libavcodec/libpostproc ---> libpostproc
[libav.git] / libpostproc / postprocess_altivec_template.c
1 /*
2 AltiVec optimizations (C) 2004 Romain Dolbeau <romain@dolbeau.org>
3
4 based on code by Copyright (C) 2001-2003 Michael Niedermayer (michaelni@gmx.at)
5
6 This program is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 2 of the License, or
9 (at your option) any later version.
10
11 This program is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with this program; if not, write to the Free Software
18 Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 */
20
21
22 #ifdef CONFIG_DARWIN
23 #define AVV(x...) (x)
24 #else
25 #define AVV(x...) {x}
26 #endif
27
28 #define ALTIVEC_TRANSPOSE_8x8_SHORT(src_a,src_b,src_c,src_d,src_e,src_f,src_g,src_h) \
29 do { \
30 __typeof__(src_a) tempA1, tempB1, tempC1, tempD1; \
31 __typeof__(src_a) tempE1, tempF1, tempG1, tempH1; \
32 __typeof__(src_a) tempA2, tempB2, tempC2, tempD2; \
33 __typeof__(src_a) tempE2, tempF2, tempG2, tempH2; \
34 tempA1 = vec_mergeh (src_a, src_e); \
35 tempB1 = vec_mergel (src_a, src_e); \
36 tempC1 = vec_mergeh (src_b, src_f); \
37 tempD1 = vec_mergel (src_b, src_f); \
38 tempE1 = vec_mergeh (src_c, src_g); \
39 tempF1 = vec_mergel (src_c, src_g); \
40 tempG1 = vec_mergeh (src_d, src_h); \
41 tempH1 = vec_mergel (src_d, src_h); \
42 tempA2 = vec_mergeh (tempA1, tempE1); \
43 tempB2 = vec_mergel (tempA1, tempE1); \
44 tempC2 = vec_mergeh (tempB1, tempF1); \
45 tempD2 = vec_mergel (tempB1, tempF1); \
46 tempE2 = vec_mergeh (tempC1, tempG1); \
47 tempF2 = vec_mergel (tempC1, tempG1); \
48 tempG2 = vec_mergeh (tempD1, tempH1); \
49 tempH2 = vec_mergel (tempD1, tempH1); \
50 src_a = vec_mergeh (tempA2, tempE2); \
51 src_b = vec_mergel (tempA2, tempE2); \
52 src_c = vec_mergeh (tempB2, tempF2); \
53 src_d = vec_mergel (tempB2, tempF2); \
54 src_e = vec_mergeh (tempC2, tempG2); \
55 src_f = vec_mergel (tempC2, tempG2); \
56 src_g = vec_mergeh (tempD2, tempH2); \
57 src_h = vec_mergel (tempD2, tempH2); \
58 } while (0)
59
60
61 static inline int vertClassify_altivec(uint8_t src[], int stride, PPContext *c) {
62 /*
63 this code makes no assumption on src or stride.
64 One could remove the recomputation of the perm
65 vector by assuming (stride % 16) == 0, unfortunately
66 this is not always true.
67 */
68 register int y;
69 short __attribute__ ((aligned(16))) data[8];
70 int numEq;
71 uint8_t *src2 = src;
72 vector signed short v_dcOffset;
73 vector signed short v2QP;
74 vector unsigned short v4QP;
75 vector unsigned short v_dcThreshold;
76 const int properStride = (stride % 16);
77 const int srcAlign = ((unsigned long)src2 % 16);
78 const int two_vectors = ((srcAlign > 8) || properStride) ? 1 : 0;
79 const vector signed int zero = vec_splat_s32(0);
80 const vector signed short mask = vec_splat_s16(1);
81 vector signed int v_numEq = vec_splat_s32(0);
82
83 data[0] = ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
84 data[1] = data[0] * 2 + 1;
85 data[2] = c->QP * 2;
86 data[3] = c->QP * 4;
87 vector signed short v_data = vec_ld(0, data);
88 v_dcOffset = vec_splat(v_data, 0);
89 v_dcThreshold = (vector unsigned short)vec_splat(v_data, 1);
90 v2QP = vec_splat(v_data, 2);
91 v4QP = (vector unsigned short)vec_splat(v_data, 3);
92
93 src2 += stride * 4;
94
95 vector signed short v_srcAss0, v_srcAss1, v_srcAss2, v_srcAss3, v_srcAss4, v_srcAss5, v_srcAss6, v_srcAss7;
96
97 #define LOAD_LINE(i) \
98 register int j##i = i * stride; \
99 vector unsigned char perm##i = vec_lvsl(j##i, src2); \
100 const vector unsigned char v_srcA1##i = vec_ld(j##i, src2); \
101 vector unsigned char v_srcA2##i; \
102 if (two_vectors) \
103 v_srcA2##i = vec_ld(j##i + 16, src2); \
104 const vector unsigned char v_srcA##i = \
105 vec_perm(v_srcA1##i, v_srcA2##i, perm##i); \
106 v_srcAss##i = \
107 (vector signed short)vec_mergeh((vector signed char)zero, \
108 (vector signed char)v_srcA##i)
109
110 #define LOAD_LINE_ALIGNED(i) \
111 register int j##i = i * stride; \
112 const vector unsigned char v_srcA##i = vec_ld(j##i, src2); \
113 v_srcAss##i = \
114 (vector signed short)vec_mergeh((vector signed char)zero, \
115 (vector signed char)v_srcA##i)
116
117 // special casing the aligned case is worthwhile, as all call from
118 // the (transposed) horizontable deblocks will be aligned, i naddition
119 // to the naturraly aligned vertical deblocks.
