34 |
#include "interpolate8x8.h" |
#include "interpolate8x8.h" |
35 |
#include "../utils/mem_align.h" |
#include "../utils/mem_align.h" |
36 |
#include "../motion/sad.h" |
#include "../motion/sad.h" |
37 |
|
#include "../utils/emms.h" |
38 |
|
|
39 |
#include "font.h" /* XXX: remove later */ |
#include "font.h" /* XXX: remove later */ |
40 |
|
|
256 |
|
|
257 |
n_ptr = (uint8_t*)refn; |
n_ptr = (uint8_t*)refn; |
258 |
h_ptr = refh; |
h_ptr = refh; |
259 |
v_ptr = refhv; |
v_ptr = refv; |
260 |
|
|
261 |
n_ptr -= offset; |
n_ptr -= offset; |
262 |
h_ptr -= offset; |
h_ptr -= offset; |
385 |
uint8_t * y_ptr, uint8_t * u_ptr, uint8_t * v_ptr, |
uint8_t * y_ptr, uint8_t * u_ptr, uint8_t * v_ptr, |
386 |
int y_stride, int uv_stride, |
int y_stride, int uv_stride, |
387 |
int width, int height, int vflip, |
int width, int height, int vflip, |
388 |
packedFunc * func_opt, packedFunc func_c, int size) |
packedFunc * func_opt, packedFunc func_c, |
389 |
|
int size, int interlacing) |
390 |
{ |
{ |
391 |
int width_opt, width_c; |
int width_opt, width_c, height_opt; |
392 |
|
|
393 |
|
if (width==1 || height==1) return; /* forget about it */ |
394 |
|
|
395 |
if (func_opt != func_c && x_stride < size*((width+15)/16)*16) |
if (func_opt != func_c && x_stride < size*((width+15)/16)*16) |
396 |
{ |
{ |
397 |
width_opt = width & (~15); |
width_opt = width & (~15); |
398 |
width_c = width - width_opt; |
width_c = (width - width_opt) & (~1); |
399 |
} |
} |
400 |
else |
else |
401 |
{ |
{ |
402 |
width_opt = width; |
/* Enforce the width to be divisable by two. */ |
403 |
|
width_opt = width & (~1); |
404 |
width_c = 0; |
width_c = 0; |
405 |
} |
} |
406 |
|
|
407 |
|
/* packed conversions require height to be divisable by 2 |
408 |
|
(or even by 4 for interlaced conversion) */ |
409 |
|
if (interlacing) |
410 |
|
height_opt = height & (~3); |
411 |
|
else |
412 |
|
height_opt = height & (~1); |
413 |
|
|
414 |
func_opt(x_ptr, x_stride, |
func_opt(x_ptr, x_stride, |
415 |
y_ptr, u_ptr, v_ptr, y_stride, uv_stride, |
y_ptr, u_ptr, v_ptr, y_stride, uv_stride, |
416 |
width_opt, height, vflip); |
width_opt, height_opt, vflip); |
417 |
|
|
418 |
if (width_c) |
if (width_c) |
419 |
{ |
{ |
420 |
func_c(x_ptr + size*width_opt, x_stride, |
func_c(x_ptr + size*width_opt, x_stride, |
421 |
y_ptr + width_opt, u_ptr + width_opt/2, v_ptr + width_opt/2, |
y_ptr + width_opt, u_ptr + width_opt/2, v_ptr + width_opt/2, |
422 |
y_stride, uv_stride, width_c, height, vflip); |
y_stride, uv_stride, width_c, height_opt, vflip); |
423 |
} |
} |
424 |
} |
} |
425 |
|
|
448 |
src[0], src_stride[0], image->y, image->u, image->v, |
src[0], src_stride[0], image->y, image->u, image->v, |
449 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
450 |
interlacing?rgb555i_to_yv12 :rgb555_to_yv12, |
interlacing?rgb555i_to_yv12 :rgb555_to_yv12, |
451 |
interlacing?rgb555i_to_yv12_c:rgb555_to_yv12_c, 2); |
interlacing?rgb555i_to_yv12_c:rgb555_to_yv12_c, 2, interlacing); |
452 |
break; |
break; |
453 |
|
|
454 |
case XVID_CSP_RGB565: |
case XVID_CSP_RGB565: |
456 |
src[0], src_stride[0], image->y, image->u, image->v, |
