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 "util_altivec.h"
29 #include "types_altivec.h"
30 #include "dsputil_altivec.h"
32 static int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
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 static int sad16_y2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
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 static int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
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 static int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
232 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
233 vector unsigned char perm1, perm2, pix1v_low, pix1v_high, pix2v_low, pix2v_high;
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_high = vec_ld( 0, pix1);
245 pix1v_low = vec_ld(15, pix1);
246 perm2 = vec_lvsl(0, pix2);
247 pix2v_high = vec_ld( 0, pix2);
248 pix2v_low = vec_ld(15, pix2);
249 t1 = vec_perm(pix1v_high, pix1v_low, perm1);
250 t2 = vec_perm(pix2v_high, pix2v_low, perm2);
252 /* Calculate a sum of abs differences vector */
253 t3 = vec_max(t1, t2);
254 t4 = vec_min(t1, t2);
255 t5 = vec_sub(t3, t4);
257 /* Add each 4 pixel group together and put 4 results into sad */
258 sad = vec_sum4s(t5, sad);
264 /* Sum up the four partial sums, and put the result into s */
265 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
266 sumdiffs = vec_splat(sumdiffs, 3);
267 vec_ste(sumdiffs, 0, &s);
272 static int sad8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
276 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
277 vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
278 vector unsigned char t1, t2, t3,t4, t5;
279 vector unsigned int sad;
280 vector signed int sumdiffs;
282 sad = (vector unsigned int)vec_splat_u32(0);
284 permclear = (vector unsigned char){255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0};
286 for (i = 0; i < h; i++) {
287 /* Read potentially unaligned pixels into t1 and t2
288 Since we're reading 16 pixels, and actually only want 8,
289 mask out the last 8 pixels. The 0s don't change the sum. */
290 perm1 = vec_lvsl(0, pix1);
291 pix1v = (vector unsigned char *) pix1;
292 perm2 = vec_lvsl(0, pix2);
293 pix2v = (vector unsigned char *) pix2;
294 t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
295 t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
297 /* Calculate a sum of abs differences vector */
298 t3 = vec_max(t1, t2);
299 t4 = vec_min(t1, t2);
300 t5 = vec_sub(t3, t4);
302 /* Add each 4 pixel group together and put 4 results into sad */
303 sad = vec_sum4s(t5, sad);
309 /* Sum up the four partial sums, and put the result into s */
310 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
311 sumdiffs = vec_splat(sumdiffs, 3);
312 vec_ste(sumdiffs, 0, &s);
317 static int pix_norm1_altivec(uint8_t *pix, int line_size)
321 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
322 vector unsigned char *tv;
323 vector unsigned char pixv;
324 vector unsigned int sv;
325 vector signed int sum;
327 sv = (vector unsigned int)vec_splat_u32(0);
330 for (i = 0; i < 16; i++) {
331 /* Read in the potentially unaligned pixels */
332 tv = (vector unsigned char *) pix;
333 pixv = vec_perm(tv[0], tv[1], vec_lvsl(0, pix));
335 /* Square the values, and add them to our sum */
336 sv = vec_msum(pixv, pixv, sv);
340 /* Sum up the four partial sums, and put the result into s */
341 sum = vec_sums((vector signed int) sv, (vector signed int) zero);
342 sum = vec_splat(sum, 3);
349 * Sum of Squared Errors for a 8x8 block.
351 * It's the sad8_altivec code above w/ squaring added.
353 static int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
357 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
358 vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
359 vector unsigned char t1, t2, t3,t4, t5;
360 vector unsigned int sum;
361 vector signed int sumsqr;
363 sum = (vector unsigned int)vec_splat_u32(0);
365 permclear = (vector unsigned char){255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0};
368 for (i = 0; i < h; i++) {
369 /* Read potentially unaligned pixels into t1 and t2
370 Since we're reading 16 pixels, and actually only want 8,
371 mask out the last 8 pixels. The 0s don't change the sum. */
372 perm1 = vec_lvsl(0, pix1);
373 pix1v = (vector unsigned char *) pix1;
374 perm2 = vec_lvsl(0, pix2);
375 pix2v = (vector unsigned char *) pix2;
376 t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
377 t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
379 /* Since we want to use unsigned chars, we can take advantage
380 of the fact that abs(a-b)^2 = (a-b)^2. */
382 /* Calculate abs differences vector */
383 t3 = vec_max(t1, t2);
384 t4 = vec_min(t1, t2);
385 t5 = vec_sub(t3, t4);
387 /* Square the values and add them to our sum */
388 sum = vec_msum(t5, t5, sum);
394 /* Sum up the four partial sums, and put the result into s */
395 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
396 sumsqr = vec_splat(sumsqr, 3);
397 vec_ste(sumsqr, 0, &s);
403 * Sum of Squared Errors for a 16x16 block.
405 * It's the sad16_altivec code above w/ squaring added.
