2 * Copyright (c) 2002 Brian Foley
3 * Copyright (c) 2002 Dieter Shirley
4 * Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
6 * This file is part of FFmpeg.
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (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 GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
27 #include "libavcodec/dsputil.h"
28 #include "dsputil_ppc.h"
29 #include "util_altivec.h"
30 #include "types_altivec.h"
32 int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
35 DECLARE_ALIGNED_16(int, s);
36 const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
37 vector unsigned char *tv;
38 vector unsigned char pix1v, pix2v, pix2iv, avgv, t5;
39 vector unsigned int sad;
40 vector signed int sumdiffs;
43 sad = (vector unsigned int)vec_splat_u32(0);
44 for (i = 0; i < h; i++) {
45 /* Read unaligned pixels into our vectors. The vectors are as follows:
46 pix1v: pix1[0]-pix1[15]
47 pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16] */
48 tv = (vector unsigned char *) pix1;
49 pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
51 tv = (vector unsigned char *) &pix2[0];
52 pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
54 tv = (vector unsigned char *) &pix2[1];
55 pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1]));
57 /* Calculate the average vector */
58 avgv = vec_avg(pix2v, pix2iv);
60 /* Calculate a sum of abs differences vector */
61 t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
63 /* Add each 4 pixel group together and put 4 results into sad */
64 sad = vec_sum4s(t5, sad);
69 /* Sum up the four partial sums, and put the result into s */
70 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
71 sumdiffs = vec_splat(sumdiffs, 3);
72 vec_ste(sumdiffs, 0, &s);
77 int sad16_y2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
80 DECLARE_ALIGNED_16(int, s);
81 const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
82 vector unsigned char *tv;
83 vector unsigned char pix1v, pix2v, pix3v, avgv, t5;
84 vector unsigned int sad;
85 vector signed int sumdiffs;
86 uint8_t *pix3 = pix2 + line_size;
89 sad = (vector unsigned int)vec_splat_u32(0);
91 /* Due to the fact that pix3 = pix2 + line_size, the pix3 of one
92 iteration becomes pix2 in the next iteration. We can use this
93 fact to avoid a potentially expensive unaligned read, each
95 Read unaligned pixels into our vectors. The vectors are as follows:
96 pix2v: pix2[0]-pix2[15]
97 Split the pixel vectors into shorts */
98 tv = (vector unsigned char *) &pix2[0];
99 pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
101 for (i = 0; i < h; i++) {
102 /* Read unaligned pixels into our vectors. The vectors are as follows:
103 pix1v: pix1[0]-pix1[15]
104 pix3v: pix3[0]-pix3[15] */
105 tv = (vector unsigned char *) pix1;
106 pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
108 tv = (vector unsigned char *) &pix3[0];
109 pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0]));
111 /* Calculate the average vector */
112 avgv = vec_avg(pix2v, pix3v);
114 /* Calculate a sum of abs differences vector */
115 t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
117 /* Add each 4 pixel group together and put 4 results into sad */
118 sad = vec_sum4s(t5, sad);
126 /* Sum up the four partial sums, and put the result into s */
127 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
128 sumdiffs = vec_splat(sumdiffs, 3);
129 vec_ste(sumdiffs, 0, &s);
133 int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
136 DECLARE_ALIGNED_16(int, s);
137 uint8_t *pix3 = pix2 + line_size;
138 const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
139 const vector unsigned short two = (const vector unsigned short)vec_splat_u16(2);
140 vector unsigned char *tv, avgv, t5;
141 vector unsigned char pix1v, pix2v, pix3v, pix2iv, pix3iv;
142 vector unsigned short pix2lv, pix2hv, pix2ilv, pix2ihv;
143 vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
144 vector unsigned short avghv, avglv;
145 vector unsigned short t1, t2, t3, t4;
146 vector unsigned int sad;
147 vector signed int sumdiffs;
149 sad = (vector unsigned int)vec_splat_u32(0);
153 /* Due to the fact that pix3 = pix2 + line_size, the pix3 of one
154 iteration becomes pix2 in the next iteration. We can use this
155 fact to avoid a potentially expensive unaligned read, as well
156 as some splitting, and vector addition each time around the loop.
157 Read unaligned pixels into our vectors. The vectors are as follows:
158 pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16]
159 Split the pixel vectors into shorts */
160 tv = (vector unsigned char *) &pix2[0];
161 pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
163 tv = (vector unsigned char *) &pix2[1];
164 pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1]));
166 pix2hv = (vector unsigned short) vec_mergeh(zero, pix2v);
167 pix2lv = (vector unsigned short) vec_mergel(zero, pix2v);
168 pix2ihv = (vector unsigned short) vec_mergeh(zero, pix2iv);
169 pix2ilv = (vector unsigned short) vec_mergel(zero, pix2iv);
170 t1 = vec_add(pix2hv, pix2ihv);
171 t2 = vec_add(pix2lv, pix2ilv);
173 for (i = 0; i < h; i++) {
174 /* Read unaligned pixels into our vectors. The vectors are as follows:
175 pix1v: pix1[0]-pix1[15]
176 pix3v: pix3[0]-pix3[15] pix3iv: pix3[1]-pix3[16] */
177 tv = (vector unsigned char *) pix1;
178 pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
180 tv = (vector unsigned char *) &pix3[0];
181 pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0]));
183 tv = (vector unsigned char *) &pix3[1];
184 pix3iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[1]));
186 /* Note that AltiVec does have vec_avg, but this works on vector pairs
187 and rounds up. We could do avg(avg(a,b),avg(c,d)), but the rounding
188 would mean that, for example, avg(3,0,0,1) = 2, when it should be 1.