src[0], src_stride[0], image->y, image->u, image->v, |
457 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
458 |
interlacing?rgb565i_to_yv12 :rgb565_to_yv12, |
interlacing?rgb565i_to_yv12 :rgb565_to_yv12, |
459 |
interlacing?rgb565i_to_yv12_c:rgb565_to_yv12_c, 2); |
interlacing?rgb565i_to_yv12_c:rgb565_to_yv12_c, 2, interlacing); |
460 |
break; |
break; |
461 |
|
|
462 |
|
|
465 |
src[0], src_stride[0], image->y, image->u, image->v, |
src[0], src_stride[0], image->y, image->u, image->v, |
466 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
467 |
interlacing?bgri_to_yv12 :bgr_to_yv12, |
interlacing?bgri_to_yv12 :bgr_to_yv12, |
468 |
interlacing?bgri_to_yv12_c:bgr_to_yv12_c, 3); |
interlacing?bgri_to_yv12_c:bgr_to_yv12_c, 3, interlacing); |
469 |
break; |
break; |
470 |
|
|
471 |
case XVID_CSP_BGRA: |
case XVID_CSP_BGRA: |
473 |
src[0], src_stride[0], image->y, image->u, image->v, |
src[0], src_stride[0], image->y, image->u, image->v, |
474 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
475 |
interlacing?bgrai_to_yv12 :bgra_to_yv12, |
interlacing?bgrai_to_yv12 :bgra_to_yv12, |
476 |
interlacing?bgrai_to_yv12_c:bgra_to_yv12_c, 4); |
interlacing?bgrai_to_yv12_c:bgra_to_yv12_c, 4, interlacing); |
477 |
break; |
break; |
478 |
|
|
479 |
case XVID_CSP_ABGR : |
case XVID_CSP_ABGR : |
481 |
src[0], src_stride[0], image->y, image->u, image->v, |
src[0], src_stride[0], image->y, image->u, image->v, |
482 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
483 |
interlacing?abgri_to_yv12 :abgr_to_yv12, |
interlacing?abgri_to_yv12 :abgr_to_yv12, |
484 |
interlacing?abgri_to_yv12_c:abgr_to_yv12_c, 4); |
interlacing?abgri_to_yv12_c:abgr_to_yv12_c, 4, interlacing); |
485 |
|
break; |
486 |
|
|
487 |
|
case XVID_CSP_RGB: |
488 |
|
safe_packed_conv( |
489 |
|
src[0], src_stride[0], image->y, image->u, image->v, |
490 |
|
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
491 |
|
interlacing?rgbi_to_yv12 :rgb_to_yv12, |
492 |
|
interlacing?rgbi_to_yv12_c:rgb_to_yv12_c, 3, interlacing); |
493 |
break; |
break; |
494 |
|
|
495 |
case XVID_CSP_RGBA : |
case XVID_CSP_RGBA : |
497 |
src[0], src_stride[0], image->y, image->u, image->v, |
src[0], src_stride[0], image->y, image->u, image->v, |
498 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
499 |
interlacing?rgbai_to_yv12 :rgba_to_yv12, |
interlacing?rgbai_to_yv12 :rgba_to_yv12, |
500 |
interlacing?rgbai_to_yv12_c:rgba_to_yv12_c, 4); |
interlacing?rgbai_to_yv12_c:rgba_to_yv12_c, 4, interlacing); |
501 |
break; |
break; |
502 |
|
|
503 |
case XVID_CSP_ARGB: |
case XVID_CSP_ARGB: |
505 |
src[0], src_stride[0], image->y, image->u, image->v, |
src[0], src_stride[0], image->y, image->u, image->v, |
506 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
507 |
interlacing?argbi_to_yv12 : argb_to_yv12, |
interlacing?argbi_to_yv12 : argb_to_yv12, |
508 |
interlacing?argbi_to_yv12_c: argb_to_yv12_c, 4); |
interlacing?argbi_to_yv12_c: argb_to_yv12_c, 4, interlacing); |
509 |
break; |
break; |
510 |
|
|
511 |
case XVID_CSP_YUY2: |
case XVID_CSP_YUY2: |
513 |
src[0], src_stride[0], image->y, image->u, image->v, |
src[0], src_stride[0], image->y, image->u, image->v, |
