Mercurial > mplayer.hg
annotate mp3lib/dct64_k7.s @ 1271:2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
author | nick |
---|---|
date | Wed, 04 Jul 2001 09:47:56 +0000 |
parents | 03b7e2955a20 |
children | 3a9699d9e7da |
rev | line source |
---|---|
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
1 # This code was taken from http://www.mpg123.org |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
2 # See ChangeLog of mpg123-0.59s-pre.1 for detail |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
3 # Applied to mplayer by Nick Kurshev <nickols_k@mail.ru> |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
4 # Partial 3dnowex-DSP! optimization by Nick Kurshev |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
5 # |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
6 # TODO: optimize scalar 3dnow! code |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
7 # Warning: Phases 7 & 8 are not tested |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
8 # |
736 | 9 |
1173 | 10 .text |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
11 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
12 .align 16 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
13 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
14 .globl dct64_MMX_3dnowex |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
15 dct64_MMX_3dnowex: |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
16 pushl %ebx |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
17 pushl %esi |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
18 pushl %edi |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
19 subl $256,%esp |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
20 movl 280(%esp),%eax |
781 | 21 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
22 leal 128(%esp),%edx |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
23 movl 272(%esp),%esi |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
24 movl 276(%esp),%edi |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
25 movl $costab_mmx,%ebx |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
26 orl %ecx,%ecx |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
27 movl %esp,%ecx |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
28 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
29 /* Phase 1*/ |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
30 movq (%eax), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
31 movq 8(%eax), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
32 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
33 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
34 movq 120(%eax), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
35 movq 112(%eax), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
36 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
37 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
38 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
39 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
40 movq %mm0, (%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
41 movq %mm4, 8(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
42 pfsub %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
43 pfsub %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
44 pfmul (%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
45 pfmul 8(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
46 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
47 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
48 movq %mm3, 120(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
49 movq %mm7, 112(%edx) |
781 | 50 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
51 movq 16(%eax), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
52 movq 24(%eax), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
53 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
54 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
55 movq 104(%eax), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
56 movq 96(%eax), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
57 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
58 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
59 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
60 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
61 movq %mm0, 16(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
62 movq %mm4, 24(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
63 pfsub %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
64 pfsub %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
65 pfmul 16(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
66 pfmul 24(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
67 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
68 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
69 movq %mm3, 104(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
70 movq %mm7, 96(%edx) |
781 | 71 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
72 movq 32(%eax), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
73 movq 40(%eax), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
74 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
75 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
76 movq 88(%eax), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
77 movq 80(%eax), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
78 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
79 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
80 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
81 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
82 movq %mm0, 32(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
83 movq %mm4, 40(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
84 pfsub %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
85 pfsub %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
86 pfmul 32(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
87 pfmul 40(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
88 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
89 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
90 movq %mm3, 88(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
91 movq %mm7, 80(%edx) |
781 | 92 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
93 movq 48(%eax), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
94 movq 56(%eax), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
95 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
96 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
97 movq 72(%eax), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
98 movq 64(%eax), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
99 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
100 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
101 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
102 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
103 movq %mm0, 48(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
104 movq %mm4, 56(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
105 pfsub %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
106 pfsub %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
107 pfmul 48(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
108 pfmul 56(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
109 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
110 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
111 movq %mm3, 72(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
112 movq %mm7, 64(%edx) |
781 | 113 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
114 /* Phase 2*/ |
781 | 115 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
116 movq (%edx), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
117 movq 8(%edx), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
118 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
119 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
120 movq 56(%edx), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
121 movq 48(%edx), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
122 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
123 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
124 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
125 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
126 movq %mm0, (%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
127 movq %mm4, 8(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
128 pfsub %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
129 pfsub %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
130 pfmul 64(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
131 pfmul 72(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
132 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
133 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
134 movq %mm3, 56(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
135 movq %mm7, 48(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
136 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
137 movq 16(%edx), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
138 movq 24(%edx), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
139 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
140 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
141 movq 40(%edx), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
142 movq 32(%edx), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
143 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
144 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
145 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
146 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
147 movq %mm0, 16(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
148 movq %mm4, 24(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
149 pfsub %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
150 pfsub %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
151 pfmul 80(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
152 pfmul 88(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
153 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
154 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
155 movq %mm3, 40(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
156 movq %mm7, 32(%ecx) |
781 | 157 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
158 /* Phase 3*/ |
781 | 159 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
160 movq 64(%edx), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
161 movq 72(%edx), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