120 if (properStride && srcAlign) {
121 LOAD_LINE_ALIGNED(0);
122 LOAD_LINE_ALIGNED(1);
123 LOAD_LINE_ALIGNED(2);
124 LOAD_LINE_ALIGNED(3);
125 LOAD_LINE_ALIGNED(4);
126 LOAD_LINE_ALIGNED(5);
127 LOAD_LINE_ALIGNED(6);
128 LOAD_LINE_ALIGNED(7);
129 } else {
130 LOAD_LINE(0);
131 LOAD_LINE(1);
132 LOAD_LINE(2);
133 LOAD_LINE(3);
134 LOAD_LINE(4);
135 LOAD_LINE(5);
136 LOAD_LINE(6);
137 LOAD_LINE(7);
138 }
139 #undef LOAD_LINE
140 #undef LOAD_LINE_ALIGNED
141
142 #define ITER(i, j) \
143 const vector signed short v_diff##i = \
144 vec_sub(v_srcAss##i, v_srcAss##j); \
145 const vector signed short v_sum##i = \
146 vec_add(v_diff##i, v_dcOffset); \
147 const vector signed short v_comp##i = \
148 (vector signed short)vec_cmplt((vector unsigned short)v_sum##i, \
149 v_dcThreshold); \
150 const vector signed short v_part##i = vec_and(mask, v_comp##i); \
151 v_numEq = vec_sum4s(v_part##i, v_numEq);
152
153 ITER(0, 1);
154 ITER(1, 2);
155 ITER(2, 3);
156 ITER(3, 4);
157 ITER(4, 5);
158 ITER(5, 6);
159 ITER(6, 7);
160 #undef ITER
161
162 v_numEq = vec_sums(v_numEq, zero);
163
164 v_numEq = vec_splat(v_numEq, 3);
165 vec_ste(v_numEq, 0, &numEq);
166
167 if (numEq > c->ppMode.flatnessThreshold)
168 {
169 const vector unsigned char mmoP1 = (const vector unsigned char)
170 AVV(0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f,
171 0x00, 0x01, 0x12, 0x13, 0x08, 0x09, 0x1A, 0x1B);
172 const vector unsigned char mmoP2 = (const vector unsigned char)
173 AVV(0x04, 0x05, 0x16, 0x17, 0x0C, 0x0D, 0x1E, 0x1F,
174 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f);
175 const vector unsigned char mmoP = (const vector unsigned char)
176 vec_lvsl(8, (unsigned char*)0);
177
178 vector signed short mmoL1 = vec_perm(v_srcAss0, v_srcAss2, mmoP1);
179 vector signed short mmoL2 = vec_perm(v_srcAss4, v_srcAss6, mmoP2);
180 vector signed short mmoL = vec_perm(mmoL1, mmoL2, mmoP);
181 vector signed short mmoR1 = vec_perm(v_srcAss5, v_srcAss7, mmoP1);
182 vector signed short mmoR2 = vec_perm(v_srcAss1, v_srcAss3, mmoP2);
183 vector signed short mmoR = vec_perm(mmoR1, mmoR2, mmoP);
184 vector signed short mmoDiff = vec_sub(mmoL, mmoR);
185 vector unsigned short mmoSum = (vector unsigned short)vec_add(mmoDiff, v2QP);
186
187 if (vec_any_gt(mmoSum, v4QP))
188 return 0;
189 else
190 return 1;
191 }
192 else return 2;
193 }
194
195 static inline void doVertLowPass_altivec(uint8_t *src, int stride, PPContext *c) {
196 /*
197 this code makes no assumption on src or stride.
198 One could remove the recomputation of the perm
199 vector by assuming (stride % 16) == 0, unfortunately
200 this is not always true. Quite a lot of load/stores
201 can be removed by assuming proper alignement of
202 src & stride :-(
203 */
204 uint8_t *src2 = src;
205 const vector signed int zero = vec_splat_s32(0);
206 const int properStride = (stride % 16);
207 const int srcAlign = ((unsigned long)src2 % 16);
208 short __attribute__ ((aligned(16))) qp[8];
209 qp[0] = c->QP;
210 vector signed short vqp = vec_ld(0, qp);
211 vqp = vec_splat(vqp, 0);
212
213 src2 += stride*3;
214
215 vector signed short vb0, vb1, vb2, vb3, vb4, vb5, vb6, vb7, vb8, vb9;
216 vector unsigned char vbA0, vbA1, vbA2, vbA3, vbA4, vbA5, vbA6, vbA7, vbA8, vbA9;
217 vector unsigned char vbB0, vbB1, vbB2, vbB3, vbB4, vbB5, vbB6, vbB7, vbB8, vbB9;
218 vector unsigned char vbT0, vbT1, vbT2, vbT3, vbT4, vbT5, vbT6, vbT7, vbT8, vbT9;
219
220 #define LOAD_LINE(i) \
221 const vector unsigned char perml##i = \
222 vec_lvsl(i * stride, src2); \
223 vbA##i = vec_ld(i * stride, src2); \
224 vbB##i = vec_ld(i * stride + 16, src2); \
225 vbT##i = vec_perm(vbA##i, vbB##i, perml##i); \
226 vb##i = \
227 (vector signed short)vec_mergeh((vector unsigned char)zero, \
228 (vector unsigned char)vbT##i)
229
230 #define LOAD_LINE_ALIGNED(i) \
231 register int j##i = i * stride; \
232 vbT##i = vec_ld(j##i, src2); \
233 vb##i = \
234 (vector signed short)vec_mergeh((vector signed char)zero, \
235 (vector signed char)vbT##i)
236
237 // special casing the aligned case is worthwhile, as all call from
238 // the (transposed) horizontable deblocks will be aligned, in addition
239 // to the naturraly aligned vertical deblocks.
