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
30 #define ALTIVEC_TRANSPOSE_8x8_SHORT(src_a,src_b,src_c,src_d,src_e,src_f,src_g,src_h) \
32 __typeof__(src_a) tempA1, tempB1, tempC1, tempD1; \
33 __typeof__(src_a) tempE1, tempF1, tempG1, tempH1; \
34 __typeof__(src_a) tempA2, tempB2, tempC2, tempD2; \
35 __typeof__(src_a) tempE2, tempF2, tempG2, tempH2; \
36 tempA1 = vec_mergeh (src_a, src_e); \
37 tempB1 = vec_mergel (src_a, src_e); \
38 tempC1 = vec_mergeh (src_b, src_f); \
39 tempD1 = vec_mergel (src_b, src_f); \
40 tempE1 = vec_mergeh (src_c, src_g); \
41 tempF1 = vec_mergel (src_c, src_g); \
42 tempG1 = vec_mergeh (src_d, src_h); \
43 tempH1 = vec_mergel (src_d, src_h); \
44 tempA2 = vec_mergeh (tempA1, tempE1); \
45 tempB2 = vec_mergel (tempA1, tempE1); \
46 tempC2 = vec_mergeh (tempB1, tempF1); \
47 tempD2 = vec_mergel (tempB1, tempF1); \
48 tempE2 = vec_mergeh (tempC1, tempG1); \
49 tempF2 = vec_mergel (tempC1, tempG1); \
50 tempG2 = vec_mergeh (tempD1, tempH1); \
51 tempH2 = vec_mergel (tempD1, tempH1); \
52 src_a = vec_mergeh (tempA2, tempE2); \
53 src_b = vec_mergel (tempA2, tempE2); \
54 src_c = vec_mergeh (tempB2, tempF2); \
55 src_d = vec_mergel (tempB2, tempF2); \
56 src_e = vec_mergeh (tempC2, tempG2); \
57 src_f = vec_mergel (tempC2, tempG2); \
58 src_g = vec_mergeh (tempD2, tempH2); \
59 src_h = vec_mergel (tempD2, tempH2); \
63 static inline int vertClassify_altivec(uint8_t src[], int stride, PPContext *c) {
65 this code makes no assumption on src or stride.
66 One could remove the recomputation of the perm
67 vector by assuming (stride % 16) == 0, unfortunately
68 this is not always true.
70 short __attribute__ ((aligned(16))) data[8];
73 vector signed short v_dcOffset;
74 vector signed short v2QP;
75 vector unsigned short v4QP;
76 vector unsigned short v_dcThreshold;
77 const int properStride = (stride % 16);
78 const int srcAlign = ((unsigned long)src2 % 16);
79 const int two_vectors = ((srcAlign > 8) || properStride) ? 1 : 0;
80 const vector signed int zero = vec_splat_s32(0);
81 const vector signed short mask = vec_splat_s16(1);
82 vector signed int v_numEq = vec_splat_s32(0);
84 data[0] = ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
85 data[1] = data[0] * 2 + 1;
88 vector signed short v_data = vec_ld(0, data);
89 v_dcOffset = vec_splat(v_data, 0);
90 v_dcThreshold = (vector unsigned short)vec_splat(v_data, 1);
91 v2QP = vec_splat(v_data, 2);
92 v4QP = (vector unsigned short)vec_splat(v_data, 3);
96 vector signed short v_srcAss0, v_srcAss1, v_srcAss2, v_srcAss3, v_srcAss4, v_srcAss5, v_srcAss6, v_srcAss7;
98 #define LOAD_LINE(i) \
99 register int j##i = i * stride; \
100 vector unsigned char perm##i = vec_lvsl(j##i, src2); \
101 const vector unsigned char v_srcA1##i = vec_ld(j##i, src2); \
102 vector unsigned char v_srcA2##i; \
104 v_srcA2##i = vec_ld(j##i + 16, src2); \
105 const vector unsigned char v_srcA##i = \
106 vec_perm(v_srcA1##i, v_srcA2##i, perm##i); \
108 (vector signed short)vec_mergeh((vector signed char)zero, \
109 (vector signed char)v_srcA##i)
111 #define LOAD_LINE_ALIGNED(i) \
112 register int j##i = i * stride; \
113 const vector unsigned char v_srcA##i = vec_ld(j##i, src2); \
115 (vector signed short)vec_mergeh((vector signed char)zero, \
116 (vector signed char)v_srcA##i)
118 // special casing the aligned case is worthwhile, as all call from
119 // the (transposed) horizontable deblocks will be aligned, i naddition
120 // to the naturraly aligned vertical deblocks.
121 if (properStride && srcAlign) {
122 LOAD_LINE_ALIGNED(0);
123 LOAD_LINE_ALIGNED(1);
124 LOAD_LINE_ALIGNED(2);
125 LOAD_LINE_ALIGNED(3);
126 LOAD_LINE_ALIGNED(4);
127 LOAD_LINE_ALIGNED(5);
128 LOAD_LINE_ALIGNED(6);
129 LOAD_LINE_ALIGNED(7);
141 #undef LOAD_LINE_ALIGNED
144 const vector signed short v_diff##i = \
145 vec_sub(v_srcAss##i, v_srcAss##j); \
146 const vector signed short v_sum##i = \
147 vec_add(v_diff##i, v_dcOffset); \
148 const vector signed short v_comp##i = \
149 (vector signed short)vec_cmplt((vector unsigned short)v_sum##i, \
151 const vector signed short v_part##i = vec_and(mask, v_comp##i); \
152 v_numEq = vec_sum4s(v_part##i, v_numEq);
163 v_numEq = vec_sums(v_numEq, zero);
165 v_numEq = vec_splat(v_numEq, 3);
166 vec_ste(v_numEq, 0, &numEq);
168 if (numEq > c->ppMode.flatnessThreshold)
170 const vector unsigned char mmoP1 = (const vector unsigned char)
171 AVV(0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f,
172 0x00, 0x01, 0x12, 0x13, 0x08, 0x09, 0x1A, 0x1B);
173 const vector unsigned char mmoP2 = (const vector unsigned char)
174 AVV(0x04, 0x05, 0x16, 0x17, 0x0C, 0x0D, 0x1E, 0x1F,
175 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f);
176 const vector unsigned char mmoP = (const vector unsigned char)
177 vec_lvsl(8, (unsigned char*)0);
179 vector signed short mmoL1 = vec_perm(v_srcAss0, v_srcAss2, mmoP1);
180 vector signed short mmoL2 = vec_perm(v_srcAss4, v_srcAss6, mmoP2);
181 vector signed short mmoL = vec_perm(mmoL1, mmoL2, mmoP);
182 vector signed short mmoR1 = vec_perm(v_srcAss5, v_srcAss7, mmoP1);
183 vector signed short mmoR2 = vec_perm(v_srcAss1, v_srcAss3, mmoP2);
184 vector signed short mmoR = vec_perm(mmoR1, mmoR2, mmoP);
185 vector signed short mmoDiff = vec_sub(mmoL, mmoR);
186 vector unsigned short mmoSum = (vector unsigned short)vec_add(mmoDiff, v2QP);
188 if (vec_any_gt(mmoSum, v4QP))
196 static inline void doVertLowPass_altivec(uint8_t *src, int stride, PPContext *c) {
198 this code makes no assumption on src or stride.
