2
* AltiVec optimizations (C) 2004 Romain Dolbeau <romain@dolbeau.org>
4
* based on code by Copyright (C) 2001-2003 Michael Niedermayer (michaelni@gmx.at)
6
* This file is part of FFmpeg.
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.
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.
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
31
#define ALTIVEC_TRANSPOSE_8x8_SHORT(src_a,src_b,src_c,src_d,src_e,src_f,src_g,src_h) \
33
__typeof__(src_a) tempA1, tempB1, tempC1, tempD1; \
34
__typeof__(src_a) tempE1, tempF1, tempG1, tempH1; \
35
__typeof__(src_a) tempA2, tempB2, tempC2, tempD2; \
36
__typeof__(src_a) tempE2, tempF2, tempG2, tempH2; \
37
tempA1 = vec_mergeh (src_a, src_e); \
38
tempB1 = vec_mergel (src_a, src_e); \
39
tempC1 = vec_mergeh (src_b, src_f); \
40
tempD1 = vec_mergel (src_b, src_f); \
41
tempE1 = vec_mergeh (src_c, src_g); \
42
tempF1 = vec_mergel (src_c, src_g); \
43
tempG1 = vec_mergeh (src_d, src_h); \
44
tempH1 = vec_mergel (src_d, src_h); \
45
tempA2 = vec_mergeh (tempA1, tempE1); \
46
tempB2 = vec_mergel (tempA1, tempE1); \
47
tempC2 = vec_mergeh (tempB1, tempF1); \
48
tempD2 = vec_mergel (tempB1, tempF1); \
49
tempE2 = vec_mergeh (tempC1, tempG1); \
50
tempF2 = vec_mergel (tempC1, tempG1); \
51
tempG2 = vec_mergeh (tempD1, tempH1); \
52
tempH2 = vec_mergel (tempD1, tempH1); \
53
src_a = vec_mergeh (tempA2, tempE2); \
54
src_b = vec_mergel (tempA2, tempE2); \
55
src_c = vec_mergeh (tempB2, tempF2); \
56
src_d = vec_mergel (tempB2, tempF2); \
57
src_e = vec_mergeh (tempC2, tempG2); \
58
src_f = vec_mergel (tempC2, tempG2); \
59
src_g = vec_mergeh (tempD2, tempH2); \
60
src_h = vec_mergel (tempD2, tempH2); \
64
static inline int vertClassify_altivec(uint8_t src[], int stride, PPContext *c) {
66
this code makes no assumption on src or stride.
67
One could remove the recomputation of the perm
68
vector by assuming (stride % 16) == 0, unfortunately
69
this is not always true.
71
DECLARE_ALIGNED(16, short, data[8]);
74
vector signed short v_dcOffset;
75
vector signed short v2QP;
76
vector unsigned short v4QP;
77
vector unsigned short v_dcThreshold;
78
const int properStride = (stride % 16);
79
const int srcAlign = ((unsigned long)src2 % 16);
80
const int two_vectors = ((srcAlign > 8) || properStride) ? 1 : 0;
81
const vector signed int zero = vec_splat_s32(0);
82
const vector signed short mask = vec_splat_s16(1);
83
vector signed int v_numEq = vec_splat_s32(0);
85
data[0] = ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
86
data[1] = data[0] * 2 + 1;
89
vector signed short v_data = vec_ld(0, data);
90
v_dcOffset = vec_splat(v_data, 0);
91
v_dcThreshold = (vector unsigned short)vec_splat(v_data, 1);
92
v2QP = vec_splat(v_data, 2);
93
v4QP = (vector unsigned short)vec_splat(v_data, 3);
97
vector signed short v_srcAss0, v_srcAss1, v_srcAss2, v_srcAss3, v_srcAss4, v_srcAss5, v_srcAss6, v_srcAss7;
99
#define LOAD_LINE(i) \
100
register int j##i = i * stride; \
101
vector unsigned char perm##i = vec_lvsl(j##i, src2); \
102
const vector unsigned char v_srcA1##i = vec_ld(j##i, src2); \
103
vector unsigned char v_srcA2##i; \
105
v_srcA2##i = vec_ld(j##i + 16, src2); \
106
const vector unsigned char v_srcA##i = \
107
vec_perm(v_srcA1##i, v_srcA2##i, perm##i); \
109
(vector signed short)vec_mergeh((vector signed char)zero, \
110
(vector signed char)v_srcA##i)
112
#define LOAD_LINE_ALIGNED(i) \
113
register int j##i = i * stride; \
114
const vector unsigned char v_srcA##i = vec_ld(j##i, src2); \
116
(vector signed short)vec_mergeh((vector signed char)zero, \
117
(vector signed char)v_srcA##i)
119
// special casing the aligned case is worthwhile, as all call from
120
// the (transposed) horizontable deblocks will be aligned, i naddition
121
// to the naturraly aligned vertical deblocks.
