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
28 #include "libavutil/attributes.h"
29 #include "libavutil/ppc/types_altivec.h"
30 #include "libavutil/ppc/util_altivec.h"
31 #include "libavcodec/avcodec.h"
32 #include "libavcodec/dsputil.h"
33 #include "dsputil_altivec.h"
35 static int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2,
39 const vector unsigned char zero =
40 (const vector unsigned char) vec_splat_u8(0);
41 vector unsigned char perm1 = vec_lvsl(0, pix2);
42 vector unsigned char perm2 = vec_add(perm1, vec_splat_u8(1));
43 vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
44 vector signed int sumdiffs;
46 for (i = 0; i < h; i++) {
47 /* Read unaligned pixels into our vectors. The vectors are as follows:
48 * pix1v: pix1[0] - pix1[15]
49 * pix2v: pix2[0] - pix2[15] pix2iv: pix2[1] - pix2[16] */
50 vector unsigned char pix1v = vec_ld(0, pix1);
51 vector unsigned char pix2l = vec_ld(0, pix2);
52 vector unsigned char pix2r = vec_ld(16, pix2);
53 vector unsigned char pix2v = vec_perm(pix2l, pix2r, perm1);
54 vector unsigned char pix2iv = vec_perm(pix2l, pix2r, perm2);
56 /* Calculate the average vector. */
57 vector unsigned char avgv = vec_avg(pix2v, pix2iv);
59 /* Calculate a sum of abs differences vector. */
60 vector unsigned char t5 = vec_sub(vec_max(pix1v, avgv),
61 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,
81 const vector unsigned char zero =
82 (const vector unsigned char) vec_splat_u8(0);
83 vector unsigned char perm = vec_lvsl(0, pix2);
84 vector unsigned char pix1v, pix3v, avgv, t5;
85 vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
86 vector signed int sumdiffs;
87 uint8_t *pix3 = pix2 + line_size;
89 /* Due to the fact that pix3 = pix2 + line_size, the pix3 of one
90 * iteration becomes pix2 in the next iteration. We can use this
91 * fact to avoid a potentially expensive unaligned read, each
92 * time around the loop.
93 * Read unaligned pixels into our vectors. The vectors are as follows:
94 * pix2v: pix2[0] - pix2[15]
95 * Split the pixel vectors into shorts. */
96 vector unsigned char pix2l = vec_ld(0, pix2);
97 vector unsigned char pix2r = vec_ld(15, pix2);
98 vector unsigned char pix2v = vec_perm(pix2l, pix2r, perm);
100 for (i = 0; i < h; i++) {
101 /* Read unaligned pixels into our vectors. The vectors are as follows:
102 * pix1v: pix1[0] - pix1[15]
103 * pix3v: pix3[0] - pix3[15] */
104 pix1v = vec_ld(0, pix1);
106 pix2l = vec_ld(0, pix3);
107 pix2r = vec_ld(15, pix3);
108 pix3v = vec_perm(pix2l, pix2r, perm);
110 /* Calculate the average vector. */
111 avgv = vec_avg(pix2v, pix3v);
113 /* Calculate a sum of abs differences vector. */
114 t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
116 /* Add each 4 pixel group together and put 4 results into sad. */
117 sad = vec_sum4s(t5, sad);
124 /* Sum up the four partial sums, and put the result into s. */
125 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
126 sumdiffs = vec_splat(sumdiffs, 3);
127 vec_ste(sumdiffs, 0, &s);
131 static int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2,
132 int line_size, int h)
135 uint8_t *pix3 = pix2 + line_size;
136 const vector unsigned char zero =
137 (const vector unsigned char) vec_splat_u8(0);
138 const vector unsigned short two =
139 (const vector unsigned short) vec_splat_u16(2);
140 vector unsigned char avgv, t5;
141 vector unsigned char perm1 = vec_lvsl(0, pix2);
142 vector unsigned char perm2 = vec_add(perm1, vec_splat_u8(1));
143 vector unsigned char pix1v, pix3v, pix3iv;
144 vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
145 vector unsigned short avghv, avglv;
146 vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
147 vector signed int sumdiffs;
149 /* Due to the fact that pix3 = pix2 + line_size, the pix3 of one
150 * iteration becomes pix2 in the next iteration. We can use this
151 * fact to avoid a potentially expensive unaligned read, as well
152 * as some splitting, and vector addition each time around the loop.