162 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
163 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
164 movq 120(%edx), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
165 movq 112(%edx), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
166 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
167 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
168 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
169 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
170 movq %mm0, 64(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
171 movq %mm4, 72(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
172 pfsubr %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
173 pfsubr %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
174 pfmul 64(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
175 pfmul 72(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
176 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
177 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
178 movq %mm3, 120(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
179 movq %mm7, 112(%ecx) |
736 | 180 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
181 movq 80(%edx), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
182 movq 88(%edx), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
183 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
184 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
185 movq 104(%edx), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
186 movq 96(%edx), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
187 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
188 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
189 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
190 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
191 movq %mm0, 80(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
192 movq %mm4, 88(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
193 pfsubr %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
194 pfsubr %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
195 pfmul 80(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
196 pfmul 88(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
197 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
198 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
199 movq %mm3, 104(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
200 movq %mm7, 96(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
201 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
202 /* Phase 4*/ |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
203 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
204 movq (%ecx), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
205 movq 8(%ecx), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
206 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
207 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
208 movq 24(%ecx), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
209 movq 16(%ecx), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
210 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
211 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
212 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
213 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
214 movq %mm0, (%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
215 movq %mm4, 8(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
216 pfsub %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
217 pfsub %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
218 pfmul 96(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
219 pfmul 104(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
220 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
221 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
222 movq %mm3, 24(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
223 movq %mm7, 16(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
224 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
225 movq 32(%ecx), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
226 movq 40(%ecx), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
227 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
228 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
229 movq 56(%ecx), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
230 movq 48(%ecx), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
231 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
232 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
233 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
234 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
235 movq %mm0, 32(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
236 movq %mm4, 40(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
237 pfsubr %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
238 pfsubr %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
239 pfmul 96(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
240 pfmul 104(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
241 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
242 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
243 movq %mm3, 56(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
244 movq %mm7, 48(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
245 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
246 movq 64(%ecx), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
247 movq 72(%ecx), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
248 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
249 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
250 movq 88(%ecx), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
251 movq 80(%ecx), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
252 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
253 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
254 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
255 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
256 movq %mm0, 64(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
257 movq %mm4, 72(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
258 pfsub %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
259 pfsub %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
260 pfmul 96(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
261 pfmul 104(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
262 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
263 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
264 movq %mm3, 88(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
265 movq %mm7, 80(%edx) |
736 | 266 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
267 movq 96(%ecx), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
268 movq 104(%ecx), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
269 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
270 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
271 movq 120(%ecx), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
272 movq 112(%ecx), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
273 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
274 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
275 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
276 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
277 movq %mm0, 96(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
278 movq %mm4, 104(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
279 pfsubr %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
280 pfsubr %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
281 pfmul 96(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
282 pfmul 104(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
283 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
284 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
285 movq %mm3, 120(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
286 movq %mm7, 112(%edx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
287 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
288 /* Phase 5 */ |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
289 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
290 movq (%edx), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
291 movq 16(%edx), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
292 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
293 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
294 movq 8(%edx), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
295 movq 24(%edx), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
296 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
297 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
298 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
299 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
300 movq %mm0, (%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
301 movq %mm4, 16(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
302 pfsub %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
303 pfsubr %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
304 pfmul 112(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
305 pfmul 112(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
306 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
307 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
308 movq %mm3, 8(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
309 movq %mm7, 24(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
310 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
311 movq 32(%edx), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
312 movq 48(%edx), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
313 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
314 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
315 movq 40(%edx), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
316 movq 56(%edx), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
317 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
318 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
319 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
320 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
321 movq %mm0, 32(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
322 movq %mm4, 48(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
323 pfsub %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
324 pfsubr %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
325 pfmul 112(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