122
if (properStride && srcAlign) {
123
LOAD_LINE_ALIGNED(0);
124
LOAD_LINE_ALIGNED(1);
125
LOAD_LINE_ALIGNED(2);
126
LOAD_LINE_ALIGNED(3);
127
LOAD_LINE_ALIGNED(4);
128
LOAD_LINE_ALIGNED(5);
129
LOAD_LINE_ALIGNED(6);
130
LOAD_LINE_ALIGNED(7);
142
#undef LOAD_LINE_ALIGNED
145
const vector signed short v_diff##i = \
146
vec_sub(v_srcAss##i, v_srcAss##j); \
147
const vector signed short v_sum##i = \
148
vec_add(v_diff##i, v_dcOffset); \
149
const vector signed short v_comp##i = \
150
(vector signed short)vec_cmplt((vector unsigned short)v_sum##i, \
152
const vector signed short v_part##i = vec_and(mask, v_comp##i); \
153
v_numEq = vec_sum4s(v_part##i, v_numEq);
164
v_numEq = vec_sums(v_numEq, zero);
166
v_numEq = vec_splat(v_numEq, 3);
167
vec_ste(v_numEq, 0, &numEq);
169
if (numEq > c->ppMode.flatnessThreshold)
171
const vector unsigned char mmoP1 = (const vector unsigned char)
172
AVV(0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f,
173
0x00, 0x01, 0x12, 0x13, 0x08, 0x09, 0x1A, 0x1B);
174
const vector unsigned char mmoP2 = (const vector unsigned char)
175
AVV(0x04, 0x05, 0x16, 0x17, 0x0C, 0x0D, 0x1E, 0x1F,
176
0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f);
177
const vector unsigned char mmoP = (const vector unsigned char)
178
vec_lvsl(8, (unsigned char*)0);
180
vector signed short mmoL1 = vec_perm(v_srcAss0, v_srcAss2, mmoP1);
181
vector signed short mmoL2 = vec_perm(v_srcAss4, v_srcAss6, mmoP2);
182
vector signed short mmoL = vec_perm(mmoL1, mmoL2, mmoP);
183
vector signed short mmoR1 = vec_perm(v_srcAss5, v_srcAss7, mmoP1);
184
vector signed short mmoR2 = vec_perm(v_srcAss1, v_srcAss3, mmoP2);
185
vector signed short mmoR = vec_perm(mmoR1, mmoR2, mmoP);
186
vector signed short mmoDiff = vec_sub(mmoL, mmoR);
187
vector unsigned short mmoSum = (vector unsigned short)vec_add(mmoDiff, v2QP);
189
if (vec_any_gt(mmoSum, v4QP))
197
static inline void doVertLowPass_altivec(uint8_t *src, int stride, PPContext *c) {
199
this code makes no assumption on src or stride.
200
One could remove the recomputation of the perm
201
vector by assuming (stride % 16) == 0, unfortunately
202
this is not always true. Quite a lot of load/stores
203
can be removed by assuming proper alignement of
207
const vector signed int zero = vec_splat_s32(0);
208
const int properStride = (stride % 16);
209
const int srcAlign = ((unsigned long)src2 % 16);
210
DECLARE_ALIGNED(16, short, qp[8]);
212
vector signed short vqp = vec_ld(0, qp);
213
vqp = vec_splat(vqp, 0);
217
vector signed short vb0, vb1, vb2, vb3, vb4, vb5, vb6, vb7, vb8, vb9;
218
vector unsigned char vbA0, vbA1, vbA2, vbA3, vbA4, vbA5, vbA6, vbA7, vbA8, vbA9;
219
vector unsigned char vbB0, vbB1, vbB2, vbB3, vbB4, vbB5, vbB6, vbB7, vbB8, vbB9;
220
vector unsigned char vbT0, vbT1, vbT2, vbT3, vbT4, vbT5, vbT6, vbT7, vbT8, vbT9;
222
#define LOAD_LINE(i) \
223
const vector unsigned char perml##i = \
224
vec_lvsl(i * stride, src2); \
225
vbA##i = vec_ld(i * stride, src2); \
226
vbB##i = vec_ld(i * stride + 16, src2); \
227
vbT##i = vec_perm(vbA##i, vbB##i, perml##i); \
229
(vector signed short)vec_mergeh((vector unsigned char)zero, \
230
(vector unsigned char)vbT##i)
232
#define LOAD_LINE_ALIGNED(i) \
233
register int j##i = i * stride; \
234
vbT##i = vec_ld(j##i, src2); \
236
(vector signed short)vec_mergeh((vector signed char)zero, \
237
(vector signed char)vbT##i)
239
// special casing the aligned case is worthwhile, as all call from
240
// the (transposed) horizontable deblocks will be aligned, in addition
241
// to the naturraly aligned vertical deblocks.
242
if (properStride && srcAlign) {
243
LOAD_LINE_ALIGNED(0);
244
LOAD_LINE_ALIGNED(1);
245
LOAD_LINE_ALIGNED(2);
246
LOAD_LINE_ALIGNED(3);
247
LOAD_LINE_ALIGNED(4);
248
LOAD_LINE_ALIGNED(5);
249
LOAD_LINE_ALIGNED(6);
250
LOAD_LINE_ALIGNED(7);
251
LOAD_LINE_ALIGNED(8);
252
LOAD_LINE_ALIGNED(9);
266
#undef LOAD_LINE_ALIGNED
268
const vector unsigned short v_2 = vec_splat_u16(2);
269
const vector unsigned short v_4 = vec_splat_u16(4);
271
const vector signed short v_diff01 = vec_sub(vb0, vb1);
272
const vector unsigned short v_cmp01 =
273
(const vector unsigned short) vec_cmplt(vec_abs(v_diff01), vqp);
274
const vector signed short v_first = vec_sel(vb1, vb0, v_cmp01);
275
const vector signed short v_diff89 = vec_sub(vb8, vb9);
276
const vector unsigned short v_cmp89 =
277
(const vector unsigned short) vec_cmplt(vec_abs(v_diff89), vqp);
278
const vector signed short v_last = vec_sel(vb8, vb9, v_cmp89);
280
const vector signed short temp01 = vec_mladd(v_first, (vector signed short)v_4, vb1);
281
const vector signed short temp02 = vec_add(vb2, vb3);
282
const vector signed short temp03 = vec_add(temp01, (vector signed short)v_4);
283
const vector signed short v_sumsB0 = vec_add(temp02, temp03);
285
const vector signed short temp11 = vec_sub(v_sumsB0, v_first);
286
const vector signed short v_sumsB1 = vec_add(temp11, vb4);
288
const vector signed short temp21 = vec_sub(v_sumsB1, v_first);
289
const vector signed short v_sumsB2 = vec_add(temp21, vb5);
291
const vector signed short temp31 = vec_sub(v_sumsB2, v_first);
292
const vector signed short v_sumsB3 = vec_add(temp31, vb6);
