_first = vec_sel(vb1, vb0, v_cmp01);
300  const vector signed short v_diff89 = vec_sub(vb8, vb9);
301  const vector unsigned short v_cmp89 =
302  (const vector unsigned short) vec_cmplt(vec_abs(v_diff89), vqp);
303  const vector signed short v_last = vec_sel(vb8, vb9, v_cmp89);
304 
305  const vector signed short temp01 = vec_mladd(v_first, (vector signed short)v_4, vb1);
306  const vector signed short temp02 = vec_add(vb2, vb3);
307  const vector signed short temp03 = vec_add(temp01, (vector signed short)v_4);
308  const vector signed short v_sumsB0 = vec_add(temp02, temp03);
309 
310  const vector signed short temp11 = vec_sub(v_sumsB0, v_first);
311  const vector signed short v_sumsB1 = vec_add(temp11, vb4);
312 
313  const vector signed short temp21 = vec_sub(v_sumsB1, v_first);
314  const vector signed short v_sumsB2 = vec_add(temp21, vb5);
315 
316  const vector signed short temp31 = vec_sub(v_sumsB2, v_first);
317  const vector signed short v_sumsB3 = vec_add(temp31, vb6);
318 
319  const vector signed short temp41 = vec_sub(v_sumsB3, v_first);
320  const vector signed short v_sumsB4 = vec_add(temp41, vb7);
321 
322  const vector signed short temp51 = vec_sub(v_sumsB4, vb1);
323  const vector signed short v_sumsB5 = vec_add(temp51, vb8);
324 
325  const vector signed short temp61 = vec_sub(v_sumsB5, vb2);
326  const vector signed short v_sumsB6 = vec_add(temp61, v_last);
327 
328  const vector signed short temp71 = vec_sub(v_sumsB6, vb3);
329  const vector signed short v_sumsB7 = vec_add(temp71, v_last);
330 
331  const vector signed short temp81 = vec_sub(v_sumsB7, vb4);
332  const vector signed short v_sumsB8 = vec_add(temp81, v_last);
333 
334  const vector signed short temp91 = vec_sub(v_sumsB8, vb5);
335  const vector signed short v_sumsB9 = vec_add(temp91, v_last);
336 
337  #define COMPUTE_VR(i, j, k) \
338  const vector signed short temps1##i = \
339  vec_add(v_sumsB##i, v_sumsB##k); \
340  const vector signed short temps2##i = \
341  vec_mladd(vb##j, (vector signed short)v_2, temps1##i); \
342  const vector signed short vr##j = vec_sra(temps2##i, v_4)
343 
344  COMPUTE_VR(0, 1, 2);
345  COMPUTE_VR(1, 2, 3);
346  COMPUTE_VR(2, 3, 4);
347  COMPUTE_VR(3, 4, 5);
348  COMPUTE_VR(4, 5, 6);
349  COMPUTE_VR(5, 6, 7);
350  COMPUTE_VR(6, 7, 8);
351  COMPUTE_VR(7, 8, 9);
352 
353  const vector signed char neg1 = vec_splat_s8(-1);
354  const vector unsigned char permHH = (const vector unsigned char){0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
355  0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F};
356 
357 #define PACK_AND_STORE(i) \
358 { const vector unsigned char perms##i = \
359  vec_lvsr(i * stride, src2); \
360  const vector unsigned char vf##i = \
361  vec_packsu(vr##i, (vector signed short)zero); \
362  const vector unsigned char vg##i = \
363  vec_perm(vf##i, vbT##i, permHH); \
364  const vector unsigned char mask##i = \
365  vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \
366  const vector unsigned char vg2##i = \
367  vec_perm(vg##i, vg##i, perms##i); \
368  const vector unsigned char svA##i = \
369  vec_sel(vbA##i, vg2##i, mask##i); \
370  const vector unsigned char svB##i = \
371  vec_sel(vg2##i, vbB##i, mask##i); \
372  vec_st(svA##i, i * stride, src2); \
373  vec_st(svB##i, i * stride + 16, src2);}
374 
375 #define PACK_AND_STORE_ALIGNED(i) \
376 { const vector unsigned char vf##i = \
377  vec_packsu(vr##i, (vector signed short)zero); \
378  const vector unsigned char vg##i = \
379  vec_perm(vf##i, vbT##i, permHH); \
380  vec_st(vg##i, i * stride, src2);}
381 
382  /* Special-casing the aligned case is worthwhile, as all calls from
383  * the (transposed) horizontable deblocks will be aligned, in addition
384  * to the naturally aligned vertical deblocks. */
385  if (properStride && srcAlign) {
394  } else {
395  PACK_AND_STORE(1)
396  PACK_AND_STORE(2)
397  PACK_AND_STORE(3)
398  PACK_AND_STORE(4)
399  PACK_AND_STORE(5)
400  PACK_AND_STORE(6)
401  PACK_AND_STORE(7)
402  PACK_AND_STORE(8)
403  }
404  #undef PACK_AND_STORE
405  #undef PACK_AND_STORE_ALIGNED
406  }
407 }
408 
409 
410 
411 static inline void doVertDefFilter_altivec(uint8_t src[], int stride, PPContext *c) {
412  /*
413  this code makes no assumption on src or stride.