514 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
515 |
interlacing?yuyvi_to_yv12 :yuyv_to_yv12, |
interlacing?yuyvi_to_yv12 :yuyv_to_yv12, |
516 |
interlacing?yuyvi_to_yv12_c:yuyv_to_yv12_c, 2); |
interlacing?yuyvi_to_yv12_c:yuyv_to_yv12_c, 2, interlacing); |
517 |
break; |
break; |
518 |
|
|
519 |
case XVID_CSP_YVYU: /* u/v swapped */ |
case XVID_CSP_YVYU: /* u/v swapped */ |
521 |
src[0], src_stride[0], image->y, image->v, image->u, |
src[0], src_stride[0], image->y, image->v, image->u, |
522 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
523 |
interlacing?yuyvi_to_yv12 :yuyv_to_yv12, |
interlacing?yuyvi_to_yv12 :yuyv_to_yv12, |
524 |
interlacing?yuyvi_to_yv12_c:yuyv_to_yv12_c, 2); |
interlacing?yuyvi_to_yv12_c:yuyv_to_yv12_c, 2, interlacing); |
525 |
break; |
break; |
526 |
|
|
527 |
case XVID_CSP_UYVY: |
case XVID_CSP_UYVY: |
529 |
src[0], src_stride[0], image->y, image->u, image->v, |
src[0], src_stride[0], image->y, image->u, image->v, |
530 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
531 |
interlacing?uyvyi_to_yv12 :uyvy_to_yv12, |
interlacing?uyvyi_to_yv12 :uyvy_to_yv12, |
532 |
interlacing?uyvyi_to_yv12_c:uyvy_to_yv12_c, 2); |
interlacing?uyvyi_to_yv12_c:uyvy_to_yv12_c, 2, interlacing); |
533 |
break; |
break; |
534 |
|
|
535 |
case XVID_CSP_I420: /* YCbCr == YUV == internal colorspace for MPEG */ |
case XVID_CSP_I420: /* YCbCr == YUV == internal colorspace for MPEG */ |
633 |
dst[0], dst_stride[0], image->y, image->u, image->v, |
dst[0], dst_stride[0], image->y, image->u, image->v, |
634 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
635 |
interlacing?yv12_to_rgb555i :yv12_to_rgb555, |
interlacing?yv12_to_rgb555i :yv12_to_rgb555, |
636 |
interlacing?yv12_to_rgb555i_c:yv12_to_rgb555_c, 2); |
interlacing?yv12_to_rgb555i_c:yv12_to_rgb555_c, 2, interlacing); |
637 |
return 0; |
return 0; |
638 |
|
|
639 |
case XVID_CSP_RGB565: |
case XVID_CSP_RGB565: |
641 |
dst[0], dst_stride[0], image->y, image->u, image->v, |
dst[0], dst_stride[0], image->y, image->u, image->v, |
642 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
643 |
interlacing?yv12_to_rgb565i :yv12_to_rgb565, |
interlacing?yv12_to_rgb565i :yv12_to_rgb565, |
644 |
interlacing?yv12_to_rgb565i_c:yv12_to_rgb565_c, 2); |
interlacing?yv12_to_rgb565i_c:yv12_to_rgb565_c, 2, interlacing); |
645 |
return 0; |
return 0; |
646 |
|
|
647 |
case XVID_CSP_BGR: |
case XVID_CSP_BGR: |
649 |
dst[0], dst_stride[0], image->y, image->u, image->v, |
dst[0], dst_stride[0], image->y, image->u, image->v, |
650 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
651 |
interlacing?yv12_to_bgri :yv12_to_bgr, |
interlacing?yv12_to_bgri :yv12_to_bgr, |
652 |
interlacing?yv12_to_bgri_c:yv12_to_bgr_c, 3); |
interlacing?yv12_to_bgri_c:yv12_to_bgr_c, 3, interlacing); |
653 |
return 0; |
return 0; |
654 |
|
|
655 |
case XVID_CSP_BGRA: |
case XVID_CSP_BGRA: |
657 |
dst[0], dst_stride[0], image->y, image->u, image->v, |
dst[0], dst_stride[0], image->y, image->u, image->v, |
658 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
659 |
interlacing?yv12_to_bgrai :yv12_to_bgra, |