199 One could remove the recomputation of the perm
200 vector by assuming (stride % 16) == 0, unfortunately
201 this is not always true. Quite a lot of load/stores
202 can be removed by assuming proper alignement of
206 const vector signed int zero = vec_splat_s32(0);
207 const int properStride = (stride % 16);
208 const int srcAlign = ((unsigned long)src2 % 16);
209 short __attribute__ ((aligned(16))) qp[8];
211 vector signed short vqp = vec_ld(0, qp);
212 vqp = vec_splat(vqp, 0);
216 vector signed short vb0, vb1, vb2, vb3, vb4, vb5, vb6, vb7, vb8, vb9;
217 vector unsigned char vbA0, vbA1, vbA2, vbA3, vbA4, vbA5, vbA6, vbA7, vbA8, vbA9;
218 vector unsigned char vbB0, vbB1, vbB2, vbB3, vbB4, vbB5, vbB6, vbB7, vbB8, vbB9;
219 vector unsigned char vbT0, vbT1, vbT2, vbT3, vbT4, vbT5, vbT6, vbT7, vbT8, vbT9;
221 #define LOAD_LINE(i) \
222 const vector unsigned char perml##i = \
223 vec_lvsl(i * stride, src2); \
224 vbA##i = vec_ld(i * stride, src2); \
225 vbB##i = vec_ld(i * stride + 16, src2); \
226 vbT##i = vec_perm(vbA##i, vbB##i, perml##i); \
228 (vector signed short)vec_mergeh((vector unsigned char)zero, \
229 (vector unsigned char)vbT##i)
231 #define LOAD_LINE_ALIGNED(i) \
232 register int j##i = i * stride; \
233 vbT##i = vec_ld(j##i, src2); \
235 (vector signed short)vec_mergeh((vector signed char)zero, \
236 (vector signed char)vbT##i)
238 // special casing the aligned case is worthwhile, as all call from
239 // the (transposed) horizontable deblocks will be aligned, in addition
240 // to the naturraly aligned vertical deblocks.
241 if (properStride && srcAlign) {
242 LOAD_LINE_ALIGNED(0);
243 LOAD_LINE_ALIGNED(1);
244 LOAD_LINE_ALIGNED(2);
245 LOAD_LINE_ALIGNED(3);
246 LOAD_LINE_ALIGNED(4);
247 LOAD_LINE_ALIGNED(5);
248 LOAD_LINE_ALIGNED(6);
249 LOAD_LINE_ALIGNED(7);
250 LOAD_LINE_ALIGNED(8);
251 LOAD_LINE_ALIGNED(9);
265 #undef LOAD_LINE_ALIGNED
267 const vector unsigned short v_2 = vec_splat_u16(2);
268 const vector unsigned short v_4 = vec_splat_u16(4);
270 const vector signed short v_diff01 = vec_sub(vb0, vb1);
271 const vector unsigned short v_cmp01 =
272 (const vector unsigned short) vec_cmplt(vec_abs(v_diff01), vqp);
273 const vector signed short v_first = vec_sel(vb1, vb0, v_cmp01);
274 const vector signed short v_diff89 = vec_sub(vb8, vb9);
275 const vector unsigned short v_cmp89 =
276 (const vector unsigned short) vec_cmplt(vec_abs(v_diff89), vqp);
277 const vector signed short v_last = vec_sel(vb8, vb9, v_cmp89);
279 const vector signed short temp01 = vec_mladd(v_first, (vector signed short)v_4, vb1);
280 const vector signed short temp02 = vec_add(vb2, vb3);
281 const vector signed short temp03 = vec_add(temp01, (vector signed short)v_4);
282 const vector signed short v_sumsB0 = vec_add(temp02, temp03);
284 const vector signed short temp11 = vec_sub(v_sumsB0, v_first);
285 const vector signed short v_sumsB1 = vec_add(temp11, vb4);
287 const vector signed short temp21 = vec_sub(v_sumsB1, v_first);
288 const vector signed short v_sumsB2 = vec_add(temp21, vb5);
290 const vector signed short temp31 = vec_sub(v_sumsB2, v_first);
291 const vector signed short v_sumsB3 = vec_add(temp31, vb6);
293 const vector signed short temp41 = vec_sub(v_sumsB3, v_first);
294 const vector signed short v_sumsB4 = vec_add(temp41, vb7);
296 const vector signed short temp51 = vec_sub(v_sumsB4, vb1);
297 const vector signed short v_sumsB5 = vec_add(temp51, vb8);
299 const vector signed short temp61 = vec_sub(v_sumsB5, vb2);
300 const vector signed short v_sumsB6 = vec_add(temp61, v_last);
302 const vector signed short temp71 = vec_sub(v_sumsB6, vb3);
303 const vector signed short v_sumsB7 = vec_add(temp71, v_last);
305 const vector signed short temp81 = vec_sub(v_sumsB7, vb4);
306 const vector signed short v_sumsB8 = vec_add(temp81, v_last);
308 const vector signed short temp91 = vec_sub(v_sumsB8, vb5);
309 const vector signed short v_sumsB9 = vec_add(temp91, v_last);
311 #define COMPUTE_VR(i, j, k) \
312 const vector signed short temps1##i = \
313 vec_add(v_sumsB##i, v_sumsB##k); \
314 const vector signed short temps2##i = \
315 vec_mladd(vb##j, (vector signed short)v_2, temps1##i); \
316 const vector signed short vr##j = vec_sra(temps2##i, v_4)
327 const vector signed char neg1 = vec_splat_s8(-1);
328 const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
329 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
331 #define PACK_AND_STORE(i) \
332 const vector unsigned char perms##i = \
333 vec_lvsr(i * stride, src2); \
334 const vector unsigned char vf##i = \
335 vec_packsu(vr##i, (vector signed short)zero); \
336 const vector unsigned char vg##i = \
337 vec_perm(vf##i, vbT##i, permHH); \
338 const vector unsigned char mask##i = \
339 vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \
340 const vector unsigned char vg2##i = \
341 vec_perm(vg##i, vg##i, perms##i); \
342 const vector unsigned char svA##i = \
343 vec_sel(vbA##i, vg2##i, mask##i); \
344 const vector unsigned char svB##i = \
345 vec_sel(vg2##i, vbB##i, mask##i); \
346 vec_st(svA##i, i * stride, src2); \
347 vec_st(svB##i, i * stride + 16, src2)
349 #define PACK_AND_STORE_ALIGNED(i) \
350 const vector unsigned char vf##i = \
351 vec_packsu(vr##i, (vector signed short)zero); \
352 const vector unsigned char vg##i = \
353 vec_perm(vf##i, vbT##i, permHH); \
354 vec_st(vg##i, i * stride, src2)
356 // special casing the aligned case is worthwhile, as all call from