414  One could remove the recomputation of the perm
415  vector by assuming (stride % 16) == 0, unfortunately
416  this is not always true. Quite a lot of load/stores
417  can be removed by assuming proper alignment of
418  src & stride :-(
419  */
420  uint8_t *src2 = src + stride*3;
421  const vector signed int zero = vec_splat_s32(0);
422  DECLARE_ALIGNED(16, short, qp)[8] = {8*c->QP};
423  vector signed short vqp = vec_splat(
424  (vector signed short)vec_ld(0, qp), 0);
425 
426 #define LOAD_LINE(i) \
427  const vector unsigned char perm##i = \
428  vec_lvsl(i * stride, src2); \
429  const vector unsigned char vbA##i = \
430  vec_ld(i * stride, src2); \
431  const vector unsigned char vbB##i = \
432  vec_ld(i * stride + 16, src2); \
433  const vector unsigned char vbT##i = \
434  vec_perm(vbA##i, vbB##i, perm##i); \
435  const vector signed short vb##i = \
436  (vector signed short)vec_mergeh((vector unsigned char)zero, \
437  (vector unsigned char)vbT##i)
438 
439  LOAD_LINE(1);
440  LOAD_LINE(2);
441  LOAD_LINE(3);
442  LOAD_LINE(4);
443  LOAD_LINE(5);
444  LOAD_LINE(6);
445  LOAD_LINE(7);
446  LOAD_LINE(8);
447 #undef LOAD_LINE
448 
449  const vector signed short v_1 = vec_splat_s16(1);
450  const vector signed short v_2 = vec_splat_s16(2);
451  const vector signed short v_5 = vec_splat_s16(5);
452  const vector signed short v_32 = vec_sl(v_1,
453  (vector unsigned short)v_5);
454  /* middle energy */
455  const vector signed short l3minusl6 = vec_sub(vb3, vb6);
456  const vector signed short l5minusl4 = vec_sub(vb5, vb4);
457  const vector signed short twotimes_l3minusl6 = vec_mladd(v_2, l3minusl6, (vector signed short)zero);
458  const vector signed short mE = vec_mladd(v_5, l5minusl4, twotimes_l3minusl6);
459  const vector signed short absmE = vec_abs(mE);
460  /* left & right energy */
461  const vector signed short l1minusl4 = vec_sub(vb1, vb4);
462  const vector signed short l3minusl2 = vec_sub(vb3, vb2);
463  const vector signed short l5minusl8 = vec_sub(vb5, vb8);
464  const vector signed short l7minusl6 = vec_sub(vb7, vb6);
465  const vector signed short twotimes_l1minusl4 = vec_mladd(v_2, l1minusl4, (vector signed short)zero);
466  const vector signed short twotimes_l5minusl8 = vec_mladd(v_2, l5minusl8, (vector signed short)zero);
467  const vector signed short lE = vec_mladd(v_5, l3minusl2, twotimes_l1minusl4);
468  const vector signed short rE = vec_mladd(v_5, l7minusl6, twotimes_l5minusl8);
469  /* d */
470  const vector signed short ddiff = vec_sub(absmE,
471  vec_min(vec_abs(lE),
472  vec_abs(rE)));
473  const vector signed short ddiffclamp = vec_max(ddiff, (vector signed short)zero);
474  const vector signed short dtimes64 = vec_mladd(v_5, ddiffclamp, v_32);
475  const vector signed short d = vec_sra(dtimes64, vec_splat_u16(6));
476  const vector signed short minusd = vec_sub((vector signed short)zero, d);
477  const vector signed short finald = vec_sel(minusd,
478  d,
479  vec_cmpgt(vec_sub((vector signed short)zero, mE),
480  (vector signed short)zero));
481  /* q */
482  const vector signed short qtimes2 = vec_sub(vb4, vb5);
483  /* for a shift right to behave like /2, we need to add one
484  to all negative integer */
485  const vector signed short rounddown = vec_sel((vector signed short)zero,
486  v_1,
487  vec_cmplt(qtimes2, (vector signed short)zero));
488  const vector signed short q = vec_sra(vec_add(qtimes2, rounddown), vec_splat_u16(1));
489  /* clamp */
490  const vector signed short dclamp_P1 = vec_max((vector signed short)zero, finald);
491  const vector signed short dclamp_P = vec_min(dclamp_P1, q);
492  const vector signed short dclamp_N1 = vec_min((vector signed short)zero, finald);
493  const vector signed short dclamp_N = vec_max(dclamp_N1, q);
494 
495  const vector signed short dclampedfinal = vec_sel(dclamp_N,
496  dclamp_P,
497  vec_cmpgt(q, (vector signed short)zero));
498  const vector signed short dornotd = vec_sel((vector signed short)zero,
499  dclampedfinal,
500  vec_cmplt(absmE, vqp));
501  /* add/subtract to l4 and l5 */
502  const vector signed short vb4minusd = vec_sub(vb4, dornotd);
503  const vector signed short vb5plusd = vec_add(vb5, dornotd);
504  /* finally, stores */
505  const vector unsigned char st4 = vec_packsu(vb4minusd, (vector signed short)zero);
506  const vector unsigned char st5 = vec_packsu(vb5plusd, (vector signed short)zero);
507 
508  const vector signed char neg1 = vec_splat_s8(-1);