interlacing?yv12_to_bgrai :yv12_to_bgra, |
660 |
interlacing?yv12_to_bgrai_c:yv12_to_bgra_c, 4); |
interlacing?yv12_to_bgrai_c:yv12_to_bgra_c, 4, interlacing); |
661 |
return 0; |
return 0; |
662 |
|
|
663 |
case XVID_CSP_ABGR: |
case XVID_CSP_ABGR: |
665 |
dst[0], dst_stride[0], image->y, image->u, image->v, |
dst[0], dst_stride[0], image->y, image->u, image->v, |
666 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
667 |
interlacing?yv12_to_abgri :yv12_to_abgr, |
interlacing?yv12_to_abgri :yv12_to_abgr, |
668 |
interlacing?yv12_to_abgri_c:yv12_to_abgr_c, 4); |
interlacing?yv12_to_abgri_c:yv12_to_abgr_c, 4, interlacing); |
669 |
|
return 0; |
670 |
|
|
671 |
|
case XVID_CSP_RGB: |
672 |
|
safe_packed_conv( |
673 |
|
dst[0], dst_stride[0], image->y, image->u, image->v, |
674 |
|
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
675 |
|
interlacing?yv12_to_rgbi :yv12_to_rgb, |
676 |
|
interlacing?yv12_to_rgbi_c:yv12_to_rgb_c, 3, interlacing); |
677 |
return 0; |
return 0; |
678 |
|
|
679 |
case XVID_CSP_RGBA: |
case XVID_CSP_RGBA: |
681 |
dst[0], dst_stride[0], image->y, image->u, image->v, |
dst[0], dst_stride[0], image->y, image->u, image->v, |
682 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
683 |
interlacing?yv12_to_rgbai :yv12_to_rgba, |
interlacing?yv12_to_rgbai :yv12_to_rgba, |
684 |
interlacing?yv12_to_rgbai_c:yv12_to_rgba_c, 4); |
interlacing?yv12_to_rgbai_c:yv12_to_rgba_c, 4, interlacing); |
685 |
return 0; |
return 0; |
686 |
|
|
687 |
case XVID_CSP_ARGB: |
case XVID_CSP_ARGB: |
689 |
dst[0], dst_stride[0], image->y, image->u, image->v, |
dst[0], dst_stride[0], image->y, image->u, image->v, |
690 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
691 |
interlacing?yv12_to_argbi :yv12_to_argb, |
interlacing?yv12_to_argbi :yv12_to_argb, |
692 |
interlacing?yv12_to_argbi_c:yv12_to_argb_c, 4); |
interlacing?yv12_to_argbi_c:yv12_to_argb_c, 4, interlacing); |
693 |
return 0; |
return 0; |
694 |
|
|
695 |
case XVID_CSP_YUY2: |
case XVID_CSP_YUY2: |
697 |
dst[0], dst_stride[0], image->y, image->u, image->v, |
dst[0], dst_stride[0], image->y, image->u, image->v, |
698 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
699 |
interlacing?yv12_to_yuyvi :yv12_to_yuyv, |
interlacing?yv12_to_yuyvi :yv12_to_yuyv, |
700 |
interlacing?yv12_to_yuyvi_c:yv12_to_yuyv_c, 2); |
interlacing?yv12_to_yuyvi_c:yv12_to_yuyv_c, 2, interlacing); |
701 |
return 0; |
return 0; |
702 |
|
|
703 |
case XVID_CSP_YVYU: /* u,v swapped */ |
case XVID_CSP_YVYU: /* u,v swapped */ |
705 |
dst[0], dst_stride[0], image->y, image->v, image->u, |
dst[0], dst_stride[0], image->y, image->v, image->u, |
706 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
707 |
interlacing?yv12_to_yuyvi :yv12_to_yuyv, |
interlacing?yv12_to_yuyvi :yv12_to_yuyv, |
708 |
interlacing?yv12_to_yuyvi_c:yv12_to_yuyv_c, 2); |
interlacing?yv12_to_yuyvi_c:yv12_to_yuyv_c, 2, interlacing); |
709 |
return 0; |
return 0; |
710 |
|
|
711 |
case XVID_CSP_UYVY: |
case XVID_CSP_UYVY: |
713 |
dst[0], dst_stride[0], image->y, image->u, image->v, |
dst[0], dst_stride[0], image->y, image->u, image->v, |
714 |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