153 * Read unaligned pixels into our vectors. The vectors are as follows:
154 * pix2v: pix2[0] - pix2[15] pix2iv: pix2[1] - pix2[16]
155 * Split the pixel vectors into shorts. */
156 vector unsigned char pix2l = vec_ld(0, pix2);
157 vector unsigned char pix2r = vec_ld(16, pix2);
158 vector unsigned char pix2v = vec_perm(pix2l, pix2r, perm1);
159 vector unsigned char pix2iv = vec_perm(pix2l, pix2r, perm2);
161 vector unsigned short pix2hv =
162 (vector unsigned short) vec_mergeh(zero, pix2v);
163 vector unsigned short pix2lv =
164 (vector unsigned short) vec_mergel(zero, pix2v);
165 vector unsigned short pix2ihv =
166 (vector unsigned short) vec_mergeh(zero, pix2iv);
167 vector unsigned short pix2ilv =
168 (vector unsigned short) vec_mergel(zero, pix2iv);
169 vector unsigned short t1 = vec_add(pix2hv, pix2ihv);
170 vector unsigned short t2 = vec_add(pix2lv, pix2ilv);
171 vector unsigned short t3, t4;
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 pix1v = vec_ld(0, pix1);
179 pix2l = vec_ld(0, pix3);
180 pix2r = vec_ld(16, pix3);
181 pix3v = vec_perm(pix2l, pix2r, perm1);
182 pix3iv = vec_perm(pix2l, pix2r, perm2);
184 /* Note that AltiVec does have vec_avg, but this works on vector pairs
185 * and rounds up. We could do avg(avg(a, b), avg(c, d)), but the
186 * rounding would mean that, for example, avg(3, 0, 0, 1) = 2, when
187 * it should be 1. Instead, we have to split the pixel vectors into
188 * vectors of shorts and do the averaging by hand. */
190 /* Split the pixel vectors into shorts. */
191 pix3hv = (vector unsigned short) vec_mergeh(zero, pix3v);
192 pix3lv = (vector unsigned short) vec_mergel(zero, pix3v);
193 pix3ihv = (vector unsigned short) vec_mergeh(zero, pix3iv);
194 pix3ilv = (vector unsigned short) vec_mergel(zero, pix3iv);
196 /* Do the averaging on them. */
197 t3 = vec_add(pix3hv, pix3ihv);
198 t4 = vec_add(pix3lv, pix3ilv);
200 avghv = vec_sr(vec_add(vec_add(t1, t3), two), two);
201 avglv = vec_sr(vec_add(vec_add(t2, t4), two), two);
203 /* Pack the shorts back into a result. */
204 avgv = vec_pack(avghv, avglv);
206 /* Calculate a sum of abs differences vector. */
207 t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
209 /* Add each 4 pixel group together and put 4 results into sad. */
210 sad = vec_sum4s(t5, sad);
214 /* Transfer the calculated values for pix3 into pix2. */
218 /* Sum up the four partial sums, and put the result into s. */
219 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
220 sumdiffs = vec_splat(sumdiffs, 3);
221 vec_ste(sumdiffs, 0, &s);
226 static int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2,
227 int line_size, int h)
230 const vector unsigned int zero =
231 (const vector unsigned int) vec_splat_u32(0);
232 vector unsigned char perm = vec_lvsl(0, pix2);
233 vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
234 vector signed int sumdiffs;
236 for (i = 0; i < h; i++) {
237 /* Read potentially unaligned pixels into t1 and t2. */
238 vector unsigned char pix2l = vec_ld(0, pix2);
239 vector unsigned char pix2r = vec_ld(15, pix2);
240 vector unsigned char t1 = vec_ld(0, pix1);
241 vector unsigned char t2 = vec_perm(pix2l, pix2r, perm);
243 /* Calculate a sum of abs differences vector. */
244 vector unsigned char t3 = vec_max(t1, t2);
245 vector unsigned char t4 = vec_min(t1, t2);
246 vector unsigned char t5 = vec_sub(t3, t4);
248 /* Add each 4 pixel group together and put 4 results into sad. */
249 sad = vec_sum4s(t5, sad);
255 /* Sum up the four partial sums, and put the result into s. */
256 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
257 sumdiffs = vec_splat(sumdiffs, 3);
258 vec_ste(sumdiffs, 0, &s);
263 static int sad8_altivec(void *v, uint8_t *pix1, uint8_t *pix2,
264 int line_size, int h)
267 const vector unsigned int zero =
268 (const vector unsigned int) vec_splat_u32(0);
269 const vector unsigned char permclear =
270 (vector unsigned char)
271 { 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0 };
272 vector unsigned char perm1 = vec_lvsl(0, pix1);
273 vector unsigned char perm2 = vec_lvsl(0, pix2);
274 vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
275 vector signed int sumdiffs;
277 for (i = 0; i < h; i++) {
278 /* Read potentially unaligned pixels into t1 and t2.