326 pfmul 112(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
327 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
328 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
329 movq %mm3, 40(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
330 movq %mm7, 56(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
331 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
332 movq 64(%edx), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
333 movq 80(%edx), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
334 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
335 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
336 movq 72(%edx), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
337 movq 88(%edx), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
338 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
339 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
340 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
341 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
342 movq %mm0, 64(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
343 movq %mm4, 80(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
344 pfsub %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
345 pfsubr %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
346 pfmul 112(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
347 pfmul 112(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
348 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
349 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
350 movq %mm3, 72(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
351 movq %mm7, 88(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
352 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
353 movq 96(%edx), %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
354 movq 112(%edx), %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
355 movq %mm0, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
356 movq %mm4, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
357 movq 104(%edx), %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
358 movq 120(%edx), %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
359 pswapd %mm1, %mm1 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
360 pswapd %mm5, %mm5 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
361 pfadd %mm1, %mm0 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
362 pfadd %mm5, %mm4 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
363 movq %mm0, 96(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
364 movq %mm4, 112(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
365 pfsub %mm1, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
366 pfsubr %mm5, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
367 pfmul 112(%ebx), %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
368 pfmul 112(%ebx), %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
369 pswapd %mm3, %mm3 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
370 pswapd %mm7, %mm7 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
371 movq %mm3, 104(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
372 movq %mm7, 120(%ecx) |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
373 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
374 /* Phase 6. This is the end of easy road. */ |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
375 /* Code below is coded in scalar mode. Should be optimized */ |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
376 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
377 movd 32(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
378 pfadd 36(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
379 movd %mm0, 32(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
380 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
381 movd 32(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
382 pfsub 36(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
383 pfmul 120(%ebx),%mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
384 movd %mm0, 36(%edx) |
736 | 385 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
386 movd 44(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
387 pfsub 40(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
388 pfmul 120(%ebx),%mm0 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
389 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
390 movd %mm0, 44(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
391 pfadd 40(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
392 pfadd 44(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
393 movd %mm0, 40(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
394 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
395 movd 48(%ecx), %mm3 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
396 pfsub 52(%ecx), %mm3 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
397 pfmul 120(%ebx), %mm3 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
398 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
399 movd 60(%ecx), %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
400 pfsub 56(%ecx), %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
401 pfmul 120(%ebx), %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
402 movq %mm2, %mm1 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
403 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
404 pfadd 56(%ecx), %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
405 pfadd 60(%ecx), %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
406 movq %mm1, %mm0 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
407 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
408 pfadd 48(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
409 pfadd 52(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
410 movd %mm0, 48(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
411 pfadd %mm3, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
412 movd %mm1, 56(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
413 movd %mm2, 60(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
414 pfadd %mm3, %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
415 movd %mm2, 52(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
416 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
417 /*---*/ |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
418 movd 64(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
419 pfadd 68(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
420 movd %mm0, 64(%edx) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
421 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
422 movd 64(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
423 pfsub 68(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
424 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
425 movd %mm0, 68(%edx) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
426 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
427 movd 76(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
428 pfsub 72(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
429 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
430 movd %mm0, 76(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
431 pfadd 72(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
432 pfadd 76(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
433 movd %mm0, 72(%edx) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
434 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
435 movd 92(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
436 pfsub 88(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
437 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
438 movd %mm0, 92(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
439 pfadd 92(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
440 pfadd 88(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
441 movq %mm0, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
442 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
443 pfadd 80(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
444 pfadd 84(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
445 movd %mm0, 80(%edx) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
446 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
447 movd 80(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
448 pfsub 84(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
449 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
450 pfadd %mm0, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
451 pfadd 92(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
452 movd %mm0, 84(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
453 movd %mm1, 88(%edx) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
454 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
455 movd 96(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
456 pfadd 100(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
457 movd %mm0, 96(%edx) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
458 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
459 movd 96(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
460 pfsub 100(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
461 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
462 movd %mm0, 100(%edx) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
463 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
464 movd 108(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
465 pfsub 104(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
466 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
467 movd %mm0, 108(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
468 pfadd 104(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
469 pfadd 108(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
470 movd %mm0, 104(%edx) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
471 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
472 movd 124(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
473 pfsub 120(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
474 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
475 movd %mm0, 124(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
476 pfadd 120(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
477 pfadd 124(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
478 movq %mm0, %mm1 |
736 | 479 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
480 pfadd 112(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
481 pfadd 116(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