509  const vector unsigned char permHH = (const vector unsigned char){0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
510  0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F};
511 
512 #define STORE(i) \
513 { const vector unsigned char perms##i = \
514  vec_lvsr(i * stride, src2); \
515  const vector unsigned char vg##i = \
516  vec_perm(st##i, vbT##i, permHH); \
517  const vector unsigned char mask##i = \
518  vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \
519  const vector unsigned char vg2##i = \
520  vec_perm(vg##i, vg##i, perms##i); \
521  const vector unsigned char svA##i = \
522  vec_sel(vbA##i, vg2##i, mask##i); \
523  const vector unsigned char svB##i = \
524  vec_sel(vg2##i, vbB##i, mask##i); \
525  vec_st(svA##i, i * stride, src2); \
526  vec_st(svB##i, i * stride + 16, src2);}
527 
528  STORE(4)
529  STORE(5)
530 }
531 
532 static inline void dering_altivec(uint8_t src[], int stride, PPContext *c) {
533  const vector signed int vsint32_8 = vec_splat_s32(8);
534  const vector unsigned int vuint32_4 = vec_splat_u32(4);
535  const vector signed char neg1 = vec_splat_s8(-1);
536 
537  const vector unsigned char permA1 = (vector unsigned char)
538  {0x00, 0x01, 0x02, 0x10, 0x11, 0x12, 0x1F, 0x1F,
539  0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F};
540  const vector unsigned char permA2 = (vector unsigned char)
541  {0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x10, 0x11,
542  0x12, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F};
543  const vector unsigned char permA1inc = (vector unsigned char)
544  {0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00,
545  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00};
546  const vector unsigned char permA2inc = (vector unsigned char)
547  {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01,
548  0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00};
549  const vector unsigned char magic = (vector unsigned char)
550  {0x01, 0x02, 0x01, 0x02, 0x04, 0x02, 0x01, 0x02,
551  0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00};
552  const vector unsigned char extractPerm = (vector unsigned char)
553  {0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01,
554  0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01};
555  const vector unsigned char extractPermInc = (vector unsigned char)
556  {0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01,
557  0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01};
558  const vector unsigned char identity = vec_lvsl(0,(unsigned char *)0);
559  const vector unsigned char tenRight = (vector unsigned char)
560  {0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
561  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00};
562  const vector unsigned char eightLeft = (vector unsigned char)
563  {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
564  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08};
565 
566  /*
567  this code makes no assumption on src or stride.
568  One could remove the recomputation of the perm
569  vector by assuming (stride % 16) == 0, unfortunately
570  this is not always true. Quite a lot of load/stores
571  can be removed by assuming proper alignment of
572  src & stride :-(
573  */
574  uint8_t *srcCopy = src;
575  DECLARE_ALIGNED(16, uint8_t, dt)[16] = { deringThreshold };
576  const vector signed int zero = vec_splat_s32(0);
577  vector unsigned char v_dt = vec_splat(vec_ld(0, dt), 0);
578 
579 #define LOAD_LINE(i) \
580  const vector unsigned char perm##i = \
581  vec_lvsl(i * stride, srcCopy); \
582  vector unsigned char sA##i = vec_ld(i * stride, srcCopy); \
583  vector unsigned char sB##i = vec_ld(i * stride + 16, srcCopy); \
584  vector unsigned char src##i = vec_perm(sA##i, sB##i, perm##i)
585 
586  LOAD_LINE(0);
587  LOAD_LINE(1);
588  LOAD_LINE(2);
589  LOAD_LINE(3);
590  LOAD_LINE(4);
591  LOAD_LINE(5);
592  LOAD_LINE(6);
593  LOAD_LINE(7);
594  LOAD_LINE(8);
595  LOAD_LINE(9);
596 #undef LOAD_LINE
597 
598  vector unsigned char v_avg;
599  DECLARE_ALIGNED(16, signed int, S)[8];
600  DECLARE_ALIGNED(16, int, tQP2)[4] = { c->QP/2 + 1 };
601  vector signed int vQP2 = vec_ld(0, tQP2);
602  vQP2 = vec_splat(vQP2, 0);
603 
604  {
605  const vector unsigned char trunc_perm = (vector unsigned char)
606  {0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08,
607  0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18};
608  const vector unsigned char trunc_src12 = vec_perm(src1, src2, trunc_perm);
