2
* Copyright (c) 2002 Brian Foley
3
* Copyright (c) 2002 Dieter Shirley
4
* Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
6
* This library is free software; you can redistribute it and/or
7
* modify it under the terms of the GNU Lesser General Public
8
* License as published by the Free Software Foundation; either
9
* version 2 of the License, or (at your option) any later version.
11
* This library is distributed in the hope that it will be useful,
12
* but WITHOUT ANY WARRANTY; without even the implied warranty of
13
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14
* Lesser General Public License for more details.
16
* You should have received a copy of the GNU Lesser General Public
17
* License along with this library; if not, write to the Free Software
18
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21
#include "../dsputil.h"
23
#include "gcc_fixes.h"
25
#include "dsputil_altivec.h"
28
#include <sys/sysctl.h>
29
#else /* CONFIG_DARWIN */
33
static sigjmp_buf jmpbuf;
34
static volatile sig_atomic_t canjump = 0;
36
static void sigill_handler (int sig)
39
signal (sig, SIG_DFL);
44
siglongjmp (jmpbuf, 1);
46
#endif /* CONFIG_DARWIN */
48
int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
51
int s __attribute__((aligned(16)));
52
const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
53
vector unsigned char *tv;
54
vector unsigned char pix1v, pix2v, pix2iv, avgv, t5;
55
vector unsigned int sad;
56
vector signed int sumdiffs;
59
sad = (vector unsigned int)vec_splat_u32(0);
62
Read unaligned pixels into our vectors. The vectors are as follows:
63
pix1v: pix1[0]-pix1[15]
64
pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16]
66
tv = (vector unsigned char *) pix1;
67
pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
69
tv = (vector unsigned char *) &pix2[0];
70
pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
72
tv = (vector unsigned char *) &pix2[1];
73
pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1]));
75
/* Calculate the average vector */
76
avgv = vec_avg(pix2v, pix2iv);
78
/* Calculate a sum of abs differences vector */
79
t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
81
/* Add each 4 pixel group together and put 4 results into sad */
82
sad = vec_sum4s(t5, sad);
87
/* Sum up the four partial sums, and put the result into s */
88
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
89
sumdiffs = vec_splat(sumdiffs, 3);
90
vec_ste(sumdiffs, 0, &s);
95
int sad16_y2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
98
int s __attribute__((aligned(16)));
99
const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
100
vector unsigned char *tv;
101
vector unsigned char pix1v, pix2v, pix3v, avgv, t5;
102
vector unsigned int sad;
103
vector signed int sumdiffs;
104
uint8_t *pix3 = pix2 + line_size;
107
sad = (vector unsigned int)vec_splat_u32(0);
110
Due to the fact that pix3 = pix2 + line_size, the pix3 of one
111
iteration becomes pix2 in the next iteration. We can use this
112
fact to avoid a potentially expensive unaligned read, each
113
time around the loop.
114
Read unaligned pixels into our vectors. The vectors are as follows:
115
pix2v: pix2[0]-pix2[15]
116
Split the pixel vectors into shorts
118
tv = (vector unsigned char *) &pix2[0];
119
pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
123
Read unaligned pixels into our vectors. The vectors are as follows:
124
pix1v: pix1[0]-pix1[15]
125
pix3v: pix3[0]-pix3[15]
127
tv = (vector unsigned char *) pix1;
128
pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
130
tv = (vector unsigned char *) &pix3[0];
131
pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0]));
133
/* Calculate the average vector */
134
avgv = vec_avg(pix2v, pix3v);
136
/* Calculate a sum of abs differences vector */
137
t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
139
/* Add each 4 pixel group together and put 4 results into sad */
140
sad = vec_sum4s(t5, sad);
148
/* Sum up the four partial sums, and put the result into s */
149
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
150
sumdiffs = vec_splat(sumdiffs, 3);
151
vec_ste(sumdiffs, 0, &s);
155
int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
158
int s __attribute__((aligned(16)));
159
uint8_t *pix3 = pix2 + line_size;
160
const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
161
const_vector unsigned short two = (const_vector unsigned short)vec_splat_u16(2);
162
vector unsigned char *tv, avgv, t5;
163
vector unsigned char pix1v, pix2v, pix3v, pix2iv, pix3iv;
164
vector unsigned short pix2lv, pix2hv, pix2ilv, pix2ihv;
165
vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
166
vector unsigned short avghv, avglv;
167
vector unsigned short t1, t2, t3, t4;
168
vector unsigned int sad;
169
vector signed int sumdiffs;
171
sad = (vector unsigned int)vec_splat_u32(0);
176
Due to the fact that pix3 = pix2 + line_size, the pix3 of one
177
iteration becomes pix2 in the next iteration. We can use this
178
fact to avoid a potentially expensive unaligned read, as well
179
as some splitting, and vector addition each time around the loop.
180
Read unaligned pixels into our vectors. The vectors are as follows:
181
pix2v: pix2[0]-pix2[15] pix2iv: pix2[1]-pix2[16]
182
Split the pixel vectors into shorts
184
tv = (vector unsigned char *) &pix2[0];
185
pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));
187
tv = (vector unsigned char *) &pix2[1];
188
pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1]));
190
pix2hv = (vector unsigned short) vec_mergeh(zero, pix2v);
191
pix2lv = (vector unsigned short) vec_mergel(zero, pix2v);
192
pix2ihv = (vector unsigned short) vec_mergeh(zero, pix2iv);
193
pix2ilv = (vector unsigned short) vec_mergel(zero, pix2iv);
194
t1 = vec_add(pix2hv, pix2ihv);
195
t2 = vec_add(pix2lv, pix2ilv);
199
Read unaligned pixels into our vectors. The vectors are as follows:
200
pix1v: pix1[0]-pix1[15]
201
pix3v: pix3[0]-pix3[15] pix3iv: pix3[1]-pix3[16]
203
tv = (vector unsigned char *) pix1;
204
pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));
206
tv = (vector unsigned char *) &pix3[0];
207
pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0]));
209
tv = (vector unsigned char *) &pix3[1];
210
pix3iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[1]));
213
Note that Altivec does have vec_avg, but this works on vector pairs
214
and rounds up. We could do avg(avg(a,b),avg(c,d)), but the rounding
215
would mean that, for example, avg(3,0,0,1) = 2, when it should be 1.
216
Instead, we have to split the pixel vectors into vectors of shorts,
217
and do the averaging by hand.