407 static int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
411 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
412 vector unsigned char perm1, perm2, *pix1v, *pix2v;
413 vector unsigned char t1, t2, t3,t4, t5;
414 vector unsigned int sum;
415 vector signed int sumsqr;
417 sum = (vector unsigned int)vec_splat_u32(0);
419 for (i = 0; i < h; i++) {
420 /* Read potentially unaligned pixels into t1 and t2 */
421 perm1 = vec_lvsl(0, pix1);
422 pix1v = (vector unsigned char *) pix1;
423 perm2 = vec_lvsl(0, pix2);
424 pix2v = (vector unsigned char *) pix2;
425 t1 = vec_perm(pix1v[0], pix1v[1], perm1);
426 t2 = vec_perm(pix2v[0], pix2v[1], perm2);
428 /* Since we want to use unsigned chars, we can take advantage
429 of the fact that abs(a-b)^2 = (a-b)^2. */
431 /* Calculate abs differences vector */
432 t3 = vec_max(t1, t2);
433 t4 = vec_min(t1, t2);
434 t5 = vec_sub(t3, t4);
436 /* Square the values and add them to our sum */
437 sum = vec_msum(t5, t5, sum);
443 /* Sum up the four partial sums, and put the result into s */
444 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
445 sumsqr = vec_splat(sumsqr, 3);
446 vec_ste(sumsqr, 0, &s);
451 static int pix_sum_altivec(uint8_t * pix, int line_size)
453 const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
454 vector unsigned char perm, *pixv;
455 vector unsigned char t1;
456 vector unsigned int sad;
457 vector signed int sumdiffs;
462 sad = (vector unsigned int)vec_splat_u32(0);
464 for (i = 0; i < 16; i++) {
465 /* Read the potentially unaligned 16 pixels into t1 */
466 perm = vec_lvsl(0, pix);
467 pixv = (vector unsigned char *) pix;
468 t1 = vec_perm(pixv[0], pixv[1], perm);
470 /* Add each 4 pixel group together and put 4 results into sad */
471 sad = vec_sum4s(t1, sad);
476 /* Sum up the four partial sums, and put the result into s */
477 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
478 sumdiffs = vec_splat(sumdiffs, 3);
479 vec_ste(sumdiffs, 0, &s);
484 static void get_pixels_altivec(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
487 vector unsigned char perm, bytes, *pixv;
488 const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
489 vector signed short shorts;
491 for (i = 0; i < 8; i++) {
492 // Read potentially unaligned pixels.
493 // We're reading 16 pixels, and actually only want 8,
494 // but we simply ignore the extras.
495 perm = vec_lvsl(0, pixels);
496 pixv = (vector unsigned char *) pixels;
497 bytes = vec_perm(pixv[0], pixv[1], perm);
499 // convert the bytes into shorts
500 shorts = (vector signed short)vec_mergeh(zero, bytes);
502 // save the data to the block, we assume the block is 16-byte aligned
503 vec_st(shorts, i*16, (vector signed short*)block);
509 static void diff_pixels_altivec(DCTELEM *restrict block, const uint8_t *s1,
510 const uint8_t *s2, int stride)
513 vector unsigned char perm, bytes, *pixv;
514 const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
515 vector signed short shorts1, shorts2;
517 for (i = 0; i < 4; i++) {
518 // Read potentially unaligned pixels
519 // We're reading 16 pixels, and actually only want 8,
520 // but we simply ignore the extras.
521 perm = vec_lvsl(0, s1);
522 pixv = (vector unsigned char *) s1;
523 bytes = vec_perm(pixv[0], pixv[1], perm);
525 // convert the bytes into shorts
526 shorts1 = (vector signed short)vec_mergeh(zero, bytes);
528 // Do the same for the second block of pixels
529 perm = vec_lvsl(0, s2);
530 pixv = (vector unsigned char *) s2;
531 bytes = vec_perm(pixv[0], pixv[1], perm);
533 // convert the bytes into shorts
534 shorts2 = (vector signed short)vec_mergeh(zero, bytes);
536 // Do the subtraction
537 shorts1 = vec_sub(shorts1, shorts2);
539 // save the data to the block, we assume the block is 16-byte aligned
540 vec_st(shorts1, 0, (vector signed short*)block);
547 // The code below is a copy of the code above... This is a manual
550 // Read potentially unaligned pixels
551 // We're reading 16 pixels, and actually only want 8,
552 // but we simply ignore the extras.
