Mercurial > mplayer.hg
annotate mp3lib/dct36_3dnow.c @ 32282:606e4157cd4c
Split alloc and init of context so that parameters can be set in the context
instead of requireing being passed through function parameters. This also
makes sws work with AVOptions.
| author | michael |
|---|---|
| date | Sun, 26 Sep 2010 19:33:57 +0000 |
| parents | d0f70692a140 |
| children |
| 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 | 17 * Modified for use with MPlayer, for details see the changelog at |
| 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 |
|
28117
bd6833421e56
Consistently include config.h before mangle.h, fixes possible compilation
reimar
parents:
27757
diff
changeset
|
30 #include "config.h" |
| 16989 | 31 #include "mangle.h" |
|
30167
347d152a5cfa
Refactor real --> float #define to a typedef in a common header.
diego
parents:
28117
diff
changeset
|
32 #include "mpg123.h" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
33 #include "libavutil/x86_cpu.h" |
|
10322
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
34 |
|
25325
7c7885350d89
Identifiers starting with __ are reserved for the system.
diego
parents:
18783
diff
changeset
|
35 #ifdef DCT36_OPTIMIZE_FOR_K7 |
|
10322
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
36 void dct36_3dnowex(real *inbuf, real *o1, |
|
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
37 real *o2, real *wintab, real *tsbuf) |
|
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
38 #else |
|
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
39 void dct36_3dnow(real *inbuf, real *o1, |
|
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
40 real *o2, real *wintab, real *tsbuf) |
|
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
41 #endif |
|
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
42 { |
|
27757
b5a46071062a
Replace all occurrences of '__volatile__' and '__volatile' by plain 'volatile'.
diego
parents:
25325
diff
changeset
|
43 __asm__ volatile( |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
44 "movq (%%"REG_a"),%%mm0\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
45 "movq 4(%%"REG_a"),%%mm1\n\t" |
| 30990 | 46 "pfadd %%mm1,%%mm0\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
47 "movq %%mm0,4(%%"REG_a")\n\t" |
| 30990 | 48 "psrlq $32,%%mm1\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
49 "movq 12(%%"REG_a"),%%mm2\n\t" |
| 30990 | 50 "punpckldq %%mm2,%%mm1\n\t" |
| 51 "pfadd %%mm2,%%mm1\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
52 "movq %%mm1,12(%%"REG_a")\n\t" |
| 30990 | 53 "psrlq $32,%%mm2\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
54 "movq 20(%%"REG_a"),%%mm3\n\t" |
| 30990 | 55 "punpckldq %%mm3,%%mm2\n\t" |
| 56 "pfadd %%mm3,%%mm2\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
57 "movq %%mm2,20(%%"REG_a")\n\t" |
| 30990 | 58 "psrlq $32,%%mm3\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
59 "movq 28(%%"REG_a"),%%mm4\n\t" |
| 30990 | 60 "punpckldq %%mm4,%%mm3\n\t" |
| 61 "pfadd %%mm4,%%mm3\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
62 "movq %%mm3,28(%%"REG_a")\n\t" |
| 30990 | 63 "psrlq $32,%%mm4\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
64 "movq 36(%%"REG_a"),%%mm5\n\t" |
| 30990 | 65 "punpckldq %%mm5,%%mm4\n\t" |
| 66 "pfadd %%mm5,%%mm4\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
67 "movq %%mm4,36(%%"REG_a")\n\t" |
| 30990 | 68 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
69 "movq 44(%%"REG_a"),%%mm6\n\t" |