189 Instead, we have to split the pixel vectors into vectors of shorts,
190 and do the averaging by hand. */
192 /* Split the pixel vectors into shorts */
193 pix3hv = (vector unsigned short) vec_mergeh(zero, pix3v);
194 pix3lv = (vector unsigned short) vec_mergel(zero, pix3v);
195 pix3ihv = (vector unsigned short) vec_mergeh(zero, pix3iv);
196 pix3ilv = (vector unsigned short) vec_mergel(zero, pix3iv);
198 /* Do the averaging on them */
199 t3 = vec_add(pix3hv, pix3ihv);
200 t4 = vec_add(pix3lv, pix3ilv);
202 avghv = vec_sr(vec_add(vec_add(t1, t3), two), two);
203 avglv = vec_sr(vec_add(vec_add(t2, t4), two), two);
205 /* Pack the shorts back into a result */
206 avgv = vec_pack(avghv, avglv);
208 /* Calculate a sum of abs differences vector */
209 t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
211 /* Add each 4 pixel group together and put 4 results into sad */
212 sad = vec_sum4s(t5, sad);
216 /* Transfer the calculated values for pix3 into pix2 */
220 /* Sum up the four partial sums, and put the result into s */
221 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
222 sumdiffs = vec_splat(sumdiffs, 3);
223 vec_ste(sumdiffs, 0, &s);
228 int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
231 DECLARE_ALIGNED_16(int, s);
232 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
233 vector unsigned char perm1, perm2, *pix1v, *pix2v;
234 vector unsigned char t1, t2, t3,t4, t5;
235 vector unsigned int sad;
236 vector signed int sumdiffs;
238 sad = (vector unsigned int)vec_splat_u32(0);
241 for (i = 0; i < h; i++) {
242 /* Read potentially unaligned pixels into t1 and t2 */
243 perm1 = vec_lvsl(0, pix1);
244 pix1v = (vector unsigned char *) pix1;
245 perm2 = vec_lvsl(0, pix2);
246 pix2v = (vector unsigned char *) pix2;
247 t1 = vec_perm(pix1v[0], pix1v[1], perm1);
248 t2 = vec_perm(pix2v[0], pix2v[1], perm2);
250 /* Calculate a sum of abs differences vector */
251 t3 = vec_max(t1, t2);
252 t4 = vec_min(t1, t2);
253 t5 = vec_sub(t3, t4);
255 /* Add each 4 pixel group together and put 4 results into sad */
256 sad = vec_sum4s(t5, sad);
262 /* Sum up the four partial sums, and put the result into s */
263 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
264 sumdiffs = vec_splat(sumdiffs, 3);
265 vec_ste(sumdiffs, 0, &s);
270 int sad8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
273 DECLARE_ALIGNED_16(int, s);
274 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
275 vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
276 vector unsigned char t1, t2, t3,t4, t5;
277 vector unsigned int sad;
278 vector signed int sumdiffs;
280 sad = (vector unsigned int)vec_splat_u32(0);
282 permclear = (vector unsigned char){255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0};
284 for (i = 0; i < h; i++) {
285 /* Read potentially unaligned pixels into t1 and t2
286 Since we're reading 16 pixels, and actually only want 8,
287 mask out the last 8 pixels. The 0s don't change the sum. */
288 perm1 = vec_lvsl(0, pix1);
289 pix1v = (vector unsigned char *) pix1;
290 perm2 = vec_lvsl(0, pix2);
291 pix2v = (vector unsigned char *) pix2;
292 t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
293 t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
295 /* Calculate a sum of abs differences vector */
296 t3 = vec_max(t1, t2);
297 t4 = vec_min(t1, t2);
298 t5 = vec_sub(t3, t4);
300 /* Add each 4 pixel group together and put 4 results into sad */
301 sad = vec_sum4s(t5, sad);
307 /* Sum up the four partial sums, and put the result into s */
308 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
309 sumdiffs = vec_splat(sumdiffs, 3);
310 vec_ste(sumdiffs, 0, &s);
315 int pix_norm1_altivec(uint8_t *pix, int line_size)
318 DECLARE_ALIGNED_16(int, s);
319 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
320 vector unsigned char *tv;
321 vector unsigned char pixv;
322 vector unsigned int sv;
323 vector signed int sum;
325 sv = (vector unsigned int)vec_splat_u32(0);
328 for (i = 0; i < 16; i++) {
329 /* Read in the potentially unaligned pixels */
330 tv = (vector unsigned char *) pix;
331 pixv = vec_perm(tv[0], tv[1], vec_lvsl(0, pix));
333 /* Square the values, and add them to our sum */
334 sv = vec_msum(pixv, pixv, sv);
338 /* Sum up the four partial sums, and put the result into s */
339 sum = vec_sums((vector signed int) sv, (vector signed int) zero);
340 sum = vec_splat(sum, 3);
347 * Sum of Squared Errors for a 8x8 block.
349 * It's the sad8_altivec code above w/ squaring added.
351 int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
354 DECLARE_ALIGNED_16(int, s);
355 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
356 vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
357 vector unsigned char t1, t2, t3,t4, t5;
358 vector unsigned int sum;
359 vector signed int sumsqr;
361 sum = (vector unsigned int)vec_splat_u32(0);
363 permclear = (vector unsigned char){255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0};
366 for (i = 0; i < h; i++) {
367 /* Read potentially unaligned pixels into t1 and t2
368 Since we're reading 16 pixels, and actually only want 8,
369 mask out the last 8 pixels. The 0s don't change the sum. */
370 perm1 = vec_lvsl(0, pix1);
371 pix1v = (vector unsigned char *) pix1;
372 perm2 = vec_lvsl(0, pix2);
373 pix2v = (vector unsigned char *) pix2;
374 t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
375 t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
377 /* Since we want to use unsigned chars, we can take advantage
378 of the fact that abs(a-b)^2 = (a-b)^2. */
380 /* Calculate abs differences vector */
381 t3 = vec_max(t1, t2);
382 t4 = vec_min(t1, t2);
383 t5 = vec_sub(t3, t4);
385 /* Square the values and add them to our sum */
386 sum = vec_msum(t5, t5, sum);
392 /* Sum up the four partial sums, and put the result into s */
393 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
394 sumsqr = vec_splat(sumsqr, 3);
395 vec_ste(sumsqr, 0, &s);
401 * Sum of Squared Errors for a 16x16 block.
403 * It's the sad16_altivec code above w/ squaring added.