357 // the (transposed) horizontable deblocks will be aligned, in addition
358 // to the naturraly aligned vertical deblocks.
359 if (properStride && srcAlign) {
360 PACK_AND_STORE_ALIGNED(1);
361 PACK_AND_STORE_ALIGNED(2);
362 PACK_AND_STORE_ALIGNED(3);
363 PACK_AND_STORE_ALIGNED(4);
364 PACK_AND_STORE_ALIGNED(5);
365 PACK_AND_STORE_ALIGNED(6);
366 PACK_AND_STORE_ALIGNED(7);
367 PACK_AND_STORE_ALIGNED(8);
378 #undef PACK_AND_STORE
379 #undef PACK_AND_STORE_ALIGNED
384 static inline void doVertDefFilter_altivec(uint8_t src[], int stride, PPContext *c) {
386 this code makes no assumption on src or stride.
387 One could remove the recomputation of the perm
388 vector by assuming (stride % 16) == 0, unfortunately
389 this is not always true. Quite a lot of load/stores
390 can be removed by assuming proper alignement of
394 const vector signed int zero = vec_splat_s32(0);
395 short __attribute__ ((aligned(16))) qp[8];
397 vector signed short vqp = vec_ld(0, qp);
398 vqp = vec_splat(vqp, 0);
400 #define LOAD_LINE(i) \
401 const vector unsigned char perm##i = \
402 vec_lvsl(i * stride, src2); \
403 const vector unsigned char vbA##i = \
404 vec_ld(i * stride, src2); \
405 const vector unsigned char vbB##i = \
406 vec_ld(i * stride + 16, src2); \
407 const vector unsigned char vbT##i = \
408 vec_perm(vbA##i, vbB##i, perm##i); \
409 const vector signed short vb##i = \
410 (vector signed short)vec_mergeh((vector unsigned char)zero, \
411 (vector unsigned char)vbT##i)
425 const vector signed short v_1 = vec_splat_s16(1);
426 const vector signed short v_2 = vec_splat_s16(2);
427 const vector signed short v_5 = vec_splat_s16(5);
428 const vector signed short v_32 = vec_sl(v_1,
429 (vector unsigned short)v_5);
431 const vector signed short l3minusl6 = vec_sub(vb3, vb6);
432 const vector signed short l5minusl4 = vec_sub(vb5, vb4);
433 const vector signed short twotimes_l3minusl6 = vec_mladd(v_2, l3minusl6, (vector signed short)zero);
434 const vector signed short mE = vec_mladd(v_5, l5minusl4, twotimes_l3minusl6);
435 const vector signed short absmE = vec_abs(mE);
436 /* left & right energy */
437 const vector signed short l1minusl4 = vec_sub(vb1, vb4);
438 const vector signed short l3minusl2 = vec_sub(vb3, vb2);
439 const vector signed short l5minusl8 = vec_sub(vb5, vb8);
440 const vector signed short l7minusl6 = vec_sub(vb7, vb6);
441 const vector signed short twotimes_l1minusl4 = vec_mladd(v_2, l1minusl4, (vector signed short)zero);
442 const vector signed short twotimes_l5minusl8 = vec_mladd(v_2, l5minusl8, (vector signed short)zero);
443 const vector signed short lE = vec_mladd(v_5, l3minusl2, twotimes_l1minusl4);
444 const vector signed short rE = vec_mladd(v_5, l7minusl6, twotimes_l5minusl8);
446 const vector signed short ddiff = vec_sub(absmE,
449 const vector signed short ddiffclamp = vec_max(ddiff, (vector signed short)zero);
450 const vector signed short dtimes64 = vec_mladd(v_5, ddiffclamp, v_32);
451 const vector signed short d = vec_sra(dtimes64, vec_splat_u16(6));
452 const vector signed short minusd = vec_sub((vector signed short)zero, d);
453 const vector signed short finald = vec_sel(minusd,
455 vec_cmpgt(vec_sub((vector signed short)zero, mE),
456 (vector signed short)zero));
458 const vector signed short qtimes2 = vec_sub(vb4, vb5);
459 /* for a shift right to behave like /2, we need to add one
460 to all negative integer */
461 const vector signed short rounddown = vec_sel((vector signed short)zero,
463 vec_cmplt(qtimes2, (vector signed short)zero));
464 const vector signed short q = vec_sra(vec_add(qtimes2, rounddown), vec_splat_u16(1));
466 const vector signed short dclamp_P1 = vec_max((vector signed short)zero, finald);
467 const vector signed short dclamp_P = vec_min(dclamp_P1, q);
468 const vector signed short dclamp_N1 = vec_min((vector signed short)zero, finald);
469 const vector signed short dclamp_N = vec_max(dclamp_N1, q);
471 const vector signed short dclampedfinal = vec_sel(dclamp_N,
473 vec_cmpgt(q, (vector signed short)zero));
474 const vector signed short dornotd = vec_sel((vector signed short)zero,
476 vec_cmplt(absmE, vqp));
477 /* add/substract to l4 and l5 */
478 const vector signed short vb4minusd = vec_sub(vb4, dornotd);
479 const vector signed short vb5plusd = vec_add(vb5, dornotd);
480 /* finally, stores */
481 const vector unsigned char st4 = vec_packsu(vb4minusd, (vector signed short)zero);
482 const vector unsigned char st5 = vec_packsu(vb5plusd, (vector signed short)zero);
484 const vector signed char neg1 = vec_splat_s8(-1);
485 const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
486 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
489 const vector unsigned char perms##i = \
490 vec_lvsr(i * stride, src2); \
491 const vector unsigned char vg##i = \
492 vec_perm(st##i, vbT##i, permHH); \
493 const vector unsigned char mask##i = \
494 vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \
495 const vector unsigned char vg2##i = \
496 vec_perm(vg##i, vg##i, perms##i); \
497 const vector unsigned char svA##i = \
498 vec_sel(vbA##i, vg2##i, mask##i); \
499 const vector unsigned char svB##i = \
500 vec_sel(vg2##i, vbB##i, mask##i); \
501 vec_st(svA##i, i * stride, src2); \
502 vec_st(svB##i, i * stride + 16, src2)
508 static inline void dering_altivec(uint8_t src[], int stride, PPContext *c) {
510 this code makes no assumption on src or stride.