279 * Since we're reading 16 pixels, and actually only want 8,
280 * mask out the last 8 pixels. The 0s don't change the sum. */
281 vector unsigned char pix1l = vec_ld(0, pix1);
282 vector unsigned char pix1r = vec_ld(7, pix1);
283 vector unsigned char pix2l = vec_ld(0, pix2);
284 vector unsigned char pix2r = vec_ld(7, pix2);
285 vector unsigned char t1 = vec_and(vec_perm(pix1l, pix1r, perm1),
287 vector unsigned char t2 = vec_and(vec_perm(pix2l, pix2r, perm2),
290 /* Calculate a sum of abs differences vector. */
291 vector unsigned char t3 = vec_max(t1, t2);
292 vector unsigned char t4 = vec_min(t1, t2);
293 vector unsigned char t5 = vec_sub(t3, t4);
295 /* Add each 4 pixel group together and put 4 results into sad. */
296 sad = vec_sum4s(t5, sad);
302 /* Sum up the four partial sums, and put the result into s. */
303 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
304 sumdiffs = vec_splat(sumdiffs, 3);
305 vec_ste(sumdiffs, 0, &s);
310 static int pix_norm1_altivec(uint8_t *pix, int line_size)
313 const vector unsigned int zero =
314 (const vector unsigned int) vec_splat_u32(0);
315 vector unsigned char perm = vec_lvsl(0, pix);
316 vector unsigned int sv = (vector unsigned int) vec_splat_u32(0);
317 vector signed int sum;
319 for (i = 0; i < 16; i++) {
320 /* Read the potentially unaligned pixels. */
321 vector unsigned char pixl = vec_ld(0, pix);
322 vector unsigned char pixr = vec_ld(15, pix);
323 vector unsigned char pixv = vec_perm(pixl, pixr, perm);
325 /* Square the values, and add them to our sum. */
326 sv = vec_msum(pixv, pixv, sv);
330 /* Sum up the four partial sums, and put the result into s. */
331 sum = vec_sums((vector signed int) sv, (vector signed int) zero);
332 sum = vec_splat(sum, 3);
338 /* Sum of Squared Errors for an 8x8 block, AltiVec-enhanced.
339 * It's the sad8_altivec code above w/ squaring added. */
340 static int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2,
341 int line_size, int h)
344 const vector unsigned int zero =
345 (const vector unsigned int) vec_splat_u32(0);
346 const vector unsigned char permclear =
347 (vector unsigned char)
348 { 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0 };
349 vector unsigned char perm1 = vec_lvsl(0, pix1);
350 vector unsigned char perm2 = vec_lvsl(0, pix2);
351 vector unsigned int sum = (vector unsigned int) vec_splat_u32(0);
352 vector signed int sumsqr;
354 for (i = 0; i < h; i++) {
355 /* Read potentially unaligned pixels into t1 and t2.
356 * Since we're reading 16 pixels, and actually only want 8,
357 * mask out the last 8 pixels. The 0s don't change the sum. */
358 vector unsigned char pix1l = vec_ld(0, pix1);
359 vector unsigned char pix1r = vec_ld(7, pix1);
360 vector unsigned char pix2l = vec_ld(0, pix2);
361 vector unsigned char pix2r = vec_ld(7, pix2);
362 vector unsigned char t1 = vec_and(vec_perm(pix1l, pix1r, perm1),
364 vector unsigned char t2 = vec_and(vec_perm(pix2l, pix2r, perm2),
367 /* Since we want to use unsigned chars, we can take advantage
368 * of the fact that abs(a - b) ^ 2 = (a - b) ^ 2. */
370 /* Calculate abs differences vector. */
371 vector unsigned char t3 = vec_max(t1, t2);
372 vector unsigned char t4 = vec_min(t1, t2);
373 vector unsigned char t5 = vec_sub(t3, t4);
375 /* Square the values and add them to our sum. */
376 sum = vec_msum(t5, t5, sum);
382 /* Sum up the four partial sums, and put the result into s. */
383 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
384 sumsqr = vec_splat(sumsqr, 3);
385 vec_ste(sumsqr, 0, &s);
390 /* Sum of Squared Errors for a 16x16 block, AltiVec-enhanced.