| 30990 | 70 "punpckldq %%mm6,%%mm5\n\t" |
| 71 "pfadd %%mm6,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
72 "movq %%mm5,44(%%"REG_a")\n\t" |
| 30990 | 73 "psrlq $32,%%mm6\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
74 "movq 52(%%"REG_a"),%%mm7\n\t" |
| 30990 | 75 "punpckldq %%mm7,%%mm6\n\t" |
| 76 "pfadd %%mm7,%%mm6\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
77 "movq %%mm6,52(%%"REG_a")\n\t" |
| 30990 | 78 "psrlq $32,%%mm7\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
79 "movq 60(%%"REG_a"),%%mm0\n\t" |
| 30990 | 80 "punpckldq %%mm0,%%mm7\n\t" |
| 81 "pfadd %%mm0,%%mm7\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
82 "movq %%mm7,60(%%"REG_a")\n\t" |
| 30990 | 83 "psrlq $32,%%mm0\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
84 "movd 68(%%"REG_a"),%%mm1\n\t" |
| 30990 | 85 "pfadd %%mm1,%%mm0\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
86 "movd %%mm0,68(%%"REG_a")\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
87 "movd 4(%%"REG_a"),%%mm0\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
88 "movd 12(%%"REG_a"),%%mm1\n\t" |
| 30990 | 89 "punpckldq %%mm1,%%mm0\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
90 "punpckldq 20(%%"REG_a"),%%mm1\n\t" |
| 30990 | 91 "pfadd %%mm1,%%mm0\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
92 "movd %%mm0,12(%%"REG_a")\n\t" |
| 30990 | 93 "psrlq $32,%%mm0\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
94 "movd %%mm0,20(%%"REG_a")\n\t" |
| 30990 | 95 "psrlq $32,%%mm1\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
96 "movd 28(%%"REG_a"),%%mm2\n\t" |
| 30990 | 97 "punpckldq %%mm2,%%mm1\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
98 "punpckldq 36(%%"REG_a"),%%mm2\n\t" |
| 30990 | 99 "pfadd %%mm2,%%mm1\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
100 "movd %%mm1,28(%%"REG_a")\n\t" |
| 30990 | 101 "psrlq $32,%%mm1\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
102 "movd %%mm1,36(%%"REG_a")\n\t" |
| 30990 | 103 "psrlq $32,%%mm2\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
104 "movd 44(%%"REG_a"),%%mm3\n\t" |
| 30990 | 105 "punpckldq %%mm3,%%mm2\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
106 "punpckldq 52(%%"REG_a"),%%mm3\n\t" |
| 30990 | 107 "pfadd %%mm3,%%mm2\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
108 "movd %%mm2,44(%%"REG_a")\n\t" |
| 30990 | 109 "psrlq $32,%%mm2\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
110 "movd %%mm2,52(%%"REG_a")\n\t" |
| 30990 | 111 "psrlq $32,%%mm3\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
112 "movd 60(%%"REG_a"),%%mm4\n\t" |
| 30990 | 113 "punpckldq %%mm4,%%mm3\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
114 "punpckldq 68(%%"REG_a"),%%mm4\n\t" |
| 30990 | 115 "pfadd %%mm4,%%mm3\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
116 "movd %%mm3,60(%%"REG_a")\n\t" |
| 30990 | 117 "psrlq $32,%%mm3\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
118 "movd %%mm3,68(%%"REG_a")\n\t" |
|
10322
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
119 |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
120 "movq 24(%%"REG_a"),%%mm0\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
121 "movq 48(%%"REG_a"),%%mm1\n\t" |
| 30990 | 122 "movd "MANGLE(COS9)"+12,%%mm2\n\t" |
| 123 "punpckldq %%mm2,%%mm2\n\t" | |
| 124 "movd "MANGLE(COS9)"+24,%%mm3\n\t" | |
| 125 "punpckldq %%mm3,%%mm3\n\t" | |
| 126 "pfmul %%mm2,%%mm0\n\t" | |