511 One could remove the recomputation of the perm
512 vector by assuming (stride % 16) == 0, unfortunately
513 this is not always true. Quite a lot of load/stores
514 can be removed by assuming proper alignement of
517 uint8_t *srcCopy = src;
518 uint8_t __attribute__((aligned(16))) dt[16];
519 const vector signed int zero = vec_splat_s32(0);
520 vector unsigned char v_dt;
521 dt[0] = deringThreshold;
522 v_dt = vec_splat(vec_ld(0, dt), 0);
524 #define LOAD_LINE(i) \
525 const vector unsigned char perm##i = \
526 vec_lvsl(i * stride, srcCopy); \
527 vector unsigned char sA##i = vec_ld(i * stride, srcCopy); \
528 vector unsigned char sB##i = vec_ld(i * stride + 16, srcCopy); \
529 vector unsigned char src##i = vec_perm(sA##i, sB##i, perm##i)
543 vector unsigned char v_avg;
545 const vector unsigned char trunc_perm = (vector unsigned char)
546 AVV(0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08,
547 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18);
548 const vector unsigned char trunc_src12 = vec_perm(src1, src2, trunc_perm);
549 const vector unsigned char trunc_src34 = vec_perm(src3, src4, trunc_perm);
550 const vector unsigned char trunc_src56 = vec_perm(src5, src6, trunc_perm);
551 const vector unsigned char trunc_src78 = vec_perm(src7, src8, trunc_perm);
553 #define EXTRACT(op) do { \
554 const vector unsigned char s##op##_1 = vec_##op(trunc_src12, trunc_src34); \
555 const vector unsigned char s##op##_2 = vec_##op(trunc_src56, trunc_src78); \
556 const vector unsigned char s##op##_6 = vec_##op(s##op##_1, s##op##_2); \
557 const vector unsigned char s##op##_8h = vec_mergeh(s##op##_6, s##op##_6); \
558 const vector unsigned char s##op##_8l = vec_mergel(s##op##_6, s##op##_6); \
559 const vector unsigned char s##op##_9 = vec_##op(s##op##_8h, s##op##_8l); \
560 const vector unsigned char s##op##_9h = vec_mergeh(s##op##_9, s##op##_9); \
561 const vector unsigned char s##op##_9l = vec_mergel(s##op##_9, s##op##_9); \
562 const vector unsigned char s##op##_10 = vec_##op(s##op##_9h, s##op##_9l); \
563 const vector unsigned char s##op##_10h = vec_mergeh(s##op##_10, s##op##_10); \
564 const vector unsigned char s##op##_10l = vec_mergel(s##op##_10, s##op##_10); \
565 const vector unsigned char s##op##_11 = vec_##op(s##op##_10h, s##op##_10l); \
566 const vector unsigned char s##op##_11h = vec_mergeh(s##op##_11, s##op##_11); \
567 const vector unsigned char s##op##_11l = vec_mergel(s##op##_11, s##op##_11); \
568 v_##op = vec_##op(s##op##_11h, s##op##_11l); } while (0)
570 vector unsigned char v_min;
571 vector unsigned char v_max;
576 if (vec_all_lt(vec_sub(v_max, v_min), v_dt))
579 v_avg = vec_avg(v_min, v_max);
582 signed int __attribute__((aligned(16))) S[8];
584 const vector unsigned short mask1 = (vector unsigned short)
585 AVV(0x0001, 0x0002, 0x0004, 0x0008,
586 0x0010, 0x0020, 0x0040, 0x0080);
587 const vector unsigned short mask2 = (vector unsigned short)
588 AVV(0x0100, 0x0200, 0x0000, 0x0000,
589 0x0000, 0x0000, 0x0000, 0x0000);
591 const vector unsigned int vuint32_16 = vec_sl(vec_splat_u32(1), vec_splat_u32(4));
592 const vector unsigned int vuint32_1 = vec_splat_u32(1);
595 vector signed int sum##i; \
597 const vector unsigned char cmp##i = \
598 (vector unsigned char)vec_cmpgt(src##i, v_avg); \
599 const vector unsigned short cmpHi##i = \
600 (vector unsigned short)vec_mergeh(cmp##i, cmp##i); \
601 const vector unsigned short cmpLi##i = \
602 (vector unsigned short)vec_mergel(cmp##i, cmp##i); \
603 const vector signed short cmpHf##i = \
604 (vector signed short)vec_and(cmpHi##i, mask1); \
605 const vector signed short cmpLf##i = \
606 (vector signed short)vec_and(cmpLi##i, mask2); \
607 const vector signed int sump##i = vec_sum4s(cmpHf##i, zero); \
608 const vector signed int sumq##i = vec_sum4s(cmpLf##i, sump##i); \
609 sum##i = vec_sums(sumq##i, zero); } while (0)
623 vector signed int sumA2;
624 vector signed int sumB2;
626 const vector signed int sump02 = vec_mergel(sum0, sum2);
627 const vector signed int sump13 = vec_mergel(sum1, sum3);
628 const vector signed int sumA = vec_mergel(sump02, sump13);
630 const vector signed int sump46 = vec_mergel(sum4, sum6);
631 const vector signed int sump57 = vec_mergel(sum5, sum7);
632 const vector signed int sumB = vec_mergel(sump46, sump57);
634 const vector signed int sump8A = vec_mergel(sum8, zero);
635 const vector signed int sump9B = vec_mergel(sum9, zero);
636 const vector signed int sumC = vec_mergel(sump8A, sump9B);
638 const vector signed int tA = vec_sl(vec_nor(zero, sumA), vuint32_16);
639 const vector signed int tB = vec_sl(vec_nor(zero, sumB), vuint32_16);