391 * It's the sad16_altivec code above w/ squaring added. */
392 static int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2,
393 int line_size, int h)
396 const vector unsigned int zero =
397 (const vector unsigned int) vec_splat_u32(0);
398 vector unsigned char perm = vec_lvsl(0, pix2);
399 vector unsigned int sum = (vector unsigned int) vec_splat_u32(0);
400 vector signed int sumsqr;
402 for (i = 0; i < h; i++) {
403 /* Read potentially unaligned pixels into t1 and t2. */
404 vector unsigned char pix2l = vec_ld(0, pix2);
405 vector unsigned char pix2r = vec_ld(15, pix2);
406 vector unsigned char t1 = vec_ld(0, pix1);
407 vector unsigned char t2 = vec_perm(pix2l, pix2r, perm);
409 /* Since we want to use unsigned chars, we can take advantage
410 * of the fact that abs(a - b) ^ 2 = (a - b) ^ 2. */
412 /* Calculate abs differences vector. */
413 vector unsigned char t3 = vec_max(t1, t2);
414 vector unsigned char t4 = vec_min(t1, t2);
415 vector unsigned char t5 = vec_sub(t3, t4);
417 /* Square the values and add them to our sum. */
418 sum = vec_msum(t5, t5, sum);
424 /* Sum up the four partial sums, and put the result into s. */
425 sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
426 sumsqr = vec_splat(sumsqr, 3);
427 vec_ste(sumsqr, 0, &s);
432 static int pix_sum_altivec(uint8_t *pix, int line_size)
435 const vector unsigned int zero =
436 (const vector unsigned int) vec_splat_u32(0);
437 vector unsigned char perm = vec_lvsl(0, pix);
438 vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
439 vector signed int sumdiffs;
441 for (i = 0; i < 16; i++) {
442 /* Read the potentially unaligned 16 pixels into t1. */
443 vector unsigned char pixl = vec_ld(0, pix);
444 vector unsigned char pixr = vec_ld(15, pix);
445 vector unsigned char t1 = vec_perm(pixl, pixr, perm);
447 /* Add each 4 pixel group together and put 4 results into sad. */
448 sad = vec_sum4s(t1, sad);
453 /* Sum up the four partial sums, and put the result into s. */
454 sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
455 sumdiffs = vec_splat(sumdiffs, 3);
456 vec_ste(sumdiffs, 0, &s);
461 static void get_pixels_altivec(int16_t *restrict block, const uint8_t *pixels,
465 vector unsigned char perm = vec_lvsl(0, pixels);
466 const vector unsigned char zero =
467 (const vector unsigned char) vec_splat_u8(0);
469 for (i = 0; i < 8; i++) {
470 /* Read potentially unaligned pixels.
471 * We're reading 16 pixels, and actually only want 8,
472 * but we simply ignore the extras. */
473 vector unsigned char pixl = vec_ld(0, pixels);
474 vector unsigned char pixr = vec_ld(7, pixels);
475 vector unsigned char bytes = vec_perm(pixl, pixr, perm);
477 // Convert the bytes into shorts.
478 vector signed short shorts = (vector signed short) vec_mergeh(zero,
481 // Save the data to the block, we assume the block is 16-byte aligned.
482 vec_st(shorts, i * 16, (vector signed short *) block);
488 static void diff_pixels_altivec(int16_t *restrict block, const uint8_t *s1,
489 const uint8_t *s2, int stride)
492 vector unsigned char perm1 = vec_lvsl(0, s1);
493 vector unsigned char perm2 = vec_lvsl(0, s2);
494 const vector unsigned char zero =
495 (const vector unsigned char) vec_splat_u8(0);
496 vector signed short shorts1, shorts2;
498 for (i = 0; i < 4; i++) {
499 /* Read potentially unaligned pixels.
500 * We're reading 16 pixels, and actually only want 8,
501 * but we simply ignore the extras. */
502 vector unsigned char pixl = vec_ld(0, s1);
503 vector unsigned char pixr = vec_ld(15, s1);
504 vector unsigned char bytes = vec_perm(pixl, pixr, perm1);
506 // Convert the bytes into shorts.
507 shorts1 = (vector signed short) vec_mergeh(zero, bytes);
509 // Do the same for the second block of pixels.
510 pixl = vec_ld(0, s2);
511 pixr = vec_ld(15, s2);
512 bytes = vec_perm(pixl, pixr, perm2);
514 // Convert the bytes into shorts.
515 shorts2 = (vector signed short) vec_mergeh(zero, bytes);
517 // Do the subtraction.
518 shorts1 = vec_sub(shorts1, shorts2);
520 // Save the data to the block, we assume the block is 16-byte aligned.
521 vec_st(shorts1, 0, (vector signed short *) block);
527 /* The code below is a copy of the code above...
528 * This is a manual unroll. */
530 /* Read potentially unaligned pixels.
531 * We're reading 16 pixels, and actually only want 8,
532 * but we simply ignore the extras. */
533 pixl = vec_ld(0, s1);
534 pixr = vec_ld(15, s1);
535 bytes = vec_perm(pixl, pixr, perm1);
537 // Convert the bytes into shorts.
538 shorts1 = (vector signed short) vec_mergeh(zero, bytes);
540 // Do the same for the second block of pixels.
541 pixl = vec_ld(0, s2);
542 pixr = vec_ld(15, s2);
543 bytes = vec_perm(pixl, pixr, perm2);
545 // Convert the bytes into shorts.