405 int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
408 DECLARE_ALIGNED_16(int, s);
409 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
410 vector unsigned char perm1, perm2, *pix1v, *pix2v;
411 vector unsigned char t1, t2, t3,t4, t5;
412 vector unsigned int sum;
413 vector signed int sumsqr;
415 sum = (vector unsigned int)vec_splat_u32(0);
417 for (i = 0; i < h; i++) {
418 /* Read potentially unaligned pixels into t1 and t2 */
419 perm1 = vec_lvsl(0, pix1);
420 pix1v = (vector unsigned char *) pix1;
421 perm2 = vec_lvsl(0, pix2);
422 pix2v = (vector unsigned char *) pix2;
423 t1 = vec_perm(pix1v[0], pix1v[1], perm1);
424 t2 = vec_perm(pix2v[0], pix2v[1], perm2);
426 /* Since we want to use unsigned chars, we can take advantage
427 of the fact that abs(a-b)^2 = (a-b)^2. */
429 /* Calculate abs differences vector */
430 t3 = vec_max(t1, t2);
431 t4 = vec_min(t1, t2);
432 t5 = vec_sub(t3, t4);
434 /* Square the values and add them to our sum */
435 sum = vec_msum(t5, t5, sum);
441 /* Sum up the four partial sums, and put the result into s */
442 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
443 sumsqr = vec_splat(sumsqr, 3);
444 vec_ste(sumsqr, 0, &s);
449 int pix_sum_altivec(uint8_t * pix, int line_size)
451 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
452 vector unsigned char perm, *pixv;
453 vector unsigned char t1;
454 vector unsigned int sad;
455 vector signed int sumdiffs;
458 DECLARE_ALIGNED_16(int, s);
460 sad = (vector unsigned int)vec_splat_u32(0);
462 for (i = 0; i < 16; i++) {
463 /* Read the potentially unaligned 16 pixels into t1 */
464 perm = vec_lvsl(0, pix);
465 pixv = (vector unsigned char *) pix;
466 t1 = vec_perm(pixv[0], pixv[1], perm);
468 /* Add each 4 pixel group together and put 4 results into sad */
469 sad = vec_sum4s(t1, sad);
474 /* Sum up the four partial sums, and put the result into s */
475 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
476 sumdiffs = vec_splat(sumdiffs, 3);
477 vec_ste(sumdiffs, 0, &s);
482 void get_pixels_altivec(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
485 vector unsigned char perm, bytes, *pixv;
486 const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
487 vector signed short shorts;
489 for (i = 0; i < 8; i++) {
490 // Read potentially unaligned pixels.
491 // We're reading 16 pixels, and actually only want 8,
492 // but we simply ignore the extras.
493 perm = vec_lvsl(0, pixels);
494 pixv = (vector unsigned char *) pixels;
495 bytes = vec_perm(pixv[0], pixv[1], perm);
497 // convert the bytes into shorts
498 shorts = (vector signed short)vec_mergeh(zero, bytes);
500 // save the data to the block, we assume the block is 16-byte aligned
501 vec_st(shorts, i*16, (vector signed short*)block);
507 void diff_pixels_altivec(DCTELEM *restrict block, const uint8_t *s1,
508 const uint8_t *s2, int stride)
511 vector unsigned char perm, bytes, *pixv;
512 const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
513 vector signed short shorts1, shorts2;
515 for (i = 0; i < 4; i++) {
516 // Read potentially unaligned pixels
517 // We're reading 16 pixels, and actually only want 8,
518 // but we simply ignore the extras.
519 perm = vec_lvsl(0, s1);
520 pixv = (vector unsigned char *) s1;
521 bytes = vec_perm(pixv[0], pixv[1], perm);
523 // convert the bytes into shorts
524 shorts1 = (vector signed short)vec_mergeh(zero, bytes);
526 // Do the same for the second block of pixels
527 perm = vec_lvsl(0, s2);
528 pixv = (vector unsigned char *) s2;
529 bytes = vec_perm(pixv[0], pixv[1], perm);
531 // convert the bytes into shorts
532 shorts2 = (vector signed short)vec_mergeh(zero, bytes);
534 // Do the subtraction
535 shorts1 = vec_sub(shorts1, shorts2);
537 // save the data to the block, we assume the block is 16-byte aligned
538 vec_st(shorts1, 0, (vector signed short*)block);
545 // The code below is a copy of the code above... This is a manual
548 // Read potentially unaligned pixels
549 // We're reading 16 pixels, and actually only want 8,
550 // but we simply ignore the extras.
551 perm = vec_lvsl(0, s1);
552 pixv = (vector unsigned char *) s1;
553 bytes = vec_perm(pixv[0], pixv[1], perm);
555 // convert the bytes into shorts
556 shorts1 = (vector signed short)vec_mergeh(zero, bytes);
558 // Do the same for the second block of pixels
559 perm = vec_lvsl(0, s2);
560 pixv = (vector unsigned char *) s2;
561 bytes = vec_perm(pixv[0], pixv[1], perm);
563 // convert the bytes into shorts
564 shorts2 = (vector signed short)vec_mergeh(zero, bytes);
566 // Do the subtraction
567 shorts1 = vec_sub(shorts1, shorts2);
569 // save the data to the block, we assume the block is 16-byte aligned
570 vec_st(shorts1, 0, (vector signed short*)block);
579 static void clear_block_altivec(DCTELEM *block) {
581 vec_st(zero_s16v, 0, block);
582 vec_st(zero_s16v, 16, block);
583 vec_st(zero_s16v, 32, block);
584 vec_st(zero_s16v, 48, block);
585 vec_st(zero_s16v, 64, block);
586 vec_st(zero_s16v, 80, block);
587 vec_st(zero_s16v, 96, block);
588 vec_st(zero_s16v, 112, block);
592 void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {
594 register vector unsigned char vdst, vsrc;
596 /* dst and src are 16 bytes-aligned (guaranteed) */
597 for (i = 0 ; (i + 15) < w ; i+=16) {
598 vdst = vec_ld(i, (unsigned char*)dst);
599 vsrc = vec_ld(i, (unsigned char*)src);
600 vdst = vec_add(vsrc, vdst);
601 vec_st(vdst, i, (unsigned char*)dst);
603 /* if w is not a multiple of 16 */
604 for (; (i < w) ; i++) {
609 /* next one assumes that ((line_size % 16) == 0) */
610 void put_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
612 POWERPC_PERF_DECLARE(altivec_put_pixels16_num, 1);
613 register vector unsigned char pixelsv1, pixelsv2;
614 register vector unsigned char pixelsv1B, pixelsv2B;
615 register vector unsigned char pixelsv1C, pixelsv2C;
616 register vector unsigned char pixelsv1D, pixelsv2D;
618 register vector unsigned char perm = vec_lvsl(0, pixels);
620 register int line_size_2 = line_size << 1;
621 register int line_size_3 = line_size + line_size_2;
622 register int line_size_4 = line_size << 2;
624 POWERPC_PERF_START_COUNT(altivec_put_pixels16_num, 1);
625 // hand-unrolling the loop by 4 gains about 15%
626 // mininum execution time goes from 74 to 60 cycles
627 // it's faster than -funroll-loops, but using
628 // -funroll-loops w/ this is bad - 74 cycles again.