240 if (properStride && srcAlign) {
241 LOAD_LINE_ALIGNED(0);
242 LOAD_LINE_ALIGNED(1);
243 LOAD_LINE_ALIGNED(2);
244 LOAD_LINE_ALIGNED(3);
245 LOAD_LINE_ALIGNED(4);
246 LOAD_LINE_ALIGNED(5);
247 LOAD_LINE_ALIGNED(6);
248 LOAD_LINE_ALIGNED(7);
249 LOAD_LINE_ALIGNED(8);
250 LOAD_LINE_ALIGNED(9);
251 } else {
252 LOAD_LINE(0);
253 LOAD_LINE(1);
254 LOAD_LINE(2);
255 LOAD_LINE(3);
256 LOAD_LINE(4);
257 LOAD_LINE(5);
258 LOAD_LINE(6);
259 LOAD_LINE(7);
260 LOAD_LINE(8);
261 LOAD_LINE(9);
262 }
263 #undef LOAD_LINE
264 #undef LOAD_LINE_ALIGNED
265
266 const vector unsigned short v_1 = vec_splat_u16(1);
267 const vector unsigned short v_2 = vec_splat_u16(2);
268 const vector unsigned short v_4 = vec_splat_u16(4);
269
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);
278
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);
283
284 const vector signed short temp11 = vec_sub(v_sumsB0, v_first);
285 const vector signed short v_sumsB1 = vec_add(temp11, vb4);
286
287 const vector signed short temp21 = vec_sub(v_sumsB1, v_first);
288 const vector signed short v_sumsB2 = vec_add(temp21, vb5);
289
290 const vector signed short temp31 = vec_sub(v_sumsB2, v_first);
291 const vector signed short v_sumsB3 = vec_add(temp31, vb6);
292
293 const vector signed short temp41 = vec_sub(v_sumsB3, v_first);
294 const vector signed short v_sumsB4 = vec_add(temp41, vb7);
295
296 const vector signed short temp51 = vec_sub(v_sumsB4, vb1);
297 const vector signed short v_sumsB5 = vec_add(temp51, vb8);
298
299 const vector signed short temp61 = vec_sub(v_sumsB5, vb2);
300 const vector signed short v_sumsB6 = vec_add(temp61, v_last);
301
302 const vector signed short temp71 = vec_sub(v_sumsB6, vb3);
303 const vector signed short v_sumsB7 = vec_add(temp71, v_last);
304
305 const vector signed short temp81 = vec_sub(v_sumsB7, vb4);
306 const vector signed short v_sumsB8 = vec_add(temp81, v_last);
307
308 const vector signed short temp91 = vec_sub(v_sumsB8, vb5);
309 const vector signed short v_sumsB9 = vec_add(temp91, v_last);
310
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)
317
318 COMPUTE_VR(0, 1, 2);
319 COMPUTE_VR(1, 2, 3);
320 COMPUTE_VR(2, 3, 4);
321 COMPUTE_VR(3, 4, 5);
322 COMPUTE_VR(4, 5, 6);
323 COMPUTE_VR(5, 6, 7);
324 COMPUTE_VR(6, 7, 8);
325 COMPUTE_VR(7, 8, 9);
326
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);
330
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)
348
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)
355
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);
368 } else {
369 PACK_AND_STORE(1);
370 PACK_AND_STORE(2);
371 PACK_AND_STORE(3);
372 PACK_AND_STORE(4);
373 PACK_AND_STORE(5);
374 PACK_AND_STORE(6);
375 PACK_AND_STORE(7);
376 PACK_AND_STORE(8);
377 }
378 #undef PACK_AND_STORE
379 #undef PACK_AND_STORE_ALIGNED
380 }
381
382
383
384 static inline void doVertDefFilter_altivec(uint8_t src[], int stride, PPContext *c) {
385 /*
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
391 src & stride :-(
392 */
393 uint8_t *src2 = src;
394 const vector signed int zero = vec_splat_s32(0);
395 short __attribute__ ((aligned(16))) qp[8];
396 qp[0] = 8*c->QP;
397 vector signed short vqp = vec_ld(0, qp);
398 vqp = vec_splat(vqp, 0);
399
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)
412
413 src2 += stride*3;
414
415 LOAD_LINE(1);
416 LOAD_LINE(2);
417 LOAD_LINE(3);
418 LOAD_LINE(4);
419 LOAD_LINE(5);
420 LOAD_LINE(6);
421 LOAD_LINE(7);
422 LOAD_LINE(8);
423 #undef LOAD_LINE
424
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);
430 /* middle energy */
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);
445 /* d */
446 const vector signed short ddiff = vec_sub(absmE,
447 vec_min(vec_abs(lE),
448 vec_abs(rE)));
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,
454 d,
455 vec_cmpgt(vec_sub((vector signed short)zero, mE),
456 (vector signed short)zero));
457 /* q */
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,
462 v_1,
463 vec_cmplt(qtimes2, (vector signed short)zero));
464 const vector signed short q = vec_sra(vec_add(qtimes2, rounddown), vec_splat_u16(1));
465 /* clamp */
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);
470
471 const vector signed short dclampedfinal = vec_sel(dclamp_N,
472 dclamp_P,
473 vec_cmpgt(q, (vector signed short)zero));
474 const vector signed short dornotd = vec_sel((vector signed short)zero,