294
const vector signed short temp41 = vec_sub(v_sumsB3, v_first);
295
const vector signed short v_sumsB4 = vec_add(temp41, vb7);
297
const vector signed short temp51 = vec_sub(v_sumsB4, vb1);
298
const vector signed short v_sumsB5 = vec_add(temp51, vb8);
300
const vector signed short temp61 = vec_sub(v_sumsB5, vb2);
301
const vector signed short v_sumsB6 = vec_add(temp61, v_last);
303
const vector signed short temp71 = vec_sub(v_sumsB6, vb3);
304
const vector signed short v_sumsB7 = vec_add(temp71, v_last);
306
const vector signed short temp81 = vec_sub(v_sumsB7, vb4);
307
const vector signed short v_sumsB8 = vec_add(temp81, v_last);
309
const vector signed short temp91 = vec_sub(v_sumsB8, vb5);
310
const vector signed short v_sumsB9 = vec_add(temp91, v_last);
312
#define COMPUTE_VR(i, j, k) \
313
const vector signed short temps1##i = \
314
vec_add(v_sumsB##i, v_sumsB##k); \
315
const vector signed short temps2##i = \
316
vec_mladd(vb##j, (vector signed short)v_2, temps1##i); \
317
const vector signed short vr##j = vec_sra(temps2##i, v_4)
328
const vector signed char neg1 = vec_splat_s8(-1);
329
const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
330
0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
332
#define PACK_AND_STORE(i) \
333
const vector unsigned char perms##i = \
334
vec_lvsr(i * stride, src2); \
335
const vector unsigned char vf##i = \
336
vec_packsu(vr##i, (vector signed short)zero); \
337
const vector unsigned char vg##i = \
338
vec_perm(vf##i, vbT##i, permHH); \
339
const vector unsigned char mask##i = \
340
vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \
341
const vector unsigned char vg2##i = \
342
vec_perm(vg##i, vg##i, perms##i); \
343
const vector unsigned char svA##i = \
344
vec_sel(vbA##i, vg2##i, mask##i); \
345
const vector unsigned char svB##i = \
346
vec_sel(vg2##i, vbB##i, mask##i); \
347
vec_st(svA##i, i * stride, src2); \
348
vec_st(svB##i, i * stride + 16, src2)
350
#define PACK_AND_STORE_ALIGNED(i) \
351
const vector unsigned char vf##i = \
352
vec_packsu(vr##i, (vector signed short)zero); \
353
const vector unsigned char vg##i = \
354
vec_perm(vf##i, vbT##i, permHH); \
355
vec_st(vg##i, i * stride, src2)
357
// special casing the aligned case is worthwhile, as all call from
358
// the (transposed) horizontable deblocks will be aligned, in addition
359
// to the naturraly aligned vertical deblocks.
360
if (properStride && srcAlign) {
361
PACK_AND_STORE_ALIGNED(1);
362
PACK_AND_STORE_ALIGNED(2);
363
PACK_AND_STORE_ALIGNED(3);
364
PACK_AND_STORE_ALIGNED(4);
365
PACK_AND_STORE_ALIGNED(5);
366
PACK_AND_STORE_ALIGNED(6);
367
PACK_AND_STORE_ALIGNED(7);
368
PACK_AND_STORE_ALIGNED(8);
379
#undef PACK_AND_STORE
380
#undef PACK_AND_STORE_ALIGNED
385
static inline void doVertDefFilter_altivec(uint8_t src[], int stride, PPContext *c) {
387
this code makes no assumption on src or stride.
388
One could remove the recomputation of the perm
389
vector by assuming (stride % 16) == 0, unfortunately
390
this is not always true. Quite a lot of load/stores
391
can be removed by assuming proper alignement of
395
const vector signed int zero = vec_splat_s32(0);
396
DECLARE_ALIGNED(16, short, qp[8]);
398
vector signed short vqp = vec_ld(0, qp);
399
vqp = vec_splat(vqp, 0);
401
#define LOAD_LINE(i) \
402
const vector unsigned char perm##i = \
403
vec_lvsl(i * stride, src2); \
404
const vector unsigned char vbA##i = \
405
vec_ld(i * stride, src2); \
406
const vector unsigned char vbB##i = \
407
vec_ld(i * stride + 16, src2); \
408
const vector unsigned char vbT##i = \
409
vec_perm(vbA##i, vbB##i, perm##i); \
410
const vector signed short vb##i = \
411
(vector signed short)vec_mergeh((vector unsigned char)zero, \
412
(vector unsigned char)vbT##i)
426
const vector signed short v_1 = vec_splat_s16(1);
427
const vector signed short v_2 = vec_splat_s16(2);
428
const vector signed short v_5 = vec_splat_s16(5);
429
const vector signed short v_32 = vec_sl(v_1,
430
(vector unsigned short)v_5);
432
const vector signed short l3minusl6 = vec_sub(vb3, vb6);
433
const vector signed short l5minusl4 = vec_sub(vb5, vb4);
434
const vector signed short twotimes_l3minusl6 = vec_mladd(v_2, l3minusl6, (vector signed short)zero);
435
const vector signed short mE = vec_mladd(v_5, l5minusl4, twotimes_l3minusl6);
436
const vector signed short absmE = vec_abs(mE);
437
/* left & right energy */
438
const vector signed short l1minusl4 = vec_sub(vb1, vb4);
439
const vector signed short l3minusl2 = vec_sub(vb3, vb2);
440
const vector signed short l5minusl8 = vec_sub(vb5, vb8);
441
const vector signed short l7minusl6 = vec_sub(vb7, vb6);
442
const vector signed short twotimes_l1minusl4 = vec_mladd(v_2, l1minusl4, (vector signed short)zero);
443
const vector signed short twotimes_l5minusl8 = vec_mladd(v_2, l5minusl8, (vector signed short)zero);
444
const vector signed short lE = vec_mladd(v_5, l3minusl2, twotimes_l1minusl4);
445
const vector signed short rE = vec_mladd(v_5, l7minusl6, twotimes_l5minusl8);
447
const vector signed short ddiff = vec_sub(absmE,
450
const vector signed short ddiffclamp = vec_max(ddiff, (vector signed short)zero);
451