609  const vector unsigned char trunc_src34 = vec_perm(src3, src4, trunc_perm);
610  const vector unsigned char trunc_src56 = vec_perm(src5, src6, trunc_perm);
611  const vector unsigned char trunc_src78 = vec_perm(src7, src8, trunc_perm);
612 
613 #define EXTRACT(op) do { \
614  const vector unsigned char s_1 = vec_##op(trunc_src12, trunc_src34); \
615  const vector unsigned char s_2 = vec_##op(trunc_src56, trunc_src78); \
616  const vector unsigned char s_6 = vec_##op(s_1, s_2); \
617  const vector unsigned char s_8h = vec_mergeh(s_6, s_6); \
618  const vector unsigned char s_8l = vec_mergel(s_6, s_6); \
619  const vector unsigned char s_9 = vec_##op(s_8h, s_8l); \
620  const vector unsigned char s_9h = vec_mergeh(s_9, s_9); \
621  const vector unsigned char s_9l = vec_mergel(s_9, s_9); \
622  const vector unsigned char s_10 = vec_##op(s_9h, s_9l); \
623  const vector unsigned char s_10h = vec_mergeh(s_10, s_10); \
624  const vector unsigned char s_10l = vec_mergel(s_10, s_10); \
625  const vector unsigned char s_11 = vec_##op(s_10h, s_10l); \
626  const vector unsigned char s_11h = vec_mergeh(s_11, s_11); \
627  const vector unsigned char s_11l = vec_mergel(s_11, s_11); \
628  v_##op = vec_##op(s_11h, s_11l); \
629 } while (0)
630 
631  vector unsigned char v_min;
632  vector unsigned char v_max;
633  EXTRACT(min);
634  EXTRACT(max);
635 #undef EXTRACT
636 
637  if (vec_all_lt(vec_sub(v_max, v_min), v_dt))
638  return;
639 
640  v_avg = vec_avg(v_min, v_max);
641  }
642 
643  {
644  const vector unsigned short mask1 = (vector unsigned short)
645  {0x0001, 0x0002, 0x0004, 0x0008,
646  0x0010, 0x0020, 0x0040, 0x0080};
647  const vector unsigned short mask2 = (vector unsigned short)
648  {0x0100, 0x0200, 0x0000, 0x0000,
649  0x0000, 0x0000, 0x0000, 0x0000};
650 
651  const vector unsigned int vuint32_16 = vec_sl(vec_splat_u32(1), vec_splat_u32(4));
652  const vector unsigned int vuint32_1 = vec_splat_u32(1);
653 
654  vector signed int sumA2;
655  vector signed int sumB2;
656  vector signed int sum0, sum1, sum2, sum3, sum4;
657  vector signed int sum5, sum6, sum7, sum8, sum9;
658 
659 #define COMPARE(i) \
660  do { \
661  const vector unsigned char cmp = \
662  (vector unsigned char)vec_cmpgt(src##i, v_avg); \
663  const vector unsigned short cmpHi = \
664  (vector unsigned short)vec_mergeh(cmp, cmp); \
665  const vector unsigned short cmpLi = \
666  (vector unsigned short)vec_mergel(cmp, cmp); \
667  const vector signed short cmpHf = \
668  (vector signed short)vec_and(cmpHi, mask1); \
669  const vector signed short cmpLf = \
670  (vector signed short)vec_and(cmpLi, mask2); \
671  const vector signed int sump = vec_sum4s(cmpHf, zero); \
672  const vector signed int sumq = vec_sum4s(cmpLf, sump); \
673  sum##i = vec_sums(sumq, zero); \
674  } while (0)
675 
676  COMPARE(0);
677  COMPARE(1);
678  COMPARE(2);
679  COMPARE(3);
680  COMPARE(4);
681  COMPARE(5);
682  COMPARE(6);
683  COMPARE(7);
684  COMPARE(8);
685  COMPARE(9);
686 #undef COMPARE
687 
688  {
689  const vector signed int sump02 = vec_mergel(sum0, sum2);
690  const vector signed int sump13 = vec_mergel(sum1, sum3);
691  const vector signed int sumA = vec_mergel(sump02, sump13);
692 
693  const vector signed int sump46 = vec_mergel(sum4, sum6);
694  const vector signed int sump57 = vec_mergel(sum5, sum7);
695  const vector signed int sumB = vec_mergel(sump46, sump57);
696 
697  const vector signed int sump8A = vec_mergel(sum8, zero);
698  const vector signed int sump9B = vec_mergel(sum9, zero);
699  const vector signed int sumC = vec_mergel(sump8A, sump9B);
700 
701  const vector signed int tA = vec_sl(vec_nor(zero, sumA), vuint32_16);
702  const vector signed int tB = vec_sl(vec_nor(zero, sumB), vuint32_16);
703  const vector signed int tC = vec_sl(vec_nor(zero, sumC), vuint32_16);
704  const vector signed int t2A = vec_or(sumA, tA);
705  const vector signed int t2B = vec_or(sumB, tB);
706  const vector signed int t2C = vec_or(sumC, tC);
707  const vector signed int t3A = vec_and(vec_sra(t2A, vuint32_1),
708  vec_sl(t2A, vuint32_1));
709  const vector signed int t3B = vec_and(vec_sra(t2B, vuint32_1),
710  vec_sl(t2B, vuint32_1));
711  const vector signed int t3C = vec_and(vec_sra(t2C, vuint32_1),
712  vec_sl(t2C, vuint32_1));
713  const vector signed int yA = vec_and(t2A, t3A);
714  const vector signed int yB = vec_and(t2B, t3B);
715  const vector signed int yC = vec_and(t2C, t3C);
716 