475 dclampedfinal,
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);
483
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);
487
488 #define STORE(i) \
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)
503
504 STORE(4);
505 STORE(5);
506 }
507
508 static inline void dering_altivec(uint8_t src[], int stride, PPContext *c) {
509 /*
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
515 src & stride :-(
516 */
517 uint8_t *srcCopy = src;
518 uint8_t __attribute__((aligned(16))) dt[16];
519 const vector unsigned char vuint8_1 = vec_splat_u8(1);
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);
524
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)
531
532 LOAD_LINE(0);
533 LOAD_LINE(1);
534 LOAD_LINE(2);
535 LOAD_LINE(3);
536 LOAD_LINE(4);
537 LOAD_LINE(5);
538 LOAD_LINE(6);
539 LOAD_LINE(7);
540 LOAD_LINE(8);
541 LOAD_LINE(9);
542 #undef LOAD_LINE
543
544 vector unsigned char v_avg;
545 {
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);
553
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)
570
571 vector unsigned char v_min;
572 vector unsigned char v_max;
573 EXTRACT(min);
574 EXTRACT(max);
575 #undef EXTRACT
576
577 if (vec_all_lt(vec_sub(v_max, v_min), v_dt))
578 return;
579
580 v_avg = vec_avg(v_min, v_max);
581 }
582
583 signed int __attribute__((aligned(16))) S[8];
584 {
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);
591
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);
594
595 #define COMPARE(i) \
596 vector signed int sum##i; \
597 do { \
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)
611
612 COMPARE(0);
613 COMPARE(1);
614 COMPARE(2);
615 COMPARE(3);
616 COMPARE(4);
617 COMPARE(5);
618 COMPARE(6);
619 COMPARE(7);
620 COMPARE(8);
621 COMPARE(9);
622 #undef COMPARE
623
624 vector signed int sumA2;
625 vector signed int sumB2;
626 {
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);
630
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);
634
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);
638
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);
654
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,
666 vec_sra(sumAp,
667 vuint32_16));
668 sumB2 = vec_or(sumBp,
669 vec_sra(sumBp,
670 vuint32_16));
671 }
672 vec_st(sumA2, 0, S);
673 vec_st(sumB2, 16, S);
674 }
675
676 /* I'm not sure the following is actually faster
677 than straight, unvectorized C code :-( */
678
679 int __attribute__((aligned(16))) 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 unsigned char vuint8_2 = vec_splat_u8(2);
684 const vector signed int vsint32_8 = vec_splat_s32(8);
685 const vector unsigned int vuint32_4 = vec_splat_u32(4);
686
687 const vector unsigned char permA1 = (vector unsigned char)
688 AVV(0x00, 0x01, 0x02, 0x10, 0x11, 0x12, 0x1F, 0x1F,
689 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F);
690 const vector unsigned char permA2 = (vector unsigned char)
691 AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x10, 0x11,
692 0x12, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F);
693 const vector unsigned char permA1inc = (vector unsigned char)
694 AVV(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00,
695 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
696 const vector unsigned char permA2inc = (vector unsigned char)
697 AVV(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01,
698 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
699 const vector unsigned char magic = (vector unsigned char)
700 AVV(0x01, 0x02, 0x01, 0x02, 0x04, 0x02, 0x01, 0x02,
701 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
702 const vector unsigned char extractPerm = (vector unsigned char)
703 AVV(0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01,
704 0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01);
705 const vector unsigned char extractPermInc = (vector unsigned char)
706 AVV(0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01,
707 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01);
708 const vector unsigned char identity = vec_lvsl(0,(unsigned char *)0);
709 const vector unsigned char tenRight = (vector unsigned char)
710 AVV(0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
711 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00);
712 const vector unsigned char eightLeft = (vector unsigned char)