| 127 "pfmul %%mm3,%%mm1\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
128 "push %%"REG_a"\n\t" |
| 30990 | 129 "movl $1,%%eax\n\t" |
| 130 "movd %%eax,%%mm7\n\t" | |
| 131 "pi2fd %%mm7,%%mm7\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
132 "pop %%"REG_a"\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
133 "movq 8(%%"REG_a"),%%mm2\n\t" |
| 30990 | 134 "movd "MANGLE(COS9)"+4,%%mm3\n\t" |
| 135 "punpckldq %%mm3,%%mm3\n\t" | |
| 136 "pfmul %%mm3,%%mm2\n\t" | |
| 137 "pfadd %%mm0,%%mm2\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
138 "movq 40(%%"REG_a"),%%mm3\n\t" |
| 30990 | 139 "movd "MANGLE(COS9)"+20,%%mm4\n\t" |
| 140 "punpckldq %%mm4,%%mm4\n\t" | |
| 141 "pfmul %%mm4,%%mm3\n\t" | |
| 142 "pfadd %%mm3,%%mm2\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
143 "movq 56(%%"REG_a"),%%mm3\n\t" |
| 30990 | 144 "movd "MANGLE(COS9)"+28,%%mm4\n\t" |
| 145 "punpckldq %%mm4,%%mm4\n\t" | |
| 146 "pfmul %%mm4,%%mm3\n\t" | |
| 147 "pfadd %%mm3,%%mm2\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
148 "movq (%%"REG_a"),%%mm3\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
149 "movq 16(%%"REG_a"),%%mm4\n\t" |
| 30990 | 150 "movd "MANGLE(COS9)"+8,%%mm5\n\t" |
| 151 "punpckldq %%mm5,%%mm5\n\t" | |
| 152 "pfmul %%mm5,%%mm4\n\t" | |
| 153 "pfadd %%mm4,%%mm3\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
154 "movq 32(%%"REG_a"),%%mm4\n\t" |
| 30990 | 155 "movd "MANGLE(COS9)"+16,%%mm5\n\t" |
| 156 "punpckldq %%mm5,%%mm5\n\t" | |
| 157 "pfmul %%mm5,%%mm4\n\t" | |
| 158 "pfadd %%mm4,%%mm3\n\t" | |
| 159 "pfadd %%mm1,%%mm3\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
160 "movq 64(%%"REG_a"),%%mm4\n\t" |
| 30990 | 161 "movd "MANGLE(COS9)"+32,%%mm5\n\t" |
| 162 "punpckldq %%mm5,%%mm5\n\t" | |
| 163 "pfmul %%mm5,%%mm4\n\t" | |
| 164 "pfadd %%mm4,%%mm3\n\t" | |
| 165 "movq %%mm2,%%mm4\n\t" | |
| 166 "pfadd %%mm3,%%mm4\n\t" | |
| 167 "movq %%mm7,%%mm5\n\t" | |
| 168 "punpckldq "MANGLE(tfcos36)"+0,%%mm5\n\t" | |
| 169 "pfmul %%mm5,%%mm4\n\t" | |
| 170 "movq %%mm4,%%mm5\n\t" | |
| 171 "pfacc %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
172 "movd 108(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
173 "punpckldq 104(%%"REG_d"),%%mm6\n\t" |
| 30990 | 174 "pfmul %%mm6,%%mm5\n\t" |
|
25325
7c7885350d89
Identifiers starting with __ are reserved for the system.
diego
parents:
18783
diff
changeset
|
175 #ifdef DCT36_OPTIMIZE_FOR_K7 |
| 30990 | 176 "pswapd %%mm5,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
177 "movq %%mm5,32(%%"REG_c")\n\t" |
|
10322
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
178 #else |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
179 "movd %%mm5,36(%%"REG_c")\n\t" |
| 30990 | 180 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
181 "movd %%mm5,32(%%"REG_c")\n\t" |
|
10322
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
182 #endif |
| 30990 | 183 "movq %%mm4,%%mm6\n\t" |
| 184 "punpckldq %%mm6,%%mm5\n\t" | |
| 185 "pfsub %%mm6,%%mm5\n\t" | |
| 186 "punpckhdq %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
187 "movd 32(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
188 "punpckldq 36(%%"REG_d"),%%mm6\n\t" |
| 30990 | 189 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
190 "movd 32(%%"REG_S"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
191 "punpckldq 36(%%"REG_S"),%%mm6\n\t" |
| 30990 | 192 "pfadd %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
193 "movd %%mm5,1024(%%"REG_D")\n\t" |
| 30990 | 194 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