717  const vector unsigned char strangeperm1 = vec_lvsl(4, (unsigned char*)0);
718  const vector unsigned char strangeperm2 = vec_lvsl(8, (unsigned char*)0);
719  const vector signed int sumAd4 = vec_perm(yA, yB, strangeperm1);
720  const vector signed int sumAd8 = vec_perm(yA, yB, strangeperm2);
721  const vector signed int sumBd4 = vec_perm(yB, yC, strangeperm1);
722  const vector signed int sumBd8 = vec_perm(yB, yC, strangeperm2);
723  const vector signed int sumAp = vec_and(yA,
724  vec_and(sumAd4,sumAd8));
725  const vector signed int sumBp = vec_and(yB,
726  vec_and(sumBd4,sumBd8));
727  sumA2 = vec_or(sumAp,
728  vec_sra(sumAp,
729  vuint32_16));
730  sumB2 = vec_or(sumBp,
731  vec_sra(sumBp,
732  vuint32_16));
733  }
734  vec_st(sumA2, 0, S);
735  vec_st(sumB2, 16, S);
736  }
737 
738  /* I'm not sure the following is actually faster
739  than straight, unvectorized C code :-( */
740 
741 #define F_INIT() \
742  vector unsigned char tenRightM = tenRight; \
743  vector unsigned char permA1M = permA1; \
744  vector unsigned char permA2M = permA2; \
745  vector unsigned char extractPermM = extractPerm
746 
747 #define F2(i, j, k, l) \
748  if (S[i] & (1 << (l+1))) { \
749  const vector unsigned char a_A = vec_perm(src##i, src##j, permA1M); \
750  const vector unsigned char a_B = vec_perm(a_A, src##k, permA2M); \
751  const vector signed int a_sump = \
752  (vector signed int)vec_msum(a_B, magic, (vector unsigned int)zero);\
753  vector signed int F = vec_sr(vec_sums(a_sump, vsint32_8), vuint32_4); \
754  const vector signed int p = \
755  (vector signed int)vec_perm(src##j, (vector unsigned char)zero, \
756  extractPermM); \
757  const vector signed int sum = vec_add(p, vQP2); \
758  const vector signed int diff = vec_sub(p, vQP2); \
759  vector signed int newpm; \
760  vector unsigned char newpm2, mask; \
761  F = vec_splat(F, 3); \
762  if (vec_all_lt(sum, F)) \
763  newpm = sum; \
764  else if (vec_all_gt(diff, F)) \
765  newpm = diff; \
766  else newpm = F; \
767  newpm2 = vec_splat((vector unsigned char)newpm, 15); \
768  mask = vec_add(identity, tenRightM); \
769  src##j = vec_perm(src##j, newpm2, mask); \
770  } \
771  permA1M = vec_add(permA1M, permA1inc); \
772  permA2M = vec_add(permA2M, permA2inc); \
773  tenRightM = vec_sro(tenRightM, eightLeft); \
774  extractPermM = vec_add(extractPermM, extractPermInc)
775 
776 #define ITER(i, j, k) do { \
777  F_INIT(); \
778  F2(i, j, k, 0); \
779  F2(i, j, k, 1); \
780  F2(i, j, k, 2); \
781  F2(i, j, k, 3); \
782  F2(i, j, k, 4); \
783  F2(i, j, k, 5); \
784  F2(i, j, k, 6); \
785  F2(i, j, k, 7); \
786 } while (0)
787 
788  ITER(0, 1, 2);
789  ITER(1, 2, 3);
790  ITER(2, 3, 4);
791  ITER(3, 4, 5);
792  ITER(4, 5, 6);
793  ITER(5, 6, 7);
794  ITER(6, 7, 8);
795  ITER(7, 8, 9);
796 
797 #define STORE_LINE(i) do { \
798  const vector unsigned char permST = \
799  vec_lvsr(i * stride, srcCopy); \
800  const vector unsigned char maskST = \
801  vec_perm((vector unsigned char)zero, \
802  (vector unsigned char)neg1, permST); \
803  src##i = vec_perm(src##i ,src##i, permST); \
804  sA##i= vec_sel(sA##i, src##i, maskST); \
805  sB##i= vec_sel(src##i, sB##i, maskST); \
806  vec_st(sA##i, i * stride, srcCopy); \
807  vec_st(sB##i, i * stride + 16, srcCopy); \
808 } while (0)
809 
810  STORE_LINE(1);
811  STORE_LINE(2);
812  STORE_LINE(3);
813  STORE_LINE(4);
814  STORE_LINE(5);
815  STORE_LINE(6);
816  STORE_LINE(7);
817  STORE_LINE(8);
818 
819 #undef STORE_LINE
820 #undef ITER
821 #undef F2
822 }
823 
824 #define doHorizLowPass_altivec(a...) doHorizLowPass_C(a)
825 #define doHorizDefFilter_altivec(a...) doHorizDefFilter_C(a)
826 #define do_a_deblock_altivec(a...) do_a_deblock_C(a)
827 
828 static inline void tempNoiseReducer_altivec(uint8_t *src, int stride,
829  uint8_t *tempBlurred, uint32_t *tempBlurredPast, int *maxNoise)
830 {
831  const vector signed char neg1 = vec_splat_s8(-1);
832  const vector unsigned char permHH = (const vector unsigned char){0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
833  0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F};
834 
835  const vector signed int zero = vec_splat_s32(0);
836  const vector signed short vsint16_1 = vec_splat_s16(1);
837  vector signed int v_dp = zero;
838  vector signed int v_sysdp = zero;
839  int d, sysd, i;
840 
841 #define LOAD_LINE(src, i) \
842  register int j##src##i = i * stride; \
843  vector unsigned char perm##src##i = vec_lvsl(j##src##i, src); \
844  const vector unsigned char v_##src##A1##i = vec_ld(j##src##i, src); \