713 AVV(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
714 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08);
715
716
717 #define F_INIT(i) \
718 vector unsigned char tenRightM##i = tenRight; \
719 vector unsigned char permA1M##i = permA1; \
720 vector unsigned char permA2M##i = permA2; \
721 vector unsigned char extractPermM##i = extractPerm
722
723 #define F2(i, j, k, l) \
724 if (S[i] & (1 << (l+1))) { \
725 const vector unsigned char a_##j##_A##l = \
726 vec_perm(src##i, src##j, permA1M##i); \
727 const vector unsigned char a_##j##_B##l = \
728 vec_perm(a_##j##_A##l, src##k, permA2M##i); \
729 const vector signed int a_##j##_sump##l = \
730 (vector signed int)vec_msum(a_##j##_B##l, magic, \
731 (vector unsigned int)zero); \
732 vector signed int F_##j##_##l = \
733 vec_sr(vec_sums(a_##j##_sump##l, vsint32_8), vuint32_4); \
734 F_##j##_##l = vec_splat(F_##j##_##l, 3); \
735 const vector signed int p_##j##_##l = \
736 (vector signed int)vec_perm(src##j, \
737 (vector unsigned char)zero, \
738 extractPermM##i); \
739 const vector signed int sum_##j##_##l = vec_add( p_##j##_##l, vQP2);\
740 const vector signed int diff_##j##_##l = vec_sub( p_##j##_##l, vQP2);\
741 vector signed int newpm_##j##_##l; \
742 if (vec_all_lt(sum_##j##_##l, F_##j##_##l)) \
743 newpm_##j##_##l = sum_##j##_##l; \
744 else if (vec_all_gt(diff_##j##_##l, F_##j##_##l)) \
745 newpm_##j##_##l = diff_##j##_##l; \
746 else newpm_##j##_##l = F_##j##_##l; \
747 const vector unsigned char newpm2_##j##_##l = \
748 vec_splat((vector unsigned char)newpm_##j##_##l, 15); \
749 const vector unsigned char mask##j##l = vec_add(identity, \
750 tenRightM##i); \
751 src##j = vec_perm(src##j, newpm2_##j##_##l, mask##j##l); \
752 } \
753 permA1M##i = vec_add(permA1M##i, permA1inc); \
754 permA2M##i = vec_add(permA2M##i, permA2inc); \
755 tenRightM##i = vec_sro(tenRightM##i, eightLeft); \
756 extractPermM##i = vec_add(extractPermM##i, extractPermInc)
757
758 #define ITER(i, j, k) \
759 F_INIT(i); \
760 F2(i, j, k, 0); \
761 F2(i, j, k, 1); \
762 F2(i, j, k, 2); \
763 F2(i, j, k, 3); \
764 F2(i, j, k, 4); \
765 F2(i, j, k, 5); \
766 F2(i, j, k, 6); \
767 F2(i, j, k, 7)
768
769 ITER(0, 1, 2);
770 ITER(1, 2, 3);
771 ITER(2, 3, 4);
772 ITER(3, 4, 5);
773 ITER(4, 5, 6);
774 ITER(5, 6, 7);
775 ITER(6, 7, 8);
776 ITER(7, 8, 9);
777
778 const vector signed char neg1 = vec_splat_s8(-1);
779
780 #define STORE_LINE(i) \
781 const vector unsigned char permST##i = \
782 vec_lvsr(i * stride, srcCopy); \
783 const vector unsigned char maskST##i = \
784 vec_perm((vector unsigned char)zero, \
785 (vector unsigned char)neg1, permST##i); \
786 src##i = vec_perm(src##i ,src##i, permST##i); \
787 sA##i= vec_sel(sA##i, src##i, maskST##i); \
788 sB##i= vec_sel(src##i, sB##i, maskST##i); \
789 vec_st(sA##i, i * stride, srcCopy); \
790 vec_st(sB##i, i * stride + 16, srcCopy)
791
792 STORE_LINE(1);
793 STORE_LINE(2);
794 STORE_LINE(3);
795 STORE_LINE(4);
796 STORE_LINE(5);
797 STORE_LINE(6);
798 STORE_LINE(7);
799 STORE_LINE(8);
800
801 #undef STORE_LINE
802 #undef ITER
803 #undef F2
804 }
805
806 #define doHorizLowPass_altivec(a...) doHorizLowPass_C(a)
807 #define doHorizDefFilter_altivec(a...) doHorizDefFilter_C(a)
808 #define do_a_deblock_altivec(a...) do_a_deblock_C(a)
809
810 static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,
811 uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)
812 {
813 const vector signed int zero = vec_splat_s32(0);
814 const vector signed short vsint16_1 = vec_splat_s16(1);
815 vector signed int v_dp = zero;
816 vector signed int v_sysdp = zero;
817 int d, sysd, i;
818
819 tempBluredPast[127]= maxNoise[0];
820 tempBluredPast[128]= maxNoise[1];
821 tempBluredPast[129]= maxNoise[2];
822
823 #define LOAD_LINE(src, i) \
824 register int j##src##i = i * stride; \
825 vector unsigned char perm##src##i = vec_lvsl(j##src##i, src); \
826 const vector unsigned char v_##src##A1##i = vec_ld(j##src##i, src); \
827 const vector unsigned char v_##src##A2##i = vec_ld(j##src##i + 16, src); \
828 const vector unsigned char v_##src##A##i = \
829 vec_perm(v_##src##A1##i, v_##src##A2##i, perm##src##i); \
830 vector signed short v_##src##Ass##i = \
831 (vector signed short)vec_mergeh((vector signed char)zero, \
832 (vector signed char)v_##src##A##i)
833
834 LOAD_LINE(src, 0);
835 LOAD_LINE(src, 1);