640 const vector signed int tC = vec_sl(vec_nor(zero, sumC), vuint32_16);
641 const vector signed int t2A = vec_or(sumA, tA);
642 const vector signed int t2B = vec_or(sumB, tB);
643 const vector signed int t2C = vec_or(sumC, tC);
644 const vector signed int t3A = vec_and(vec_sra(t2A, vuint32_1),
645 vec_sl(t2A, vuint32_1));
646 const vector signed int t3B = vec_and(vec_sra(t2B, vuint32_1),
647 vec_sl(t2B, vuint32_1));
648 const vector signed int t3C = vec_and(vec_sra(t2C, vuint32_1),
649 vec_sl(t2C, vuint32_1));
650 const vector signed int yA = vec_and(t2A, t3A);
651 const vector signed int yB = vec_and(t2B, t3B);
652 const vector signed int yC = vec_and(t2C, t3C);
654 const vector unsigned char strangeperm1 = vec_lvsl(4, (unsigned char*)0);
655 const vector unsigned char strangeperm2 = vec_lvsl(8, (unsigned char*)0);
656 const vector signed int sumAd4 = vec_perm(yA, yB, strangeperm1);
657 const vector signed int sumAd8 = vec_perm(yA, yB, strangeperm2);
658 const vector signed int sumBd4 = vec_perm(yB, yC, strangeperm1);
659 const vector signed int sumBd8 = vec_perm(yB, yC, strangeperm2);
660 const vector signed int sumAp = vec_and(yA,
661 vec_and(sumAd4,sumAd8));
662 const vector signed int sumBp = vec_and(yB,
663 vec_and(sumBd4,sumBd8));
664 sumA2 = vec_or(sumAp,
667 sumB2 = vec_or(sumBp,
672 vec_st(sumB2, 16, S);
675 /* I'm not sure the following is actually faster
676 than straight, unvectorized C code :-( */
678 int __attribute__((aligned(16))) tQP2[4];
679 tQP2[0]= c->QP/2 + 1;
680 vector signed int vQP2 = vec_ld(0, tQP2);
681 vQP2 = vec_splat(vQP2, 0);
682 const vector signed int vsint32_8 = vec_splat_s32(8);
683 const vector unsigned int vuint32_4 = vec_splat_u32(4);
685 const vector unsigned char permA1 = (vector unsigned char)
686 AVV(0x00, 0x01, 0x02, 0x10, 0x11, 0x12, 0x1F, 0x1F,
687 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F);
688 const vector unsigned char permA2 = (vector unsigned char)
689 AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x10, 0x11,
690 0x12, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F);
691 const vector unsigned char permA1inc = (vector unsigned char)
692 AVV(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00,
693 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
694 const vector unsigned char permA2inc = (vector unsigned char)
695 AVV(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01,
696 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
697 const vector unsigned char magic = (vector unsigned char)
698 AVV(0x01, 0x02, 0x01, 0x02, 0x04, 0x02, 0x01, 0x02,
699 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
700 const vector unsigned char extractPerm = (vector unsigned char)
701 AVV(0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01,
702 0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01);
703 const vector unsigned char extractPermInc = (vector unsigned char)
704 AVV(0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01,
705 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01);
706 const vector unsigned char identity = vec_lvsl(0,(unsigned char *)0);
707 const vector unsigned char tenRight = (vector unsigned char)
708 AVV(0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
709 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
710 const vector unsigned char eightLeft = (vector unsigned char)
711 AVV(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
712 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08);
716 vector unsigned char tenRightM##i = tenRight; \
717 vector unsigned char permA1M##i = permA1; \
718 vector unsigned char permA2M##i = permA2; \
719 vector unsigned char extractPermM##i = extractPerm
721 #define F2(i, j, k, l) \
722 if (S[i] & (1 << (l+1))) { \
723 const vector unsigned char a_##j##_A##l = \
724 vec_perm(src##i, src##j, permA1M##i); \
725 const vector unsigned char a_##j##_B##l = \
726 vec_perm(a_##j##_A##l, src##k, permA2M##i); \
727 const vector signed int a_##j##_sump##l = \
728 (vector signed int)vec_msum(a_##j##_B##l, magic, \
729 (vector unsigned int)zero); \
730 vector signed int F_##j##_##l = \
731 vec_sr(vec_sums(a_##j##_sump##l, vsint32_8), vuint32_4); \
732 F_##j##_##l = vec_splat(F_##j##_##l, 3); \
733 const vector signed int p_##j##_##l = \
734 (vector signed int)vec_perm(src##j, \
735 (vector unsigned char)zero, \
737 const vector signed int sum_##j##_##l = vec_add( p_##j##_##l, vQP2);\
738 const vector signed int diff_##j##_##l = vec_sub( p_##j##_##l, vQP2);\
739 vector signed int newpm_##j##_##l; \
740 if (vec_all_lt(sum_##j##_##l, F_##j##_##l)) \
741 newpm_##j##_##l = sum_##j##_##l; \
742 else if (vec_all_gt(diff_##j##_##l, F_##j##_##l)) \