845  const vector unsigned char v_##src##A2##i = vec_ld(j##src##i + 16, src); \
846  const vector unsigned char v_##src##A##i = \
847  vec_perm(v_##src##A1##i, v_##src##A2##i, perm##src##i); \
848  vector signed short v_##src##Ass##i = \
849  (vector signed short)vec_mergeh((vector signed char)zero, \
850  (vector signed char)v_##src##A##i)
851 
852  LOAD_LINE(src, 0);
853  LOAD_LINE(src, 1);
854  LOAD_LINE(src, 2);
855  LOAD_LINE(src, 3);
856  LOAD_LINE(src, 4);
857  LOAD_LINE(src, 5);
858  LOAD_LINE(src, 6);
859  LOAD_LINE(src, 7);
860 
861  LOAD_LINE(tempBlurred, 0);
862  LOAD_LINE(tempBlurred, 1);
863  LOAD_LINE(tempBlurred, 2);
864  LOAD_LINE(tempBlurred, 3);
865  LOAD_LINE(tempBlurred, 4);
866  LOAD_LINE(tempBlurred, 5);
867  LOAD_LINE(tempBlurred, 6);
868  LOAD_LINE(tempBlurred, 7);
869 #undef LOAD_LINE
870 
871 #define ACCUMULATE_DIFFS(i) do { \
872  vector signed short v_d = vec_sub(v_tempBlurredAss##i, \
873  v_srcAss##i); \
874  v_dp = vec_msums(v_d, v_d, v_dp); \
875  v_sysdp = vec_msums(v_d, vsint16_1, v_sysdp); \
876  } while (0)
877 
878  ACCUMULATE_DIFFS(0);
879  ACCUMULATE_DIFFS(1);
880  ACCUMULATE_DIFFS(2);
881  ACCUMULATE_DIFFS(3);
882  ACCUMULATE_DIFFS(4);
883  ACCUMULATE_DIFFS(5);
884  ACCUMULATE_DIFFS(6);
885  ACCUMULATE_DIFFS(7);
886 #undef ACCUMULATE_DIFFS
887 
888  tempBlurredPast[127]= maxNoise[0];
889  tempBlurredPast[128]= maxNoise[1];
890  tempBlurredPast[129]= maxNoise[2];
891 
892  v_dp = vec_sums(v_dp, zero);
893  v_sysdp = vec_sums(v_sysdp, zero);
894 
895  v_dp = vec_splat(v_dp, 3);
896  v_sysdp = vec_splat(v_sysdp, 3);
897 
898  vec_ste(v_dp, 0, &d);
899  vec_ste(v_sysdp, 0, &sysd);
900 
901  i = d;
902  d = (4*d
903  +(*(tempBlurredPast-256))
904  +(*(tempBlurredPast-1))+ (*(tempBlurredPast+1))
905  +(*(tempBlurredPast+256))
906  +4)>>3;
907 
908  *tempBlurredPast=i;
909 
910  if (d > maxNoise[1]) {
911  if (d < maxNoise[2]) {
912 #define OP(i) v_tempBlurredAss##i = vec_avg(v_tempBlurredAss##i, v_srcAss##i);
913 
914  OP(0);
915  OP(1);
916  OP(2);
917  OP(3);
918  OP(4);
919  OP(5);
920  OP(6);
921  OP(7);
922 #undef OP
923  } else {
924 #define OP(i) v_tempBlurredAss##i = v_srcAss##i;
925 
926  OP(0);
927  OP(1);
928  OP(2);
929  OP(3);
930  OP(4);
931  OP(5);
932  OP(6);
933  OP(7);
934 #undef OP
935  }
936  } else {
937  if (d < maxNoise[0]) {
938  const vector signed short vsint16_7 = vec_splat_s16(7);
939  const vector signed short vsint16_4 = vec_splat_s16(4);
940  const vector unsigned short vuint16_3 = vec_splat_u16(3);
941 
942 #define OP(i) do { \
943  const vector signed short v_temp = \
944  vec_mladd(v_tempBlurredAss##i, vsint16_7, v_srcAss##i); \
945  const vector signed short v_temp2 = vec_add(v_temp, vsint16_4); \
946  v_tempBlurredAss##i = vec_sr(v_temp2, vuint16_3); \
947  } while (0)
948 
949  OP(0);
950  OP(1);
951  OP(2);
952  OP(3);
953  OP(4);
954  OP(5);
955  OP(6);
956  OP(7);
957 #undef OP
958  } else {
959  const vector signed short vsint16_3 = vec_splat_s16(3);
960  const vector signed short vsint16_2 = vec_splat_s16(2);
961 
962 #define OP(i) do { \
963  const vector signed short v_temp = \
964  vec_mladd(v_tempBlurredAss##i, vsint16_3, v_srcAss##i); \
965  const vector signed short v_temp2 = vec_add(v_temp, vsint16_2); \
966  v_tempBlurredAss##i = \
967  vec_sr(v_temp2, (vector unsigned short)vsint16_2); \
968  } while (0)
969 
970  OP(0);
971  OP(1);
972  OP(2);
973  OP(3);
974  OP(4);
975  OP(5);
976  OP(6);
977  OP(7);
978 #undef OP
979  }
980  }
981 
982 #define PACK_AND_STORE(src, i) do { \
983  const vector unsigned char perms = vec_lvsr(i * stride, src); \
984  const vector unsigned char vf = \
985  vec_packsu(v_tempBlurredAss##1, (vector signed short)zero); \
986  const vector unsigned char vg = vec_perm(vf, v_##src##A##i, permHH); \
987  const vector unsigned char mask = \
988  vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms); \
989  const vector unsigned char vg2 = vec_perm(vg, vg, perms); \
990  const vector unsigned char svA = vec_sel(v_##src##A1##i, vg2, mask); \
991  const vector unsigned char svB = vec_sel(vg2, v_##src##A2##i, mask); \
992  vec_st(svA, i * stride, src); \
993  vec_st(svB, i * stride + 16, src); \
994 } while (0)
995 
996  PACK_AND_STORE(src, 0);
997  PACK_AND_STORE(src, 1);
998  PACK_AND_STORE(src, 2);
999  PACK_AND_STORE(src, 3);
1000  PACK_AND_STORE(src, 4);
1001  PACK_AND_STORE(src, 5);
1002  PACK_AND_STORE(src, 6);
1003  PACK_AND_STORE(src, 7);