553 perm = vec_lvsl(0, s1);
554 pixv = (vector unsigned char *) s1;
555 bytes = vec_perm(pixv[0], pixv[1], perm);
557 // convert the bytes into shorts
558 shorts1 = (vector signed short)vec_mergeh(zero, bytes);
560 // Do the same for the second block of pixels
561 perm = vec_lvsl(0, s2);
562 pixv = (vector unsigned char *) s2;
563 bytes = vec_perm(pixv[0], pixv[1], perm);
565 // convert the bytes into shorts
566 shorts2 = (vector signed short)vec_mergeh(zero, bytes);
568 // Do the subtraction
569 shorts1 = vec_sub(shorts1, shorts2);
571 // save the data to the block, we assume the block is 16-byte aligned
572 vec_st(shorts1, 0, (vector signed short*)block);
581 static void clear_block_altivec(DCTELEM *block) {
583 vec_st(zero_s16v, 0, block);
584 vec_st(zero_s16v, 16, block);
585 vec_st(zero_s16v, 32, block);
586 vec_st(zero_s16v, 48, block);
587 vec_st(zero_s16v, 64, block);
588 vec_st(zero_s16v, 80, block);
589 vec_st(zero_s16v, 96, block);
590 vec_st(zero_s16v, 112, block);
594 static void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {
596 register vector unsigned char vdst, vsrc;
598 /* dst and src are 16 bytes-aligned (guaranteed) */
599 for (i = 0 ; (i + 15) < w ; i+=16) {
600 vdst = vec_ld(i, (unsigned char*)dst);
601 vsrc = vec_ld(i, (unsigned char*)src);
602 vdst = vec_add(vsrc, vdst);
603 vec_st(vdst, i, (unsigned char*)dst);
605 /* if w is not a multiple of 16 */
606 for (; (i < w) ; i++) {
611 /* next one assumes that ((line_size % 16) == 0) */
612 void put_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
614 register vector unsigned char pixelsv1, pixelsv2;
615 register vector unsigned char pixelsv1B, pixelsv2B;
616 register vector unsigned char pixelsv1C, pixelsv2C;
617 register vector unsigned char pixelsv1D, pixelsv2D;
619 register vector unsigned char perm = vec_lvsl(0, pixels);
621 register int line_size_2 = line_size << 1;
622 register int line_size_3 = line_size + line_size_2;
623 register int line_size_4 = line_size << 2;
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, pixels);
633 pixelsv2 = vec_ld(16, pixels);
634 vec_st(vec_perm(pixelsv1, pixelsv2, perm),
640 for (i = 0; i < h; i += 4) {
641 pixelsv1 = vec_ld( 0, pixels);
642 pixelsv2 = vec_ld(15, pixels);
643 pixelsv1B = vec_ld(line_size, pixels);
644 pixelsv2B = vec_ld(15 + line_size, pixels);
645 pixelsv1C = vec_ld(line_size_2, pixels);
646 pixelsv2C = vec_ld(15 + line_size_2, pixels);
647 pixelsv1D = vec_ld(line_size_3, pixels);
648 pixelsv2D = vec_ld(15 + line_size_3, 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);
663 /* next one assumes that ((line_size % 16) == 0) */
664 #define op_avg(a,b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEUL)>>1) )
665 void avg_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
667 register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
668 register vector unsigned char perm = vec_lvsl(0, pixels);
671 for (i = 0; i < h; i++) {
672 pixelsv1 = vec_ld( 0, pixels);
673 pixelsv2 = vec_ld(16,pixels);
674 blockv = vec_ld(0, block);
675 pixelsv = vec_perm(pixelsv1, pixelsv2, perm);
676 blockv = vec_avg(blockv,pixelsv);
677 vec_st(blockv, 0, (unsigned char*)block);
683 /* next one assumes that ((line_size % 8) == 0) */
684 static void avg_pixels8_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
686 register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
689 for (i = 0; i < h; i++) {
690 /* block is 8 bytes-aligned, so we're either in the
691 left block (16 bytes-aligned) or in the right block (not) */
692 int rightside = ((unsigned long)block & 0x0000000F);
694 blockv = vec_ld(0, block);
695 pixelsv1 = vec_ld( 0, pixels);
696 pixelsv2 = vec_ld(16, pixels);
697 pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels));
700 pixelsv = vec_perm(blockv, pixelsv, vcprm(0,1,s0,s1));
702 pixelsv = vec_perm(blockv, pixelsv, vcprm(s0,s1,2,3));
705 blockv = vec_avg(blockv, pixelsv);
707 vec_st(blockv, 0, block);
714 /* next one assumes that ((line_size % 8) == 0) */
715 static void put_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
718 register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
719 register vector unsigned char blockv, temp1, temp2;
720 register vector unsigned short pixelssum1, pixelssum2, temp3;
721 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
722 register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
724 temp1 = vec_ld(0, pixels);
725 temp2 = vec_ld(16, pixels);
726 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
727 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
730 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
732 pixelsv1 = vec_mergeh(vczero, pixelsv1);
733 pixelsv2 = vec_mergeh(vczero, pixelsv2);
734 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
735 (vector unsigned short)pixelsv2);
736 pixelssum1 = vec_add(pixelssum1, vctwo);
738 for (i = 0; i < h ; i++) {
739 int rightside = ((unsigned long)block & 0x0000000F);
740 blockv = vec_ld(0, block);
742 temp1 = vec_ld(line_size, pixels);
743 temp2 = vec_ld(line_size + 16, pixels);
744 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
745 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
748 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
751 pixelsv1 = vec_mergeh(vczero, pixelsv1);
752 pixelsv2 = vec_mergeh(vczero, pixelsv2);
753 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
754 (vector unsigned short)pixelsv2);
755 temp3 = vec_add(pixelssum1, pixelssum2);
756 temp3 = vec_sra(temp3, vctwo);
757 pixelssum1 = vec_add(pixelssum2, vctwo);
758 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
761 blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
763 blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