546 shorts2 = (vector signed short) vec_mergeh(zero, bytes);
548 // Do the subtraction.
549 shorts1 = vec_sub(shorts1, shorts2);
551 // Save the data to the block, we assume the block is 16-byte aligned.
552 vec_st(shorts1, 0, (vector signed short *) block);
560 static void clear_block_altivec(int16_t *block)
563 vec_st(zero_s16v, 0, block);
564 vec_st(zero_s16v, 16, block);
565 vec_st(zero_s16v, 32, block);
566 vec_st(zero_s16v, 48, block);
567 vec_st(zero_s16v, 64, block);
568 vec_st(zero_s16v, 80, block);
569 vec_st(zero_s16v, 96, block);
570 vec_st(zero_s16v, 112, block);
573 static void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w)
576 register vector unsigned char vdst, vsrc;
578 /* dst and src are 16 bytes-aligned (guaranteed). */
579 for (i = 0; i + 15 < w; i += 16) {
580 vdst = vec_ld(i, (unsigned char *) dst);
581 vsrc = vec_ld(i, (unsigned char *) src);
582 vdst = vec_add(vsrc, vdst);
583 vec_st(vdst, i, (unsigned char *) dst);
585 /* If w is not a multiple of 16. */
590 static int hadamard8_diff8x8_altivec(/* MpegEncContext */ void *s, uint8_t *dst,
591 uint8_t *src, int stride, int h)
594 register const vector unsigned char vzero =
595 (const vector unsigned char) vec_splat_u8(0);
596 register vector signed short temp0, temp1, temp2, temp3, temp4,
599 register const vector signed short vprod1 =
600 (const vector signed short) { 1, -1, 1, -1, 1, -1, 1, -1 };
601 register const vector signed short vprod2 =
602 (const vector signed short) { 1, 1, -1, -1, 1, 1, -1, -1 };
603 register const vector signed short vprod3 =
604 (const vector signed short) { 1, 1, 1, 1, -1, -1, -1, -1 };
605 register const vector unsigned char perm1 =
606 (const vector unsigned char)
607 { 0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
608 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D };
609 register const vector unsigned char perm2 =
610 (const vector unsigned char)
611 { 0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
612 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B };
613 register const vector unsigned char perm3 =
614 (const vector unsigned char)
615 { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
616 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07 };
618 #define ONEITERBUTTERFLY(i, res) \
620 register vector unsigned char src1 = vec_ld(stride * i, src); \
621 register vector unsigned char src2 = vec_ld(stride * i + 15, src); \
622 register vector unsigned char srcO = \
623 vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
624 register vector unsigned char dst1 = vec_ld(stride * i, dst); \
625 register vector unsigned char dst2 = vec_ld(stride * i + 15, dst); \
626 register vector unsigned char dstO = \
627 vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
629 /* Promote the unsigned chars to signed shorts. */ \
630 /* We're in the 8x8 function, we only care for the first 8. */ \
631 register vector signed short srcV = \
632 (vector signed short) vec_mergeh((vector signed char) vzero, \
633 (vector signed char) srcO); \
634 register vector signed short dstV = \
635 (vector signed short) vec_mergeh((vector signed char) vzero, \
636 (vector signed char) dstO); \
638 /* subtractions inside the first butterfly */ \
639 register vector signed short but0 = vec_sub(srcV, dstV); \
640 register vector signed short op1 = vec_perm(but0, but0, perm1); \
641 register vector signed short but1 = vec_mladd(but0, vprod1, op1); \
642 register vector signed short op2 = vec_perm(but1, but1, perm2); \
643 register vector signed short but2 = vec_mladd(but1, vprod2, op2); \
644 register vector signed short op3 = vec_perm(but2, but2, perm3); \
645 res = vec_mladd(but2, vprod3, op3); \
647 ONEITERBUTTERFLY(0, temp0);
648 ONEITERBUTTERFLY(1, temp1);
649 ONEITERBUTTERFLY(2, temp2);
650 ONEITERBUTTERFLY(3, temp3);
651 ONEITERBUTTERFLY(4, temp4);
652 ONEITERBUTTERFLY(5, temp5);
653 ONEITERBUTTERFLY(6, temp6);
654 ONEITERBUTTERFLY(7, temp7);
656 #undef ONEITERBUTTERFLY
658 register vector signed int vsum;
659 register vector signed short line0 = vec_add(temp0, temp1);
660 register vector signed short line1 = vec_sub(temp0, temp1);
661 register vector signed short line2 = vec_add(temp2, temp3);