1004  PACK_AND_STORE(tempBlurred, 0);
1005  PACK_AND_STORE(tempBlurred, 1);
1006  PACK_AND_STORE(tempBlurred, 2);
1007  PACK_AND_STORE(tempBlurred, 3);
1008  PACK_AND_STORE(tempBlurred, 4);
1009  PACK_AND_STORE(tempBlurred, 5);
1010  PACK_AND_STORE(tempBlurred, 6);
1011  PACK_AND_STORE(tempBlurred, 7);
1012 #undef PACK_AND_STORE
1013 }
1014 
1015 static inline void transpose_16x8_char_toPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1016  const vector unsigned char zero = vec_splat_u8(0);
1017 
1018 #define LOAD_DOUBLE_LINE(i, j) \
1019  vector unsigned char perm1##i = vec_lvsl(i * stride, src); \
1020  vector unsigned char perm2##i = vec_lvsl(j * stride, src); \
1021  vector unsigned char srcA##i = vec_ld(i * stride, src); \
1022  vector unsigned char srcB##i = vec_ld(i * stride + 16, src); \
1023  vector unsigned char srcC##i = vec_ld(j * stride, src); \
1024  vector unsigned char srcD##i = vec_ld(j * stride+ 16, src); \
1025  vector unsigned char src##i = vec_perm(srcA##i, srcB##i, perm1##i); \
1026  vector unsigned char src##j = vec_perm(srcC##i, srcD##i, perm2##i)
1027 
1028  LOAD_DOUBLE_LINE(0, 1);
1029  LOAD_DOUBLE_LINE(2, 3);
1030  LOAD_DOUBLE_LINE(4, 5);
1031  LOAD_DOUBLE_LINE(6, 7);
1032 #undef LOAD_DOUBLE_LINE
1033 
1034  vector unsigned char tempA = vec_mergeh(src0, zero);
1035  vector unsigned char tempB = vec_mergel(src0, zero);
1036  vector unsigned char tempC = vec_mergeh(src1, zero);
1037  vector unsigned char tempD = vec_mergel(src1, zero);
1038  vector unsigned char tempE = vec_mergeh(src2, zero);
1039  vector unsigned char tempF = vec_mergel(src2, zero);
1040  vector unsigned char tempG = vec_mergeh(src3, zero);
1041  vector unsigned char tempH = vec_mergel(src3, zero);
1042  vector unsigned char tempI = vec_mergeh(src4, zero);
1043  vector unsigned char tempJ = vec_mergel(src4, zero);
1044  vector unsigned char tempK = vec_mergeh(src5, zero);
1045  vector unsigned char tempL = vec_mergel(src5, zero);
1046  vector unsigned char tempM = vec_mergeh(src6, zero);
1047  vector unsigned char tempN = vec_mergel(src6, zero);
1048  vector unsigned char tempO = vec_mergeh(src7, zero);
1049  vector unsigned char tempP = vec_mergel(src7, zero);
1050 
1051  vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1052  vector unsigned char temp1 = vec_mergel(tempA, tempI);
1053  vector unsigned char temp2 = vec_mergeh(tempB, tempJ);
1054  vector unsigned char temp3 = vec_mergel(tempB, tempJ);
1055  vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1056  vector unsigned char temp5 = vec_mergel(tempC, tempK);
1057  vector unsigned char temp6 = vec_mergeh(tempD, tempL);
1058  vector unsigned char temp7 = vec_mergel(tempD, tempL);
1059  vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1060  vector unsigned char temp9 = vec_mergel(tempE, tempM);
1061  vector unsigned char temp10 = vec_mergeh(tempF, tempN);
1062  vector unsigned char temp11 = vec_mergel(tempF, tempN);
1063  vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1064  vector unsigned char temp13 = vec_mergel(tempG, tempO);
1065  vector unsigned char temp14 = vec_mergeh(tempH, tempP);
1066  vector unsigned char temp15 = vec_mergel(tempH, tempP);
1067 
1068  tempA = vec_mergeh(temp0, temp8);
1069  tempB = vec_mergel(temp0, temp8);
1070  tempC = vec_mergeh(temp1, temp9);
1071  tempD = vec_mergel(temp1, temp9);
1072  tempE = vec_mergeh(temp2, temp10);
1073  tempF = vec_mergel(temp2, temp10);
1074  tempG = vec_mergeh(temp3, temp11);
1075  tempH = vec_mergel(temp3, temp11);
1076  tempI = vec_mergeh(temp4, temp12);
1077  tempJ = vec_mergel(temp4, temp12);
1078  tempK = vec_mergeh(temp5, temp13);
1079  tempL = vec_mergel(temp5, temp13);
1080  tempM = vec_mergeh(temp6, temp14);
1081  tempN = vec_mergel(temp6, temp14);
1082  tempO = vec_mergeh(temp7, temp15);
1083  tempP = vec_mergel(temp7, temp15);
1084 
1085  temp0 = vec_mergeh(tempA, tempI);
1086  temp1 = vec_mergel(tempA, tempI);
1087  temp2 = vec_mergeh(tempB, tempJ);
1088  temp3 = vec_mergel(tempB, tempJ);
1089  temp4 = vec_mergeh(tempC, tempK);
1090  temp5 = vec_mergel(tempC, tempK);
1091  temp6 = vec_mergeh(tempD, tempL);
1092  temp7 = vec_mergel(tempD, tempL);
1093  temp8 = vec_mergeh(tempE, tempM);
1094  temp9 = vec_mergel(tempE, tempM);
1095  temp10 = vec_mergeh(tempF, tempN);
1096  temp11 = vec_mergel(tempF, tempN);
1097  temp12 = vec_mergeh(tempG, tempO);
1098  temp13 = vec_mergel(tempG, tempO);
1099  temp14 = vec_mergeh(tempH, tempP);
1100  temp15 = vec_mergel(tempH, tempP);