482 movd %mm0, 112(%edx) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
483 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
484 movd 112(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
485 pfsub 116(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
486 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
487 pfadd %mm0,%mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
488 pfadd 124(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
489 movd %mm0, 116(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
490 movd %mm1, 120(%edx) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
491 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
492 jnz .L01 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
493 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
494 /* Phase 7*/ |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
495 /* Code below is coded in scalar mode. Should be optimized */ |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
496 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
497 movd (%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
498 pfadd 4(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
499 movd %mm0, 1024(%esi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
500 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
501 movd (%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
502 pfsub 4(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
503 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
504 movd %mm0, (%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
505 movd %mm0, (%edi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
506 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
507 movd 12(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
508 pfsub 8(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
509 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
510 movd %mm0, 512(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
511 pfadd 12(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
512 pfadd 8(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
513 movd %mm0, 512(%esi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
514 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
515 movd 16(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
516 pfsub 20(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
517 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
518 movq %mm0, %mm3 |
781 | 519 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
520 movd 28(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
521 pfsub 24(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
522 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
523 movd %mm0, 768(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
524 movq %mm0, %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
525 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
526 pfadd 24(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
527 pfadd 28(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
528 movq %mm0, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
529 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
530 pfadd 16(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
531 pfadd 20(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
532 movd %mm0, 768(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
533 pfadd %mm3, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
534 movd %mm1, 256(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
535 pfadd %mm3, %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
536 movd %mm2, 256(%edi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
537 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
538 /* Phase 8*/ |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
539 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
540 movq 32(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
541 movq 48(%edx), %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
542 pfadd 48(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
543 pfadd 40(%edx), %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
544 movd %mm0, 896(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
545 movd %mm1, 640(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
546 psrlq $32, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
547 psrlq $32, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
548 movd %mm0, 128(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
549 movd %mm1, 384(%edi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
550 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
551 movd 40(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
552 pfadd 56(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
553 movd %mm0, 384(%esi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
554 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
555 movd 56(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
556 pfadd 36(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
557 movd %mm0, 128(%esi) |
781 | 558 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
559 movd 60(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
560 movd %mm0, 896(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
561 pfadd 44(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
562 movd %mm0, 640(%edi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
563 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
564 movq 96(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
565 movq 112(%edx), %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
566 movq 104(%edx), %mm4 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
567 pfadd 112(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
568 pfadd 104(%edx), %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
569 pfadd 120(%edx), %mm4 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
570 movq %mm0, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
571 movq %mm2, %mm3 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
572 movq %mm4, %mm5 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
573 pfadd 64(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
574 pfadd 80(%edx), %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
575 pfadd 72(%edx), %mm4 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
576 movd %mm0, 960(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
577 movd %mm2, 704(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
578 movd %mm4, 448(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
579 psrlq $32, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
580 psrlq $32, %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
581 psrlq $32, %mm4 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
582 movd %mm0, 64(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
583 movd %mm2, 320(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
584 movd %mm4, 576(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
585 pfadd 80(%edx), %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
586 pfadd 72(%edx), %mm3 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
587 pfadd 88(%edx), %mm5 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
588 movd %mm1, 832(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
589 movd %mm3, 576(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
590 movd %mm5, 320(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
591 psrlq $32, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
592 psrlq $32, %mm3 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
593 psrlq $32, %mm5 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
594 movd %mm1, 192(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
595 movd %mm3, 448(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
596 movd %mm5, 704(%edi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
597 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
598 movd 120(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
599 pfadd 100(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
600 movq %mm0, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
601 pfadd 88(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
602 movd %mm0, 192(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
603 pfadd 68(%edx), %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
604 movd %mm1, 64(%esi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
605 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
606 movd 124(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
607 movd %mm0, 960(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
608 pfadd 92(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
609 movd %mm0, 832(%edi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
610 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
611 jmp .L_bye |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
612 .L01: |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
613 /* Phase 9*/ |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
614 movd (%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
615 pfadd 4(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
616 pf2id %mm0, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
617 movd %mm0, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
618 movw %ax, 512(%esi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
619 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
620 movd (%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
621 pfsub 4(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
622 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
623 pf2id %mm0, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
624 movd %mm0, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
625 movw %ax, (%esi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
626 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
627 movd 12(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
628 pfsub 8(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
629 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
630 pf2id %mm0, %mm7 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
631 movd %mm7, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
632 movw %ax, 256(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
633 pfadd 12(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
634 pfadd 8(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
635 pf2id %mm0, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
636 movd %mm0, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