const vector signed short dtimes64 = vec_mladd(v_5, ddiffclamp, v_32);
452
const vector signed short d = vec_sra(dtimes64, vec_splat_u16(6));
453
const vector signed short minusd = vec_sub((vector signed short)zero, d);
454
const vector signed short finald = vec_sel(minusd,
456
vec_cmpgt(vec_sub((vector signed short)zero, mE),
457
(vector signed short)zero));
459
const vector signed short qtimes2 = vec_sub(vb4, vb5);
460
/* for a shift right to behave like /2, we need to add one
461
to all negative integer */
462
const vector signed short rounddown = vec_sel((vector signed short)zero,
464
vec_cmplt(qtimes2, (vector signed short)zero));
465
const vector signed short q = vec_sra(vec_add(qtimes2, rounddown), vec_splat_u16(1));
467
const vector signed short dclamp_P1 = vec_max((vector signed short)zero, finald);
468
const vector signed short dclamp_P = vec_min(dclamp_P1, q);
469
const vector signed short dclamp_N1 = vec_min((vector signed short)zero, finald);
470
const vector signed short dclamp_N = vec_max(dclamp_N1, q);
472
const vector signed short dclampedfinal = vec_sel(dclamp_N,
474
vec_cmpgt(q, (vector signed short)zero));
475
const vector signed short dornotd = vec_sel((vector signed short)zero,
477
vec_cmplt(absmE, vqp));
478
/* add/substract to l4 and l5 */
479
const vector signed short vb4minusd = vec_sub(vb4, dornotd);
480
const vector signed short vb5plusd = vec_add(vb5, dornotd);
481
/* finally, stores */
482
const vector unsigned char st4 = vec_packsu(vb4minusd, (vector signed short)zero);
483
const vector unsigned char st5 = vec_packsu(vb5plusd, (vector signed short)zero);
485
const vector signed char neg1 = vec_splat_s8(-1);
486
const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
487
0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
490
const vector unsigned char perms##i = \
491
vec_lvsr(i * stride, src2); \
492
const vector unsigned char vg##i = \
493
vec_perm(st##i, vbT##i, permHH); \
494
const vector unsigned char mask##i = \
495
vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \
496
const vector unsigned char vg2##i = \
497
vec_perm(vg##i, vg##i, perms##i); \
498
const vector unsigned char svA##i = \
499
vec_sel(vbA##i, vg2##i, mask##i); \
500
const vector unsigned char svB##i = \
501
vec_sel(vg2##i, vbB##i, mask##i); \
502
vec_st(svA##i, i * stride, src2); \
503
vec_st(svB##i, i * stride + 16, src2)
509
static inline void dering_altivec(uint8_t src[], int stride, PPContext *c) {
511
this code makes no assumption on src or stride.
512
One could remove the recomputation of the perm
513
vector by assuming (stride % 16) == 0, unfortunately
514
this is not always true. Quite a lot of load/stores
515
can be removed by assuming proper alignement of
518
uint8_t *srcCopy = src;
519
DECLARE_ALIGNED(16, uint8_t, dt[16]);
520
const vector signed int zero = vec_splat_s32(0);
521
vector unsigned char v_dt;
522
dt[0] = deringThreshold;
523
v_dt = vec_splat(vec_ld(0, dt), 0);
525
#define LOAD_LINE(i) \
526
const vector unsigned char perm##i = \
527
vec_lvsl(i * stride, srcCopy); \
528
vector unsigned char sA##i = vec_ld(i * stride, srcCopy); \
529
vector unsigned char sB##i = vec_ld(i * stride + 16, srcCopy); \
530
vector unsigned char src##i = vec_perm(sA##i, sB##i, perm##i)
544
vector unsigned char v_avg;
546
const vector unsigned char trunc_perm = (vector unsigned char)
547
AVV(0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08,
548
0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18);
549
const vector unsigned char trunc_src12 = vec_perm(src1, src2, trunc_perm);
550
const vector unsigned char trunc_src34 = vec_perm(src3, src4, trunc_perm);
551
const vector unsigned char trunc_src56 = vec_perm(src5, src6, trunc_perm);
552
const vector unsigned char trunc_src78 = vec_perm(src7, src8, trunc_perm);
554
#define EXTRACT(op) do { \
555
const vector unsigned char s##op##_1 = vec_##op(trunc_src12, trunc_src34); \
556
const vector unsigned char s##op##_2 = vec_##op(trunc_src56, trunc_src78); \
557
const vector unsigned char s##op##_6 = vec_##op(s##op##_1, s##op##_2); \
558
const vector unsigned char s##op##_8h = vec_mergeh(s##op##_6, s##op##_6); \
559
const vector unsigned char s##op##_8l = vec_mergel(s##op##_6, s##op##_6); \
560
const vector unsigned char s##op##_9 = vec_##op(s##op##_8h, s##op##_8l); \
561
const vector unsigned char s##op##_9h = vec_mergeh(s##op##_9, s##op##_9); \
562
const vector unsigned char s##op##_9l = vec_mergel(s##op##_9, s##op##_9); \
563
const vector unsigned char s##op##_10 = vec_##op(s##op##_9h, s##op##_9l); \
564
const vector unsigned char s##op##_10h = vec_mergeh(s##op##_10, s##op##_10); \
565
const vector unsigned char s##op##_10l = vec_mergel(s##op##_10, s##op##_10); \
566
const vector unsigned char s##op##_11 = vec_##op(s##op##_10h, s##op##_10l); \
567
const vector unsigned char s##op##_11h = vec_mergeh(s##op##_11, s##op##_11); \
568
const vector unsigned char s##op##_11l = vec_mergel(s##op##_11, s##op##_11); \
569
v_##op = vec_##op(s##op##_11h, s##op##_11l); } while (0)
571
vector unsigned char v_min;
572
vector unsigned char v_max;
577
if (vec_all_lt(vec_sub(v_max, v_min), v_dt))
580