662 register vector signed short line3 = vec_sub(temp2, temp3);
663 register vector signed short line4 = vec_add(temp4, temp5);
664 register vector signed short line5 = vec_sub(temp4, temp5);
665 register vector signed short line6 = vec_add(temp6, temp7);
666 register vector signed short line7 = vec_sub(temp6, temp7);
668 register vector signed short line0B = vec_add(line0, line2);
669 register vector signed short line2B = vec_sub(line0, line2);
670 register vector signed short line1B = vec_add(line1, line3);
671 register vector signed short line3B = vec_sub(line1, line3);
672 register vector signed short line4B = vec_add(line4, line6);
673 register vector signed short line6B = vec_sub(line4, line6);
674 register vector signed short line5B = vec_add(line5, line7);
675 register vector signed short line7B = vec_sub(line5, line7);
677 register vector signed short line0C = vec_add(line0B, line4B);
678 register vector signed short line4C = vec_sub(line0B, line4B);
679 register vector signed short line1C = vec_add(line1B, line5B);
680 register vector signed short line5C = vec_sub(line1B, line5B);
681 register vector signed short line2C = vec_add(line2B, line6B);
682 register vector signed short line6C = vec_sub(line2B, line6B);
683 register vector signed short line3C = vec_add(line3B, line7B);
684 register vector signed short line7C = vec_sub(line3B, line7B);
686 vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
687 vsum = vec_sum4s(vec_abs(line1C), vsum);
688 vsum = vec_sum4s(vec_abs(line2C), vsum);
689 vsum = vec_sum4s(vec_abs(line3C), vsum);
690 vsum = vec_sum4s(vec_abs(line4C), vsum);
691 vsum = vec_sum4s(vec_abs(line5C), vsum);
692 vsum = vec_sum4s(vec_abs(line6C), vsum);
693 vsum = vec_sum4s(vec_abs(line7C), vsum);
694 vsum = vec_sums(vsum, (vector signed int) vzero);
695 vsum = vec_splat(vsum, 3);
696 vec_ste(vsum, 0, &sum);
702 * 16x8 works with 16 elements; it allows to avoid replicating loads, and
703 * gives the compiler more room for scheduling. It's only used from
704 * inside hadamard8_diff16_altivec.
706 * Unfortunately, it seems gcc-3.3 is a bit dumb, and the compiled code has
707 * a LOT of spill code, it seems gcc (unlike xlc) cannot keep everything in
708 * registers by itself. The following code includes hand-made register
709 * allocation. It's not clean, but on a 7450 the resulting code is much faster
710 * (best case falls from 700+ cycles to 550).
712 * xlc doesn't add spill code, but it doesn't know how to schedule for the
713 * 7450, and its code isn't much faster than gcc-3.3 on the 7450 (but uses
714 * 25% fewer instructions...)
716 * On the 970, the hand-made RA is still a win (around 690 vs. around 780),
717 * but xlc goes to around 660 on the regular C code...
719 static int hadamard8_diff16x8_altivec(/* MpegEncContext */ void *s, uint8_t *dst,
720 uint8_t *src, int stride, int h)
723 register vector signed short
724 temp0 __asm__ ("v0"),
725 temp1 __asm__ ("v1"),
726 temp2 __asm__ ("v2"),
727 temp3 __asm__ ("v3"),
728 temp4 __asm__ ("v4"),
729 temp5 __asm__ ("v5"),
730 temp6 __asm__ ("v6"),
731 temp7 __asm__ ("v7");
732 register vector signed short
733 temp0S __asm__ ("v8"),
734 temp1S __asm__ ("v9"),
735 temp2S __asm__ ("v10"),
736 temp3S __asm__ ("v11"),
737 temp4S __asm__ ("v12"),
738 temp5S __asm__ ("v13"),
739 temp6S __asm__ ("v14"),
740 temp7S __asm__ ("v15");
741 register const vector unsigned char vzero __asm__ ("v31") =
742 (const vector unsigned char) vec_splat_u8(0);
744 register const vector signed short vprod1 __asm__ ("v16") =
745 (const vector signed short) { 1, -1, 1, -1, 1, -1, 1, -1 };
747 register const vector signed short vprod2 __asm__ ("v17") =
748 (const vector signed short) { 1, 1, -1, -1, 1, 1, -1, -1 };
750 register const vector signed short vprod3 __asm__ ("v18") =
751 (const vector signed short) { 1, 1, 1, 1, -1, -1, -1, -1 };
753 register const vector unsigned char perm1 __asm__ ("v19") =
754 (const vector unsigned char)
755 { 0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
756 0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D };
758 register const vector unsigned char perm2 __asm__ ("v20") =
759 (const vector unsigned char)
760 { 0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