220
/* Split the pixel vectors into shorts */
221
pix3hv = (vector unsigned short) vec_mergeh(zero, pix3v);
222
pix3lv = (vector unsigned short) vec_mergel(zero, pix3v);
223
pix3ihv = (vector unsigned short) vec_mergeh(zero, pix3iv);
224
pix3ilv = (vector unsigned short) vec_mergel(zero, pix3iv);
226
/* Do the averaging on them */
227
t3 = vec_add(pix3hv, pix3ihv);
228
t4 = vec_add(pix3lv, pix3ilv);
230
avghv = vec_sr(vec_add(vec_add(t1, t3), two), two);
231
avglv = vec_sr(vec_add(vec_add(t2, t4), two), two);
233
/* Pack the shorts back into a result */
234
avgv = vec_pack(avghv, avglv);
236
/* Calculate a sum of abs differences vector */
237
t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
239
/* Add each 4 pixel group together and put 4 results into sad */
240
sad = vec_sum4s(t5, sad);
244
/* Transfer the calculated values for pix3 into pix2 */
248
/* Sum up the four partial sums, and put the result into s */
249
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
250
sumdiffs = vec_splat(sumdiffs, 3);
251
vec_ste(sumdiffs, 0, &s);
256
int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
259
int s __attribute__((aligned(16)));
260
const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
261
vector unsigned char perm1, perm2, *pix1v, *pix2v;
262
vector unsigned char t1, t2, t3,t4, t5;
263
vector unsigned int sad;
264
vector signed int sumdiffs;
266
sad = (vector unsigned int)vec_splat_u32(0);
270
/* Read potentially unaligned pixels into t1 and t2 */
271
perm1 = vec_lvsl(0, pix1);
272
pix1v = (vector unsigned char *) pix1;
273
perm2 = vec_lvsl(0, pix2);
274
pix2v = (vector unsigned char *) pix2;
275
t1 = vec_perm(pix1v[0], pix1v[1], perm1);
276
t2 = vec_perm(pix2v[0], pix2v[1], perm2);
278
/* Calculate a sum of abs differences vector */
279
t3 = vec_max(t1, t2);
280
t4 = vec_min(t1, t2);
281
t5 = vec_sub(t3, t4);
283
/* Add each 4 pixel group together and put 4 results into sad */
284
sad = vec_sum4s(t5, sad);
290
/* Sum up the four partial sums, and put the result into s */
291
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
292
sumdiffs = vec_splat(sumdiffs, 3);
293
vec_ste(sumdiffs, 0, &s);
298
int sad8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
301
int s __attribute__((aligned(16)));
302
const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
303
vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
304
vector unsigned char t1, t2, t3,t4, t5;
305
vector unsigned int sad;
306
vector signed int sumdiffs;
308
sad = (vector unsigned int)vec_splat_u32(0);
310
permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0);
313
/* Read potentially unaligned pixels into t1 and t2
314
Since we're reading 16 pixels, and actually only want 8,
315
mask out the last 8 pixels. The 0s don't change the sum. */
316
perm1 = vec_lvsl(0, pix1);
317
pix1v = (vector unsigned char *) pix1;
318
perm2 = vec_lvsl(0, pix2);
319
pix2v = (vector unsigned char *) pix2;
320
t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
321
t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
323
/* Calculate a sum of abs differences vector */
324
t3 = vec_max(t1, t2);
325
t4 = vec_min(t1, t2);
326
t5 = vec_sub(t3, t4);
328
/* Add each 4 pixel group together and put 4 results into sad */
329
sad = vec_sum4s(t5, sad);
335
/* Sum up the four partial sums, and put the result into s */
336
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
337
sumdiffs = vec_splat(sumdiffs, 3);
338
vec_ste(sumdiffs, 0, &s);
343
int pix_norm1_altivec(uint8_t *pix, int line_size)
346
int s __attribute__((aligned(16)));
347
const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
348
vector unsigned char *tv;
349
vector unsigned char pixv;
350
vector unsigned int sv;
351
vector signed int sum;
353
sv = (vector unsigned int)vec_splat_u32(0);
356
for (i = 0; i < 16; i++) {
357
/* Read in the potentially unaligned pixels */
358
tv = (vector unsigned char *) pix;
359
pixv = vec_perm(tv[0], tv[1], vec_lvsl(0, pix));
361
/* Square the values, and add them to our sum */
362
sv = vec_msum(pixv, pixv, sv);
366
/* Sum up the four partial sums, and put the result into s */
367
sum = vec_sums((vector signed int) sv, (vector signed int) zero);
368
sum = vec_splat(sum, 3);
375
* Sum of Squared Errors for a 8x8 block.
377
* It's the sad8_altivec code above w/ squaring added.
379
int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
382
int s __attribute__((aligned(16)));
383
const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
384
vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
385
vector unsigned char t1, t2, t3,t4, t5;
386
vector unsigned int sum;
387
vector signed int sumsqr;
389
sum = (vector unsigned int)vec_splat_u32(0);
391
permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0);
395
/* Read potentially unaligned pixels into t1 and t2
396
Since we're reading 16 pixels, and actually only want 8,
397
mask out the last 8 pixels. The 0s don't change the sum. */
398
perm1 = vec_lvsl(0, pix1);
399
pix1v = (vector unsigned char *) pix1;
400
perm2 = vec_lvsl(0, pix2);
401
pix2v = (vector unsigned char *) pix2;
402
t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
403
t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);
406
Since we want to use unsigned chars, we can take advantage
407
of the fact that abs(a-b)^2 = (a-b)^2.
410
/* Calculate abs differences vector */
411
t3 = vec_max(t1, t2);
412
t4 = vec_min(t1, t2);
413
t5 = vec_sub(t3, t4);
415
/* Square the values and add them to our sum */
416
sum = vec_msum(t5, t5, sum);
422
/* Sum up the four partial sums, and put the result into s */
423
sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
424
sumsqr = vec_splat(sumsqr, 3);
425
vec_ste(sumsqr, 0, &s);
431
* Sum of Squared Errors for a 16x16 block.
433
* It's the sad16_altivec code above w/ squaring added.
435
int sse16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
438
int s __attribute__((aligned(16)));
439
const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
440
vector unsigned char perm1, perm2, *pix1v, *pix2v;
441
vector unsigned char t1, t2, t3,t4, t5;
442
vector unsigned int sum;
443
vector signed int sumsqr;
445
sum = (vector unsigned int)vec_splat_u32(0);
448
/* Read potentially unaligned pixels into t1 and t2 */
449
perm1 = vec_lvsl(0, pix1);
450
pix1v = (vector unsigned char *) pix1;
451
perm2 = vec_lvsl(0, pix2);
452
pix2v = (vector unsigned char *) pix2;
453
t1 = vec_perm(pix1v[0], pix1v[1], perm1);
454
t2 = vec_perm(pix2v[0], pix2v[1], perm2);
457
Since we want to use unsigned chars, we can take advantage
458
of the fact that abs(a-b)^2 = (a-b)^2.