836 LOAD_LINE(src, 2);
837 LOAD_LINE(src, 3);
838 LOAD_LINE(src, 4);
839 LOAD_LINE(src, 5);
840 LOAD_LINE(src, 6);
841 LOAD_LINE(src, 7);
842
843 LOAD_LINE(tempBlured, 0);
844 LOAD_LINE(tempBlured, 1);
845 LOAD_LINE(tempBlured, 2);
846 LOAD_LINE(tempBlured, 3);
847 LOAD_LINE(tempBlured, 4);
848 LOAD_LINE(tempBlured, 5);
849 LOAD_LINE(tempBlured, 6);
850 LOAD_LINE(tempBlured, 7);
851 #undef LOAD_LINE
852
853 #define ACCUMULATE_DIFFS(i) \
854 vector signed short v_d##i = vec_sub(v_tempBluredAss##i, \
855 v_srcAss##i); \
856 v_dp = vec_msums(v_d##i, v_d##i, v_dp); \
857 v_sysdp = vec_msums(v_d##i, vsint16_1, v_sysdp)
858
859 ACCUMULATE_DIFFS(0);
860 ACCUMULATE_DIFFS(1);
861 ACCUMULATE_DIFFS(2);
862 ACCUMULATE_DIFFS(3);
863 ACCUMULATE_DIFFS(4);
864 ACCUMULATE_DIFFS(5);
865 ACCUMULATE_DIFFS(6);
866 ACCUMULATE_DIFFS(7);
867 #undef ACCUMULATE_DIFFS
868
869 v_dp = vec_sums(v_dp, zero);
870 v_sysdp = vec_sums(v_sysdp, zero);
871
872 v_dp = vec_splat(v_dp, 3);
873 v_sysdp = vec_splat(v_sysdp, 3);
874
875 vec_ste(v_dp, 0, &d);
876 vec_ste(v_sysdp, 0, &sysd);
877
878 i = d;
879 d = (4*d
880 +(*(tempBluredPast-256))
881 +(*(tempBluredPast-1))+ (*(tempBluredPast+1))
882 +(*(tempBluredPast+256))
883 +4)>>3;
884
885 *tempBluredPast=i;
886
887 if (d > maxNoise[1]) {
888 if (d < maxNoise[2]) {
889 #define OP(i) v_tempBluredAss##i = vec_avg(v_tempBluredAss##i, v_srcAss##i);
890
891 OP(0);
892 OP(1);
893 OP(2);
894 OP(3);
895 OP(4);
896 OP(5);
897 OP(6);
898 OP(7);
899 #undef OP
900 } else {
901 #define OP(i) v_tempBluredAss##i = v_srcAss##i;
902
903 OP(0);
904 OP(1);
905 OP(2);
906 OP(3);
907 OP(4);
908 OP(5);
909 OP(6);
910 OP(7);
911 #undef OP
912 }
913 } else {
914 if (d < maxNoise[0]) {
915 const vector signed short vsint16_7 = vec_splat_s16(7);
916 const vector signed short vsint16_4 = vec_splat_s16(4);
917 const vector unsigned short vuint16_3 = vec_splat_u16(3);
918
919 #define OP(i) \
920 const vector signed short v_temp##i = \
921 vec_mladd(v_tempBluredAss##i, \
922 vsint16_7, v_srcAss##i); \
923 const vector signed short v_temp2##i = \
924 vec_add(v_temp##i, vsint16_4); \
925 v_tempBluredAss##i = vec_sr(v_temp2##i, vuint16_3)
926
927 OP(0);
928 OP(1);
929 OP(2);
930 OP(3);
931 OP(4);
932 OP(5);
933 OP(6);
934 OP(7);
935 #undef OP
936 } else {
937 const vector signed short vsint16_3 = vec_splat_s16(3);
938 const vector signed short vsint16_2 = vec_splat_s16(2);
939
940 #define OP(i) \
941 const vector signed short v_temp##i = \
942 vec_mladd(v_tempBluredAss##i, \
943 vsint16_3, v_srcAss##i); \
944 const vector signed short v_temp2##i = \
945 vec_add(v_temp##i, vsint16_2); \
946 v_tempBluredAss##i = vec_sr(v_temp2##i, (vector unsigned short)vsint16_2)
947
948 OP(0);
949 OP(1);
950 OP(2);
951 OP(3);
952 OP(4);
953 OP(5);
954 OP(6);
955 OP(7);
956 #undef OP
957 }
958 }
959
960 const vector signed char neg1 = vec_splat_s8(-1);
961 const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
962 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
963
964 #define PACK_AND_STORE(src, i) \
965 const vector unsigned char perms##src##i = \
966 vec_lvsr(i * stride, src); \
967 const vector unsigned char vf##src##i = \
968 vec_packsu(v_tempBluredAss##i, (vector signed short)zero); \
969 const vector unsigned char vg##src##i = \
970 vec_perm(vf##src##i, v_##src##A##i, permHH); \
971 const vector unsigned char mask##src##i = \
972 vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##src##i); \
973 const vector unsigned char vg2##src##i = \
974 vec_perm(vg##src##i, vg##src##i, perms##src##i); \
975 const vector unsigned char svA##src##i = \
976 vec_sel(v_##src##A1##i, vg2##src##i, mask##src##i); \
977 const vector unsigned char svB##src##i = \
978 vec_sel(vg2##src##i, v_##src##A2##i, mask##src##i); \
979 vec_st(svA##src##i, i * stride, src); \
980 vec_st(svB##src##i, i * stride + 16, src)
981
982 PACK_AND_STORE(src, 0);
983 PACK_AND_STORE(src, 1);
984 PACK_AND_STORE(src, 2);
985 PACK_AND_STORE(src, 3);
986 PACK_AND_STORE(src, 4);
987 PACK_AND_STORE(src, 5);
988 PACK_AND_STORE(src, 6);
989 PACK_AND_STORE(src, 7);
990 PACK_AND_STORE(tempBlured, 0);
991 PACK_AND_STORE(tempBlured, 1);
992 PACK_AND_STORE(tempBlured, 2);
993 PACK_AND_STORE(tempBlured, 3);
994 PACK_AND_STORE(tempBlured, 4);
995 PACK_AND_STORE(tempBlured, 5);
996 PACK_AND_STORE(tempBlured, 6);
997 PACK_AND_STORE(tempBlured, 7);