629 // all this is on a 7450, tuning for the 7450
631 for (i = 0; i < h; i++) {
632 pixelsv1 = vec_ld(0, (unsigned char*)pixels);
633 pixelsv2 = vec_ld(16, (unsigned char*)pixels);
634 vec_st(vec_perm(pixelsv1, pixelsv2, perm),
635 0, (unsigned char*)block);
640 for (i = 0; i < h; i += 4) {
641 pixelsv1 = vec_ld(0, (unsigned char*)pixels);
642 pixelsv2 = vec_ld(15, (unsigned char*)pixels);
643 pixelsv1B = vec_ld(line_size, (unsigned char*)pixels);
644 pixelsv2B = vec_ld(15 + line_size, (unsigned char*)pixels);
645 pixelsv1C = vec_ld(line_size_2, (unsigned char*)pixels);
646 pixelsv2C = vec_ld(15 + line_size_2, (unsigned char*)pixels);
647 pixelsv1D = vec_ld(line_size_3, (unsigned char*)pixels);
648 pixelsv2D = vec_ld(15 + line_size_3, (unsigned char*)pixels);
649 vec_st(vec_perm(pixelsv1, pixelsv2, perm),
650 0, (unsigned char*)block);
651 vec_st(vec_perm(pixelsv1B, pixelsv2B, perm),
652 line_size, (unsigned char*)block);
653 vec_st(vec_perm(pixelsv1C, pixelsv2C, perm),
654 line_size_2, (unsigned char*)block);
655 vec_st(vec_perm(pixelsv1D, pixelsv2D, perm),
656 line_size_3, (unsigned char*)block);
661 POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_num, 1);
664 /* next one assumes that ((line_size % 16) == 0) */
665 #define op_avg(a,b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEUL)>>1) )
666 void avg_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
668 POWERPC_PERF_DECLARE(altivec_avg_pixels16_num, 1);
669 register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
670 register vector unsigned char perm = vec_lvsl(0, pixels);
673 POWERPC_PERF_START_COUNT(altivec_avg_pixels16_num, 1);
675 for (i = 0; i < h; i++) {
676 pixelsv1 = vec_ld(0, (unsigned char*)pixels);
677 pixelsv2 = vec_ld(16, (unsigned char*)pixels);
678 blockv = vec_ld(0, block);
679 pixelsv = vec_perm(pixelsv1, pixelsv2, perm);
680 blockv = vec_avg(blockv,pixelsv);
681 vec_st(blockv, 0, (unsigned char*)block);
686 POWERPC_PERF_STOP_COUNT(altivec_avg_pixels16_num, 1);
689 /* next one assumes that ((line_size % 8) == 0) */
690 void avg_pixels8_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
692 POWERPC_PERF_DECLARE(altivec_avg_pixels8_num, 1);
693 register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
696 POWERPC_PERF_START_COUNT(altivec_avg_pixels8_num, 1);
698 for (i = 0; i < h; i++) {
699 /* block is 8 bytes-aligned, so we're either in the
700 left block (16 bytes-aligned) or in the right block (not) */
701 int rightside = ((unsigned long)block & 0x0000000F);
703 blockv = vec_ld(0, block);
704 pixelsv1 = vec_ld(0, (unsigned char*)pixels);
705 pixelsv2 = vec_ld(16, (unsigned char*)pixels);
706 pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels));
709 pixelsv = vec_perm(blockv, pixelsv, vcprm(0,1,s0,s1));
711 pixelsv = vec_perm(blockv, pixelsv, vcprm(s0,s1,2,3));
714 blockv = vec_avg(blockv, pixelsv);
716 vec_st(blockv, 0, block);
722 POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_num, 1);
725 /* next one assumes that ((line_size % 8) == 0) */
726 void put_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
728 POWERPC_PERF_DECLARE(altivec_put_pixels8_xy2_num, 1);
730 register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
731 register vector unsigned char blockv, temp1, temp2;
732 register vector unsigned short pixelssum1, pixelssum2, temp3;
733 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
734 register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
736 temp1 = vec_ld(0, pixels);
737 temp2 = vec_ld(16, pixels);
738 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
739 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
742 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
744 pixelsv1 = vec_mergeh(vczero, pixelsv1);
745 pixelsv2 = vec_mergeh(vczero, pixelsv2);
746 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
747 (vector unsigned short)pixelsv2);
748 pixelssum1 = vec_add(pixelssum1, vctwo);
750 POWERPC_PERF_START_COUNT(altivec_put_pixels8_xy2_num, 1);
751 for (i = 0; i < h ; i++) {
752 int rightside = ((unsigned long)block & 0x0000000F);
753 blockv = vec_ld(0, block);
755 temp1 = vec_ld(line_size, pixels);
756 temp2 = vec_ld(line_size + 16, pixels);
757 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
758 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
761 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
764 pixelsv1 = vec_mergeh(vczero, pixelsv1);
765 pixelsv2 = vec_mergeh(vczero, pixelsv2);
766 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
767 (vector unsigned short)pixelsv2);
768 temp3 = vec_add(pixelssum1, pixelssum2);
769 temp3 = vec_sra(temp3, vctwo);
770 pixelssum1 = vec_add(pixelssum2, vctwo);
771 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
774 blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
776 blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
779 vec_st(blockv, 0, block);
785 POWERPC_PERF_STOP_COUNT(altivec_put_pixels8_xy2_num, 1);
788 /* next one assumes that ((line_size % 8) == 0) */
789 void put_no_rnd_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
791 POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels8_xy2_num, 1);
793 register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
794 register vector unsigned char blockv, temp1, temp2;
795 register vector unsigned short pixelssum1, pixelssum2, temp3;
796 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
797 register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
798 register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
800 temp1 = vec_ld(0, pixels);
801 temp2 = vec_ld(16, pixels);
802 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
803 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
806 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
808 pixelsv1 = vec_mergeh(vczero, pixelsv1);
809 pixelsv2 = vec_mergeh(vczero, pixelsv2);
810 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
811 (vector unsigned short)pixelsv2);
812 pixelssum1 = vec_add(pixelssum1, vcone);
814 POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
815 for (i = 0; i < h ; i++) {