v_avg = vec_avg(v_min, v_max);
583
DECLARE_ALIGNED(16, signed int, S[8]);
585
const vector unsigned short mask1 = (vector unsigned short)
586
AVV(0x0001, 0x0002, 0x0004, 0x0008,
587
0x0010, 0x0020, 0x0040, 0x0080);
588
const vector unsigned short mask2 = (vector unsigned short)
589
AVV(0x0100, 0x0200, 0x0000, 0x0000,
590
0x0000, 0x0000, 0x0000, 0x0000);
592
const vector unsigned int vuint32_16 = vec_sl(vec_splat_u32(1), vec_splat_u32(4));
593
const vector unsigned int vuint32_1 = vec_splat_u32(1);
596
vector signed int sum##i; \
598
const vector unsigned char cmp##i = \
599
(vector unsigned char)vec_cmpgt(src##i, v_avg); \
600
const vector unsigned short cmpHi##i = \
601
(vector unsigned short)vec_mergeh(cmp##i, cmp##i); \
602
const vector unsigned short cmpLi##i = \
603
(vector unsigned short)vec_mergel(cmp##i, cmp##i); \
604
const vector signed short cmpHf##i = \
605
(vector signed short)vec_and(cmpHi##i, mask1); \
606
const vector signed short cmpLf##i = \
607
(vector signed short)vec_and(cmpLi##i, mask2); \
608
const vector signed int sump##i = vec_sum4s(cmpHf##i, zero); \
609
const vector signed int sumq##i = vec_sum4s(cmpLf##i, sump##i); \
610
sum##i = vec_sums(sumq##i, zero); } while (0)
624
vector signed int sumA2;
625
vector signed int sumB2;
627
const vector signed int sump02 = vec_mergel(sum0, sum2);
628
const vector signed int sump13 = vec_mergel(sum1, sum3);
629
const vector signed int sumA = vec_mergel(sump02, sump13);
631
const vector signed int sump46 = vec_mergel(sum4, sum6);
632
const vector signed int sump57 = vec_mergel(sum5, sum7);
633
const vector signed int sumB = vec_mergel(sump46, sump57);
635
const vector signed int sump8A = vec_mergel(sum8, zero);
636
const vector signed int sump9B = vec_mergel(sum9, zero);
637
const vector signed int sumC = vec_mergel(sump8A, sump9B);
639
const vector signed int tA = vec_sl(vec_nor(zero, sumA), vuint32_16);
640
const vector signed int tB = vec_sl(vec_nor(zero, sumB), vuint32_16);
641
const vector signed int tC = vec_sl(vec_nor(zero, sumC), vuint32_16);
642
const vector signed int t2A = vec_or(sumA, tA);
643
const vector signed int t2B = vec_or(sumB, tB);
644
const vector signed int t2C = vec_or(sumC, tC);
645
const vector signed int t3A = vec_and(vec_sra(t2A, vuint32_1),
646
vec_sl(t2A, vuint32_1));
647
const vector signed int t3B = vec_and(vec_sra(t2B, vuint32_1),
648
vec_sl(t2B, vuint32_1));
649
const vector signed int t3C = vec_and(vec_sra(t2C, vuint32_1),
650
vec_sl(t2C, vuint32_1));
651
const vector signed int yA = vec_and(t2A, t3A);
652
const vector signed int yB = vec_and(t2B, t3B);
653
const vector signed int yC = vec_and(t2C, t3C);
655
const vector unsigned char strangeperm1 = vec_lvsl(4, (unsigned char*)0);
656
const vector unsigned char strangeperm2 = vec_lvsl(8, (unsigned char*)0);
657
const vector signed int sumAd4 = vec_perm(yA, yB, strangeperm1);
658
const vector signed int sumAd8 = vec_perm(yA, yB, strangeperm2);
659
const vector signed int sumBd4 = vec_perm(yB, yC, strangeperm1);
660
const vector signed int sumBd8 = vec_perm(yB, yC, strangeperm2);
661
const vector signed int sumAp = vec_and(yA,
662
vec_and(sumAd4,sumAd8));
663
const vector signed int sumBp = vec_and(yB,
664
vec_and(sumBd4,sumBd8));
665
sumA2 = vec_or(sumAp,
668
sumB2 = vec_or(sumBp,
673
vec_st(sumB2, 16, S);
676
/* I'm not sure the following is actually faster
677
than straight, unvectorized C code :-( */
679
DECLARE_ALIGNED(16, int, tQP2[4]);
680
tQP2[0]= c->QP/2 + 1;
681
vector signed int vQP2 = vec_ld(0, tQP2);
682
vQP2 = vec_splat(vQP2, 0);
683
const vector signed int vsint32_8 = vec_splat_s32(8);
684
const vector unsigned int vuint32_4 = vec_splat_u32(4);
686
const vector unsigned char permA1 = (vector unsigned char)
687
AVV(0x00, 0x01, 0x02, 0x10, 0x11, 0x12, 0x1F, 0x1F,
688
0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F);
689
const vector unsigned char permA2 = (vector unsigned char)
690
AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x10, 0x11,
691
0x12, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F);
692
const vector unsigned char permA1inc = (vector unsigned char)
693
AVV(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00,
694
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
695
const vector unsigned char permA2inc = (vector unsigned char)
696
AVV(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01,
697
0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
698
const vector unsigned char magic = (vector unsigned char)
699
AVV(0x01, 0x02, 0x01, 0x02, 0x04, 0x02, 0x01, 0x02,
700
0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
701
const vector unsigned char extractPerm = (vector unsigned char)
702
AVV(0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01,
703
0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01);
704
const vector unsigned char extractPermInc = (vector unsigned char)
705
AVV(0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01,
706
0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01);
707
const vector unsigned char identity = vec_lvsl(0,(unsigned char *)0);
708
const vector unsigned char tenRight = (vector unsigned char)
709