743 newpm_##j##_##l = diff_##j##_##l; \
744 else newpm_##j##_##l = F_##j##_##l; \
745 const vector unsigned char newpm2_##j##_##l = \
746 vec_splat((vector unsigned char)newpm_##j##_##l, 15); \
747 const vector unsigned char mask##j##l = vec_add(identity, \
749 src##j = vec_perm(src##j, newpm2_##j##_##l, mask##j##l); \
751 permA1M##i = vec_add(permA1M##i, permA1inc); \
752 permA2M##i = vec_add(permA2M##i, permA2inc); \
753 tenRightM##i = vec_sro(tenRightM##i, eightLeft); \
754 extractPermM##i = vec_add(extractPermM##i, extractPermInc)
756 #define ITER(i, j, k) \
776 const vector signed char neg1 = vec_splat_s8(-1);
778 #define STORE_LINE(i) \
779 const vector unsigned char permST##i = \
780 vec_lvsr(i * stride, srcCopy); \
781 const vector unsigned char maskST##i = \
782 vec_perm((vector unsigned char)zero, \
783 (vector unsigned char)neg1, permST##i); \
784 src##i = vec_perm(src##i ,src##i, permST##i); \
785 sA##i= vec_sel(sA##i, src##i, maskST##i); \
786 sB##i= vec_sel(src##i, sB##i, maskST##i); \
787 vec_st(sA##i, i * stride, srcCopy); \
788 vec_st(sB##i, i * stride + 16, srcCopy)
804 #define doHorizLowPass_altivec(a...) doHorizLowPass_C(a)
805 #define doHorizDefFilter_altivec(a...) doHorizDefFilter_C(a)
806 #define do_a_deblock_altivec(a...) do_a_deblock_C(a)
808 static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,
809 uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)
811 const vector signed int zero = vec_splat_s32(0);
812 const vector signed short vsint16_1 = vec_splat_s16(1);
813 vector signed int v_dp = zero;
814 vector signed int v_sysdp = zero;
817 tempBluredPast[127]= maxNoise[0];
818 tempBluredPast[128]= maxNoise[1];
819 tempBluredPast[129]= maxNoise[2];
821 #define LOAD_LINE(src, i) \
822 register int j##src##i = i * stride; \
823 vector unsigned char perm##src##i = vec_lvsl(j##src##i, src); \
824 const vector unsigned char v_##src##A1##i = vec_ld(j##src##i, src); \
825 const vector unsigned char v_##src##A2##i = vec_ld(j##src##i + 16, src); \
826 const vector unsigned char v_##src##A##i = \
827 vec_perm(v_##src##A1##i, v_##src##A2##i, perm##src##i); \
828 vector signed short v_##src##Ass##i = \
829 (vector signed short)vec_mergeh((vector signed char)zero, \
830 (vector signed char)v_##src##A##i)
841 LOAD_LINE(tempBlured, 0);
842 LOAD_LINE(tempBlured, 1);
843 LOAD_LINE(tempBlured, 2);
844 LOAD_LINE(tempBlured, 3);
845 LOAD_LINE(tempBlured, 4);
846 LOAD_LINE(tempBlured, 5);
847 LOAD_LINE(tempBlured, 6);
848 LOAD_LINE(tempBlured, 7);
851 #define ACCUMULATE_DIFFS(i) \
852 vector signed short v_d##i = vec_sub(v_tempBluredAss##i, \
854 v_dp = vec_msums(v_d##i, v_d##i, v_dp); \
855 v_sysdp = vec_msums(v_d##i, vsint16_1, v_sysdp)
865 #undef ACCUMULATE_DIFFS
867 v_dp = vec_sums(v_dp, zero);
868 v_sysdp = vec_sums(v_sysdp, zero);
870 v_dp = vec_splat(v_dp, 3);
871 v_sysdp = vec_splat(v_sysdp, 3);
873 vec_ste(v_dp, 0, &d);
874 vec_ste(v_sysdp, 0, &sysd);
878 +(*(tempBluredPast-256))
879 +(*(tempBluredPast-1))+ (*(tempBluredPast+1))
880 +(*(tempBluredPast+256))
885 if (d > maxNoise[1]) {
886 if (d < maxNoise[2]) {
887 #define OP(i) v_tempBluredAss##i = vec_avg(v_tempBluredAss##i, v_srcAss##i);
899 #define OP(i) v_tempBluredAss##i = v_srcAss##i;
912 if (d < maxNoise[0]) {
913 const vector signed short vsint16_7 = vec_splat_s16(7);
914 const vector signed short vsint16_4 = vec_splat_s16(4);
915 const vector unsigned short vuint16_3 = vec_splat_u16(3);
918 const vector signed short v_temp##i = \
919 vec_mladd(v_tempBluredAss##i, \
920 vsint16_7, v_srcAss##i); \
921 const vector signed short v_temp2##i = \
922 vec_add(v_temp##i, vsint16_4); \
923 v_tempBluredAss##i = vec_sr(v_temp2##i, vuint16_3)
935 const vector signed short vsint16_3 = vec_splat_s16(3);
936 const vector signed short vsint16_2 = vec_splat_s16(2);
939 const vector signed short v_temp##i = \
940 vec_mladd(v_tempBluredAss##i, \
941 vsint16_3, v_srcAss##i); \
942 const vector signed short v_temp2##i = \
943 vec_add(v_temp##i, vsint16_2); \
944 v_tempBluredAss##i = vec_sr(v_temp2##i, (vector unsigned short)vsint16_2)
958 const vector signed char neg1 = vec_splat_s8(-1);
959 const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
960 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
962 #define PACK_AND_STORE(src, i) \
963 const vector unsigned char perms##src##i = \
964 vec_lvsr(i * stride, src); \
965 const vector unsigned char vf##src##i = \
966 vec_packsu(v_tempBluredAss##i, (vector signed short)zero); \
967 const vector unsigned char vg##src##i = \
968 vec_perm(vf##src##i, v_##src##A##i, permHH); \
969 const vector unsigned char mask##src##i = \
970 vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##src##i); \