816 int rightside = ((unsigned long)block & 0x0000000F);
817 blockv = vec_ld(0, block);
819 temp1 = vec_ld(line_size, pixels);
820 temp2 = vec_ld(line_size + 16, pixels);
821 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
822 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
825 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
828 pixelsv1 = vec_mergeh(vczero, pixelsv1);
829 pixelsv2 = vec_mergeh(vczero, pixelsv2);
830 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
831 (vector unsigned short)pixelsv2);
832 temp3 = vec_add(pixelssum1, pixelssum2);
833 temp3 = vec_sra(temp3, vctwo);
834 pixelssum1 = vec_add(pixelssum2, vcone);
835 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
838 blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
840 blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
843 vec_st(blockv, 0, block);
849 POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
852 /* next one assumes that ((line_size % 16) == 0) */
853 void put_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
855 POWERPC_PERF_DECLARE(altivec_put_pixels16_xy2_num, 1);
857 register vector unsigned char pixelsv1, pixelsv2, pixelsv3, pixelsv4;
858 register vector unsigned char blockv, temp1, temp2;
859 register vector unsigned short temp3, temp4,
860 pixelssum1, pixelssum2, pixelssum3, pixelssum4;
861 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
862 register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
864 POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1);
866 temp1 = vec_ld(0, pixels);
867 temp2 = vec_ld(16, pixels);
868 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
869 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
872 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
874 pixelsv3 = vec_mergel(vczero, pixelsv1);
875 pixelsv4 = vec_mergel(vczero, pixelsv2);
876 pixelsv1 = vec_mergeh(vczero, pixelsv1);
877 pixelsv2 = vec_mergeh(vczero, pixelsv2);
878 pixelssum3 = vec_add((vector unsigned short)pixelsv3,
879 (vector unsigned short)pixelsv4);
880 pixelssum3 = vec_add(pixelssum3, vctwo);
881 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
882 (vector unsigned short)pixelsv2);
883 pixelssum1 = vec_add(pixelssum1, vctwo);
885 for (i = 0; i < h ; i++) {
886 blockv = vec_ld(0, block);
888 temp1 = vec_ld(line_size, pixels);
889 temp2 = vec_ld(line_size + 16, pixels);
890 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
891 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
894 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
897 pixelsv3 = vec_mergel(vczero, pixelsv1);
898 pixelsv4 = vec_mergel(vczero, pixelsv2);
899 pixelsv1 = vec_mergeh(vczero, pixelsv1);
900 pixelsv2 = vec_mergeh(vczero, pixelsv2);
902 pixelssum4 = vec_add((vector unsigned short)pixelsv3,
903 (vector unsigned short)pixelsv4);
904 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
905 (vector unsigned short)pixelsv2);
906 temp4 = vec_add(pixelssum3, pixelssum4);
907 temp4 = vec_sra(temp4, vctwo);
908 temp3 = vec_add(pixelssum1, pixelssum2);
909 temp3 = vec_sra(temp3, vctwo);
911 pixelssum3 = vec_add(pixelssum4, vctwo);
912 pixelssum1 = vec_add(pixelssum2, vctwo);
914 blockv = vec_packsu(temp3, temp4);
916 vec_st(blockv, 0, block);
922 POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_xy2_num, 1);
925 /* next one assumes that ((line_size % 16) == 0) */
926 void put_no_rnd_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
928 POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels16_xy2_num, 1);
930 register vector unsigned char pixelsv1, pixelsv2, pixelsv3, pixelsv4;
931 register vector unsigned char blockv, temp1, temp2;
932 register vector unsigned short temp3, temp4,
933 pixelssum1, pixelssum2, pixelssum3, pixelssum4;
934 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
935 register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
936 register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
938 POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
940 temp1 = vec_ld(0, pixels);
941 temp2 = vec_ld(16, pixels);
942 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
943 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
946 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
948 pixelsv3 = vec_mergel(vczero, pixelsv1);
949 pixelsv4 = vec_mergel(vczero, pixelsv2);
950 pixelsv1 = vec_mergeh(vczero, pixelsv1);
951 pixelsv2 = vec_mergeh(vczero, pixelsv2);
952 pixelssum3 = vec_add((vector unsigned short)pixelsv3,
953 (vector unsigned short)pixelsv4);
954 pixelssum3 = vec_add(pixelssum3, vcone);
955 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
956 (vector unsigned short)pixelsv2);
957 pixelssum1 = vec_add(pixelssum1, vcone);
959 for (i = 0; i < h ; i++) {
960 blockv = vec_ld(0, block);
962 temp1 = vec_ld(line_size, pixels);
963 temp2 = vec_ld(line_size + 16, pixels);
964 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
965 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
968 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
971 pixelsv3 = vec_mergel(vczero, pixelsv1);
972 pixelsv4 = vec_mergel(vczero, pixelsv2);
973 pixelsv1 = vec_mergeh(vczero, pixelsv1);
974 pixelsv2 = vec_mergeh(vczero, pixelsv2);
976 pixelssum4 = vec_add((vector unsigned short)pixelsv3,
977 (vector unsigned short)pixelsv4);
978 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
979 (vector unsigned short)pixelsv2);
980 temp4 = vec_add(pixelssum3, pixelssum4);
981 temp4 = vec_sra(temp4, vctwo);
982 temp3 = vec_add(pixelssum1, pixelssum2);
983 temp3 = vec_sra(temp3, vctwo);
985 pixelssum3 = vec_add(pixelssum4, vcone);
986 pixelssum1 = vec_add(pixelssum2, vcone);
988 blockv = vec_packsu(temp3, temp4);
990 vec_st(blockv, 0, block);
996 POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
999 int hadamard8_diff8x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
1000 POWERPC_PERF_DECLARE(altivec_hadamard8_diff8x8_num, 1);
1002 register const vector unsigned char vzero =
1003 (const vector unsigned char)vec_splat_u8(0);
1004 register vector signed short temp0, temp1, temp2, temp3, temp4,
1005 temp5, temp6, temp7;