766 vec_st(blockv, 0, block);
773 /* next one assumes that ((line_size % 8) == 0) */
774 static void put_no_rnd_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
777 register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
778 register vector unsigned char blockv, temp1, temp2;
779 register vector unsigned short pixelssum1, pixelssum2, temp3;
780 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
781 register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
782 register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
784 temp1 = vec_ld(0, pixels);
785 temp2 = vec_ld(16, pixels);
786 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
787 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
790 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
792 pixelsv1 = vec_mergeh(vczero, pixelsv1);
793 pixelsv2 = vec_mergeh(vczero, pixelsv2);
794 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
795 (vector unsigned short)pixelsv2);
796 pixelssum1 = vec_add(pixelssum1, vcone);
798 for (i = 0; i < h ; i++) {
799 int rightside = ((unsigned long)block & 0x0000000F);
800 blockv = vec_ld(0, block);
802 temp1 = vec_ld(line_size, pixels);
803 temp2 = vec_ld(line_size + 16, pixels);
804 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
805 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
808 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
811 pixelsv1 = vec_mergeh(vczero, pixelsv1);
812 pixelsv2 = vec_mergeh(vczero, pixelsv2);
813 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
814 (vector unsigned short)pixelsv2);
815 temp3 = vec_add(pixelssum1, pixelssum2);
816 temp3 = vec_sra(temp3, vctwo);
817 pixelssum1 = vec_add(pixelssum2, vcone);
818 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
821 blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
823 blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
826 vec_st(blockv, 0, block);
833 /* next one assumes that ((line_size % 16) == 0) */
834 static void put_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
837 register vector unsigned char pixelsv1, pixelsv2, pixelsv3, pixelsv4;
838 register vector unsigned char blockv, temp1, temp2;
839 register vector unsigned short temp3, temp4,
840 pixelssum1, pixelssum2, pixelssum3, pixelssum4;
841 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
842 register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
844 temp1 = vec_ld(0, pixels);
845 temp2 = vec_ld(16, pixels);
846 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
847 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
850 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
852 pixelsv3 = vec_mergel(vczero, pixelsv1);
853 pixelsv4 = vec_mergel(vczero, pixelsv2);
854 pixelsv1 = vec_mergeh(vczero, pixelsv1);
855 pixelsv2 = vec_mergeh(vczero, pixelsv2);
856 pixelssum3 = vec_add((vector unsigned short)pixelsv3,
857 (vector unsigned short)pixelsv4);
858 pixelssum3 = vec_add(pixelssum3, vctwo);
859 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
860 (vector unsigned short)pixelsv2);
861 pixelssum1 = vec_add(pixelssum1, vctwo);
863 for (i = 0; i < h ; i++) {
864 blockv = vec_ld(0, block);
866 temp1 = vec_ld(line_size, pixels);
867 temp2 = vec_ld(line_size + 16, pixels);
868 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
869 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
872 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
875 pixelsv3 = vec_mergel(vczero, pixelsv1);
876 pixelsv4 = vec_mergel(vczero, pixelsv2);
877 pixelsv1 = vec_mergeh(vczero, pixelsv1);
878 pixelsv2 = vec_mergeh(vczero, pixelsv2);
880 pixelssum4 = vec_add((vector unsigned short)pixelsv3,
881 (vector unsigned short)pixelsv4);
882 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
883 (vector unsigned short)pixelsv2);
884 temp4 = vec_add(pixelssum3, pixelssum4);
885 temp4 = vec_sra(temp4, vctwo);
886 temp3 = vec_add(pixelssum1, pixelssum2);
887 temp3 = vec_sra(temp3, vctwo);
889 pixelssum3 = vec_add(pixelssum4, vctwo);
890 pixelssum1 = vec_add(pixelssum2, vctwo);
892 blockv = vec_packsu(temp3, temp4);
894 vec_st(blockv, 0, block);
901 /* next one assumes that ((line_size % 16) == 0) */
902 static void put_no_rnd_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
905 register vector unsigned char pixelsv1, pixelsv2, pixelsv3, pixelsv4;
906 register vector unsigned char blockv, temp1, temp2;
907 register vector unsigned short temp3, temp4,
908 pixelssum1, pixelssum2, pixelssum3, pixelssum4;
909 register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
910 register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
911 register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
913 temp1 = vec_ld(0, pixels);
914 temp2 = vec_ld(16, pixels);
915 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
916 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
919 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
921 pixelsv3 = vec_mergel(vczero, pixelsv1);
922 pixelsv4 = vec_mergel(vczero, pixelsv2);
923 pixelsv1 = vec_mergeh(vczero, pixelsv1);
924 pixelsv2 = vec_mergeh(vczero, pixelsv2);
925 pixelssum3 = vec_add((vector unsigned short)pixelsv3,
926 (vector unsigned short)pixelsv4);
927 pixelssum3 = vec_add(pixelssum3, vcone);
928 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
929 (vector unsigned short)pixelsv2);
930 pixelssum1 = vec_add(pixelssum1, vcone);
932 for (i = 0; i < h ; i++) {
933 blockv = vec_ld(0, block);
935 temp1 = vec_ld(line_size, pixels);
936 temp2 = vec_ld(line_size + 16, pixels);
937 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
938 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
941 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
