Mercurial > mplayer.hg

annotate mp3lib/decode_MMX.s @ 1271:2864e32cd267

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Finished 3dnow optimization (in scalar mode) and minor improvements
author nick
date Wed, 04 Jul 2001 09:47:56 +0000
parents 9bf97b404134
children 3a9699d9e7da
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
1245
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
1 # this code comes under GPL
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
2 # 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:
diff changeset
3 # 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:
diff changeset
4 # 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:
diff changeset
5 #
1259
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
6 # Local ChangeLog:
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
7 # - Partial loops unrolling and removing MOVW insn from loops
1245
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
8 #
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
9
1259
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
10 .data
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
11 .align 8
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
12 null_one: .long 0x0000ffff, 0x0000ffff
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
13 one_null: .long 0xffff0000, 0xffff0000
1271
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
14 .globl costab_mmx
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
15 costab_mmx:
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
16 .long 1056974725
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
17 .long 1057056395
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
18 .long 1057223771
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
19 .long 1057485416
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
20 .long 1057855544
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
21 .long 1058356026
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
22 .long 1059019886
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
23 .long 1059897405
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
24 .long 1061067246
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
25 .long 1062657950
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
26 .long 1064892987
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
27 .long 1066774581
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
28 .long 1069414683
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
29 .long 1073984175
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
30 .long 1079645762
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
31 .long 1092815430
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
32 .long 1057005197
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
33 .long 1057342072
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
34 .long 1058087743
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
35 .long 1059427869
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
36 .long 1061799040
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
37 .long 1065862217
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
38 .long 1071413542
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
39 .long 1084439708
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
40 .long 1057128951
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
41 .long 1058664893
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
42 .long 1063675095
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
43 .long 1076102863
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
44 .long 1057655764
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
45 .long 1067924853
2864e32cd267 Finished 3dnow optimization (in scalar mode) and minor improvements
nick
parents: 1259
diff changeset
46 .long 1060439283
1259
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
47
1245
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
48 .text
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
49
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
50 .globl synth_1to1_MMX_s
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
51
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
52 synth_1to1_MMX_s:
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
53 pushl %ebp
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
54 pushl %edi
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
55 pushl %esi
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
56 pushl %ebx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
57 movl 24(%esp),%ecx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
58 movl 28(%esp),%edi
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
59 movl $15,%ebx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
60 movl 36(%esp),%edx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
61 leal (%edi,%ecx,2),%edi
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
62 decl %ecx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
63 movl 32(%esp),%esi
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
64 movl (%edx),%eax
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
65 jecxz .L1
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
66 decl %eax
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
67 andl %ebx,%eax
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
68 leal 1088(%esi),%esi
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
69 movl %eax,(%edx)
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
70 .L1:
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
71 leal (%esi,%eax,2),%edx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
72 movl %eax,%ebp
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
73 incl %eax
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
74 pushl 20(%esp)
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
75 andl %ebx,%eax
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
76 leal 544(%esi,%eax,2),%ecx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
77 incl %ebx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
78 testl $1, %eax
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
79 jnz .L2
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
80 xchgl %edx,%ecx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
81 incl %ebp
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
82 leal 544(%esi),%esi
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
83 .L2:
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
84 emms
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
85 pushl %edx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
86 pushl %ecx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
87 call *dct64_MMX_func
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
88 addl $12,%esp
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
89 leal 1(%ebx), %ecx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
90 subl %ebp,%ebx
1259
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
91 pushl %ecx
1245
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
92 leal decwins(%ebx,%ebx,1), %edx
1259
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
93 shrl $1, %ecx
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
94 .align 16
1245
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
95 .L3:
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
96 movq (%edx),%mm0
1259
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
97 movq 64(%edx),%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
98 pmaddwd (%esi),%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
99 pmaddwd 32(%esi),%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
100 movq 8(%edx),%mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
101 movq 72(%edx),%mm5
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
102 pmaddwd 8(%esi),%mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
103 pmaddwd 40(%esi),%mm5
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
104 movq 16(%edx),%mm2
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
105 movq 80(%edx),%mm6
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
106 pmaddwd 16(%esi),%mm2
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
107 pmaddwd 48(%esi),%mm6
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
108 movq 24(%edx),%mm3
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
109 movq 88(%edx),%mm7
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
110 pmaddwd 24(%esi),%mm3
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
111 pmaddwd 56(%esi),%mm7
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
112 paddd %mm1,%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
113 paddd %mm5,%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
114 paddd %mm2,%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
115 paddd %mm6,%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
116 paddd %mm3,%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
117 paddd %mm7,%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
118 movq %mm0,%mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
119 movq %mm4,%mm5
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
120 psrlq $32,%mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
121 psrlq $32,%mm5
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
122 paddd %mm1,%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