461
/* Calculate abs differences vector */
462
t3 = vec_max(t1, t2);
463
t4 = vec_min(t1, t2);
464
t5 = vec_sub(t3, t4);
466
/* Square the values and add them to our sum */
467
sum = vec_msum(t5, t5, sum);
473
/* Sum up the four partial sums, and put the result into s */
474
sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
475
sumsqr = vec_splat(sumsqr, 3);
476
vec_ste(sumsqr, 0, &s);
481
int pix_sum_altivec(uint8_t * pix, int line_size)
483
const_vector unsigned int zero = (const_vector unsigned int)vec_splat_u32(0);
484
vector unsigned char perm, *pixv;
485
vector unsigned char t1;
486
vector unsigned int sad;
487
vector signed int sumdiffs;
490
int s __attribute__((aligned(16)));
492
sad = (vector unsigned int)vec_splat_u32(0);
494
for (i = 0; i < 16; i++) {
495
/* Read the potentially unaligned 16 pixels into t1 */
496
perm = vec_lvsl(0, pix);
497
pixv = (vector unsigned char *) pix;
498
t1 = vec_perm(pixv[0], pixv[1], perm);
500
/* Add each 4 pixel group together and put 4 results into sad */
501
sad = vec_sum4s(t1, sad);
506
/* Sum up the four partial sums, and put the result into s */
507
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
508
sumdiffs = vec_splat(sumdiffs, 3);
509
vec_ste(sumdiffs, 0, &s);
514
void get_pixels_altivec(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
517
vector unsigned char perm, bytes, *pixv;
518
const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
519
vector signed short shorts;
523
// Read potentially unaligned pixels.
524
// We're reading 16 pixels, and actually only want 8,
525
// but we simply ignore the extras.
526
perm = vec_lvsl(0, pixels);
527
pixv = (vector unsigned char *) pixels;
528
bytes = vec_perm(pixv[0], pixv[1], perm);
530
// convert the bytes into shorts
531
shorts = (vector signed short)vec_mergeh(zero, bytes);
533
// save the data to the block, we assume the block is 16-byte aligned
534
vec_st(shorts, i*16, (vector signed short*)block);
540
void diff_pixels_altivec(DCTELEM *restrict block, const uint8_t *s1,
541
const uint8_t *s2, int stride)
544
vector unsigned char perm, bytes, *pixv;
545
const_vector unsigned char zero = (const_vector unsigned char)vec_splat_u8(0);
546
vector signed short shorts1, shorts2;
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);
579
// The code below is a copy of the code above... This is a manual
582
// Read potentially unaligned pixels
583
// We're reading 16 pixels, and actually only want 8,
584
// but we simply ignore the extras.
585
perm = vec_lvsl(0, s1);
586
pixv = (vector unsigned char *) s1;
587
bytes = vec_perm(pixv[0], pixv[1], perm);
589
// convert the bytes into shorts
590
shorts1 = (vector signed short)vec_mergeh(zero, bytes);
592
// Do the same for the second block of pixels
593
perm = vec_lvsl(0, s2);
594
pixv = (vector unsigned char *) s2;
595
bytes = vec_perm(pixv[0], pixv[1], perm);
597
// convert the bytes into shorts
598
shorts2 = (vector signed short)vec_mergeh(zero, bytes);
600
// Do the subtraction
601
shorts1 = vec_sub(shorts1, shorts2);
603
// save the data to the block, we assume the block is 16-byte aligned
604
vec_st(shorts1, 0, (vector signed short*)block);
612
void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {
613
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
615
for(i=0; i+7<w; i++){
616
dst[i+0] += src[i+0];
617
dst[i+1] += src[i+1];
618
dst[i+2] += src[i+2];
619
dst[i+3] += src[i+3];
620
dst[i+4] += src[i+4];
621
dst[i+5] += src[i+5];
622
dst[i+6] += src[i+6];
623
dst[i+7] += src[i+7];
626
dst[i+0] += src[i+0];
627
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
629
register vector unsigned char vdst, vsrc;
631
/* dst and src are 16 bytes-aligned (guaranteed) */
632
for(i = 0 ; (i + 15) < w ; i++)
634
vdst = vec_ld(i << 4, (unsigned char*)dst);
635
vsrc = vec_ld(i << 4, (unsigned char*)src);
636
vdst = vec_add(vsrc, vdst);
637
vec_st(vdst, i << 4, (unsigned char*)dst);
639
/* if w is not a multiple of 16 */
640
for (; (i < w) ; i++)
644
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
647
/* next one assumes that ((line_size % 16) == 0) */
648
void put_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
650
POWERPC_PERF_DECLARE(altivec_put_pixels16_num, 1);
651
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
654
POWERPC_PERF_START_COUNT(altivec_put_pixels16_num, 1);
657
*((uint32_t*)(block)) = LD32(pixels);
658
*((uint32_t*)(block+4)) = LD32(pixels+4);
659
*((uint32_t*)(block+8)) = LD32(pixels+8);
660
*((uint32_t*)(block+12)) = LD32(pixels+12);
665
POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_num, 1);
667
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
668
register vector unsigned char pixelsv1, pixelsv2;
669
register vector unsigned char pixelsv1B, pixelsv2B;
670
register vector unsigned char pixelsv1C, pixelsv2C;
671
register vector unsigned char pixelsv1D, pixelsv2D;
673
register vector unsigned char perm = vec_lvsl(0, pixels);
675
register int line_size_2 = line_size << 1;
676
register int line_size_3 = line_size + line_size_2;
677
register int line_size_4 = line_size << 2;
679
POWERPC_PERF_START_COUNT(altivec_put_pixels16_num, 1);
680
// hand-unrolling the loop by 4 gains about 15%
681
// mininum execution time goes from 74 to 60 cycles
682
// it's faster than -funroll-loops, but using
683
// -funroll-loops w/ this is bad - 74 cycles again.