1006 POWERPC_PERF_START_COUNT(altivec_hadamard8_diff8x8_num, 1);
1008 register const vector signed short vprod1 =(const vector signed short)
1009 { 1,-1, 1,-1, 1,-1, 1,-1 };
1010 register const vector signed short vprod2 =(const vector signed short)
1011 { 1, 1,-1,-1, 1, 1,-1,-1 };
1012 register const vector signed short vprod3 =(const vector signed short)
1013 { 1, 1, 1, 1,-1,-1,-1,-1 };
1014 register const vector unsigned char perm1 = (const vector unsigned char)
1015 {0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
1016 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D};
1017 register const vector unsigned char perm2 = (const vector unsigned char)
1018 {0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
1019 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B};
1020 register const vector unsigned char perm3 = (const vector unsigned char)
1021 {0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
1022 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07};
1024 #define ONEITERBUTTERFLY(i, res) \
1026 register vector unsigned char src1, src2, srcO; \
1027 register vector unsigned char dst1, dst2, dstO; \
1028 register vector signed short srcV, dstV; \
1029 register vector signed short but0, but1, but2, op1, op2, op3; \
1030 src1 = vec_ld(stride * i, src); \
1031 src2 = vec_ld((stride * i) + 15, src); \
1032 srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
1033 dst1 = vec_ld(stride * i, dst); \
1034 dst2 = vec_ld((stride * i) + 15, dst); \
1035 dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
1036 /* promote the unsigned chars to signed shorts */ \
1037 /* we're in the 8x8 function, we only care for the first 8 */ \
1038 srcV = (vector signed short)vec_mergeh((vector signed char)vzero, \
1039 (vector signed char)srcO); \
1040 dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \
1041 (vector signed char)dstO); \
1042 /* subtractions inside the first butterfly */ \
1043 but0 = vec_sub(srcV, dstV); \
1044 op1 = vec_perm(but0, but0, perm1); \
1045 but1 = vec_mladd(but0, vprod1, op1); \
1046 op2 = vec_perm(but1, but1, perm2); \
1047 but2 = vec_mladd(but1, vprod2, op2); \
1048 op3 = vec_perm(but2, but2, perm3); \
1049 res = vec_mladd(but2, vprod3, op3); \
1051 ONEITERBUTTERFLY(0, temp0);
1052 ONEITERBUTTERFLY(1, temp1);
1053 ONEITERBUTTERFLY(2, temp2);
1054 ONEITERBUTTERFLY(3, temp3);
1055 ONEITERBUTTERFLY(4, temp4);
1056 ONEITERBUTTERFLY(5, temp5);
1057 ONEITERBUTTERFLY(6, temp6);
1058 ONEITERBUTTERFLY(7, temp7);
1060 #undef ONEITERBUTTERFLY
1062 register vector signed int vsum;
1063 register vector signed short line0 = vec_add(temp0, temp1);
1064 register vector signed short line1 = vec_sub(temp0, temp1);
1065 register vector signed short line2 = vec_add(temp2, temp3);
1066 register vector signed short line3 = vec_sub(temp2, temp3);
1067 register vector signed short line4 = vec_add(temp4, temp5);
1068 register vector signed short line5 = vec_sub(temp4, temp5);
1069 register vector signed short line6 = vec_add(temp6, temp7);
1070 register vector signed short line7 = vec_sub(temp6, temp7);
1072 register vector signed short line0B = vec_add(line0, line2);
1073 register vector signed short line2B = vec_sub(line0, line2);
1074 register vector signed short line1B = vec_add(line1, line3);
1075 register vector signed short line3B = vec_sub(line1, line3);
1076 register vector signed short line4B = vec_add(line4, line6);
1077 register vector signed short line6B = vec_sub(line4, line6);
1078 register vector signed short line5B = vec_add(line5, line7);
1079 register vector signed short line7B = vec_sub(line5, line7);
1081 register vector signed short line0C = vec_add(line0B, line4B);
1082 register vector signed short line4C = vec_sub(line0B, line4B);
1083 register vector signed short line1C = vec_add(line1B, line5B);
1084 register vector signed short line5C = vec_sub(line1B, line5B);
1085 register vector signed short line2C = vec_add(line2B, line6B);
1086 register vector signed short line6C = vec_sub(line2B, line6B);
1087 register vector signed short line3C = vec_add(line3B, line7B);
1088 register vector signed short line7C = vec_sub(line3B, line7B);
1090 vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
1091 vsum = vec_sum4s(vec_abs(line1C), vsum);
1092 vsum = vec_sum4s(vec_abs(line2C), vsum);
1093 vsum = vec_sum4s(vec_abs(line3C), vsum);
1094 vsum = vec_sum4s(vec_abs(line4C), vsum);
1095 vsum = vec_sum4s(vec_abs(line5C), vsum);
1096 vsum = vec_sum4s(vec_abs(line6C), vsum);
1097 vsum = vec_sum4s(vec_abs(line7C), vsum);
1098 vsum = vec_sums(vsum, (vector signed int)vzero);
1099 vsum = vec_splat(vsum, 3);
1100 vec_ste(vsum, 0, &sum);
1102 POWERPC_PERF_STOP_COUNT(altivec_hadamard8_diff8x8_num, 1);
1107 16x8 works with 16 elements; it allows to avoid replicating loads, and
1108 give the compiler more rooms for scheduling. It's only used from
1109 inside hadamard8_diff16_altivec.
1111 Unfortunately, it seems gcc-3.3 is a bit dumb, and the compiled code has a LOT
1112 of spill code, it seems gcc (unlike xlc) cannot keep everything in registers
1113 by itself. The following code include hand-made registers allocation. It's not
1114 clean, but on a 7450 the resulting code is much faster (best case fall from
1115 700+ cycles to 550).
1117 xlc doesn't add spill code, but it doesn't know how to schedule for the 7450,
1118 and its code isn't much faster than gcc-3.3 on the 7450 (but uses 25% less
1121 On the 970, the hand-made RA is still a win (around 690 vs. around 780), but
1122 xlc goes to around 660 on the regular C code...
1125 static int hadamard8_diff16x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h) {
1127 register vector signed short
1128 temp0 __asm__ ("v0"),
1129 temp1 __asm__ ("v1"),
1130 temp2 __asm__ ("v2"),
1131 temp3 __asm__ ("v3"),
1132 temp4 __asm__ ("v4"),
1133 temp5 __asm__ ("v5"),
1134 temp6 __asm__ ("v6"),
1135 temp7 __asm__ ("v7");
1136 register vector signed short
1137 temp0S __asm__ ("v8"),
1138 temp1S __asm__ ("v9"),
1139 temp2S __asm__ ("v10"),
1140 temp3S __asm__ ("v11"),
1141 temp4S __asm__ ("v12"),
1142 temp5S __asm__ ("v13"),