AVV(0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
710
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
711
const vector unsigned char eightLeft = (vector unsigned char)
712
AVV(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
713
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08);
717
vector unsigned char tenRightM##i = tenRight; \
718
vector unsigned char permA1M##i = permA1; \
719
vector unsigned char permA2M##i = permA2; \
720
vector unsigned char extractPermM##i = extractPerm
722
#define F2(i, j, k, l) \
723
if (S[i] & (1 << (l+1))) { \
724
const vector unsigned char a_##j##_A##l = \
725
vec_perm(src##i, src##j, permA1M##i); \
726
const vector unsigned char a_##j##_B##l = \
727
vec_perm(a_##j##_A##l, src##k, permA2M##i); \
728
const vector signed int a_##j##_sump##l = \
729
(vector signed int)vec_msum(a_##j##_B##l, magic, \
730
(vector unsigned int)zero); \
731
vector signed int F_##j##_##l = \
732
vec_sr(vec_sums(a_##j##_sump##l, vsint32_8), vuint32_4); \
733
F_##j##_##l = vec_splat(F_##j##_##l, 3); \
734
const vector signed int p_##j##_##l = \
735
(vector signed int)vec_perm(src##j, \
736
(vector unsigned char)zero, \
738
const vector signed int sum_##j##_##l = vec_add( p_##j##_##l, vQP2);\
739
const vector signed int diff_##j##_##l = vec_sub( p_##j##_##l, vQP2);\
740
vector signed int newpm_##j##_##l; \
741
if (vec_all_lt(sum_##j##_##l, F_##j##_##l)) \
742
newpm_##j##_##l = sum_##j##_##l; \
743
else if (vec_all_gt(diff_##j##_##l, F_##j##_##l)) \
744
newpm_##j##_##l = diff_##j##_##l; \
745
else newpm_##j##_##l = F_##j##_##l; \
746
const vector unsigned char newpm2_##j##_##l = \
747
vec_splat((vector unsigned char)newpm_##j##_##l, 15); \
748
const vector unsigned char mask##j##l = vec_add(identity, \
750
src##j = vec_perm(src##j, newpm2_##j##_##l, mask##j##l); \
752
permA1M##i = vec_add(permA1M##i, permA1inc); \
753
permA2M##i = vec_add(permA2M##i, permA2inc); \
754
tenRightM##i = vec_sro(tenRightM##i, eightLeft); \
755
extractPermM##i = vec_add(extractPermM##i, extractPermInc)
757
#define ITER(i, j, k) \
777
const vector signed char neg1 = vec_splat_s8(-1);
779
#define STORE_LINE(i) \
780
const vector unsigned char permST##i = \
781
vec_lvsr(i * stride, srcCopy); \
782
const vector unsigned char maskST##i = \
783
vec_perm((vector unsigned char)zero, \
784
(vector unsigned char)neg1, permST##i); \
785
src##i = vec_perm(src##i ,src##i, permST##i); \
786
sA##i= vec_sel(sA##i, src##i, maskST##i); \
787
sB##i= vec_sel(src##i, sB##i, maskST##i); \
788
vec_st(sA##i, i * stride, srcCopy); \
789
vec_st(sB##i, i * stride + 16, srcCopy)
805
#define doHorizLowPass_altivec(a...) doHorizLowPass_C(a)
806
#define doHorizDefFilter_altivec(a...) doHorizDefFilter_C(a)
807
#define do_a_deblock_altivec(a...) do_a_deblock_C(a)
809
static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,
810
uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)
812
const vector signed int zero = vec_splat_s32(0);
813
const vector signed short vsint16_1 = vec_splat_s16(1);
814
vector signed int v_dp = zero;
815
vector signed int v_sysdp = zero;
818
tempBluredPast[127]= maxNoise[0];
819
tempBluredPast[128]= maxNoise[1];
820
tempBluredPast[129]= maxNoise[2];
822
#define LOAD_LINE(src, i) \
823
register int j##src##i = i * stride; \
824
vector unsigned char perm##src##i = vec_lvsl(j##src##i, src); \
825
const vector unsigned char v_##src##A1##i = vec_ld(j##src##i, src); \
826
const vector unsigned char v_##src##A2##i = vec_ld(j##src##i + 16, src); \
827
const vector unsigned char v_##src##A##i = \
828
vec_perm(v_##src##A1##i, v_##src##A2##i, perm##src##i); \
829
vector signed short v_##src##Ass##i = \
830
(vector signed short)vec_mergeh((vector signed char)zero, \
831
(vector signed char)v_##src##A##i)
842
LOAD_LINE(tempBlured, 0);
843
LOAD_LINE(tempBlured, 1);
844
LOAD_LINE(tempBlured, 2);
845
LOAD_LINE(tempBlured, 3);
846
LOAD_LINE(tempBlured, 4);
847
LOAD_LINE(tempBlured, 5);
848
LOAD_LINE(tempBlured, 6);
849
LOAD_LINE(tempBlured, 7);
852
#define ACCUMULATE_DIFFS(i) \
853
vector signed short v_d##i = vec_sub(v_tempBluredAss##i, \
855
v_dp = vec_msums(v_d##i, v_d##i, v_dp); \
856
v_sysdp = vec_msums(v_d##i, vsint16_1, v_sysdp)
866
#undef ACCUMULATE_DIFFS
868
v_dp = vec_sums(v_dp, zero);
869
v_sysdp = vec_sums(v_sysdp, zero);
871
v_dp = vec_splat(v_dp, 3);
872
v_sysdp = vec_splat(v_sysdp, 3);
874
vec_ste(v_dp, 0, &d);
875
vec_ste(v_sysdp, 0, &sysd);
879
+(*(tempBluredPast-256))
880
+(*(tempBluredPast-1))+ (*(tempBluredPast+1))
881
+(*(tempBluredPast+256))
886
if (d > maxNoise[1]) {
887
if (d < maxNoise[2]) {
888
#define OP(i) v_tempBluredAss##i = vec_avg(v_tempBluredAss##i, v_srcAss##i);
900
#define OP(i) v_tempBluredAss##i = v_srcAss##i;
913
if (d < maxNoise[0]) {
914
const vector signed short vsint16_7 = vec_splat_s16(7);
915
const vector signed short vsint16_4 = vec_splat_s16(4);
916
const vector unsigned short vuint16_3 = vec_splat_u16(3);
919
const vector signed short v_temp##i = \
920
vec_mladd(v_tempBluredAss##i, \
921
vsint16_7, v_srcAss##i); \
922
const vector signed short v_temp2##i = \
923
vec_add(v_temp##i, vsint16_4); \
924
v_tempBluredAss##i = vec_sr(v_temp2##i, vuint16_3)