684
// all this is on a 7450, tuning for the 7450
687
pixelsv1 = vec_ld(0, (unsigned char*)pixels);
688
pixelsv2 = vec_ld(16, (unsigned char*)pixels);
689
vec_st(vec_perm(pixelsv1, pixelsv2, perm),
690
0, (unsigned char*)block);
695
for(i=0; i<h; i+=4) {
696
pixelsv1 = vec_ld(0, (unsigned char*)pixels);
697
pixelsv2 = vec_ld(16, (unsigned char*)pixels);
698
pixelsv1B = vec_ld(line_size, (unsigned char*)pixels);
699
pixelsv2B = vec_ld(16 + line_size, (unsigned char*)pixels);
700
pixelsv1C = vec_ld(line_size_2, (unsigned char*)pixels);
701
pixelsv2C = vec_ld(16 + line_size_2, (unsigned char*)pixels);
702
pixelsv1D = vec_ld(line_size_3, (unsigned char*)pixels);
703
pixelsv2D = vec_ld(16 + line_size_3, (unsigned char*)pixels);
704
vec_st(vec_perm(pixelsv1, pixelsv2, perm),
705
0, (unsigned char*)block);
706
vec_st(vec_perm(pixelsv1B, pixelsv2B, perm),
707
line_size, (unsigned char*)block);
708
vec_st(vec_perm(pixelsv1C, pixelsv2C, perm),
709
line_size_2, (unsigned char*)block);
710
vec_st(vec_perm(pixelsv1D, pixelsv2D, perm),
711
line_size_3, (unsigned char*)block);
716
POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_num, 1);
718
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
721
/* next one assumes that ((line_size % 16) == 0) */
722
#define op_avg(a,b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEUL)>>1) )
723
void avg_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
725
POWERPC_PERF_DECLARE(altivec_avg_pixels16_num, 1);
726
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
729
POWERPC_PERF_START_COUNT(altivec_avg_pixels16_num, 1);
732
op_avg(*((uint32_t*)(block)),LD32(pixels));
733
op_avg(*((uint32_t*)(block+4)),LD32(pixels+4));
734
op_avg(*((uint32_t*)(block+8)),LD32(pixels+8));
735
op_avg(*((uint32_t*)(block+12)),LD32(pixels+12));
740
POWERPC_PERF_STOP_COUNT(altivec_avg_pixels16_num, 1);
742
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
743
register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
744
register vector unsigned char perm = vec_lvsl(0, pixels);
747
POWERPC_PERF_START_COUNT(altivec_avg_pixels16_num, 1);
750
pixelsv1 = vec_ld(0, (unsigned char*)pixels);
751
pixelsv2 = vec_ld(16, (unsigned char*)pixels);
752
blockv = vec_ld(0, block);
753
pixelsv = vec_perm(pixelsv1, pixelsv2, perm);
754
blockv = vec_avg(blockv,pixelsv);
755
vec_st(blockv, 0, (unsigned char*)block);
760
POWERPC_PERF_STOP_COUNT(altivec_avg_pixels16_num, 1);
762
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
765
/* next one assumes that ((line_size % 8) == 0) */
766
void avg_pixels8_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
768
POWERPC_PERF_DECLARE(altivec_avg_pixels8_num, 1);
769
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
771
POWERPC_PERF_START_COUNT(altivec_avg_pixels8_num, 1);
772
for (i = 0; i < h; i++) {
773
*((uint32_t *) (block)) =
774
(((*((uint32_t *) (block))) |
775
((((const struct unaligned_32 *) (pixels))->l))) -
776
((((*((uint32_t *) (block))) ^
777
((((const struct unaligned_32 *) (pixels))->
778
l))) & 0xFEFEFEFEUL) >> 1));
779
*((uint32_t *) (block + 4)) =
780
(((*((uint32_t *) (block + 4))) |
781
((((const struct unaligned_32 *) (pixels + 4))->l))) -
782
((((*((uint32_t *) (block + 4))) ^
783
((((const struct unaligned_32 *) (pixels +
785
l))) & 0xFEFEFEFEUL) >> 1));
789
POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_num, 1);
791
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
792
register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
795
POWERPC_PERF_START_COUNT(altivec_avg_pixels8_num, 1);
797
for (i = 0; i < h; i++) {
799
block is 8 bytes-aligned, so we're either in the
800
left block (16 bytes-aligned) or in the right block (not)
802
int rightside = ((unsigned long)block & 0x0000000F);
804
blockv = vec_ld(0, block);
805
pixelsv1 = vec_ld(0, (unsigned char*)pixels);
806
pixelsv2 = vec_ld(16, (unsigned char*)pixels);
807
pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels));
811
pixelsv = vec_perm(blockv, pixelsv, vcprm(0,1,s0,s1));
815
pixelsv = vec_perm(blockv, pixelsv, vcprm(s0,s1,2,3));
818
blockv = vec_avg(blockv, pixelsv);
820
vec_st(blockv, 0, block);
826
POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_num, 1);
828
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
831
/* next one assumes that ((line_size % 8) == 0) */
832
void put_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
834
POWERPC_PERF_DECLARE(altivec_put_pixels8_xy2_num, 1);
835
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
837
POWERPC_PERF_START_COUNT(altivec_put_pixels8_xy2_num, 1);
838
for (j = 0; j < 2; j++) {
840
const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
842
(((const struct unaligned_32 *) (pixels + 1))->l);
844
(a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
846
((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
849
for (i = 0; i < h; i += 2) {
850
uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
851
uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
852
l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
853
h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
854
*((uint32_t *) block) =
855
h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
858
a = (((const struct unaligned_32 *) (pixels))->l);
859
b = (((const struct unaligned_32 *) (pixels + 1))->l);
860
l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
861
h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
862
*((uint32_t *) block) =
863
h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
866
} pixels += 4 - line_size * (h + 1);
867
block += 4 - line_size * h;
870
POWERPC_PERF_STOP_COUNT(altivec_put_pixels8_xy2_num, 1);
872
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
874
register vector unsigned char
877
register vector unsigned char
878
blockv, temp1, temp2;
879
register vector unsigned short
880
pixelssum1, pixelssum2, temp3;
881
register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
882
register const_vector unsigned short vctwo = (const_vector unsigned short)vec_splat_u16(2);
884
temp1 = vec_ld(0, pixels);
885
temp2 = vec_ld(16, pixels);
886
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
887
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F)
893
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
895
pixelsv1 = vec_mergeh(vczero, pixelsv1);
896
pixelsv2 = vec_mergeh(vczero, pixelsv2);
897
pixelssum1 = vec_add((vector unsigned short)pixelsv1,
898
(vector unsigned short)pixelsv2);
899
pixelssum1 = vec_add(pixelssum1, vctwo);
901
POWERPC_PERF_START_COUNT(altivec_put_pixels8_xy2_num, 1);
902
for (i = 0; i < h ; i++) {
903
int rightside = ((unsigned long)block & 0x0000000F);
904
blockv = vec_ld(0, block);
906
temp1 = vec_ld(line_size, pixels);
907
temp2 = vec_ld(line_size + 16, pixels);
908
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
909
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
915
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
918
pixelsv1 = vec_mergeh(vczero, pixelsv1);
919
pixelsv2 = vec_mergeh(vczero, pixelsv2);
920
pixelssum2 = vec_add((vector unsigned short)pixelsv1,
921
(vector unsigned short)pixelsv2);
922
temp3 = vec_add(pixelssum1, pixelssum2);
923
temp3 = vec_sra(temp3, vctwo);
924
pixelssum1 = vec_add(pixelssum2, vctwo);
925
pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
929
blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
933
blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
936
vec_st(blockv, 0, block);
942
POWERPC_PERF_STOP_COUNT(altivec_put_pixels8_xy2_num, 1);
943
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
946
/* next one assumes that ((line_size % 8) == 0) */
947
void put_no_rnd_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
949
POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels8_xy2_num, 1);
950
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
952
POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
953
for (j = 0; j < 2; j++) {
955
const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
957
(((const struct unaligned_32 *) (pixels + 1))->l);
959
(a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
961
((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
964
for (i = 0; i < h; i += 2) {