971 const vector unsigned char vg2##src##i = \
972 vec_perm(vg##src##i, vg##src##i, perms##src##i); \
973 const vector unsigned char svA##src##i = \
974 vec_sel(v_##src##A1##i, vg2##src##i, mask##src##i); \
975 const vector unsigned char svB##src##i = \
976 vec_sel(vg2##src##i, v_##src##A2##i, mask##src##i); \
977 vec_st(svA##src##i, i * stride, src); \
978 vec_st(svB##src##i, i * stride + 16, src)
980 PACK_AND_STORE(src, 0);
981 PACK_AND_STORE(src, 1);
982 PACK_AND_STORE(src, 2);
983 PACK_AND_STORE(src, 3);
984 PACK_AND_STORE(src, 4);
985 PACK_AND_STORE(src, 5);
986 PACK_AND_STORE(src, 6);
987 PACK_AND_STORE(src, 7);
988 PACK_AND_STORE(tempBlured, 0);
989 PACK_AND_STORE(tempBlured, 1);
990 PACK_AND_STORE(tempBlured, 2);
991 PACK_AND_STORE(tempBlured, 3);
992 PACK_AND_STORE(tempBlured, 4);
993 PACK_AND_STORE(tempBlured, 5);
994 PACK_AND_STORE(tempBlured, 6);
995 PACK_AND_STORE(tempBlured, 7);
996 #undef PACK_AND_STORE
999 static inline void transpose_16x8_char_toPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1000 const vector unsigned char zero = vec_splat_u8(0);
1002 #define LOAD_DOUBLE_LINE(i, j) \
1003 vector unsigned char perm1##i = vec_lvsl(i * stride, src); \
1004 vector unsigned char perm2##i = vec_lvsl(j * stride, src); \
1005 vector unsigned char srcA##i = vec_ld(i * stride, src); \
1006 vector unsigned char srcB##i = vec_ld(i * stride + 16, src); \
1007 vector unsigned char srcC##i = vec_ld(j * stride, src); \
1008 vector unsigned char srcD##i = vec_ld(j * stride+ 16, src); \
1009 vector unsigned char src##i = vec_perm(srcA##i, srcB##i, perm1##i); \
1010 vector unsigned char src##j = vec_perm(srcC##i, srcD##i, perm2##i)
1012 LOAD_DOUBLE_LINE(0, 1);
1013 LOAD_DOUBLE_LINE(2, 3);
1014 LOAD_DOUBLE_LINE(4, 5);
1015 LOAD_DOUBLE_LINE(6, 7);
1016 #undef LOAD_DOUBLE_LINE
1018 vector unsigned char tempA = vec_mergeh(src0, zero);
1019 vector unsigned char tempB = vec_mergel(src0, zero);
1020 vector unsigned char tempC = vec_mergeh(src1, zero);
1021 vector unsigned char tempD = vec_mergel(src1, zero);
1022 vector unsigned char tempE = vec_mergeh(src2, zero);
1023 vector unsigned char tempF = vec_mergel(src2, zero);
1024 vector unsigned char tempG = vec_mergeh(src3, zero);
1025 vector unsigned char tempH = vec_mergel(src3, zero);
1026 vector unsigned char tempI = vec_mergeh(src4, zero);
1027 vector unsigned char tempJ = vec_mergel(src4, zero);
1028 vector unsigned char tempK = vec_mergeh(src5, zero);
1029 vector unsigned char tempL = vec_mergel(src5, zero);
1030 vector unsigned char tempM = vec_mergeh(src6, zero);
1031 vector unsigned char tempN = vec_mergel(src6, zero);
1032 vector unsigned char tempO = vec_mergeh(src7, zero);
1033 vector unsigned char tempP = vec_mergel(src7, zero);
1035 vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1036 vector unsigned char temp1 = vec_mergel(tempA, tempI);
1037 vector unsigned char temp2 = vec_mergeh(tempB, tempJ);
1038 vector unsigned char temp3 = vec_mergel(tempB, tempJ);
1039 vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1040 vector unsigned char temp5 = vec_mergel(tempC, tempK);
1041 vector unsigned char temp6 = vec_mergeh(tempD, tempL);
1042 vector unsigned char temp7 = vec_mergel(tempD, tempL);
1043 vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1044 vector unsigned char temp9 = vec_mergel(tempE, tempM);
1045 vector unsigned char temp10 = vec_mergeh(tempF, tempN);
1046 vector unsigned char temp11 = vec_mergel(tempF, tempN);
1047 vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1048 vector unsigned char temp13 = vec_mergel(tempG, tempO);
1049 vector unsigned char temp14 = vec_mergeh(tempH, tempP);
1050 vector unsigned char temp15 = vec_mergel(tempH, tempP);
1052 tempA = vec_mergeh(temp0, temp8);
1053 tempB = vec_mergel(temp0, temp8);
1054 tempC = vec_mergeh(temp1, temp9);
1055 tempD = vec_mergel(temp1, temp9);
1056 tempE = vec_mergeh(temp2, temp10);
1057 tempF = vec_mergel(temp2, temp10);
1058 tempG = vec_mergeh(temp3, temp11);
1059 tempH = vec_mergel(temp3, temp11);
1060 tempI = vec_mergeh(temp4, temp12);
1061 tempJ = vec_mergel(temp4, temp12);
1062 tempK = vec_mergeh(temp5, temp13);
1063 tempL = vec_mergel(temp5, temp13);
1064 tempM = vec_mergeh(temp6, temp14);
1065 tempN = vec_mergel(temp6, temp14);
1066 tempO = vec_mergeh(temp7, temp15);
1067 tempP = vec_mergel(temp7, temp15);
1069 temp0 = vec_mergeh(tempA, tempI);
1070 temp1 = vec_mergel(tempA, tempI);
1071 temp2 = vec_mergeh(tempB, tempJ);
1072 temp3 = vec_mergel(tempB, tempJ);
1073 temp4 = vec_mergeh(tempC, tempK);
1074 temp5 = vec_mergel(tempC, tempK);
1075 temp6 = vec_mergeh(tempD, tempL);
