Mercurial > mplayer.hg

annotate mp3lib/dct36_3dnow.c @ 26981:f3f553c03044

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
No need to set LC_ALL=C for individual shell commands, it is already set from config.mak.
author diego
date Sat, 07 Jun 2008 12:01:29 +0000
parents 7c7885350d89
children b5a46071062a
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
10322
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
1 /*
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
2 * dct36_3dnow.c - 3DNow! optimized dct36()
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
3 *
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
4 * This code based 'dct36_3dnow.s' by Syuuhei Kashiyama
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
5 * <squash@mb.kcom.ne.jp>, only two types of changes have been made:
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
6 *
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
7 * - removed PREFETCH instruction for speedup
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
8 * - changed function name for support 3DNow! automatic detection
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
9 *
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
10 * You can find Kashiyama's original 3dnow! support patch
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
11 * (for mpg123-0.59o) at
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
12 * http://user.ecc.u-tokyo.ac.jp/~g810370/linux-simd/ (Japanese).
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
13 *
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
14 * by KIMURA Takuhiro <kim@hannah.ipc.miyakyo-u.ac.jp> - until 31.Mar.1999
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
15 * <kim@comtec.co.jp> - after 1.Apr.1999
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
16 *
18783
0783dd397f74 CVS --> Subversion in copyright notices
diego
parents: 16989
diff changeset
17 * Modified for use with MPlayer, for details see the changelog at
0783dd397f74 CVS --> Subversion in copyright notices
diego
parents: 16989
diff changeset
18 * http://svn.mplayerhq.hu/mplayer/trunk/
15167
07e7a572bd84 Mark modified imported files as such to comply with (L)GPL ¡ø2a.
diego
parents: 10322
diff changeset
19 * $Id$
07e7a572bd84 Mark modified imported files as such to comply with (L)GPL ¡ø2a.
diego
parents: 10322
diff changeset
20 *
10322
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
21 * Original disclaimer:
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
22 * The author of this program disclaim whole expressed or implied
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
23 * warranties with regard to this program, and in no event shall the
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
24 * author of this program liable to whatever resulted from the use of
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
25 * this program. Use it at your own risk.
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
26 *
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
27 * 2003/06/21: Moved to GCC inline assembly - Alex Beregszaszi
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
28 */
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
29
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
30 #define real float /* ugly - but only way */
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
31
16989
e7a129082fda Unify include paths, -I.. is in CFLAGS.
diego
parents: 15167
diff changeset
32 #include "mangle.h"
10322
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
33
25325
7c7885350d89 Identifiers starting with __ are reserved for the system.
diego
parents: 18783
diff changeset
34 #ifdef DCT36_OPTIMIZE_FOR_K7
10322
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
35 void dct36_3dnowex(real *inbuf, real *o1,
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
36 real *o2, real *wintab, real *tsbuf)
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
37 #else
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
38 void dct36_3dnow(real *inbuf, real *o1,
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
39 real *o2, real *wintab, real *tsbuf)
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
40 #endif
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
41 {
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
42 __asm__ __volatile__(
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
43 "movq (%%eax),%%mm0\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
44 "movq 4(%%eax),%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
45 "pfadd %%mm1,%%mm0\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
46 "movq %%mm0,4(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
47 "psrlq $32,%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
48 "movq 12(%%eax),%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
49 "punpckldq %%mm2,%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
50 "pfadd %%mm2,%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
51 "movq %%mm1,12(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
52 "psrlq $32,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
53 "movq 20(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
54 "punpckldq %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
55 "pfadd %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
56 "movq %%mm2,20(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
57 "psrlq $32,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
58 "movq 28(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
59 "punpckldq %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
60 "pfadd %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
61 "movq %%mm3,28(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
62 "psrlq $32,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
63 "movq 36(%%eax),%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
64 "punpckldq %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
65 "pfadd %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
66 "movq %%mm4,36(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
67 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
68 "movq 44(%%eax),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
69 "punpckldq %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
70 "pfadd %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
71 "movq %%mm5,44(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
72 "psrlq $32,%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
73 "movq 52(%%eax),%%mm7\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
74 "punpckldq %%mm7,%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
75 "pfadd %%mm7,%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
76 "movq %%mm6,52(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
77 "psrlq $32,%%mm7\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