965
uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
966
uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
967
l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
968
h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
969
*((uint32_t *) block) =
970
h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
973
a = (((const struct unaligned_32 *) (pixels))->l);
974
b = (((const struct unaligned_32 *) (pixels + 1))->l);
975
l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
976
h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
977
*((uint32_t *) block) =
978
h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
981
} pixels += 4 - line_size * (h + 1);
982
block += 4 - line_size * h;
985
POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
987
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
989
register vector unsigned char
992
register vector unsigned char
993
blockv, temp1, temp2;
994
register vector unsigned short
995
pixelssum1, pixelssum2, temp3;
996
register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
997
register const_vector unsigned short vcone = (const_vector unsigned short)vec_splat_u16(1);
998
register const_vector unsigned short vctwo = (const_vector unsigned short)vec_splat_u16(2);
1000
temp1 = vec_ld(0, pixels);
1001
temp2 = vec_ld(16, pixels);
1002
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
1003
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F)
1009
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
1011
pixelsv1 = vec_mergeh(vczero, pixelsv1);
1012
pixelsv2 = vec_mergeh(vczero, pixelsv2);
1013
pixelssum1 = vec_add((vector unsigned short)pixelsv1,
1014
(vector unsigned short)pixelsv2);
1015
pixelssum1 = vec_add(pixelssum1, vcone);
1017
POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
1018
for (i = 0; i < h ; i++) {
1019
int rightside = ((unsigned long)block & 0x0000000F);
1020
blockv = vec_ld(0, block);
1022
temp1 = vec_ld(line_size, pixels);
1023
temp2 = vec_ld(line_size + 16, pixels);
1024
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
1025
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
1031
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
1034
pixelsv1 = vec_mergeh(vczero, pixelsv1);
1035
pixelsv2 = vec_mergeh(vczero, pixelsv2);
1036
pixelssum2 = vec_add((vector unsigned short)pixelsv1,
1037
(vector unsigned short)pixelsv2);
1038
temp3 = vec_add(pixelssum1, pixelssum2);
1039
temp3 = vec_sra(temp3, vctwo);
1040
pixelssum1 = vec_add(pixelssum2, vcone);
1041
pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
1045
blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
1049
blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
1052
vec_st(blockv, 0, block);
1055
pixels += line_size;
1058
POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
1059
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
1062
/* next one assumes that ((line_size % 16) == 0) */
1063
void put_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
1065
POWERPC_PERF_DECLARE(altivec_put_pixels16_xy2_num, 1);
1066
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
1068
POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1);
1069
for (j = 0; j < 4; j++) {
1071
const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
1073
(((const struct unaligned_32 *) (pixels + 1))->l);
1075
(a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
1077
((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
1079
pixels += line_size;
1080
for (i = 0; i < h; i += 2) {
1081
uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
1082
uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
1083
l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
1084
h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
1085
*((uint32_t *) block) =
1086
h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
1087
pixels += line_size;
1089
a = (((const struct unaligned_32 *) (pixels))->l);
1090
b = (((const struct unaligned_32 *) (pixels + 1))->l);
1091
l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
1092
h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
1093
*((uint32_t *) block) =
1094
h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
1095
pixels += line_size;
1097
} pixels += 4 - line_size * (h + 1);
1098
block += 4 - line_size * h;
1101
POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_xy2_num, 1);
1103
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
1105
register vector unsigned char
1106
pixelsv1, pixelsv2, pixelsv3, pixelsv4;
1107
register vector unsigned char
1108
blockv, temp1, temp2;
1109
register vector unsigned short
1110
pixelssum1, pixelssum2, temp3,
1111
pixelssum3, pixelssum4, temp4;
1112
register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
1113
register const_vector unsigned short vctwo = (const_vector unsigned short)vec_splat_u16(2);
1115
POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1);
1117
temp1 = vec_ld(0, pixels);
1118
temp2 = vec_ld(16, pixels);
1119
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
1120
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F)
1126
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
1128
pixelsv3 = vec_mergel(vczero, pixelsv1);
1129
pixelsv4 = vec_mergel(vczero, pixelsv2);
1130
pixelsv1 = vec_mergeh(vczero, pixelsv1);
1131
pixelsv2 = vec_mergeh(vczero, pixelsv2);
1132
pixelssum3 = vec_add((vector unsigned short)pixelsv3,
1133
(vector unsigned short)pixelsv4);
1134
pixelssum3 = vec_add(pixelssum3, vctwo);
1135
pixelssum1 = vec_add((vector unsigned short)pixelsv1,
1136
(vector unsigned short)pixelsv2);
1137
pixelssum1 = vec_add(pixelssum1, vctwo);
1139
for (i = 0; i < h ; i++) {
1140
blockv = vec_ld(0, block);
1142
temp1 = vec_ld(line_size, pixels);
1143
temp2 = vec_ld(line_size + 16, pixels);
1144
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
1145
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
1151
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
1154
pixelsv3 = vec_mergel(vczero, pixelsv1);
1155
pixelsv4 = vec_mergel(vczero, pixelsv2);
1156
pixelsv1 = vec_mergeh(vczero, pixelsv1);
1157
pixelsv2 = vec_mergeh(vczero, pixelsv2);
1159
pixelssum4 = vec_add((vector unsigned short)pixelsv3,
1160
(vector unsigned short)pixelsv4);
1161
pixelssum2 = vec_add((vector unsigned short)pixelsv1,
1162
(vector unsigned short)pixelsv2);
1163
temp4 = vec_add(pixelssum3, pixelssum4);
1164
temp4 = vec_sra(temp4, vctwo);
1165
temp3 = vec_add(pixelssum1, pixelssum2);
1166
temp3 = vec_sra(temp3, vctwo);
1168
pixelssum3 = vec_add(pixelssum4, vctwo);
1169
pixelssum1 = vec_add(pixelssum2, vctwo);
1171
blockv = vec_packsu(temp3, temp4);
1173
vec_st(blockv, 0, block);
1176
pixels += line_size;
1179
POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_xy2_num, 1);
1180
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
1183
/* next one assumes that ((line_size % 16) == 0) */
1184
void put_no_rnd_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
1186
POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels16_xy2_num, 1);
1187
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
1189
POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
1190
for (j = 0; j < 4; j++) {
1192
const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
1194
(((const struct unaligned_32 *) (pixels + 1))->l);
1196
(a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
1198
((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
1200
pixels += line_size;
1201
for (i = 0; i < h; i += 2) {
1202
uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
1203
uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
1204
l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
1205
h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
1206
*((uint32_t *) block) =
1207
h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
1208
pixels += line_size;
1210
a = (((const struct unaligned_32 *) (pixels))->l);
1211
b = (((const struct unaligned_32 *) (pixels + 1))->l);
1212
l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
1213
h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
1214
*((uint32_t *) block) =
1215
h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
1216
pixels += line_size;