1143 temp6S __asm__ ("v14"),
1144 temp7S __asm__ ("v15");
1145 register const vector unsigned char vzero __asm__ ("v31") =
1146 (const vector unsigned char)vec_splat_u8(0);
1148 register const vector signed short vprod1 __asm__ ("v16") =
1149 (const vector signed short){ 1,-1, 1,-1, 1,-1, 1,-1 };
1150 register const vector signed short vprod2 __asm__ ("v17") =
1151 (const vector signed short){ 1, 1,-1,-1, 1, 1,-1,-1 };
1152 register const vector signed short vprod3 __asm__ ("v18") =
1153 (const vector signed short){ 1, 1, 1, 1,-1,-1,-1,-1 };
1154 register const vector unsigned char perm1 __asm__ ("v19") =
1155 (const vector unsigned char)
1156 {0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
1157 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D};
1158 register const vector unsigned char perm2 __asm__ ("v20") =
1159 (const vector unsigned char)
1160 {0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
1161 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B};
1162 register const vector unsigned char perm3 __asm__ ("v21") =
1163 (const vector unsigned char)
1164 {0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
1165 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07};
1167 #define ONEITERBUTTERFLY(i, res1, res2) \
1169 register vector unsigned char src1 __asm__ ("v22"), \
1170 src2 __asm__ ("v23"), \
1171 dst1 __asm__ ("v24"), \
1172 dst2 __asm__ ("v25"), \
1173 srcO __asm__ ("v22"), \
1174 dstO __asm__ ("v23"); \
1176 register vector signed short srcV __asm__ ("v24"), \
1177 dstV __asm__ ("v25"), \
1178 srcW __asm__ ("v26"), \
1179 dstW __asm__ ("v27"), \
1180 but0 __asm__ ("v28"), \
1181 but0S __asm__ ("v29"), \
1182 op1 __asm__ ("v30"), \
1183 but1 __asm__ ("v22"), \
1184 op1S __asm__ ("v23"), \
1185 but1S __asm__ ("v24"), \
1186 op2 __asm__ ("v25"), \
1187 but2 __asm__ ("v26"), \
1188 op2S __asm__ ("v27"), \
1189 but2S __asm__ ("v28"), \
1190 op3 __asm__ ("v29"), \
1191 op3S __asm__ ("v30"); \
1193 src1 = vec_ld(stride * i, src); \
1194 src2 = vec_ld((stride * i) + 16, src); \
1195 srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
1196 dst1 = vec_ld(stride * i, dst); \
1197 dst2 = vec_ld((stride * i) + 16, dst); \
1198 dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
1199 /* promote the unsigned chars to signed shorts */ \
1200 srcV = (vector signed short)vec_mergeh((vector signed char)vzero, \
1201 (vector signed char)srcO); \
1202 dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \
1203 (vector signed char)dstO); \
1204 srcW = (vector signed short)vec_mergel((vector signed char)vzero, \
1205 (vector signed char)srcO); \
1206 dstW = (vector signed short)vec_mergel((vector signed char)vzero, \
1207 (vector signed char)dstO); \
1208 /* subtractions inside the first butterfly */ \
1209 but0 = vec_sub(srcV, dstV); \
1210 but0S = vec_sub(srcW, dstW); \
1211 op1 = vec_perm(but0, but0, perm1); \
1212 but1 = vec_mladd(but0, vprod1, op1); \
1213 op1S = vec_perm(but0S, but0S, perm1); \
1214 but1S = vec_mladd(but0S, vprod1, op1S); \
1215 op2 = vec_perm(but1, but1, perm2); \
1216 but2 = vec_mladd(but1, vprod2, op2); \
1217 op2S = vec_perm(but1S, but1S, perm2); \
1218 but2S = vec_mladd(but1S, vprod2, op2S); \
1219 op3 = vec_perm(but2, but2, perm3); \
1220 res1 = vec_mladd(but2, vprod3, op3); \
1221 op3S = vec_perm(but2S, but2S, perm3); \
1222 res2 = vec_mladd(but2S, vprod3, op3S); \
1224 ONEITERBUTTERFLY(0, temp0, temp0S);
1225 ONEITERBUTTERFLY(1, temp1, temp1S);
1226 ONEITERBUTTERFLY(2, temp2, temp2S);
1227 ONEITERBUTTERFLY(3, temp3, temp3S);
1228 ONEITERBUTTERFLY(4, temp4, temp4S);
1229 ONEITERBUTTERFLY(5, temp5, temp5S);
1230 ONEITERBUTTERFLY(6, temp6, temp6S);
1231 ONEITERBUTTERFLY(7, temp7, temp7S);
1233 #undef ONEITERBUTTERFLY
1235 register vector signed int vsum;
1236 register vector signed short line0S, line1S, line2S, line3S, line4S,
1237 line5S, line6S, line7S, line0BS,line2BS,
1238 line1BS,line3BS,line4BS,line6BS,line5BS,
1239 line7BS,line0CS,line4CS,line1CS,line5CS,
1240 line2CS,line6CS,line3CS,line7CS;
1242 register vector signed short line0 = vec_add(temp0, temp1);
1243 register vector signed short line1 = vec_sub(temp0, temp1);
1244 register vector signed short line2 = vec_add(temp2, temp3);
1245 register vector signed short line3 = vec_sub(temp2, temp3);
1246 register vector signed short line4 = vec_add(temp4, temp5);
1247 register vector signed short line5 = vec_sub(temp4, temp5);
1248 register vector signed short line6 = vec_add(temp6, temp7);
1249 register vector signed short line7 = vec_sub(temp6, temp7);
1251 register vector signed short line0B = vec_add(line0, line2);
1252 register vector signed short line2B = vec_sub(line0, line2);
1253 register vector signed short line1B = vec_add(line1, line3);
1254 register vector signed short line3B = vec_sub(line1, line3);
1255 register vector signed short line4B = vec_add(line4, line6);
1256 register vector signed short line6B = vec_sub(line4, line6);
1257 register vector signed short line5B = vec_add(line5, line7);
1258 register vector signed short line7B = vec_sub(line5, line7);
1260 register vector signed short line0C = vec_add(line0B, line4B);
1261 register vector signed short line4C = vec_sub(line0B, line4B);
1262 register vector signed short line1C = vec_add(line1B, line5B);
1263 register vector signed short line5C = vec_sub(line1B, line5B);
1264 register vector signed short line2C = vec_add(line2B, line6B);
1265 register vector signed short line6C = vec_sub(line2B, line6B);
1266 register vector signed short line3C = vec_add(line3B, line7B);
1267 register vector signed short line7C = vec_sub(line3B, line7B);
1269 vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
1270 vsum = vec_sum4s(vec_abs(line1C), vsum);
1271 vsum = vec_sum4s(vec_abs(line2C), vsum);
1272 vsum = vec_sum4s(vec_abs(line3C), vsum);
1273 vsum = vec_sum4s(vec_abs(line4C), vsum);
1274 vsum = vec_sum4s(vec_abs(line5C), vsum);
1275 vsum = vec_sum4s(vec_abs(line6C), vsum);
1276 vsum = vec_sum4s(vec_abs(line7C), vsum);
1278 line0S = vec_add(temp0S, temp1S);
1279 line1S = vec_sub(temp0S, temp1S);