637 movw %ax, 256(%esi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
638 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
639 movd 16(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
640 pfsub 20(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
641 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
642 movq %mm0, %mm3 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
643 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
644 movd 28(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
645 pfsub 24(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
646 pfmul 120(%ebx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
647 pf2id %mm0, %mm7 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
648 movd %mm7, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
649 movw %ax, 384(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
650 movq %mm0, %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
651 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
652 pfadd 24(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
653 pfadd 28(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
654 movq %mm0, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
655 pfadd 16(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
656 pfadd 20(%ecx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
657 pf2id %mm0, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
658 movd %mm0, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
659 movw %ax, 384(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
660 pfadd %mm3, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
661 pf2id %mm1, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
662 movd %mm1, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
663 movw %ax, 128(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
664 pfadd %mm3, %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
665 pf2id %mm2, %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
666 movd %mm2, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
667 movw %ax, 128(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
668 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
669 |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
670 /* Phase 10*/ |
736 | 671 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
672 movq 32(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
673 movq 48(%edx), %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
674 pfadd 48(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
675 pfadd 40(%edx), %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
676 pf2id %mm0, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
677 pf2id %mm1, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
678 movd %mm0, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
679 movd %mm1, %ecx |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
680 movw %ax, 448(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
681 movw %cx, 320(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
682 psrlq $32, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
683 psrlq $32, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
684 movd %mm0, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
685 movd %mm1, %ecx |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
686 movw %ax, 64(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
687 movw %cx, 192(%edi) |
736 | 688 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
689 movd 40(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
690 pfadd 56(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
691 pf2id %mm0, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
692 movd %mm0, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
693 movw %ax, 192(%esi) |
736 | 694 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
695 movd 56(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
696 pfadd 36(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
697 pf2id %mm0, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
698 movd %mm0, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
699 movw %ax, 64(%esi) |
736 | 700 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
701 movd 60(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
702 pf2id %mm0, %mm7 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
703 movd %mm7, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
704 movw %ax, 448(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
705 pfadd 44(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
706 pf2id %mm0, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
707 movd %mm0, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
708 movw %ax, 320(%edi) |
736 | 709 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
710 movq 96(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
711 movq 112(%edx), %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
712 movq 104(%edx), %mm4 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
713 pfadd 112(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
714 pfadd 104(%edx), %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
715 pfadd 120(%edx), %mm4 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
716 movq %mm0, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
717 movq %mm2, %mm3 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
718 movq %mm4, %mm5 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
719 pfadd 64(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
720 pfadd 80(%edx), %mm2 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
721 pfadd 72(%edx), %mm4 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
722 pf2id %mm0, %mm7 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
723 pf2id %mm2, %mm6 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
724 pf2id %mm4, %mm4 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
725 movd %mm7, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
726 movd %mm6, %ecx |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
727 movd %mm4, %ebx |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
728 movw %ax, 480(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
729 movw %cx, 352(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
730 movw %bx, 224(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
731 psrlq $32, %mm7 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
732 psrlq $32, %mm6 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
733 psrlq $32, %mm4 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
734 movd %mm7, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
735 movd %mm6, %ecx |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
736 movd %mm4, %ebx |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
737 movw %ax, 32(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
738 movw %cx, 160(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
739 movw %bx, 288(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
740 pfadd 80(%edx), %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
741 pfadd 72(%edx), %mm3 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
742 pfadd 88(%edx), %mm5 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
743 pf2id %mm1, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
744 pf2id %mm3, %mm3 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
745 pf2id %mm5, %mm5 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
746 movd %mm1, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
747 movd %mm3, %ecx |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
748 movd %mm5, %ebx |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
749 movw %ax, 416(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
750 movw %cx, 288(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
751 movw %bx, 160(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
752 psrlq $32, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
753 psrlq $32, %mm3 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
754 psrlq $32, %mm5 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
755 movd %mm1, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
756 movd %mm3, %ecx |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
757 movd %mm5, %ebx |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
758 movw %ax, 96(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
759 movw %cx, 224(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
760 movw %bx, 352(%edi) |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
761 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
762 movd 120(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
763 pfadd 100(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
764 movq %mm0, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
765 pfadd 88(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
766 pf2id %mm0, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
767 movd %mm0, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
768 movw %ax, 96(%esi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
769 pfadd 68(%edx), %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
770 pf2id %mm1, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
771 movd %mm1, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
772 movw %ax, 32(%esi) |
736 | 773 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
774 movq 124(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
775 pf2id %mm0, %mm1 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
776 movd %mm1, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
777 movw %ax, 480(%edi) |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
778 pfadd 92(%edx), %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
779 pf2id %mm0, %mm0 |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
780 movd %mm0, %eax |
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
781 movw %ax, 416(%edi) |
736 | 782 |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
783 movsw |
736 | 784 |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
785 .L_bye: |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
786 addl $256,%esp |
1271
2864e32cd267
Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents:
1245
diff
changeset
|
787 femms |
1245
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
788 popl %edi |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
789 popl %esi |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
790 popl %ebx |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
791 ret |
03b7e2955a20
Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
1173
diff
changeset
|
792 |
736 | 793 |