761 0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B };
763 register const vector unsigned char perm3 __asm__ ("v21") =
764 (const vector unsigned char)
765 { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
766 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07 };
768 #define ONEITERBUTTERFLY(i, res1, res2) \
770 register vector unsigned char src1 __asm__ ("v22") = \
771 vec_ld(stride * i, src); \
772 register vector unsigned char src2 __asm__ ("v23") = \
773 vec_ld(stride * i + 16, src); \
774 register vector unsigned char srcO __asm__ ("v22") = \
775 vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
776 register vector unsigned char dst1 __asm__ ("v24") = \
777 vec_ld(stride * i, dst); \
778 register vector unsigned char dst2 __asm__ ("v25") = \
779 vec_ld(stride * i + 16, dst); \
780 register vector unsigned char dstO __asm__ ("v23") = \
781 vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
783 /* Promote the unsigned chars to signed shorts. */ \
784 register vector signed short srcV __asm__ ("v24") = \
785 (vector signed short) vec_mergeh((vector signed char) vzero, \
786 (vector signed char) srcO); \
787 register vector signed short dstV __asm__ ("v25") = \
788 (vector signed short) vec_mergeh((vector signed char) vzero, \
789 (vector signed char) dstO); \
790 register vector signed short srcW __asm__ ("v26") = \
791 (vector signed short) vec_mergel((vector signed char) vzero, \
792 (vector signed char) srcO); \
793 register vector signed short dstW __asm__ ("v27") = \
794 (vector signed short) vec_mergel((vector signed char) vzero, \
795 (vector signed char) dstO); \
797 /* subtractions inside the first butterfly */ \
798 register vector signed short but0 __asm__ ("v28") = \
799 vec_sub(srcV, dstV); \
800 register vector signed short but0S __asm__ ("v29") = \
801 vec_sub(srcW, dstW); \
802 register vector signed short op1 __asm__ ("v30") = \
803 vec_perm(but0, but0, perm1); \
804 register vector signed short but1 __asm__ ("v22") = \
805 vec_mladd(but0, vprod1, op1); \
806 register vector signed short op1S __asm__ ("v23") = \
807 vec_perm(but0S, but0S, perm1); \
808 register vector signed short but1S __asm__ ("v24") = \
809 vec_mladd(but0S, vprod1, op1S); \
810 register vector signed short op2 __asm__ ("v25") = \
811 vec_perm(but1, but1, perm2); \
812 register vector signed short but2 __asm__ ("v26") = \
813 vec_mladd(but1, vprod2, op2); \
814 register vector signed short op2S __asm__ ("v27") = \
815 vec_perm(but1S, but1S, perm2); \
816 register vector signed short but2S __asm__ ("v28") = \
817 vec_mladd(but1S, vprod2, op2S); \
818 register vector signed short op3 __asm__ ("v29") = \
819 vec_perm(but2, but2, perm3); \
820 register vector signed short op3S __asm__ ("v30") = \
821 vec_perm(but2S, but2S, perm3); \
822 res1 = vec_mladd(but2, vprod3, op3); \
823 res2 = vec_mladd(but2S, vprod3, op3S); \
825 ONEITERBUTTERFLY(0, temp0, temp0S);
826 ONEITERBUTTERFLY(1, temp1, temp1S);
827 ONEITERBUTTERFLY(2, temp2, temp2S);
828 ONEITERBUTTERFLY(3, temp3, temp3S);
829 ONEITERBUTTERFLY(4, temp4, temp4S);
830 ONEITERBUTTERFLY(5, temp5, temp5S);
831 ONEITERBUTTERFLY(6, temp6, temp6S);
832 ONEITERBUTTERFLY(7, temp7, temp7S);
834 #undef ONEITERBUTTERFLY
836 register vector signed int vsum;
838 register vector signed short line0 = vec_add(temp0, temp1);
839 register vector signed short line1 = vec_sub(temp0, temp1);
840 register vector signed short line2 = vec_add(temp2, temp3);
841 register vector signed short line3 = vec_sub(temp2, temp3);
842 register vector signed short line4 = vec_add(temp4, temp5);
843 register vector signed short line5 = vec_sub(temp4, temp5);
844 register vector signed short line6 = vec_add(temp6, temp7);
845 register vector signed short line7 = vec_sub(temp6, temp7);
847 register vector signed short line0B = vec_add(line0, line2);
848 register vector signed short line2B = vec_sub(line0, line2);
849 register vector signed short line1B = vec_add(line1, line3);
850 register vector signed short line3B = vec_sub(line1, line3);
851 register vector signed short line4B = vec_add(line4, line6);
852 register vector signed short line6B = vec_sub(line4, line6);