1218
} pixels += 4 - line_size * (h + 1);
1219
block += 4 - line_size * h;
1222
POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
1224
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
1226
register vector unsigned char
1227
pixelsv1, pixelsv2, pixelsv3, pixelsv4;
1228
register vector unsigned char
1229
blockv, temp1, temp2;
1230
register vector unsigned short
1231
pixelssum1, pixelssum2, temp3,
1232
pixelssum3, pixelssum4, temp4;
1233
register const_vector unsigned char vczero = (const_vector unsigned char)vec_splat_u8(0);
1234
register const_vector unsigned short vcone = (const_vector unsigned short)vec_splat_u16(1);
1235
register const_vector unsigned short vctwo = (const_vector unsigned short)vec_splat_u16(2);
1237
POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
1239
temp1 = vec_ld(0, pixels);
1240
temp2 = vec_ld(16, pixels);
1241
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
1242
if ((((unsigned long)pixels) & 0x0000000F) == 0x0000000F)
1248
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
1250
pixelsv3 = vec_mergel(vczero, pixelsv1);
1251
pixelsv4 = vec_mergel(vczero, pixelsv2);
1252
pixelsv1 = vec_mergeh(vczero, pixelsv1);
1253
pixelsv2 = vec_mergeh(vczero, pixelsv2);
1254
pixelssum3 = vec_add((vector unsigned short)pixelsv3,
1255
(vector unsigned short)pixelsv4);
1256
pixelssum3 = vec_add(pixelssum3, vcone);
1257
pixelssum1 = vec_add((vector unsigned short)pixelsv1,
1258
(vector unsigned short)pixelsv2);
1259
pixelssum1 = vec_add(pixelssum1, vcone);
1261
for (i = 0; i < h ; i++) {
1262
blockv = vec_ld(0, block);
1264
temp1 = vec_ld(line_size, pixels);
1265
temp2 = vec_ld(line_size + 16, pixels);
1266
pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
1267
if (((((unsigned long)pixels) + line_size) & 0x0000000F) == 0x0000000F)
1273
pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
1276
pixelsv3 = vec_mergel(vczero, pixelsv1);
1277
pixelsv4 = vec_mergel(vczero, pixelsv2);
1278
pixelsv1 = vec_mergeh(vczero, pixelsv1);
1279
pixelsv2 = vec_mergeh(vczero, pixelsv2);
1281
pixelssum4 = vec_add((vector unsigned short)pixelsv3,
1282
(vector unsigned short)pixelsv4);
1283
pixelssum2 = vec_add((vector unsigned short)pixelsv1,
1284
(vector unsigned short)pixelsv2);
1285
temp4 = vec_add(pixelssum3, pixelssum4);
1286
temp4 = vec_sra(temp4, vctwo);
1287
temp3 = vec_add(pixelssum1, pixelssum2);
1288
temp3 = vec_sra(temp3, vctwo);
1290
pixelssum3 = vec_add(pixelssum4, vcone);
1291
pixelssum1 = vec_add(pixelssum2, vcone);
1293
blockv = vec_packsu(temp3, temp4);
1295
vec_st(blockv, 0, block);
1298
pixels += line_size;
1301
POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
1302
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
1305
#if (__GNUC__ * 100 + __GNUC_MINOR__ >= 330)
1306
int hadamard8_diff8x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
1307
POWERPC_PERF_DECLARE(altivec_hadamard8_diff8x8_num, 1);
1309
POWERPC_PERF_START_COUNT(altivec_hadamard8_diff8x8_num, 1);
1310
register const_vector unsigned char vzero = (const_vector unsigned char)vec_splat_u8(0);
1311
register vector signed short temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
1313
register const_vector signed short vprod1 = (const_vector signed short)AVV( 1,-1, 1,-1, 1,-1, 1,-1);
1314
register const_vector signed short vprod2 = (const_vector signed short)AVV( 1, 1,-1,-1, 1, 1,-1,-1);
1315
register const_vector signed short vprod3 = (const_vector signed short)AVV( 1, 1, 1, 1,-1,-1,-1,-1);
1316
register const_vector unsigned char perm1 = (const_vector unsigned char)
1317
AVV(0x02, 0x03, 0x00, 0x01,
1318
0x06, 0x07, 0x04, 0x05,
1319
0x0A, 0x0B, 0x08, 0x09,
1320
0x0E, 0x0F, 0x0C, 0x0D);
1321
register const_vector unsigned char perm2 = (const_vector unsigned char)
1322
AVV(0x04, 0x05, 0x06, 0x07,
1323
0x00, 0x01, 0x02, 0x03,
1324
0x0C, 0x0D, 0x0E, 0x0F,
1325
0x08, 0x09, 0x0A, 0x0B);
1326
register const_vector unsigned char perm3 = (const_vector unsigned char)
1327
AVV(0x08, 0x09, 0x0A, 0x0B,
1328
0x0C, 0x0D, 0x0E, 0x0F,
1329
0x00, 0x01, 0x02, 0x03,
1330
0x04, 0x05, 0x06, 0x07);
1332
#define ONEITERBUTTERFLY(i, res) \
1334
register vector unsigned char src1, src2, srcO; \
1335
register vector unsigned char dst1, dst2, dstO; \
1336
src1 = vec_ld(stride * i, src); \
1337
if ((((stride * i) + (unsigned long)src) & 0x0000000F) > 8) \
1338
src2 = vec_ld((stride * i) + 16, src); \
1339
srcO = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
1340
dst1 = vec_ld(stride * i, dst); \
1341
if ((((stride * i) + (unsigned long)dst) & 0x0000000F) > 8) \
1342
dst2 = vec_ld((stride * i) + 16, dst); \
1343
dstO = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
1344
/* promote the unsigned chars to signed shorts */ \
1345
/* we're in the 8x8 function, we only care for the first 8 */ \
1346
register vector signed short srcV = \
1347
(vector signed short)vec_mergeh((vector signed char)vzero, (vector signed char)srcO); \
1348
register vector signed short dstV = \
1349
(vector signed short)vec_mergeh((vector signed char)vzero, (vector signed char)dstO); \
1350
/* substractions inside the first butterfly */ \
1351
register vector signed short but0 = vec_sub(srcV, dstV); \
1352
register vector signed short op1 = vec_perm(but0, but0, perm1); \
1353
register vector signed short but1 = vec_mladd(but0, vprod1, op1); \
1354
register vector signed short op2 = vec_perm(but1, but1, perm2); \
1355
register vector signed short but2 = vec_mladd(but1, vprod2, op2); \
1356
register vector signed short op3 = vec_perm(but2, but2, perm3); \
1357
res = vec_mladd(but2, vprod3, op3); \
1359
ONEITERBUTTERFLY(0, temp0);
1360
ONEITERBUTTERFLY(1, temp1);
1361
ONEITERBUTTERFLY(2, temp2);
1362
ONEITERBUTTERFLY(3, temp3);
1363
ONEITERBUTTERFLY(4, temp4);
1364
ONEITERBUTTERFLY(5, temp5);
1365
ONEITERBUTTERFLY(6, temp6);
1366
ONEITERBUTTERFLY(7, temp7);
1368
#undef ONEITERBUTTERFLY
1370
register vector signed int vsum;
1371
register vector signed short line0 = vec_add(temp0, temp1);
1372
register vector signed short line1 = vec_sub(temp0, temp1);
1373
register vector signed short line2 = vec_add(temp2, temp3);
1374
register vector signed short line3 = vec_sub(temp2, temp3);
1375
register vector signed short line4 = vec_add(temp4, temp5);
1376
register vector signed short line5 = vec_sub(temp4, temp5);
1377
register vector signed short line6 = vec_add(temp6, temp7);
1378
register vector signed short line7 = vec_sub(temp6, temp7);
1380
register vector signed short line0B = vec_add(line0, line2);
1381
register vector signed short line2B = vec_sub(line0, line2);
1382
register vector signed short line1B = vec_add(line1, line3);
1383
register vector signed short line3B = vec_sub(line1, line3);
1384
register vector signed short line4B = vec_add(line4, line6);
1385
register vector signed short line6B = vec_sub(line4, line6);
1386
register vector signed short line5B = vec_add(line5, line7);
1387
register vector signed short line7B = vec_sub(line5, line7);
1389
register vector signed short line0C = vec_add(line0B, line4B);
1390
register vector signed short line4C = vec_sub(line0B, line4B);
1391
register vector signed short line1C = vec_add(line1B, line5B);
1392
register vector signed short line5C = vec_sub(line1B, line5B);
1393
register vector signed short line2C = vec_add(line2B, line6B);
1394
register vector signed short line6C = vec_sub(line2B, line6B);
1395
register vector signed short line3C = vec_add(line3B, line7B);
1396
register vector signed short line7C = vec_sub(line3B, line7B);
1398
vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
1399
vsum = vec_sum4s(vec_abs(line1C), vsum);
1400
vsum = vec_sum4s(vec_abs(line2C), vsum);
1401
vsum = vec_sum4s(vec_abs(line3C), vsum);
1402
vsum = vec_sum4s(vec_abs(line4C), vsum);
1403
vsum = vec_sum4s(vec_abs(line5C), vsum);
1404
vsum = vec_sum4s(vec_abs(line6C), vsum);
1405
vsum = vec_sum4s(vec_abs(line7C), vsum);
1406
vsum = vec_sums(vsum, (vector signed int)vzero);
1407
vsum = vec_splat(vsum, 3);
1408
vec_ste(vsum, 0, &sum);
1410
POWERPC_PERF_STOP_COUNT(altivec_hadamard8_diff8x8_num, 1);
1415
16x8 works with 16 elements ; it allows to avoid replicating
1416
loads, and give the compiler more rooms for scheduling.