944 pixelsv3 = vec_mergel(vczero, pixelsv1);
945 pixelsv4 = vec_mergel(vczero, pixelsv2);
946 pixelsv1 = vec_mergeh(vczero, pixelsv1);
947 pixelsv2 = vec_mergeh(vczero, pixelsv2);
949 pixelssum4 = vec_add((vector unsigned short)pixelsv3,
950 (vector unsigned short)pixelsv4);
951 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
952 (vector unsigned short)pixelsv2);
953 temp4 = vec_add(pixelssum3, pixelssum4);
954 temp4 = vec_sra(temp4, vctwo);
955 temp3 = vec_add(pixelssum1, pixelssum2);
956 temp3 = vec_sra(temp3, vctwo);
958 pixelssum3 = vec_add(pixelssum4, vcone);
959 pixelssum1 = vec_add(pixelssum2, vcone);
961 blockv = vec_packsu(temp3, temp4);
963 vec_st(blockv, 0, block);
970 static int hadamard8_diff8x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
972 register const vector unsigned char vzero =
973 (const vector unsigned char)vec_splat_u8(0);
974 register vector signed short temp0, temp1, temp2, temp3, temp4,
977 register const vector signed short vprod1 =(const vector signed short)
978 { 1,-1, 1,-1, 1,-1, 1,-1 };
979 register const vector signed short vprod2 =(const vector signed short)
980 { 1, 1,-1,-1, 1, 1,-1,-1 };
981 register const vector signed short vprod3 =(const vector signed short)
982 { 1, 1, 1, 1,-1,-1,-1,-1 };
983 register const vector unsigned char perm1 = (const vector unsigned char)
984 {0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
985 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D};
986 register const vector unsigned char perm2 = (const vector unsigned char)
987 {0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
988 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B};
989 register const vector unsigned char perm3 = (const vector unsigned char)
990 {0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
991 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07};
993 #define ONEITERBUTTERFLY(i, res) \
995 register vector unsigned char src1, src2, srcO; \
996 register vector unsigned char dst1, dst2, dstO; \
997 register vector signed short srcV, dstV; \
998 register vector signed short but0, but1, but2, op1, op2, op3; \
999 src1 = vec_ld(stride * i, src); \
1000 src2 = vec_ld((stride * i) + 15, src); \
1001 srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
1002 dst1 = vec_ld(stride * i, dst); \
1003 dst2 = vec_ld((stride * i) + 15, dst); \
1004 dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
1005 /* promote the unsigned chars to signed shorts */ \
1006 /* we're in the 8x8 function, we only care for the first 8 */ \
1007 srcV = (vector signed short)vec_mergeh((vector signed char)vzero, \
1008 (vector signed char)srcO); \
1009 dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \
1010 (vector signed char)dstO); \
1011 /* subtractions inside the first butterfly */ \
1012 but0 = vec_sub(srcV, dstV); \
1013 op1 = vec_perm(but0, but0, perm1); \
1014 but1 = vec_mladd(but0, vprod1, op1); \
1015 op2 = vec_perm(but1, but1, perm2); \
1016 but2 = vec_mladd(but1, vprod2, op2); \
1017 op3 = vec_perm(but2, but2, perm3); \
1018 res = vec_mladd(but2, vprod3, op3); \
1020 ONEITERBUTTERFLY(0, temp0);
1021 ONEITERBUTTERFLY(1, temp1);
1022 ONEITERBUTTERFLY(2, temp2);
1023 ONEITERBUTTERFLY(3, temp3);
1024 ONEITERBUTTERFLY(4, temp4);
1025 ONEITERBUTTERFLY(5, temp5);
1026 ONEITERBUTTERFLY(6, temp6);
1027 ONEITERBUTTERFLY(7, temp7);
1029 #undef ONEITERBUTTERFLY
1031 register vector signed int vsum;
1032 register vector signed short line0 = vec_add(temp0, temp1);
1033 register vector signed short line1 = vec_sub(temp0, temp1);
1034 register vector signed short line2 = vec_add(temp2, temp3);
1035 register vector signed short line3 = vec_sub(temp2, temp3);
1036 register vector signed short line4 = vec_add(temp4, temp5);
1037 register vector signed short line5 = vec_sub(temp4, temp5);
1038 register vector signed short line6 = vec_add(temp6, temp7);
1039 register vector signed short line7 = vec_sub(temp6, temp7);
1041 register vector signed short line0B = vec_add(line0, line2);
1042 register vector signed short line2B = vec_sub(line0, line2);
1043 register vector signed short line1B = vec_add(line1, line3);
1044 register vector signed short line3B = vec_sub(line1, line3);
1045 register vector signed short line4B = vec_add(line4, line6);
1046 register vector signed short line6B = vec_sub(line4, line6);
1047 register vector signed short line5B = vec_add(line5, line7);
1048 register vector signed short line7B = vec_sub(line5, line7);
1050 register vector signed short line0C = vec_add(line0B, line4B);
1051 register vector signed short line4C = vec_sub(line0B, line4B);
1052 register vector signed short line1C = vec_add(line1B, line5B);
1053 register vector signed short line5C = vec_sub(line1B, line5B);
1054 register vector signed short line2C = vec_add(line2B, line6B);
1055 register vector signed short line6C = vec_sub(line2B, line6B);
1056 register vector signed short line3C = vec_add(line3B, line7B);
1057 register vector signed short line7C = vec_sub(line3B, line7B);
1059 vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
1060 vsum = vec_sum4s(vec_abs(line1C), vsum);
1061 vsum = vec_sum4s(vec_abs(line2C), vsum);
1062 vsum = vec_sum4s(vec_abs(line3C), vsum);
1063 vsum = vec_sum4s(vec_abs(line4C), vsum);
1064 vsum = vec_sum4s(vec_abs(line5C), vsum);
1065 vsum = vec_sum4s(vec_abs(line6C), vsum);
1066 vsum = vec_sum4s(vec_abs(line7C), vsum);
1067 vsum = vec_sums(vsum, (vector signed int)vzero);
1068 vsum = vec_splat(vsum, 3);
1069 vec_ste(vsum, 0, &sum);
1075 16x8 works with 16 elements; it allows to avoid replicating loads, and
1076 give the compiler more rooms for scheduling. It's only used from
1077 inside hadamard8_diff16_altivec.