936
const vector signed short vsint16_3 = vec_splat_s16(3);
937
const vector signed short vsint16_2 = vec_splat_s16(2);
940
const vector signed short v_temp##i = \
941
vec_mladd(v_tempBluredAss##i, \
942
vsint16_3, v_srcAss##i); \
943
const vector signed short v_temp2##i = \
944
vec_add(v_temp##i, vsint16_2); \
945
v_tempBluredAss##i = vec_sr(v_temp2##i, (vector unsigned short)vsint16_2)
959
const vector signed char neg1 = vec_splat_s8(-1);
960
const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
961
0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
963
#define PACK_AND_STORE(src, i) \
964
const vector unsigned char perms##src##i = \
965
vec_lvsr(i * stride, src); \
966
const vector unsigned char vf##src##i = \
967
vec_packsu(v_tempBluredAss##i, (vector signed short)zero); \
968
const vector unsigned char vg##src##i = \
969
vec_perm(vf##src##i, v_##src##A##i, permHH); \
970
const vector unsigned char mask##src##i = \
971
vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##src##i); \
972
const vector unsigned char vg2##src##i = \
973
vec_perm(vg##src##i, vg##src##i, perms##src##i); \
974
const vector unsigned char svA##src##i = \
975
vec_sel(v_##src##A1##i, vg2##src##i, mask##src##i); \
976
const vector unsigned char svB##src##i = \
977
vec_sel(vg2##src##i, v_##src##A2##i, mask##src##i); \
978
vec_st(svA##src##i, i * stride, src); \
979
vec_st(svB##src##i, i * stride + 16, src)
981
PACK_AND_STORE(src, 0);
982
PACK_AND_STORE(src, 1);
983
PACK_AND_STORE(src, 2);
984
PACK_AND_STORE(src, 3);
985
PACK_AND_STORE(src, 4);
986
PACK_AND_STORE(src, 5);
987
PACK_AND_STORE(src, 6);
988
PACK_AND_STORE(src, 7);
989
PACK_AND_STORE(tempBlured, 0);
990
PACK_AND_STORE(tempBlured, 1);
991
PACK_AND_STORE(tempBlured, 2);
992
PACK_AND_STORE(tempBlured, 3);
993
PACK_AND_STORE(tempBlured, 4);
994
PACK_AND_STORE(tempBlured, 5);
995
PACK_AND_STORE(tempBlured, 6);
996
PACK_AND_STORE(tempBlured, 7);
997
#undef PACK_AND_STORE
1000
static inline void transpose_16x8_char_toPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1001
const vector unsigned char zero = vec_splat_u8(0);
1003
#define LOAD_DOUBLE_LINE(i, j) \
1004
vector unsigned char perm1##i = vec_lvsl(i * stride, src); \
1005
vector unsigned char perm2##i = vec_lvsl(j * stride, src); \
1006
vector unsigned char srcA##i = vec_ld(i * stride, src); \
1007
vector unsigned char srcB##i = vec_ld(i * stride + 16, src); \
1008
vector unsigned char srcC##i = vec_ld(j * stride, src); \
1009
vector unsigned char srcD##i = vec_ld(j * stride+ 16, src); \
1010
vector unsigned char src##i = vec_perm(srcA##i, srcB##i, perm1##i); \
1011
vector unsigned char src##j = vec_perm(srcC##i, srcD##i, perm2##i)
1013
LOAD_DOUBLE_LINE(0, 1);
1014
LOAD_DOUBLE_LINE(2, 3);
1015
LOAD_DOUBLE_LINE(4, 5);
1016
LOAD_DOUBLE_LINE(6, 7);
1017
#undef LOAD_DOUBLE_LINE
1019
vector unsigned char tempA = vec_mergeh(src0, zero);
1020
vector unsigned char tempB = vec_mergel(src0, zero);
1021
vector unsigned char tempC = vec_mergeh(src1, zero);
1022
vector unsigned char tempD = vec_mergel(src1, zero);
1023
vector unsigned char tempE = vec_mergeh(src2, zero);
1024
vector unsigned char tempF = vec_mergel(src2, zero);
1025
vector unsigned char tempG = vec_mergeh(src3, zero);
1026
vector unsigned char tempH = vec_mergel(src3, zero);
1027
vector unsigned char tempI = vec_mergeh(src4, zero);
1028
vector unsigned char tempJ = vec_mergel(src4, zero);
1029
vector unsigned char tempK = vec_mergeh(src5, zero);
1030
vector unsigned char tempL = vec_mergel(src5, zero);
1031
vector unsigned char tempM = vec_mergeh(src6, zero);
1032
vector unsigned char tempN = vec_mergel(src6, zero);
1033
vector unsigned char tempO = vec_mergeh(src7, zero);
1034
vector unsigned char tempP = vec_mergel(src7, zero);
1036
vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1037
vector unsigned char temp1 = vec_mergel(tempA, tempI);
1038
vector unsigned char temp2 = vec_mergeh(tempB, tempJ);
1039
vector unsigned char temp3 = vec_mergel(tempB, tempJ);
1040
vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1041
vector unsigned char temp5 = vec_mergel(tempC, tempK);
1042
vector unsigned char temp6 = vec_mergeh(tempD, tempL);
1043
vector unsigned char temp7 = vec_mergel(tempD, tempL);
1044
vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1045
vector unsigned char temp9 = vec_mergel(tempE, tempM);
1046
vector unsigned char temp10 = vec_mergeh(tempF, tempN);
1047
vector unsigned char temp11 = vec_mergel(tempF, tempN);
1048
vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1049
vector unsigned char temp13 = vec_mergel(tempG, tempO);
1050
vector unsigned char temp14 = vec_mergeh(tempH, tempP);
1051
vector unsigned char temp15 = vec_mergel(tempH, tempP);
1053
tempA = vec_mergeh(temp0, temp8);
1054
tempB = vec_mergel(temp0, temp8);
1055
tempC = vec_mergeh(temp1, temp9);
1056
tempD = vec_mergel(temp1, temp9);
1057
tempE = vec_mergeh(temp2, temp10);
1058
tempF = vec_mergel(temp2, temp10);
1059
tempG = vec_mergeh(temp3, temp11);
1060
tempH = vec_mergel(temp3, temp11);
1061
tempI = vec_mergeh(temp4, temp12);
1062
tempJ = vec_mergel(temp4, temp12);
1063
tempK = vec_mergeh(temp5, temp13);