998 #undef PACK_AND_STORE
999 }
1000
1001 static inline void transpose_16x8_char_toPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1002 const vector unsigned char zero = vec_splat_u8(0);
1003
1004 #define LOAD_DOUBLE_LINE(i, j) \
1005 vector unsigned char perm1##i = vec_lvsl(i * stride, src); \
1006 vector unsigned char perm2##i = vec_lvsl(j * stride, src); \
1007 vector unsigned char srcA##i = vec_ld(i * stride, src); \
1008 vector unsigned char srcB##i = vec_ld(i * stride + 16, src); \
1009 vector unsigned char srcC##i = vec_ld(j * stride, src); \
1010 vector unsigned char srcD##i = vec_ld(j * stride+ 16, src); \
1011 vector unsigned char src##i = vec_perm(srcA##i, srcB##i, perm1##i); \
1012 vector unsigned char src##j = vec_perm(srcC##i, srcD##i, perm2##i)
1013
1014 LOAD_DOUBLE_LINE(0, 1);
1015 LOAD_DOUBLE_LINE(2, 3);
1016 LOAD_DOUBLE_LINE(4, 5);
1017 LOAD_DOUBLE_LINE(6, 7);
1018 #undef LOAD_DOUBLE_LINE
1019
1020 vector unsigned char tempA = vec_mergeh(src0, zero);
1021 vector unsigned char tempB = vec_mergel(src0, zero);
1022 vector unsigned char tempC = vec_mergeh(src1, zero);
1023 vector unsigned char tempD = vec_mergel(src1, zero);
1024 vector unsigned char tempE = vec_mergeh(src2, zero);
1025 vector unsigned char tempF = vec_mergel(src2, zero);
1026 vector unsigned char tempG = vec_mergeh(src3, zero);
1027 vector unsigned char tempH = vec_mergel(src3, zero);
1028 vector unsigned char tempI = vec_mergeh(src4, zero);
1029 vector unsigned char tempJ = vec_mergel(src4, zero);
1030 vector unsigned char tempK = vec_mergeh(src5, zero);
1031 vector unsigned char tempL = vec_mergel(src5, zero);
1032 vector unsigned char tempM = vec_mergeh(src6, zero);
1033 vector unsigned char tempN = vec_mergel(src6, zero);
1034 vector unsigned char tempO = vec_mergeh(src7, zero);
1035 vector unsigned char tempP = vec_mergel(src7, zero);
1036
1037 vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1038 vector unsigned char temp1 = vec_mergel(tempA, tempI);
1039 vector unsigned char temp2 = vec_mergeh(tempB, tempJ);
1040 vector unsigned char temp3 = vec_mergel(tempB, tempJ);
1041 vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1042 vector unsigned char temp5 = vec_mergel(tempC, tempK);
1043 vector unsigned char temp6 = vec_mergeh(tempD, tempL);
1044 vector unsigned char temp7 = vec_mergel(tempD, tempL);
1045 vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1046 vector unsigned char temp9 = vec_mergel(tempE, tempM);
1047 vector unsigned char temp10 = vec_mergeh(tempF, tempN);
1048 vector unsigned char temp11 = vec_mergel(tempF, tempN);
1049 vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1050 vector unsigned char temp13 = vec_mergel(tempG, tempO);
1051 vector unsigned char temp14 = vec_mergeh(tempH, tempP);
1052 vector unsigned char temp15 = vec_mergel(tempH, tempP);
1053
1054 tempA = vec_mergeh(temp0, temp8);
1055 tempB = vec_mergel(temp0, temp8);
1056 tempC = vec_mergeh(temp1, temp9);
1057 tempD = vec_mergel(temp1, temp9);
1058 tempE = vec_mergeh(temp2, temp10);
1059 tempF = vec_mergel(temp2, temp10);
1060 tempG = vec_mergeh(temp3, temp11);
1061 tempH = vec_mergel(temp3, temp11);
1062 tempI = vec_mergeh(temp4, temp12);
1063 tempJ = vec_mergel(temp4, temp12);
1064 tempK = vec_mergeh(temp5, temp13);
1065 tempL = vec_mergel(temp5, temp13);
1066 tempM = vec_mergeh(temp6, temp14);
1067 tempN = vec_mergel(temp6, temp14);
1068 tempO = vec_mergeh(temp7, temp15);
1069 tempP = vec_mergel(temp7, temp15);
1070
1071 temp0 = vec_mergeh(tempA, tempI);
1072 temp1 = vec_mergel(tempA, tempI);
1073 temp2 = vec_mergeh(tempB, tempJ);
1074 temp3 = vec_mergel(tempB, tempJ);
1075 temp4 = vec_mergeh(tempC, tempK);
1076 temp5 = vec_mergel(tempC, tempK);
1077 temp6 = vec_mergeh(tempD, tempL);
1078 temp7 = vec_mergel(tempD, tempL);
1079 temp8 = vec_mergeh(tempE, tempM);
1080 temp9 = vec_mergel(tempE, tempM);
1081 temp10 = vec_mergeh(tempF, tempN);
1082 temp11 = vec_mergel(tempF, tempN);
1083 temp12 = vec_mergeh(tempG, tempO);
1084 temp13 = vec_mergel(tempG, tempO);
1085 temp14 = vec_mergeh(tempH, tempP);
1086 temp15 = vec_mergel(tempH, tempP);
1087
1088 vec_st(temp0, 0, dst);
1089 vec_st(temp1, 16, dst);
1090 vec_st(temp2, 32, dst);
1091 vec_st(temp3, 48, dst);
1092 vec_st(temp4, 64, dst);
1093 vec_st(temp5, 80, dst);
1094 vec_st(temp6, 96, dst);
1095 vec_st(temp7, 112, dst);