1280 line2S = vec_add(temp2S, temp3S);
1281 line3S = vec_sub(temp2S, temp3S);
1282 line4S = vec_add(temp4S, temp5S);
1283 line5S = vec_sub(temp4S, temp5S);
1284 line6S = vec_add(temp6S, temp7S);
1285 line7S = vec_sub(temp6S, temp7S);
1287 line0BS = vec_add(line0S, line2S);
1288 line2BS = vec_sub(line0S, line2S);
1289 line1BS = vec_add(line1S, line3S);
1290 line3BS = vec_sub(line1S, line3S);
1291 line4BS = vec_add(line4S, line6S);
1292 line6BS = vec_sub(line4S, line6S);
1293 line5BS = vec_add(line5S, line7S);
1294 line7BS = vec_sub(line5S, line7S);
1296 line0CS = vec_add(line0BS, line4BS);
1297 line4CS = vec_sub(line0BS, line4BS);
1298 line1CS = vec_add(line1BS, line5BS);
1299 line5CS = vec_sub(line1BS, line5BS);
1300 line2CS = vec_add(line2BS, line6BS);
1301 line6CS = vec_sub(line2BS, line6BS);
1302 line3CS = vec_add(line3BS, line7BS);
1303 line7CS = vec_sub(line3BS, line7BS);
1305 vsum = vec_sum4s(vec_abs(line0CS), vsum);
1306 vsum = vec_sum4s(vec_abs(line1CS), vsum);
1307 vsum = vec_sum4s(vec_abs(line2CS), vsum);
1308 vsum = vec_sum4s(vec_abs(line3CS), vsum);
1309 vsum = vec_sum4s(vec_abs(line4CS), vsum);
1310 vsum = vec_sum4s(vec_abs(line5CS), vsum);
1311 vsum = vec_sum4s(vec_abs(line6CS), vsum);
1312 vsum = vec_sum4s(vec_abs(line7CS), vsum);
1313 vsum = vec_sums(vsum, (vector signed int)vzero);
1314 vsum = vec_splat(vsum, 3);
1315 vec_ste(vsum, 0, &sum);
1320 int hadamard8_diff16_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
1321 POWERPC_PERF_DECLARE(altivec_hadamard8_diff16_num, 1);
1323 POWERPC_PERF_START_COUNT(altivec_hadamard8_diff16_num, 1);
1324 score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
1328 score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
1330 POWERPC_PERF_STOP_COUNT(altivec_hadamard8_diff16_num, 1);
1334 static void vorbis_inverse_coupling_altivec(float *mag, float *ang,
1339 vector bool int t0, t1;
1340 const vector unsigned int v_31 = //XXX
1341 vec_add(vec_add(vec_splat_u32(15),vec_splat_u32(15)),vec_splat_u32(1));
1342 for (i = 0; i < blocksize; i += 4) {
1343 m = vec_ld(0, mag+i);
1344 a = vec_ld(0, ang+i);
1345 t0 = vec_cmple(m, (vector float)vec_splat_u32(0));
1346 t1 = vec_cmple(a, (vector float)vec_splat_u32(0));
1347 a = vec_xor(a, (vector float) vec_sl((vector unsigned int)t0, v_31));
1348 t0 = (vector bool int)vec_and(a, t1);
1349 t1 = (vector bool int)vec_andc(a, t1);
1350 a = vec_sub(m, (vector float)t1);
1351 m = vec_add(m, (vector float)t0);
1352 vec_stl(a, 0, ang+i);
1353 vec_stl(m, 0, mag+i);
1357 /* next one assumes that ((line_size % 8) == 0) */
1358 void avg_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
1360 POWERPC_PERF_DECLARE(altivec_avg_pixels8_xy2_num, 1);
1362 register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
1363 register vector unsigned char blockv, temp1, temp2, blocktemp;
1364 register vector unsigned short pixelssum1, pixelssum2, temp3;
1366 register const vector unsigned char vczero = (const vector unsigned char)
1368 register const vector unsigned short vctwo = (const vector unsigned short)
1371 temp1 = vec_ld(0, pixels);
1372 temp2 = vec_ld(16, pixels);
1373 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
1374 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
1377 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
1379 pixelsv1 = vec_mergeh(vczero, pixelsv1);
1380 pixelsv2 = vec_mergeh(vczero, pixelsv2);
1381 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
1382 (vector unsigned short)pixelsv2);
1383 pixelssum1 = vec_add(pixelssum1, vctwo);
1385 POWERPC_PERF_START_COUNT(altivec_avg_pixels8_xy2_num, 1);
1386 for (i = 0; i < h ; i++) {
1387 int rightside = ((unsigned long)block & 0x0000000F);
1388 blockv = vec_ld(0, block);
1390 temp1 = vec_ld(line_size, pixels);
1391 temp2 = vec_ld(line_size + 16, pixels);
1392 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
1393 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
1396 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
1399 pixelsv1 = vec_mergeh(vczero, pixelsv1);
1400 pixelsv2 = vec_mergeh(vczero, pixelsv2);
1401 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
1402 (vector unsigned short)pixelsv2);
1403 temp3 = vec_add(pixelssum1, pixelssum2);
1404 temp3 = vec_sra(temp3, vctwo);
1405 pixelssum1 = vec_add(pixelssum2, vctwo);
1406 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
1409 blocktemp = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
1411 blocktemp = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
1414 blockv = vec_avg(blocktemp, blockv);
1415 vec_st(blockv, 0, block);
1418 pixels += line_size;
1421 POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_xy2_num, 1);
1424 void dsputil_init_altivec(DSPContext* c, AVCodecContext *avctx)
1426 c->pix_abs[0][1] = sad16_x2_altivec;
1427 c->pix_abs[0][2] = sad16_y2_altivec;
1428 c->pix_abs[0][3] = sad16_xy2_altivec;
1429 c->pix_abs[0][0] = sad16_altivec;
1430 c->pix_abs[1][0] = sad8_altivec;
1431 c->sad[0]= sad16_altivec;
1432 c->sad[1]= sad8_altivec;
1433 c->pix_norm1 = pix_norm1_altivec;
1434 c->sse[1]= sse8_altivec;
1435 c->sse[0]= sse16_altivec;
1436 c->pix_sum = pix_sum_altivec;
1437 c->diff_pixels = diff_pixels_altivec;
1438 c->get_pixels = get_pixels_altivec;
1439 c->clear_block = clear_block_altivec;
1440 c->add_bytes= add_bytes_altivec;
1441 c->put_pixels_tab[0][0] = put_pixels16_altivec;
1442 /* the two functions do the same thing, so use the same code */
1443 c->put_no_rnd_pixels_tab[0][0] = put_pixels16_altivec;
1444 c->avg_pixels_tab[0][0] = avg_pixels16_altivec;
1445 c->avg_pixels_tab[1][0] = avg_pixels8_altivec;
1446 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_altivec;
1447 c->put_pixels_tab[1][3] = put_pixels8_xy2_altivec;
1448 c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_altivec;
1449 c->put_pixels_tab[0][3] = put_pixels16_xy2_altivec;
1450 c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_altivec;
1452 c->hadamard8_diff[0] = hadamard8_diff16_altivec;
1453 c->hadamard8_diff[1] = hadamard8_diff8x8_altivec;
1454 if (CONFIG_VORBIS_DECODER)
1455 c->vorbis_inverse_coupling = vorbis_inverse_coupling_altivec;