78 "movq 60(%%eax),%%mm0\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
79 "punpckldq %%mm0,%%mm7\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
80 "pfadd %%mm0,%%mm7\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
81 "movq %%mm7,60(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
82 "psrlq $32,%%mm0\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
83 "movd 68(%%eax),%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
84 "pfadd %%mm1,%%mm0\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
85 "movd %%mm0,68(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
86 "movd 4(%%eax),%%mm0\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
87 "movd 12(%%eax),%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
88 "punpckldq %%mm1,%%mm0\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
89 "punpckldq 20(%%eax),%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
90 "pfadd %%mm1,%%mm0\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
91 "movd %%mm0,12(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
92 "psrlq $32,%%mm0\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
93 "movd %%mm0,20(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
94 "psrlq $32,%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
95 "movd 28(%%eax),%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
96 "punpckldq %%mm2,%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
97 "punpckldq 36(%%eax),%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
98 "pfadd %%mm2,%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
99 "movd %%mm1,28(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
100 "psrlq $32,%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
101 "movd %%mm1,36(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
102 "psrlq $32,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
103 "movd 44(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
104 "punpckldq %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
105 "punpckldq 52(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
106 "pfadd %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
107 "movd %%mm2,44(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
108 "psrlq $32,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
109 "movd %%mm2,52(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
110 "psrlq $32,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
111 "movd 60(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
112 "punpckldq %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
113 "punpckldq 68(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
114 "pfadd %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
115 "movd %%mm3,60(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
116 "psrlq $32,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
117 "movd %%mm3,68(%%eax)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
118
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
119 "movq 24(%%eax),%%mm0\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
120 "movq 48(%%eax),%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
121 "movd "MANGLE(COS9)"+12,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
122 "punpckldq %%mm2,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
123 "movd "MANGLE(COS9)"+24,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
124 "punpckldq %%mm3,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
125 "pfmul %%mm2,%%mm0\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
126 "pfmul %%mm3,%%mm1\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
127 "pushl %%eax\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
128 "movl $1,%%eax\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
129 "movd %%eax,%%mm7\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
130 "pi2fd %%mm7,%%mm7\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
131 "popl %%eax\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
132 "movq 8(%%eax),%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
133 "movd "MANGLE(COS9)"+4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
134 "punpckldq %%mm3,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
135 "pfmul %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
136 "pfadd %%mm0,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
137 "movq 40(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
138 "movd "MANGLE(COS9)"+20,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
139 "punpckldq %%mm4,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
140 "pfmul %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
141 "pfadd %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
142 "movq 56(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
143 "movd "MANGLE(COS9)"+28,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
144 "punpckldq %%mm4,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
145 "pfmul %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
146 "pfadd %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
147 "movq (%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
148 "movq 16(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
149 "movd "MANGLE(COS9)"+8,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
150 "punpckldq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
151 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
152 "pfadd %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
153 "movq 32(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
154 "movd "MANGLE(COS9)"+16,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
155 "punpckldq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
156 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
157 "pfadd %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
158 "pfadd %%mm1,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
159 "movq 64(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
160 "movd "MANGLE(COS9)"+32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
161 "punpckldq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
162 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
163 "pfadd %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
164 "movq %%mm2,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
165 "pfadd %%mm3,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
166 "movq %%mm7,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
167 "punpckldq "MANGLE(tfcos36)"+0,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
168 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
169 "movq %%mm4,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
170 "pfacc %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
171 "movd 108(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
172 "punpckldq 104(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
173 "pfmul %%mm6,%%mm5\n\t"
25325
7c7885350d89 Identifiers starting with __ are reserved for the system.