1096 vec_st(temp8, 128, dst);
1097 vec_st(temp9, 144, dst);
1098 vec_st(temp10, 160, dst);
1099 vec_st(temp11, 176, dst);
1100 vec_st(temp12, 192, dst);
1101 vec_st(temp13, 208, dst);
1102 vec_st(temp14, 224, dst);
1103 vec_st(temp15, 240, dst);
1104 }
1105
1106 static inline void transpose_8x16_char_fromPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) {
1107 const vector unsigned char zero = vec_splat_u8(0);
1108 const vector unsigned char magic_perm = (const vector unsigned char)
1109 AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
1110 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F);
1111
1112 #define LOAD_DOUBLE_LINE(i, j) \
1113 vector unsigned char src##i = vec_ld(i * 16, src); \
1114 vector unsigned char src##j = vec_ld(j * 16, src)
1115
1116 LOAD_DOUBLE_LINE(0, 1);
1117 LOAD_DOUBLE_LINE(2, 3);
1118 LOAD_DOUBLE_LINE(4, 5);
1119 LOAD_DOUBLE_LINE(6, 7);
1120 LOAD_DOUBLE_LINE(8, 9);
1121 LOAD_DOUBLE_LINE(10, 11);
1122 LOAD_DOUBLE_LINE(12, 13);
1123 LOAD_DOUBLE_LINE(14, 15);
1124 #undef LOAD_DOUBLE_LINE
1125
1126 vector unsigned char tempA = vec_mergeh(src0, src8);
1127 vector unsigned char tempB;
1128 vector unsigned char tempC = vec_mergeh(src1, src9);
1129 vector unsigned char tempD;
1130 vector unsigned char tempE = vec_mergeh(src2, src10);
1131 vector unsigned char tempG = vec_mergeh(src3, src11);
1132 vector unsigned char tempI = vec_mergeh(src4, src12);
1133 vector unsigned char tempJ;
1134 vector unsigned char tempK = vec_mergeh(src5, src13);
1135 vector unsigned char tempL;
1136 vector unsigned char tempM = vec_mergeh(src6, src14);
1137 vector unsigned char tempO = vec_mergeh(src7, src15);
1138
1139 vector unsigned char temp0 = vec_mergeh(tempA, tempI);
1140 vector unsigned char temp1 = vec_mergel(tempA, tempI);
1141 vector unsigned char temp2;
1142 vector unsigned char temp3;
1143 vector unsigned char temp4 = vec_mergeh(tempC, tempK);
1144 vector unsigned char temp5 = vec_mergel(tempC, tempK);
1145 vector unsigned char temp6;
1146 vector unsigned char temp7;
1147 vector unsigned char temp8 = vec_mergeh(tempE, tempM);
1148 vector unsigned char temp9 = vec_mergel(tempE, tempM);
1149 vector unsigned char temp12 = vec_mergeh(tempG, tempO);
1150 vector unsigned char temp13 = vec_mergel(tempG, tempO);
1151
1152 tempA = vec_mergeh(temp0, temp8);
1153 tempB = vec_mergel(temp0, temp8);
1154 tempC = vec_mergeh(temp1, temp9);
1155 tempD = vec_mergel(temp1, temp9);
1156 tempI = vec_mergeh(temp4, temp12);
1157 tempJ = vec_mergel(temp4, temp12);
1158 tempK = vec_mergeh(temp5, temp13);
1159 tempL = vec_mergel(temp5, temp13);
1160
1161 temp0 = vec_mergeh(tempA, tempI);
1162 temp1 = vec_mergel(tempA, tempI);
1163 temp2 = vec_mergeh(tempB, tempJ);
1164 temp3 = vec_mergel(tempB, tempJ);
1165 temp4 = vec_mergeh(tempC, tempK);
1166 temp5 = vec_mergel(tempC, tempK);
1167 temp6 = vec_mergeh(tempD, tempL);
1168 temp7 = vec_mergel(tempD, tempL);
1169
1170
1171 const vector signed char neg1 = vec_splat_s8(-1);
1172 #define STORE_DOUBLE_LINE(i, j) \
1173 vector unsigned char dstA##i = vec_ld(i * stride, dst); \
1174 vector unsigned char dstB##i = vec_ld(i * stride + 16, dst); \
1175 vector unsigned char dstA##j = vec_ld(j * stride, dst); \
1176 vector unsigned char dstB##j = vec_ld(j * stride+ 16, dst); \
1177 vector unsigned char align##i = vec_lvsr(i * stride, dst); \
1178 vector unsigned char align##j = vec_lvsr(j * stride, dst); \
1179 vector unsigned char mask##i = vec_perm(zero, (vector unsigned char)neg1, align##i); \
1180 vector unsigned char mask##j = vec_perm(zero, (vector unsigned char)neg1, align##j); \
1181 vector unsigned char dstR##i = vec_perm(temp##i, temp##i, align##i); \
1182 vector unsigned char dstR##j = vec_perm(temp##j, temp##j, align##j); \
1183 vector unsigned char dstAF##i = vec_sel(dstA##i, dstR##i, mask##i); \
1184 vector unsigned char dstBF##i = vec_sel(dstR##i, dstB##i, mask##i); \
1185 vector unsigned char dstAF##j = vec_sel(dstA##j, dstR##j, mask##j); \
1186 vector unsigned char dstBF##j = vec_sel(dstR##j, dstB##j, mask##j); \
1187 vec_st(dstAF##i, i * stride, dst); \
1188 vec_st(dstBF##i, i * stride + 16, dst); \
1189 vec_st(dstAF##j, j * stride, dst); \
1190 vec_st(dstBF##j, j * stride + 16, dst)
1191
1192 STORE_DOUBLE_LINE(0,1);
1193 STORE_DOUBLE_LINE(2,3);
1194 STORE_DOUBLE_LINE(4,5);
1195 STORE_DOUBLE_LINE(6,7);
1196 }