1079 Unfortunately, it seems gcc-3.3 is a bit dumb, and the compiled code has a LOT
1080 of spill code, it seems gcc (unlike xlc) cannot keep everything in registers
1081 by itself. The following code include hand-made registers allocation. It's not
1082 clean, but on a 7450 the resulting code is much faster (best case fall from
1083 700+ cycles to 550).
1085 xlc doesn't add spill code, but it doesn't know how to schedule for the 7450,
1086 and its code isn't much faster than gcc-3.3 on the 7450 (but uses 25% less
1089 On the 970, the hand-made RA is still a win (around 690 vs. around 780), but
1090 xlc goes to around 660 on the regular C code...
1093 static int hadamard8_diff16x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h) {
1095 register vector signed short
1096 temp0 __asm__ ("v0"),
1097 temp1 __asm__ ("v1"),
1098 temp2 __asm__ ("v2"),
1099 temp3 __asm__ ("v3"),
1100 temp4 __asm__ ("v4"),
1101 temp5 __asm__ ("v5"),
1102 temp6 __asm__ ("v6"),
1103 temp7 __asm__ ("v7");
1104 register vector signed short
1105 temp0S __asm__ ("v8"),
1106 temp1S __asm__ ("v9"),
1107 temp2S __asm__ ("v10"),
1108 temp3S __asm__ ("v11"),
1109 temp4S __asm__ ("v12"),
1110 temp5S __asm__ ("v13"),
1111 temp6S __asm__ ("v14"),
1112 temp7S __asm__ ("v15");
1113 register const vector unsigned char vzero __asm__ ("v31") =
1114 (const vector unsigned char)vec_splat_u8(0);
1116 register const vector signed short vprod1 __asm__ ("v16") =
1117 (const vector signed short){ 1,-1, 1,-1, 1,-1, 1,-1 };
1118 register const vector signed short vprod2 __asm__ ("v17") =
1119 (const vector signed short){ 1, 1,-1,-1, 1, 1,-1,-1 };
1120 register const vector signed short vprod3 __asm__ ("v18") =
1121 (const vector signed short){ 1, 1, 1, 1,-1,-1,-1,-1 };
1122 register const vector unsigned char perm1 __asm__ ("v19") =
1123 (const vector unsigned char)
1124 {0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
1125 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D};
1126 register const vector unsigned char perm2 __asm__ ("v20") =
1127 (const vector unsigned char)
1128 {0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
1129 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B};
1130 register const vector unsigned char perm3 __asm__ ("v21") =
1131 (const vector unsigned char)
1132 {0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
1133 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07};
1135 #define ONEITERBUTTERFLY(i, res1, res2) \
1137 register vector unsigned char src1 __asm__ ("v22"), \
1138 src2 __asm__ ("v23"), \
1139 dst1 __asm__ ("v24"), \
1140 dst2 __asm__ ("v25"), \
1141 srcO __asm__ ("v22"), \
1142 dstO __asm__ ("v23"); \
1144 register vector signed short srcV __asm__ ("v24"), \
1145 dstV __asm__ ("v25"), \
1146 srcW __asm__ ("v26"), \
1147 dstW __asm__ ("v27"), \
1148 but0 __asm__ ("v28"), \
1149 but0S __asm__ ("v29"), \
1150 op1 __asm__ ("v30"), \
1151 but1 __asm__ ("v22"), \
1152 op1S __asm__ ("v23"), \
1153 but1S __asm__ ("v24"), \
1154 op2 __asm__ ("v25"), \
1155 but2 __asm__ ("v26"), \
1156 op2S __asm__ ("v27"), \
1157 but2S __asm__ ("v28"), \
1158 op3 __asm__ ("v29"), \
1159 op3S __asm__ ("v30"); \
1161 src1 = vec_ld(stride * i, src); \
1162 src2 = vec_ld((stride * i) + 16, src); \
1163 srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
1164 dst1 = vec_ld(stride * i, dst); \
1165 dst2 = vec_ld((stride * i) + 16, dst); \
1166 dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
1167 /* promote the unsigned chars to signed shorts */ \
1168 srcV = (vector signed short)vec_mergeh((vector signed char)vzero, \
1169 (vector signed char)srcO); \
1170 dstV = (vector signed short)vec_mergeh((vector signed char)vzero, \
1171 (vector signed char)dstO); \
1172 srcW = (vector signed short)vec_mergel((vector signed char)vzero, \
1173 (vector signed char)srcO); \
1174 dstW = (vector signed short)vec_mergel((vector signed char)vzero, \
1175 (vector signed char)dstO); \
1176 /* subtractions inside the first butterfly */ \
1177 but0 = vec_sub(srcV, dstV); \
1178 but0S = vec_sub(srcW, dstW); \
1179 op1 = vec_perm(but0, but0, perm1); \
1180 but1 = vec_mladd(but0, vprod1, op1); \
1181 op1S = vec_perm(but0S, but0S, perm1); \
1182 but1S = vec_mladd(but0S, vprod1, op1S); \
1183 op2 = vec_perm(but1, but1, perm2); \
1184 but2 = vec_mladd(but1, vprod2, op2); \