195 "movd %%mm5,1152(%%"REG_D")\n\t" |
| 30990 | 196 "movq %%mm3,%%mm4\n\t" |
| 197 "pfsub %%mm2,%%mm4\n\t" | |
| 198 "movq %%mm7,%%mm5\n\t" | |
| 199 "punpckldq "MANGLE(tfcos36)"+32,%%mm5\n\t" | |
| 200 "pfmul %%mm5,%%mm4\n\t" | |
| 201 "movq %%mm4,%%mm5\n\t" | |
| 202 "pfacc %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
203 "movd 140(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
204 "punpckldq 72(%%"REG_d"),%%mm6\n\t" |
| 30990 | 205 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
206 "movd %%mm5,68(%%"REG_c")\n\t" |
| 30990 | 207 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
208 "movd %%mm5,0(%%"REG_c")\n\t" |
| 30990 | 209 "movq %%mm4,%%mm6\n\t" |
| 210 "punpckldq %%mm6,%%mm5\n\t" | |
| 211 "pfsub %%mm6,%%mm5\n\t" | |
| 212 "punpckhdq %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
213 "movd 0(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
214 "punpckldq 68(%%"REG_d"),%%mm6\n\t" |
| 30990 | 215 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
216 "movd 0(%%"REG_S"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
217 "punpckldq 68(%%"REG_S"),%%mm6\n\t" |
| 30990 | 218 "pfadd %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
219 "movd %%mm5,0(%%"REG_D")\n\t" |
| 30990 | 220 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
221 "movd %%mm5,2176(%%"REG_D")\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
222 "movq 8(%%"REG_a"),%%mm2\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
223 "movq 40(%%"REG_a"),%%mm3\n\t" |
| 30990 | 224 "pfsub %%mm3,%%mm2\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
225 "movq 56(%%"REG_a"),%%mm3\n\t" |
| 30990 | 226 "pfsub %%mm3,%%mm2\n\t" |
| 227 "movd "MANGLE(COS9)"+12,%%mm3\n\t" | |
| 228 "punpckldq %%mm3,%%mm3\n\t" | |
| 229 "pfmul %%mm3,%%mm2\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
230 "movq 16(%%"REG_a"),%%mm3\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
231 "movq 32(%%"REG_a"),%%mm4\n\t" |
| 30990 | 232 "pfsub %%mm4,%%mm3\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
233 "movq 64(%%"REG_a"),%%mm4\n\t" |
| 30990 | 234 "pfsub %%mm4,%%mm3\n\t" |
| 235 "movd "MANGLE(COS9)"+24,%%mm4\n\t" | |
| 236 "punpckldq %%mm4,%%mm4\n\t" | |
| 237 "pfmul %%mm4,%%mm3\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
238 "movq 48(%%"REG_a"),%%mm4\n\t" |
| 30990 | 239 "pfsub %%mm4,%%mm3\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
240 "movq (%%"REG_a"),%%mm4\n\t" |
| 30990 | 241 "pfadd %%mm4,%%mm3\n\t" |
| 242 "movq %%mm2,%%mm4\n\t" | |
| 243 "pfadd %%mm3,%%mm4\n\t" | |
| 244 "movq %%mm7,%%mm5\n\t" | |
| 245 "punpckldq "MANGLE(tfcos36)"+4,%%mm5\n\t" | |
| 246 "pfmul %%mm5,%%mm4\n\t" | |
| 247 "movq %%mm4,%%mm5\n\t" | |
| 248 "pfacc %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
249 "movd 112(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
250 "punpckldq 100(%%"REG_d"),%%mm6\n\t" |
| 30990 | 251 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
252 "movd %%mm5,40(%%"REG_c")\n\t" |
| 30990 | 253 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
254 "movd %%mm5,28(%%"REG_c")\n\t" |
| 30990 | 255 "movq %%mm4,%%mm6\n\t" |
| 256 "punpckldq %%mm6,%%mm5\n\t" | |
| 257 "pfsub %%mm6,%%mm5\n\t" | |
| 258 "punpckhdq %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
259 "movd 28(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
260 "punpckldq 40(%%"REG_d"),%%mm6\n\t" |
| 30990 | 261 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