1064
tempL = vec_mergel(temp5, temp13);
1065
tempM = vec_mergeh(temp6, temp14);
1066
tempN = vec_mergel(temp6, temp14);
1067
tempO = vec_mergeh(temp7, temp15);
1068
tempP = vec_mergel(temp7, temp15);
1070
temp0 = vec_mergeh(tempA, tempI);
1071
temp1 = vec_mergel(tempA, tempI);
1072
temp2 = vec_mergeh(tempB, tempJ);
1073
temp3 = vec_mergel(tempB, tempJ);
1074
temp4 = vec_mergeh(tempC, tempK);
1075
temp5 = vec_mergel(tempC, tempK);
1076
temp6 = vec_mergeh(tempD, tempL);
1077
temp7 = vec_mergel(tempD, tempL);
1078
temp8 = vec_mergeh(tempE, tempM);
1079
temp9 = vec_mergel(tempE, tempM);
1080
temp10 = vec_mergeh(tempF, tempN);
1081
temp11 = vec_mergel(tempF, tempN);
1082
temp12 = vec_mergeh(tempG, tempO);
1083
temp13 = vec_mergel(tempG, tempO);
1084
temp14 = vec_mergeh(tempH, tempP);
1085
temp15 = vec_mergel(tempH, tempP);
1087
vec_st(temp0, 0, dst);
1088
vec_st(temp1, 16, dst);
1089
vec_st(temp2, 32, dst);
1090
vec_st(temp3, 48, dst);
1091
vec_st(temp4, 64, dst);
1092
vec_st(temp5, 80, dst);
1093
vec_st(temp6, 96, dst);
1094
vec_st(temp7, 112, dst);
1095
vec_st(temp8, 128, dst);
1096
vec_st(temp9, 144, dst);
1097
vec_st(temp10, 160, dst);
1098
vec_st(temp11, 176, dst);
1099
vec_st(temp12, 192, dst);
1100
vec_st(temp13, 208, dst);
1101
vec_st(temp14, 224, dst);
1102
vec_st(temp15, 240, dst);
1105
static inline void transpose_8x16_char_fromPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1106
const vector unsigned char zero = vec_splat_u8(0);
1108
#define LOAD_DOUBLE_LINE(i, j) \
1109
vector unsigned char src##i = vec_ld(i * 16, src); \
1110
vector unsigned char src##j = vec_ld(j * 16, src)
1112
LOAD_DOUBLE_LINE(0, 1);
1113
LOAD_DOUBLE_LINE(2, 3);
1114
LOAD_DOUBLE_LINE(4, 5);
1115
LOAD_DOUBLE_LINE(6, 7);
1116
LOAD_DOUBLE_LINE(8, 9);
1117
LOAD_DOUBLE_LINE(10, 11);
1118
LOAD_DOUBLE_LINE(12, 13);
1119
LOAD_DOUBLE_LINE(14, 15);
1120
#undef LOAD_DOUBLE_LINE
1122
vector unsigned char tempA = vec_mergeh(src0, src8);
1123
vector unsigned char tempB;
1124
vector unsigned char tempC = vec_mergeh(src1, src9);
1125
vector unsigned char tempD;
1126
vector unsigned char tempE = vec_mergeh(src2, src10);
1127
vector unsigned char tempG = vec_mergeh(src3, src11);
1128
vector unsigned char tempI = vec_mergeh(src4, src12);
1129
vector unsigned char tempJ;
1130
vector unsigned char tempK = vec_mergeh(src5, src13);
1131
vector unsigned char tempL;
1132
vector unsigned char tempM = vec_mergeh(src6, src14);
1133
vector unsigned char tempO = vec_mergeh(src7, src15);
1135
vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1136
vector unsigned char temp1 = vec_mergel(tempA, tempI);
1137
vector unsigned char temp2;
1138
vector unsigned char temp3;
1139
vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1140
vector unsigned char temp5 = vec_mergel(tempC, tempK);
1141
vector unsigned char temp6;
1142
vector unsigned char temp7;
1143
vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1144
vector unsigned char temp9 = vec_mergel(tempE, tempM);
1145
vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1146
vector unsigned char temp13 = vec_mergel(tempG, tempO);
1148
tempA = vec_mergeh(temp0, temp8);
1149
tempB = vec_mergel(temp0, temp8);
1150
tempC = vec_mergeh(temp1, temp9);
1151
tempD = vec_mergel(temp1, temp9);
1152
tempI = vec_mergeh(temp4, temp12);
1153
tempJ = vec_mergel(temp4, temp12);
1154
tempK = vec_mergeh(temp5, temp13);
1155
tempL = vec_mergel(temp5, temp13);
1157
temp0 = vec_mergeh(tempA, tempI);
1158
temp1 = vec_mergel(tempA, tempI);
1159
temp2 = vec_mergeh(tempB, tempJ);
1160
temp3 = vec_mergel(tempB, tempJ);
1161
temp4 = vec_mergeh(tempC, tempK);
1162
temp5 = vec_mergel(tempC, tempK);
1163
temp6 = vec_mergeh(tempD, tempL);
1164
temp7 = vec_mergel(tempD, tempL);
1167
const vector signed char neg1 = vec_splat_s8(-1);
1168
#define STORE_DOUBLE_LINE(i, j) \
1169
vector unsigned char dstA##i = vec_ld(i * stride, dst); \
1170
vector unsigned char dstB##i = vec_ld(i * stride + 16, dst); \
1171
vector unsigned char dstA##j = vec_ld(j * stride, dst); \
1172
vector unsigned char dstB##j = vec_ld(j * stride+ 16, dst); \
1173
vector unsigned char align##i = vec_lvsr(i * stride, dst); \
1174
vector unsigned char align##j = vec_lvsr(j * stride, dst); \
1175
vector unsigned char mask##i = vec_perm(zero, (vector unsigned char)neg1, align##i); \
1176
vector unsigned char mask##j = vec_perm(zero, (vector unsigned char)neg1, align##j); \
1177
vector unsigned char dstR##i = vec_perm(temp##i, temp##i, align##i); \
1178
vector unsigned char dstR##j = vec_perm(temp##j, temp##j, align##j); \
1179
vector unsigned char dstAF##i = vec_sel(dstA##i, dstR##i, mask##i); \
1180
vector unsigned char dstBF##i = vec_sel(dstR##i, dstB##i, mask##i); \
1181
vector unsigned char dstAF##j = vec_sel(dstA##j, dstR##j, mask##j); \
1182
vector unsigned char dstBF##j = vec_sel(dstR##j, dstB##j, mask##j); \
1183
vec_st(dstAF##i, i * stride, dst); \
1184
vec_st(dstBF##i, i * stride + 16, dst); \
1185
vec_st(dstAF##j, j * stride, dst); \
1186
vec_st(dstBF##j, j * stride + 16, dst)
1188
STORE_DOUBLE_LINE(0,1);
1189
STORE_DOUBLE_LINE(2,3);
1190
STORE_DOUBLE_LINE(4,5);
1191
STORE_DOUBLE_LINE(6,7);