1076 temp7 = vec_mergel(tempD, tempL);
1077 temp8 = vec_mergeh(tempE, tempM);
1078 temp9 = vec_mergel(tempE, tempM);
1079 temp10 = vec_mergeh(tempF, tempN);
1080 temp11 = vec_mergel(tempF, tempN);
1081 temp12 = vec_mergeh(tempG, tempO);
1082 temp13 = vec_mergel(tempG, tempO);
1083 temp14 = vec_mergeh(tempH, tempP);
1084 temp15 = vec_mergel(tempH, tempP);
1086 vec_st(temp0, 0, dst);
1087 vec_st(temp1, 16, dst);
1088 vec_st(temp2, 32, dst);
1089 vec_st(temp3, 48, dst);
1090 vec_st(temp4, 64, dst);
1091 vec_st(temp5, 80, dst);
1092 vec_st(temp6, 96, dst);
1093 vec_st(temp7, 112, dst);
1094 vec_st(temp8, 128, dst);
1095 vec_st(temp9, 144, dst);
1096 vec_st(temp10, 160, dst);
1097 vec_st(temp11, 176, dst);
1098 vec_st(temp12, 192, dst);
1099 vec_st(temp13, 208, dst);
1100 vec_st(temp14, 224, dst);
1101 vec_st(temp15, 240, dst);
1104 static inline void transpose_8x16_char_fromPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1105 const vector unsigned char zero = vec_splat_u8(0);
1107 #define LOAD_DOUBLE_LINE(i, j) \
1108 vector unsigned char src##i = vec_ld(i * 16, src); \
1109 vector unsigned char src##j = vec_ld(j * 16, src)
1111 LOAD_DOUBLE_LINE(0, 1);
1112 LOAD_DOUBLE_LINE(2, 3);
1113 LOAD_DOUBLE_LINE(4, 5);
1114 LOAD_DOUBLE_LINE(6, 7);
1115 LOAD_DOUBLE_LINE(8, 9);
1116 LOAD_DOUBLE_LINE(10, 11);
1117 LOAD_DOUBLE_LINE(12, 13);
1118 LOAD_DOUBLE_LINE(14, 15);
1119 #undef LOAD_DOUBLE_LINE
1121 vector unsigned char tempA = vec_mergeh(src0, src8);
1122 vector unsigned char tempB;
1123 vector unsigned char tempC = vec_mergeh(src1, src9);
1124 vector unsigned char tempD;
1125 vector unsigned char tempE = vec_mergeh(src2, src10);
1126 vector unsigned char tempG = vec_mergeh(src3, src11);
1127 vector unsigned char tempI = vec_mergeh(src4, src12);
1128 vector unsigned char tempJ;
1129 vector unsigned char tempK = vec_mergeh(src5, src13);
1130 vector unsigned char tempL;
1131 vector unsigned char tempM = vec_mergeh(src6, src14);
1132 vector unsigned char tempO = vec_mergeh(src7, src15);
1134 vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1135 vector unsigned char temp1 = vec_mergel(tempA, tempI);
1136 vector unsigned char temp2;
1137 vector unsigned char temp3;
1138 vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1139 vector unsigned char temp5 = vec_mergel(tempC, tempK);
1140 vector unsigned char temp6;
1141 vector unsigned char temp7;
1142 vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1143 vector unsigned char temp9 = vec_mergel(tempE, tempM);
1144 vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1145 vector unsigned char temp13 = vec_mergel(tempG, tempO);
1147 tempA = vec_mergeh(temp0, temp8);
1148 tempB = vec_mergel(temp0, temp8);
1149 tempC = vec_mergeh(temp1, temp9);
1150 tempD = vec_mergel(temp1, temp9);
1151 tempI = vec_mergeh(temp4, temp12);
1152 tempJ = vec_mergel(temp4, temp12);
1153 tempK = vec_mergeh(temp5, temp13);
1154 tempL = vec_mergel(temp5, temp13);
1156 temp0 = vec_mergeh(tempA, tempI);
1157 temp1 = vec_mergel(tempA, tempI);
1158 temp2 = vec_mergeh(tempB, tempJ);
1159 temp3 = vec_mergel(tempB, tempJ);
1160 temp4 = vec_mergeh(tempC, tempK);
1161 temp5 = vec_mergel(tempC, tempK);
1162 temp6 = vec_mergeh(tempD, tempL);
1163 temp7 = vec_mergel(tempD, tempL);
1166 const vector signed char neg1 = vec_splat_s8(-1);
1167 #define STORE_DOUBLE_LINE(i, j) \
1168 vector unsigned char dstA##i = vec_ld(i * stride, dst); \
1169 vector unsigned char dstB##i = vec_ld(i * stride + 16, dst); \
1170 vector unsigned char dstA##j = vec_ld(j * stride, dst); \
1171 vector unsigned char dstB##j = vec_ld(j * stride+ 16, dst); \
1172 vector unsigned char align##i = vec_lvsr(i * stride, dst); \
1173 vector unsigned char align##j = vec_lvsr(j * stride, dst); \
1174 vector unsigned char mask##i = vec_perm(zero, (vector unsigned char)neg1, align##i); \
1175 vector unsigned char mask##j = vec_perm(zero, (vector unsigned char)neg1, align##j); \
1176 vector unsigned char dstR##i = vec_perm(temp##i, temp##i, align##i); \
1177 vector unsigned char dstR##j = vec_perm(temp##j, temp##j, align##j); \
1178 vector unsigned char dstAF##i = vec_sel(dstA##i, dstR##i, mask##i); \
1179 vector unsigned char dstBF##i = vec_sel(dstR##i, dstB##i, mask##i); \
1180 vector unsigned char dstAF##j = vec_sel(dstA##j, dstR##j, mask##j); \
1181 vector unsigned char dstBF##j = vec_sel(dstR##j, dstB##j, mask##j); \
1182 vec_st(dstAF##i, i * stride, dst); \
1183 vec_st(dstBF##i, i * stride + 16, dst); \
1184 vec_st(dstAF##j, j * stride, dst); \
1185 vec_st(dstBF##j, j * stride + 16, dst)
1187 STORE_DOUBLE_LINE(0,1);
1188 STORE_DOUBLE_LINE(2,3);
1189 STORE_DOUBLE_LINE(4,5);
1190 STORE_DOUBLE_LINE(6,7);