view release on metacpan or  search on metacpan

clib/libmpeg2/motion_comp_mmx.c  view on Meta::CPAN

33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
#include "mmx.h"
 
#define CPU_MMXEXT 0
#define CPU_3DNOW 1
 
 
/* MMX code - needs a rewrite */
 
/*
 * Motion Compensation frequently needs to average values using the
 * formula (x+y+1)>>1. Both MMXEXT and 3Dnow include one instruction
 * to compute this, but it's been left out of classic MMX.
 *
 * We need to be careful of overflows when doing this computation.
 * Rather than unpacking data to 16-bits, which reduces parallelism,
 * we use the following formulas:
 *
 * (x+y)>>1 == (x&y)+((x^y)>>1)
 * (x+y+1)>>1 == (x|y)-((x^y)>>1)
 */
 
/* some rounding constants */
static mmx_t mask1 = {0xfefefefefefefefeLL};
static mmx_t round4 = {0x0002000200020002LL};
 
/*

clib/libmpeg2/motion_comp_vis.c  view on Meta::CPAN

24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
 
#ifdef ARCH_SPARC
 
#include <inttypes.h>
 
#include "mpeg2.h"
#include "attributes.h"
#include "mpeg2_internal.h"
#include "vis.h"
 
/* The trick used in some of this file is the formula from the MMX
 * motion comp code, which is:
 *
 * (x+y+1)>>1 == (x|y)-((x^y)>>1)
 *
 * This allows us to average 8 bytes at a time in a 64-bit FPU reg.
 * We avoid overflows by masking before we do the shift, and we
 * implement the shift by multiplying by 1/2 using mul8x16.  So in
 * VIS this is (assume 'x' is in f0, 'y' is in f2, a repeating mask
 * of '0xfe' is in f4, a repeating mask of '0x7f' is in f6, and
 * the value 0x80808080 is in f8):