diego
parents: 18783
diff changeset
174 #ifdef DCT36_OPTIMIZE_FOR_K7
10322
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
175 "pswapd %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
176 "movq %%mm5,32(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
177 #else
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
178 "movd %%mm5,36(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
179 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
180 "movd %%mm5,32(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
181 #endif
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
182 "movq %%mm4,%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
183 "punpckldq %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
184 "pfsub %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
185 "punpckhdq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
186 "movd 32(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
187 "punpckldq 36(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
188 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
189 "movd 32(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
190 "punpckldq 36(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
191 "pfadd %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
192 "movd %%mm5,1024(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
193 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
194 "movd %%mm5,1152(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
195 "movq %%mm3,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
196 "pfsub %%mm2,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
197 "movq %%mm7,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
198 "punpckldq "MANGLE(tfcos36)"+32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
199 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
200 "movq %%mm4,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
201 "pfacc %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
202 "movd 140(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
203 "punpckldq 72(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
204 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
205 "movd %%mm5,68(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
206 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
207 "movd %%mm5,0(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
208 "movq %%mm4,%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
209 "punpckldq %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
210 "pfsub %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
211 "punpckhdq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
212 "movd 0(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
213 "punpckldq 68(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
214 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
215 "movd 0(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
216 "punpckldq 68(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
217 "pfadd %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
218 "movd %%mm5,0(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
219 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
220 "movd %%mm5,2176(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
221 "movq 8(%%eax),%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
222 "movq 40(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
223 "pfsub %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
224 "movq 56(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
225 "pfsub %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
226 "movd "MANGLE(COS9)"+12,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
227 "punpckldq %%mm3,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
228 "pfmul %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
229 "movq 16(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
230 "movq 32(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
231 "pfsub %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
232 "movq 64(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
233 "pfsub %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
234 "movd "MANGLE(COS9)"+24,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
235 "punpckldq %%mm4,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
236 "pfmul %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
237 "movq 48(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
238 "pfsub %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
239 "movq (%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
240 "pfadd %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
241 "movq %%mm2,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
242 "pfadd %%mm3,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
243 "movq %%mm7,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
244 "punpckldq "MANGLE(tfcos36)"+4,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
245 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
246 "movq %%mm4,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
247 "pfacc %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
248 "movd 112(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
249 "punpckldq 100(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
250 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
251 "movd %%mm5,40(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
252 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
253 "movd %%mm5,28(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
254 "movq %%mm4,%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
255 "punpckldq %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
256 "pfsub %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
257 "punpckhdq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
258 "movd 28(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
259 "punpckldq 40(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
260 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
261 "movd 28(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
262 "punpckldq 40(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
263 "pfadd %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
264 "movd %%mm5,896(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
265 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
266 "movd %%mm5,1280(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
267 "movq %%mm3,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
268 "pfsub %%mm2,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
269 "movq %%mm7,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
270 "punpckldq "MANGLE(tfcos36)"+28,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
271 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
272 "movq %%mm4,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
273 "pfacc %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
274 "movd 136(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
275 "punpckldq 76(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
276 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
277 "movd %%mm5,64(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
278 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
279 "movd %%mm5,4(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
280 "movq %%mm4,%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
281 "punpckldq %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
282 "pfsub %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
283 "punpckhdq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
284 "movd 4(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
285 "punpckldq 64(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
286 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
287 "movd 4(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
288 "punpckldq 64(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
289 "pfadd %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
290 "movd %%mm5,128(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
291 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
292 "movd %%mm5,2048(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
293
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
294 "movq 8(%%eax),%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
295 "movd "MANGLE(COS9)"+20,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
296 "punpckldq %%mm3,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
297 "pfmul %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
298 "pfsub %%mm0,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
299 "movq 40(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
300 "movd "MANGLE(COS9)"+28,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
301 "punpckldq %%mm4,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
302 "pfmul %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
303 "pfsub %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
304 "movq 56(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
305 "movd "MANGLE(COS9)"+4,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
306 "punpckldq %%mm4,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
307 "pfmul %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
308 "pfadd %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
309 "movq (%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
310 "movq 16(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
311 "movd "MANGLE(COS9)"+32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
312 "punpckldq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
313 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
314 "pfsub %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
315 "movq 32(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
316 "movd "MANGLE(COS9)"+8,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
317 "punpckldq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
318 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
319 "pfsub %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
320 "pfadd %%mm1,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
321 "movq 64(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
322 "movd "MANGLE(COS9)"+16,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
323 "punpckldq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
324 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
325 "pfadd %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
326 "movq %%mm2,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
327 "pfadd %%mm3,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
328 "movq %%mm7,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
329 "punpckldq "MANGLE(tfcos36)"+8,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
330 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
331 "movq %%mm4,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
332 "pfacc %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
333 "movd 116(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
334 "punpckldq 96(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
335 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
336 "movd %%mm5,44(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
337 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