edged_width, edged_width2, width, height, (csp & XVID_CSP_VFLIP), |
715 |
interlacing?yv12_to_uyvyi :yv12_to_uyvy, |
interlacing?yv12_to_uyvyi :yv12_to_uyvy, |
716 |
interlacing?yv12_to_uyvyi_c:yv12_to_uyvy_c, 2); |
interlacing?yv12_to_uyvyi_c:yv12_to_uyvy_c, 2, interlacing); |
717 |
return 0; |
return 0; |
718 |
|
|
719 |
case XVID_CSP_I420: /* YCbCr == YUV == internal colorspace for MPEG */ |
case XVID_CSP_I420: /* YCbCr == YUV == internal colorspace for MPEG */ |
1042 |
p += edged_width/2; |
p += edged_width/2; |
1043 |
} |
} |
1044 |
} |
} |
1045 |
|
|
1046 |
|
/****************************************************************************/ |
1047 |
|
|
1048 |
|
static void (*deintl_core)(uint8_t *, int width, int height, const int stride) = 0; |
1049 |
|
extern void xvid_deinterlace_sse(uint8_t *, int width, int height, const int stride); |
1050 |
|
|
1051 |
|
#define CLIP_255(x) ( ((x)&~255) ? ((-(x)) >> (8*sizeof((x))-1))&0xff : (x) ) |
1052 |
|
|
1053 |
|
static void deinterlace_c(uint8_t *pix, int width, int height, const int bps) |
1054 |
|
{ |
1055 |
|
pix += bps; |
1056 |
|
while(width-->0) |
1057 |
|
{ |
1058 |
|
int p1 = pix[-bps]; |
1059 |
|
int p2 = pix[0]; |
1060 |
|
int p0 = p2; |
1061 |
|
int j = (height>>1) - 1; |
1062 |
|
int V; |
1063 |
|
unsigned char *P = pix++; |
1064 |
|
while(j-->0) |
1065 |
|
{ |
1066 |
|
const int p3 = P[ bps]; |
1067 |
|
const int p4 = P[2*bps]; |
1068 |
|
V = ((p1+p3+1)>>1) + ((p2 - ((p0+p4+1)>>1)) >> 2); |
1069 |
|
P[0] = CLIP_255( V ); |
1070 |
|
p0 = p2; |
1071 |
|
p1 = p3; |
1072 |
|
p2 = p4; |
1073 |
|
P += 2*bps; |
1074 |
|
} |
1075 |
|
V = ((p1+p1+1)>>1) + ((p2 - ((p0+p2+1)>>1)) >> 2); |
1076 |
|
P[0] = CLIP_255( V ); |
1077 |
|
} |
1078 |
|
} |
1079 |
|
#undef CLIP_255 |
1080 |
|
|
1081 |
|
int xvid_image_deinterlace(xvid_image_t* img, int width, int height, int bottom_first) |
1082 |
|
{ |
1083 |
|
if (height&1) |
1084 |
|
return 0; |
1085 |
|
if (img->csp!=XVID_CSP_PLANAR && img->csp!=XVID_CSP_I420 && img->csp!=XVID_CSP_YV12) |
1086 |
|
return 0; /* not yet supported */ |
1087 |
|
if (deintl_core==0) { |
1088 |
|
deintl_core = deinterlace_c; |
1089 |
|
#if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) |
1090 |
|
{ |
1091 |
|
int cpu_flags = check_cpu_features(); |
1092 |
|
if (cpu_flags & XVID_CPU_MMX) |
1093 |
|
deintl_core = xvid_deinterlace_sse; |
1094 |
|
} |
1095 |
|
#endif |
1096 |
|
} |
1097 |
|
if (!bottom_first) { |
1098 |
|
deintl_core(img->plane[0], width, height, img->stride[0]); |
1099 |
|
deintl_core(img->plane[1], width>>1, height>>1, img->stride[1]); |
1100 |
|
deintl_core(img->plane[2], width>>1, height>>1, img->stride[2]); |
1101 |
|
} |
1102 |
|
else { |
1103 |
|
deintl_core((uint8_t *)img->plane[0] + ( height -1)*img->stride[0], width, height, -img->stride[0]); |
1104 |
|
deintl_core((uint8_t *)img->plane[1] + ((height>>1)-1)*img->stride[1], width>>1, height>>1, -img->stride[1]); |
1105 |
|
deintl_core((uint8_t *)img->plane[2] + ((height>>1)-1)*img->stride[2], width>>1, height>>1, -img->stride[2]); |
1106 |
|
} |
1107 |
|
emms(); |
1108 |
|
|
1109 |
|
return 1; |
1110 |
|
} |
1111 |
|
|