1417
It's only used from inside hadamard8_diff16_altivec.
1419
Unfortunately, it seems gcc-3.3 is a bit dumb, and
1420
the compiled code has a LOT of spill code, it seems
1421
gcc (unlike xlc) cannot keep everything in registers
1422
by itself. The following code include hand-made
1423
registers allocation. It's not clean, but on
1424
a 7450 the resulting code is much faster (best case
1425
fall from 700+ cycles to 550).
1427
xlc doesn't add spill code, but it doesn't know how to
1428
schedule for the 7450, and its code isn't much faster than
1429
gcc-3.3 on the 7450 (but uses 25% less instructions...)
1431
On the 970, the hand-made RA is still a win (arount 690
1432
vs. around 780), but xlc goes to around 660 on the
1436
static int hadamard8_diff16x8_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h) {
1438
register vector signed short
1447
register vector signed short
1456
register const_vector unsigned char vzero asm ("v31")= (const_vector unsigned char)vec_splat_u8(0);
1458
register const_vector signed short vprod1 asm ("v16")= (const_vector signed short)AVV( 1,-1, 1,-1, 1,-1, 1,-1);
1459
register const_vector signed short vprod2 asm ("v17")= (const_vector signed short)AVV( 1, 1,-1,-1, 1, 1,-1,-1);
1460
register const_vector signed short vprod3 asm ("v18")= (const_vector signed short)AVV( 1, 1, 1, 1,-1,-1,-1,-1);
1461
register const_vector unsigned char perm1 asm ("v19")= (const_vector unsigned char)
1462
AVV(0x02, 0x03, 0x00, 0x01,
1463
0x06, 0x07, 0x04, 0x05,
1464
0x0A, 0x0B, 0x08, 0x09,
1465
0x0E, 0x0F, 0x0C, 0x0D);
1466
register const_vector unsigned char perm2 asm ("v20")= (const_vector unsigned char)
1467
AVV(0x04, 0x05, 0x06, 0x07,
1468
0x00, 0x01, 0x02, 0x03,
1469
0x0C, 0x0D, 0x0E, 0x0F,
1470
0x08, 0x09, 0x0A, 0x0B);
1471
register const_vector unsigned char perm3 asm ("v21")= (const_vector unsigned char)
1472
AVV(0x08, 0x09, 0x0A, 0x0B,
1473
0x0C, 0x0D, 0x0E, 0x0F,
1474
0x00, 0x01, 0x02, 0x03,
1475
0x04, 0x05, 0x06, 0x07);
1477
#define ONEITERBUTTERFLY(i, res1, res2) \
1479
register vector unsigned char src1 asm ("v22"), src2 asm ("v23"); \
1480
register vector unsigned char dst1 asm ("v24"), dst2 asm ("v25"); \
1481
src1 = vec_ld(stride * i, src); \
1482
src2 = vec_ld((stride * i) + 16, src); \
1483
register vector unsigned char srcO asm ("v22") = vec_perm(src1, src2, vec_lvsl(stride * i, src)); \
1484
dst1 = vec_ld(stride * i, dst); \
1485
dst2 = vec_ld((stride * i) + 16, dst); \
1486
register vector unsigned char dstO asm ("v23") = vec_perm(dst1, dst2, vec_lvsl(stride * i, dst)); \
1487
/* promote the unsigned chars to signed shorts */ \
1488
register vector signed short srcV asm ("v24") = \
1489
(vector signed short)vec_mergeh((vector signed char)vzero, (vector signed char)srcO); \
1490
register vector signed short dstV asm ("v25") = \
1491
(vector signed short)vec_mergeh((vector signed char)vzero, (vector signed char)dstO); \
1492
register vector signed short srcW asm ("v26") = \
1493
(vector signed short)vec_mergel((vector signed char)vzero, (vector signed char)srcO); \
1494
register vector signed short dstW asm ("v27") = \
1495
(vector signed short)vec_mergel((vector signed char)vzero, (vector signed char)dstO); \
1496
/* substractions inside the first butterfly */ \
1497
register vector signed short but0 asm ("v28") = vec_sub(srcV, dstV); \
1498
register vector signed short but0S asm ("v29") = vec_sub(srcW, dstW); \
1499
register vector signed short op1 asm ("v30") = vec_perm(but0, but0, perm1); \
1500
register vector signed short but1 asm ("v22") = vec_mladd(but0, vprod1, op1); \
1501
register vector signed short op1S asm ("v23") = vec_perm(but0S, but0S, perm1); \
1502
register vector signed short but1S asm ("v24") = vec_mladd(but0S, vprod1, op1S); \
1503
register vector signed short op2 asm ("v25") = vec_perm(but1, but1, perm2); \
1504
register vector signed short but2 asm ("v26") = vec_mladd(but1, vprod2, op2); \
1505
register vector signed short op2S asm ("v27") = vec_perm(but1S, but1S, perm2); \
1506
register vector signed short but2S asm ("v28") = vec_mladd(but1S, vprod2, op2S); \
1507
register vector signed short op3 asm ("v29") = vec_perm(but2, but2, perm3); \
1508
res1 = vec_mladd(but2, vprod3, op3); \
1509
register vector signed short op3S asm ("v30") = vec_perm(but2S, but2S, perm3); \
1510
res2 = vec_mladd(but2S, vprod3, op3S); \
1512
ONEITERBUTTERFLY(0, temp0, temp0S);
1513
ONEITERBUTTERFLY(1, temp1, temp1S);
1514
ONEITERBUTTERFLY(2, temp2, temp2S);
1515
ONEITERBUTTERFLY(3, temp3, temp3S);
1516
ONEITERBUTTERFLY(4, temp4, temp4S);
1517
ONEITERBUTTERFLY(5, temp5, temp5S);
1518
ONEITERBUTTERFLY(6, temp6, temp6S);
1519
ONEITERBUTTERFLY(7, temp7, temp7S);
1521
#undef ONEITERBUTTERFLY
1523
register vector signed int vsum;
1524
register vector signed short line0 = vec_add(temp0, temp1);
1525
register vector signed short line1 = vec_sub(temp0, temp1);
1526
register vector signed short line2 = vec_add(temp2, temp3);
1527