123 paddd %mm5,%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
124 psrad $13,%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
125 psrad $13,%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
126 packssdw %mm0,%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
127 packssdw %mm4,%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
128
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
129 movq (%edi), %mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
130 punpckldq %mm4, %mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
131 pand one_null, %mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
132 pand null_one, %mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
133 por %mm0, %mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
134 movq %mm1,(%edi)
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
135
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
136 leal 64(%esi),%esi
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
137 leal 128(%edx),%edx
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
138 leal 8(%edi),%edi
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
139
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
140 decl %ecx
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
141 jnz .L3
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
142
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
143 popl %ecx
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
144 andl $1, %ecx
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
145 jecxz .next_loop
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
146
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
147 movq (%edx),%mm0
1245
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
148 pmaddwd (%esi),%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
149 movq 8(%edx),%mm1
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
150 pmaddwd 8(%esi),%mm1
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
151 movq 16(%edx),%mm2
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
152 pmaddwd 16(%esi),%mm2
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
153 movq 24(%edx),%mm3
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
154 pmaddwd 24(%esi),%mm3
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
155 paddd %mm1,%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
156 paddd %mm2,%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
157 paddd %mm3,%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
158 movq %mm0,%mm1
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
159 psrlq $32,%mm1
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
160 paddd %mm1,%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
161 psrad $13,%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
162 packssdw %mm0,%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
163 movd %mm0,%eax
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
164 movw %ax, (%edi)
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
165 leal 32(%esi),%esi
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
166 leal 64(%edx),%edx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
167 leal 4(%edi),%edi
1259
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
168
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
169 .next_loop:
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
170 subl $64,%esi
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
171 movl $7,%ecx
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
172 .align 16
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
173 .L4:
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
174 movq (%edx),%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
175 movq 64(%edx),%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
176 pmaddwd (%esi),%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
177 pmaddwd -32(%esi),%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
178 movq 8(%edx),%mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
179 movq 72(%edx),%mm5
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
180 pmaddwd 8(%esi),%mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
181 pmaddwd -24(%esi),%mm5
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
182 movq 16(%edx),%mm2
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
183 movq 80(%edx),%mm6
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
184 pmaddwd 16(%esi),%mm2
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
185 pmaddwd -16(%esi),%mm6
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
186 movq 24(%edx),%mm3
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
187 movq 88(%edx),%mm7
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
188 pmaddwd 24(%esi),%mm3
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
189 pmaddwd -8(%esi),%mm7
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
190 paddd %mm1,%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
191 paddd %mm5,%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
192 paddd %mm2,%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
193 paddd %mm6,%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
194 paddd %mm3,%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
195 paddd %mm7,%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
196 movq %mm0,%mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
197 movq %mm4,%mm5
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
198 psrlq $32,%mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
199 psrlq $32,%mm5
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
200 paddd %mm0,%mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
201 paddd %mm4,%mm5
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
202 psrad $13,%mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
203 psrad $13,%mm5
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
204 packssdw %mm1,%mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
205 packssdw %mm5,%mm5
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
206 psubd %mm0,%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
207 psubd %mm4,%mm4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
208 psubsw %mm1,%mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
209 psubsw %mm5,%mm4
1245
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
210
1259
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
211 movq (%edi), %mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
212 punpckldq %mm4, %mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
213 pand one_null, %mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
214 pand null_one, %mm0
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
215 por %mm0, %mm1
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
216 movq %mm1,(%edi)
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
217
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
218 subl $64,%esi
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
219 addl $128,%edx
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
220 leal 8(%edi),%edi
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
221 decl %ecx
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
222 jnz .L4
9bf97b404134 Partial loops unrolling
nick
parents: 1245
diff changeset
223
1245
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
224 movq (%edx),%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
225 pmaddwd (%esi),%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
226 movq 8(%edx),%mm1
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
227 pmaddwd 8(%esi),%mm1
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
228 movq 16(%edx),%mm2
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
229 pmaddwd 16(%esi),%mm2
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
230 movq 24(%edx),%mm3
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
231 pmaddwd 24(%esi),%mm3
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
232 paddd %mm1,%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
233 paddd %mm2,%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
234 paddd %mm3,%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
235 movq %mm0,%mm1
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
236 psrlq $32,%mm1
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
237 paddd %mm0,%mm1
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
238 psrad $13,%mm1
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
239 packssdw %mm1,%mm1
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
240 psubd %mm0,%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
241 psubsw %mm1,%mm0
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
242 movd %mm0,%eax
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
243 movw %ax,(%edi)
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
244
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
245 emms
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
246 popl %ebx
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
247 popl %esi
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
248 popl %edi
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
249 popl %ebp
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
250 ret
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
251
03b7e2955a20 Added newest MMX-optimized decore which speedups decoding at least on 13% for any cpu.
nick
parents:
diff changeset
252