1101 
1102  vec_st(temp0, 0, dst);
1103  vec_st(temp1, 16, dst);
1104  vec_st(temp2, 32, dst);
1105  vec_st(temp3, 48, dst);
1106  vec_st(temp4, 64, dst);
1107  vec_st(temp5, 80, dst);
1108  vec_st(temp6, 96, dst);
1109  vec_st(temp7, 112, dst);
1110  vec_st(temp8, 128, dst);
1111  vec_st(temp9, 144, dst);
1112  vec_st(temp10, 160, dst);
1113  vec_st(temp11, 176, dst);
1114  vec_st(temp12, 192, dst);
1115  vec_st(temp13, 208, dst);
1116  vec_st(temp14, 224, dst);
1117  vec_st(temp15, 240, dst);
1118 }
1119 
1120 static inline void transpose_8x16_char_fromPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1121  const vector unsigned char zero = vec_splat_u8(0);
1122  const vector signed char neg1 = vec_splat_s8(-1);
1123 
1124 #define LOAD_DOUBLE_LINE(i, j) \
1125  vector unsigned char src##i = vec_ld(i * 16, src); \
1126  vector unsigned char src##j = vec_ld(j * 16, src)
1127 
1128  LOAD_DOUBLE_LINE(0, 1);
1129  LOAD_DOUBLE_LINE(2, 3);
1130  LOAD_DOUBLE_LINE(4, 5);
1131  LOAD_DOUBLE_LINE(6, 7);
1132  LOAD_DOUBLE_LINE(8, 9);
1133  LOAD_DOUBLE_LINE(10, 11);
1134  LOAD_DOUBLE_LINE(12, 13);
1135  LOAD_DOUBLE_LINE(14, 15);
1136 #undef LOAD_DOUBLE_LINE
1137 
1138  vector unsigned char tempA = vec_mergeh(src0, src8);
1139  vector unsigned char tempB;
1140  vector unsigned char tempC = vec_mergeh(src1, src9);
1141  vector unsigned char tempD;
1142  vector unsigned char tempE = vec_mergeh(src2, src10);
1143  vector unsigned char tempG = vec_mergeh(src3, src11);
1144  vector unsigned char tempI = vec_mergeh(src4, src12);
1145  vector unsigned char tempJ;
1146  vector unsigned char tempK = vec_mergeh(src5, src13);
1147  vector unsigned char tempL;
1148  vector unsigned char tempM = vec_mergeh(src6, src14);
1149  vector unsigned char tempO = vec_mergeh(src7, src15);
1150 
1151  vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1152  vector unsigned char temp1 = vec_mergel(tempA, tempI);
1153  vector unsigned char temp2;
1154  vector unsigned char temp3;
1155  vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1156  vector unsigned char temp5 = vec_mergel(tempC, tempK);
1157  vector unsigned char temp6;
1158  vector unsigned char temp7;
1159  vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1160  vector unsigned char temp9 = vec_mergel(tempE, tempM);
1161  vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1162  vector unsigned char temp13 = vec_mergel(tempG, tempO);
1163 
1164  tempA = vec_mergeh(temp0, temp8);
1165  tempB = vec_mergel(temp0, temp8);
1166  tempC = vec_mergeh(temp1, temp9);
1167  tempD = vec_mergel(temp1, temp9);
1168  tempI = vec_mergeh(temp4, temp12);
1169  tempJ = vec_mergel(temp4, temp12);
1170  tempK = vec_mergeh(temp5, temp13);
1171  tempL = vec_mergel(temp5, temp13);
1172 
1173  temp0 = vec_mergeh(tempA, tempI);
1174  temp1 = vec_mergel(tempA, tempI);
1175  temp2 = vec_mergeh(tempB, tempJ);
1176  temp3 = vec_mergel(tempB, tempJ);
1177  temp4 = vec_mergeh(tempC, tempK);
1178  temp5 = vec_mergel(tempC, tempK);
1179  temp6 = vec_mergeh(tempD, tempL);
1180  temp7 = vec_mergel(tempD, tempL);
1181 
1182 
1183 #define STORE_DOUBLE_LINE(i, j) do { \
1184  vector unsigned char dstAi = vec_ld(i * stride, dst); \
1185  vector unsigned char dstBi = vec_ld(i * stride + 16, dst); \
1186  vector unsigned char dstAj = vec_ld(j * stride, dst); \
1187  vector unsigned char dstBj = vec_ld(j * stride+ 16, dst); \
1188  vector unsigned char aligni = vec_lvsr(i * stride, dst); \
1189  vector unsigned char alignj = vec_lvsr(j * stride, dst); \
1190  vector unsigned char maski = \
1191  vec_perm(zero, (vector unsigned char)neg1, aligni); \
1192  vector unsigned char maskj = \
1193  vec_perm(zero, (vector unsigned char)neg1, alignj); \
1194  vector unsigned char dstRi = vec_perm(temp##i, temp##i, aligni); \
1195  vector unsigned char dstRj = vec_perm(temp##j, temp##j, alignj); \
1196  vector unsigned char dstAFi = vec_sel(dstAi, dstRi, maski); \
1197  vector unsigned char dstBFi = vec_sel(dstRi, dstBi, maski); \
1198  vector unsigned char dstAFj = vec_sel(dstAj, dstRj, maskj); \
1199  vector unsigned char dstBFj = vec_sel(dstRj, dstBj, maskj); \
1200  vec_st(dstAFi, i * stride, dst); \
1201  vec_st(dstBFi, i * stride + 16, dst); \
1202  vec_st(dstAFj, j * stride, dst); \
1203  vec_st(dstBFj, j * stride + 16, dst); \
1204 } while (0)
1205 
1206  STORE_DOUBLE_LINE(0,1);
1207  STORE_DOUBLE_LINE(2,3);
1208  STORE_DOUBLE_LINE(4,5);
1209  STORE_DOUBLE_LINE(6,7);
1210 }