338 "movd %%mm5,24(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
339 "movq %%mm4,%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
340 "punpckldq %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
341 "pfsub %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
342 "punpckhdq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
343 "movd 24(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
344 "punpckldq 44(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
345 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
346 "movd 24(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
347 "punpckldq 44(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
348 "pfadd %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
349 "movd %%mm5,768(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
350 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
351 "movd %%mm5,1408(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
352 "movq %%mm3,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
353 "pfsub %%mm2,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
354 "movq %%mm7,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
355 "punpckldq "MANGLE(tfcos36)"+24,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
356 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
357 "movq %%mm4,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
358 "pfacc %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
359 "movd 132(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
360 "punpckldq 80(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
361 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
362 "movd %%mm5,60(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
363 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
364 "movd %%mm5,8(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
365 "movq %%mm4,%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
366 "punpckldq %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
367 "pfsub %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
368 "punpckhdq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
369 "movd 8(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
370 "punpckldq 60(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
371 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
372 "movd 8(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
373 "punpckldq 60(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
374 "pfadd %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
375 "movd %%mm5,256(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
376 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
377 "movd %%mm5,1920(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
378 "movq 8(%%eax),%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
379 "movd "MANGLE(COS9)"+28,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
380 "punpckldq %%mm3,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
381 "pfmul %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
382 "pfsub %%mm0,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
383 "movq 40(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
384 "movd "MANGLE(COS9)"+4,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
385 "punpckldq %%mm4,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
386 "pfmul %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
387 "pfadd %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
388 "movq 56(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
389 "movd "MANGLE(COS9)"+20,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
390 "punpckldq %%mm4,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
391 "pfmul %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
392 "pfsub %%mm3,%%mm2\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
393 "movq (%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
394 "movq 16(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
395 "movd "MANGLE(COS9)"+16,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
396 "punpckldq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
397 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
398 "pfsub %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
399 "movq 32(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
400 "movd "MANGLE(COS9)"+32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
401 "punpckldq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
402 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
403 "pfadd %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
404 "pfadd %%mm1,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
405 "movq 64(%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
406 "movd "MANGLE(COS9)"+8,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
407 "punpckldq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
408 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
409 "pfsub %%mm4,%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
410 "movq %%mm2,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
411 "pfadd %%mm3,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
412 "movq %%mm7,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
413 "punpckldq "MANGLE(tfcos36)"+12,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
414 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
415 "movq %%mm4,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
416 "pfacc %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
417 "movd 120(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
418 "punpckldq 92(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
419 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
420 "movd %%mm5,48(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
421 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
422 "movd %%mm5,20(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
423 "movq %%mm4,%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
424 "punpckldq %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
425 "pfsub %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
426 "punpckhdq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
427 "movd 20(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
428 "punpckldq 48(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
429 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
430 "movd 20(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
431 "punpckldq 48(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
432 "pfadd %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
433 "movd %%mm5,640(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
434 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
435 "movd %%mm5,1536(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
436 "movq %%mm3,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
437 "pfsub %%mm2,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
438 "movq %%mm7,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
439 "punpckldq "MANGLE(tfcos36)"+20,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
440 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
441 "movq %%mm4,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
442 "pfacc %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
443 "movd 128(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
444 "punpckldq 84(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
445 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
446 "movd %%mm5,56(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
447 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
448 "movd %%mm5,12(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
449 "movq %%mm4,%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
450 "punpckldq %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
451 "pfsub %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
452 "punpckhdq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
453 "movd 12(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
454 "punpckldq 56(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
455 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
456 "movd 12(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
457 "punpckldq 56(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
458 "pfadd %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
459 "movd %%mm5,384(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
460 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
461 "movd %%mm5,1792(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
462
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
463 "movq (%%eax),%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
464 "movq 16(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
465 "pfsub %%mm3,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
466 "movq 32(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
467 "pfadd %%mm3,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
468 "movq 48(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
469 "pfsub %%mm3,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
470 "movq 64(%%eax),%%mm3\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
471 "pfadd %%mm3,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
472 "movq %%mm7,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
473 "punpckldq "MANGLE(tfcos36)"+16,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
474 "pfmul %%mm5,%%mm4\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
475 "movq %%mm4,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
476 "pfacc %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
477 "movd 124(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
478 "punpckldq 88(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
479 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
480 "movd %%mm5,52(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
481 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
482 "movd %%mm5,16(%%ecx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
483 "movq %%mm4,%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
484 "punpckldq %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
485 "pfsub %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
486 "punpckhdq %%mm5,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
487 "movd 16(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
488 "punpckldq 52(%%edx),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
489 "pfmul %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
490 "movd 16(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
491 "punpckldq 52(%%esi),%%mm6\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
492 "pfadd %%mm6,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
493 "movd %%mm5,512(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
494 "psrlq $32,%%mm5\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
495 "movd %%mm5,1664(%%ebx)\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
496
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
497 "femms\n\t"
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
498 :
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
499 : "a" (inbuf), "S" (o1), "c" (o2), "d" (wintab), "b" (tsbuf)
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
500 : "memory");
9163bdb578a6 moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff changeset
501 }