853 register vector signed short line5B = vec_add(line5, line7);
854 register vector signed short line7B = vec_sub(line5, line7);
856 register vector signed short line0C = vec_add(line0B, line4B);
857 register vector signed short line4C = vec_sub(line0B, line4B);
858 register vector signed short line1C = vec_add(line1B, line5B);
859 register vector signed short line5C = vec_sub(line1B, line5B);
860 register vector signed short line2C = vec_add(line2B, line6B);
861 register vector signed short line6C = vec_sub(line2B, line6B);
862 register vector signed short line3C = vec_add(line3B, line7B);
863 register vector signed short line7C = vec_sub(line3B, line7B);
865 register vector signed short line0S = vec_add(temp0S, temp1S);
866 register vector signed short line1S = vec_sub(temp0S, temp1S);
867 register vector signed short line2S = vec_add(temp2S, temp3S);
868 register vector signed short line3S = vec_sub(temp2S, temp3S);
869 register vector signed short line4S = vec_add(temp4S, temp5S);
870 register vector signed short line5S = vec_sub(temp4S, temp5S);
871 register vector signed short line6S = vec_add(temp6S, temp7S);
872 register vector signed short line7S = vec_sub(temp6S, temp7S);
874 register vector signed short line0BS = vec_add(line0S, line2S);
875 register vector signed short line2BS = vec_sub(line0S, line2S);
876 register vector signed short line1BS = vec_add(line1S, line3S);
877 register vector signed short line3BS = vec_sub(line1S, line3S);
878 register vector signed short line4BS = vec_add(line4S, line6S);
879 register vector signed short line6BS = vec_sub(line4S, line6S);
880 register vector signed short line5BS = vec_add(line5S, line7S);
881 register vector signed short line7BS = vec_sub(line5S, line7S);
883 register vector signed short line0CS = vec_add(line0BS, line4BS);
884 register vector signed short line4CS = vec_sub(line0BS, line4BS);
885 register vector signed short line1CS = vec_add(line1BS, line5BS);
886 register vector signed short line5CS = vec_sub(line1BS, line5BS);
887 register vector signed short line2CS = vec_add(line2BS, line6BS);
888 register vector signed short line6CS = vec_sub(line2BS, line6BS);
889 register vector signed short line3CS = vec_add(line3BS, line7BS);
890 register vector signed short line7CS = vec_sub(line3BS, line7BS);
892 vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
893 vsum = vec_sum4s(vec_abs(line1C), vsum);
894 vsum = vec_sum4s(vec_abs(line2C), vsum);
895 vsum = vec_sum4s(vec_abs(line3C), vsum);
896 vsum = vec_sum4s(vec_abs(line4C), vsum);
897 vsum = vec_sum4s(vec_abs(line5C), vsum);
898 vsum = vec_sum4s(vec_abs(line6C), vsum);
899 vsum = vec_sum4s(vec_abs(line7C), vsum);
901 vsum = vec_sum4s(vec_abs(line0CS), vsum);
902 vsum = vec_sum4s(vec_abs(line1CS), vsum);
903 vsum = vec_sum4s(vec_abs(line2CS), vsum);
904 vsum = vec_sum4s(vec_abs(line3CS), vsum);
905 vsum = vec_sum4s(vec_abs(line4CS), vsum);
906 vsum = vec_sum4s(vec_abs(line5CS), vsum);
907 vsum = vec_sum4s(vec_abs(line6CS), vsum);
908 vsum = vec_sum4s(vec_abs(line7CS), vsum);
909 vsum = vec_sums(vsum, (vector signed int) vzero);
910 vsum = vec_splat(vsum, 3);
911 vec_ste(vsum, 0, &sum);
916 static int hadamard8_diff16_altivec(/* MpegEncContext */ void *s, uint8_t *dst,
917 uint8_t *src, int stride, int h)
919 int score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
924 score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
929 av_cold void ff_dsputil_init_altivec(DSPContext *c, AVCodecContext *avctx,
930 unsigned high_bit_depth)
932 c->pix_abs[0][1] = sad16_x2_altivec;
933 c->pix_abs[0][2] = sad16_y2_altivec;
934 c->pix_abs[0][3] = sad16_xy2_altivec;
935 c->pix_abs[0][0] = sad16_altivec;
936 c->pix_abs[1][0] = sad8_altivec;
938 c->sad[0] = sad16_altivec;
939 c->sad[1] = sad8_altivec;
940 c->sse[0] = sse16_altivec;
941 c->sse[1] = sse8_altivec;
943 c->pix_norm1 = pix_norm1_altivec;
944 c->pix_sum = pix_sum_altivec;
946 c->diff_pixels = diff_pixels_altivec;
947 c->add_bytes = add_bytes_altivec;
949 if (!high_bit_depth) {
950 c->get_pixels = get_pixels_altivec;
951 c->clear_block = clear_block_altivec;
954 c->hadamard8_diff[0] = hadamard8_diff16_altivec;
955 c->hadamard8_diff[1] = hadamard8_diff8x8_altivec;