1185 op2S = vec_perm(but1S, but1S, perm2); \
1186 but2S = vec_mladd(but1S, vprod2, op2S); \
1187 op3 = vec_perm(but2, but2, perm3); \
1188 res1 = vec_mladd(but2, vprod3, op3); \
1189 op3S = vec_perm(but2S, but2S, perm3); \
1190 res2 = vec_mladd(but2S, vprod3, op3S); \
1192 ONEITERBUTTERFLY(0, temp0, temp0S);
1193 ONEITERBUTTERFLY(1, temp1, temp1S);
1194 ONEITERBUTTERFLY(2, temp2, temp2S);
1195 ONEITERBUTTERFLY(3, temp3, temp3S);
1196 ONEITERBUTTERFLY(4, temp4, temp4S);
1197 ONEITERBUTTERFLY(5, temp5, temp5S);
1198 ONEITERBUTTERFLY(6, temp6, temp6S);
1199 ONEITERBUTTERFLY(7, temp7, temp7S);
1201 #undef ONEITERBUTTERFLY
1203 register vector signed int vsum;
1204 register vector signed short line0S, line1S, line2S, line3S, line4S,
1205 line5S, line6S, line7S, line0BS,line2BS,
1206 line1BS,line3BS,line4BS,line6BS,line5BS,
1207 line7BS,line0CS,line4CS,line1CS,line5CS,
1208 line2CS,line6CS,line3CS,line7CS;
1210 register vector signed short line0 = vec_add(temp0, temp1);
1211 register vector signed short line1 = vec_sub(temp0, temp1);
1212 register vector signed short line2 = vec_add(temp2, temp3);
1213 register vector signed short line3 = vec_sub(temp2, temp3);
1214 register vector signed short line4 = vec_add(temp4, temp5);
1215 register vector signed short line5 = vec_sub(temp4, temp5);
1216 register vector signed short line6 = vec_add(temp6, temp7);
1217 register vector signed short line7 = vec_sub(temp6, temp7);
1219 register vector signed short line0B = vec_add(line0, line2);
1220 register vector signed short line2B = vec_sub(line0, line2);
1221 register vector signed short line1B = vec_add(line1, line3);
1222 register vector signed short line3B = vec_sub(line1, line3);
1223 register vector signed short line4B = vec_add(line4, line6);
1224 register vector signed short line6B = vec_sub(line4, line6);
1225 register vector signed short line5B = vec_add(line5, line7);
1226 register vector signed short line7B = vec_sub(line5, line7);
1228 register vector signed short line0C = vec_add(line0B, line4B);
1229 register vector signed short line4C = vec_sub(line0B, line4B);
1230 register vector signed short line1C = vec_add(line1B, line5B);
1231 register vector signed short line5C = vec_sub(line1B, line5B);
1232 register vector signed short line2C = vec_add(line2B, line6B);
1233 register vector signed short line6C = vec_sub(line2B, line6B);
1234 register vector signed short line3C = vec_add(line3B, line7B);
1235 register vector signed short line7C = vec_sub(line3B, line7B);
1237 vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
1238 vsum = vec_sum4s(vec_abs(line1C), vsum);
1239 vsum = vec_sum4s(vec_abs(line2C), vsum);
1240 vsum = vec_sum4s(vec_abs(line3C), vsum);
1241 vsum = vec_sum4s(vec_abs(line4C), vsum);
1242 vsum = vec_sum4s(vec_abs(line5C), vsum);
1243 vsum = vec_sum4s(vec_abs(line6C), vsum);
1244 vsum = vec_sum4s(vec_abs(line7C), vsum);
1246 line0S = vec_add(temp0S, temp1S);
1247 line1S = vec_sub(temp0S, temp1S);
1248 line2S = vec_add(temp2S, temp3S);
1249 line3S = vec_sub(temp2S, temp3S);
1250 line4S = vec_add(temp4S, temp5S);
1251 line5S = vec_sub(temp4S, temp5S);
1252 line6S = vec_add(temp6S, temp7S);
1253 line7S = vec_sub(temp6S, temp7S);
1255 line0BS = vec_add(line0S, line2S);
1256 line2BS = vec_sub(line0S, line2S);
1257 line1BS = vec_add(line1S, line3S);
1258 line3BS = vec_sub(line1S, line3S);
1259 line4BS = vec_add(line4S, line6S);
1260 line6BS = vec_sub(line4S, line6S);
1261 line5BS = vec_add(line5S, line7S);
1262 line7BS = vec_sub(line5S, line7S);
1264 line0CS = vec_add(line0BS, line4BS);
1265 line4CS = vec_sub(line0BS, line4BS);
1266 line1CS = vec_add(line1BS, line5BS);
1267 line5CS = vec_sub(line1BS, line5BS);
1268 line2CS = vec_add(line2BS, line6BS);
1269 line6CS = vec_sub(line2BS, line6BS);
1270 line3CS = vec_add(line3BS, line7BS);
1271 line7CS = vec_sub(line3BS, line7BS);
1273 vsum = vec_sum4s(vec_abs(line0CS), vsum);
1274 vsum = vec_sum4s(vec_abs(line1CS), vsum);
1275 vsum = vec_sum4s(vec_abs(line2CS), vsum);
1276 vsum = vec_sum4s(vec_abs(line3CS), vsum);
1277 vsum = vec_sum4s(vec_abs(line4CS), vsum);
1278 vsum = vec_sum4s(vec_abs(line5CS), vsum);
1279 vsum = vec_sum4s(vec_abs(line6CS), vsum);
1280 vsum = vec_sum4s(vec_abs(line7CS), vsum);
1281 vsum = vec_sums(vsum, (vector signed int)vzero);
1282 vsum = vec_splat(vsum, 3);
1283 vec_ste(vsum, 0, &sum);