register vector signed short line3 = vec_sub(temp2, temp3);
1528
register vector signed short line4 = vec_add(temp4, temp5);
1529
register vector signed short line5 = vec_sub(temp4, temp5);
1530
register vector signed short line6 = vec_add(temp6, temp7);
1531
register vector signed short line7 = vec_sub(temp6, temp7);
1533
register vector signed short line0B = vec_add(line0, line2);
1534
register vector signed short line2B = vec_sub(line0, line2);
1535
register vector signed short line1B = vec_add(line1, line3);
1536
register vector signed short line3B = vec_sub(line1, line3);
1537
register vector signed short line4B = vec_add(line4, line6);
1538
register vector signed short line6B = vec_sub(line4, line6);
1539
register vector signed short line5B = vec_add(line5, line7);
1540
register vector signed short line7B = vec_sub(line5, line7);
1542
register vector signed short line0C = vec_add(line0B, line4B);
1543
register vector signed short line4C = vec_sub(line0B, line4B);
1544
register vector signed short line1C = vec_add(line1B, line5B);
1545
register vector signed short line5C = vec_sub(line1B, line5B);
1546
register vector signed short line2C = vec_add(line2B, line6B);
1547
register vector signed short line6C = vec_sub(line2B, line6B);
1548
register vector signed short line3C = vec_add(line3B, line7B);
1549
register vector signed short line7C = vec_sub(line3B, line7B);
1551
vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
1552
vsum = vec_sum4s(vec_abs(line1C), vsum);
1553
vsum = vec_sum4s(vec_abs(line2C), vsum);
1554
vsum = vec_sum4s(vec_abs(line3C), vsum);
1555
vsum = vec_sum4s(vec_abs(line4C), vsum);
1556
vsum = vec_sum4s(vec_abs(line5C), vsum);
1557
vsum = vec_sum4s(vec_abs(line6C), vsum);
1558
vsum = vec_sum4s(vec_abs(line7C), vsum);
1560
register vector signed short line0S = vec_add(temp0S, temp1S);
1561
register vector signed short line1S = vec_sub(temp0S, temp1S);
1562
register vector signed short line2S = vec_add(temp2S, temp3S);
1563
register vector signed short line3S = vec_sub(temp2S, temp3S);
1564
register vector signed short line4S = vec_add(temp4S, temp5S);
1565
register vector signed short line5S = vec_sub(temp4S, temp5S);
1566
register vector signed short line6S = vec_add(temp6S, temp7S);
1567
register vector signed short line7S = vec_sub(temp6S, temp7S);
1569
register vector signed short line0BS = vec_add(line0S, line2S);
1570
register vector signed short line2BS = vec_sub(line0S, line2S);
1571
register vector signed short line1BS = vec_add(line1S, line3S);
1572
register vector signed short line3BS = vec_sub(line1S, line3S);
1573
register vector signed short line4BS = vec_add(line4S, line6S);
1574
register vector signed short line6BS = vec_sub(line4S, line6S);
1575
register vector signed short line5BS = vec_add(line5S, line7S);
1576
register vector signed short line7BS = vec_sub(line5S, line7S);
1578
register vector signed short line0CS = vec_add(line0BS, line4BS);
1579
register vector signed short line4CS = vec_sub(line0BS, line4BS);
1580
register vector signed short line1CS = vec_add(line1BS, line5BS);
1581
register vector signed short line5CS = vec_sub(line1BS, line5BS);
1582
register vector signed short line2CS = vec_add(line2BS, line6BS);
1583
register vector signed short line6CS = vec_sub(line2BS, line6BS);
1584
register vector signed short line3CS = vec_add(line3BS, line7BS);
1585
register vector signed short line7CS = vec_sub(line3BS, line7BS);
1587
vsum = vec_sum4s(vec_abs(line0CS), vsum);
1588
vsum = vec_sum4s(vec_abs(line1CS), vsum);
1589
vsum = vec_sum4s(vec_abs(line2CS), vsum);
1590
vsum = vec_sum4s(vec_abs(line3CS), vsum);
1591
vsum = vec_sum4s(vec_abs(line4CS), vsum);
1592
vsum = vec_sum4s(vec_abs(line5CS), vsum);
1593
vsum = vec_sum4s(vec_abs(line6CS), vsum);
1594
vsum = vec_sum4s(vec_abs(line7CS), vsum);
1595
vsum = vec_sums(vsum, (vector signed int)vzero);
1596
vsum = vec_splat(vsum, 3);
1597
vec_ste(vsum, 0, &sum);
1602
int hadamard8_diff16_altivec(/*MpegEncContext*/ void *s, uint8_t *dst, uint8_t *src, int stride, int h){
1603
POWERPC_PERF_DECLARE(altivec_hadamard8_diff16_num, 1);
1605
POWERPC_PERF_START_COUNT(altivec_hadamard8_diff16_num, 1);
1606
score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
1610
score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
1612
POWERPC_PERF_STOP_COUNT(altivec_hadamard8_diff16_num, 1);
1617
int has_altivec(void)
1619
#ifdef CONFIG_DARWIN
1620
int sels[2] = {CTL_HW, HW_VECTORUNIT};
1622
size_t len = sizeof(has_vu);
1625
err = sysctl(sels, 2, &has_vu, &len, NULL, 0);
1627
if (err == 0) return (has_vu != 0);
1628
#else /* CONFIG_DARWIN */
1629
/* no Darwin, do it the brute-force way */
1630
/* this is borrowed from the libmpeg2 library */
1632
signal (SIGILL, sigill_handler);
1633
if (sigsetjmp (jmpbuf, 1)) {
1634
signal (SIGILL, SIG_DFL);
1638
asm volatile ("mtspr 256, %0\n\t"
1639
"vand %%v0, %%v0, %%v0"
1643
signal (SIGILL, SIG_DFL);
1647
#endif /* CONFIG_DARWIN */