1288 static int hadamard8_diff16_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
1290 score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
1294 score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
1299 static void vorbis_inverse_coupling_altivec(float *mag, float *ang,
1304 vector bool int t0, t1;
1305 const vector unsigned int v_31 = //XXX
1306 vec_add(vec_add(vec_splat_u32(15),vec_splat_u32(15)),vec_splat_u32(1));
1307 for (i = 0; i < blocksize; i += 4) {
1308 m = vec_ld(0, mag+i);
1309 a = vec_ld(0, ang+i);
1310 t0 = vec_cmple(m, (vector float)vec_splat_u32(0));
1311 t1 = vec_cmple(a, (vector float)vec_splat_u32(0));
1312 a = vec_xor(a, (vector float) vec_sl((vector unsigned int)t0, v_31));
1313 t0 = (vector bool int)vec_and(a, t1);
1314 t1 = (vector bool int)vec_andc(a, t1);
1315 a = vec_sub(m, (vector float)t1);
1316 m = vec_add(m, (vector float)t0);
1317 vec_stl(a, 0, ang+i);
1318 vec_stl(m, 0, mag+i);
1322 /* next one assumes that ((line_size % 8) == 0) */
1323 static void avg_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
1326 register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
1327 register vector unsigned char blockv, temp1, temp2, blocktemp;
1328 register vector unsigned short pixelssum1, pixelssum2, temp3;
1330 register const vector unsigned char vczero = (const vector unsigned char)
1332 register const vector unsigned short vctwo = (const vector unsigned short)
1335 temp1 = vec_ld(0, pixels);
1336 temp2 = vec_ld(16, pixels);
1337 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
1338 if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F) {
1341 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
1343 pixelsv1 = vec_mergeh(vczero, pixelsv1);
1344 pixelsv2 = vec_mergeh(vczero, pixelsv2);
1345 pixelssum1 = vec_add((vector unsigned short)pixelsv1,
1346 (vector unsigned short)pixelsv2);
1347 pixelssum1 = vec_add(pixelssum1, vctwo);
1349 for (i = 0; i < h ; i++) {
1350 int rightside = ((unsigned long)block & 0x0000000F);
1351 blockv = vec_ld(0, block);
1353 temp1 = vec_ld(line_size, pixels);
1354 temp2 = vec_ld(line_size + 16, pixels);
1355 pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
1356 if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F) {
1359 pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
1362 pixelsv1 = vec_mergeh(vczero, pixelsv1);
1363 pixelsv2 = vec_mergeh(vczero, pixelsv2);
1364 pixelssum2 = vec_add((vector unsigned short)pixelsv1,
1365 (vector unsigned short)pixelsv2);
1366 temp3 = vec_add(pixelssum1, pixelssum2);
1367 temp3 = vec_sra(temp3, vctwo);
1368 pixelssum1 = vec_add(pixelssum2, vctwo);
1369 pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
1372 blocktemp = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
1374 blocktemp = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
1377 blockv = vec_avg(blocktemp, blockv);
1378 vec_st(blockv, 0, block);
1381 pixels += line_size;
1385 void dsputil_init_altivec(DSPContext* c, AVCodecContext *avctx)
1387 c->pix_abs[0][1] = sad16_x2_altivec;
1388 c->pix_abs[0][2] = sad16_y2_altivec;
1389 c->pix_abs[0][3] = sad16_xy2_altivec;
1390 c->pix_abs[0][0] = sad16_altivec;
1391 c->pix_abs[1][0] = sad8_altivec;
1392 c->sad[0]= sad16_altivec;
1393 c->sad[1]= sad8_altivec;
1394 c->pix_norm1 = pix_norm1_altivec;
1395 c->sse[1]= sse8_altivec;
1396 c->sse[0]= sse16_altivec;
1397 c->pix_sum = pix_sum_altivec;
1398 c->diff_pixels = diff_pixels_altivec;
1399 c->get_pixels = get_pixels_altivec;
1400 c->clear_block = clear_block_altivec;
1401 c->add_bytes= add_bytes_altivec;
1402 c->put_pixels_tab[0][0] = put_pixels16_altivec;
1403 /* the two functions do the same thing, so use the same code */
1404 c->put_no_rnd_pixels_tab[0][0] = put_pixels16_altivec;
1405 c->avg_pixels_tab[0][0] = avg_pixels16_altivec;
1406 c->avg_pixels_tab[1][0] = avg_pixels8_altivec;
1407 c->avg_pixels_tab[1][3] = avg_pixels8_xy2_altivec;
1408 c->put_pixels_tab[1][3] = put_pixels8_xy2_altivec;
1409 c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_altivec;
1410 c->put_pixels_tab[0][3] = put_pixels16_xy2_altivec;
1411 c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_altivec;
1413 c->hadamard8_diff[0] = hadamard8_diff16_altivec;
1414 c->hadamard8_diff[1] = hadamard8_diff8x8_altivec;
1415 if (CONFIG_VORBIS_DECODER)
1416 c->vorbis_inverse_coupling = vorbis_inverse_coupling_altivec;