262 "movd 28(%%"REG_S"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
263 "punpckldq 40(%%"REG_S"),%%mm6\n\t" |
| 30990 | 264 "pfadd %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
265 "movd %%mm5,896(%%"REG_D")\n\t" |
| 30990 | 266 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
267 "movd %%mm5,1280(%%"REG_D")\n\t" |
| 30990 | 268 "movq %%mm3,%%mm4\n\t" |
| 269 "pfsub %%mm2,%%mm4\n\t" | |
| 270 "movq %%mm7,%%mm5\n\t" | |
| 271 "punpckldq "MANGLE(tfcos36)"+28,%%mm5\n\t" | |
| 272 "pfmul %%mm5,%%mm4\n\t" | |
| 273 "movq %%mm4,%%mm5\n\t" | |
| 274 "pfacc %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
275 "movd 136(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
276 "punpckldq 76(%%"REG_d"),%%mm6\n\t" |
| 30990 | 277 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
278 "movd %%mm5,64(%%"REG_c")\n\t" |
| 30990 | 279 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
280 "movd %%mm5,4(%%"REG_c")\n\t" |
| 30990 | 281 "movq %%mm4,%%mm6\n\t" |
| 282 "punpckldq %%mm6,%%mm5\n\t" | |
| 283 "pfsub %%mm6,%%mm5\n\t" | |
| 284 "punpckhdq %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
285 "movd 4(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
286 "punpckldq 64(%%"REG_d"),%%mm6\n\t" |
| 30990 | 287 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
288 "movd 4(%%"REG_S"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
289 "punpckldq 64(%%"REG_S"),%%mm6\n\t" |
| 30990 | 290 "pfadd %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
291 "movd %%mm5,128(%%"REG_D")\n\t" |
| 30990 | 292 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
293 "movd %%mm5,2048(%%"REG_D")\n\t" |
|
10322
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
294 |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
295 "movq 8(%%"REG_a"),%%mm2\n\t" |
| 30990 | 296 "movd "MANGLE(COS9)"+20,%%mm3\n\t" |
| 297 "punpckldq %%mm3,%%mm3\n\t" | |
| 298 "pfmul %%mm3,%%mm2\n\t" | |
| 299 "pfsub %%mm0,%%mm2\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
300 "movq 40(%%"REG_a"),%%mm3\n\t" |
| 30990 | 301 "movd "MANGLE(COS9)"+28,%%mm4\n\t" |
| 302 "punpckldq %%mm4,%%mm4\n\t" | |
| 303 "pfmul %%mm4,%%mm3\n\t" | |
| 304 "pfsub %%mm3,%%mm2\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
305 "movq 56(%%"REG_a"),%%mm3\n\t" |
| 30990 | 306 "movd "MANGLE(COS9)"+4,%%mm4\n\t" |
| 307 "punpckldq %%mm4,%%mm4\n\t" | |
| 308 "pfmul %%mm4,%%mm3\n\t" | |
| 309 "pfadd %%mm3,%%mm2\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
310 "movq (%%"REG_a"),%%mm3\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
311 "movq 16(%%"REG_a"),%%mm4\n\t" |
| 30990 | 312 "movd "MANGLE(COS9)"+32,%%mm5\n\t" |
| 313 "punpckldq %%mm5,%%mm5\n\t" | |
| 314 "pfmul %%mm5,%%mm4\n\t" | |
| 315 "pfsub %%mm4,%%mm3\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
316 "movq 32(%%"REG_a"),%%mm4\n\t" |
| 30990 | 317 "movd "MANGLE(COS9)"+8,%%mm5\n\t" |
| 318 "punpckldq %%mm5,%%mm5\n\t" | |
| 319 "pfmul %%mm5,%%mm4\n\t" | |
| 320 "pfsub %%mm4,%%mm3\n\t" | |
| 321 "pfadd %%mm1,%%mm3\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
322 "movq 64(%%"REG_a"),%%mm4\n\t" |
| 30990 | 323 "movd "MANGLE(COS9)"+16,%%mm5\n\t" |
| 324 "punpckldq %%mm5,%%mm5\n\t" | |
| 325 "pfmul %%mm5,%%mm4\n\t" | |
| 326 "pfadd %%mm4,%%mm3\n\t" | |
| 327 "movq %%mm2,%%mm4\n\t" | |
| 328 "pfadd %%mm3,%%mm4\n\t" | |
| 329 "movq %%mm7,%%mm5\n\t" | |
| 330 "punpckldq "MANGLE(tfcos36)"+8,%%mm5\n\t" | |
| 331 "pfmul %%mm5,%%mm4\n\t" | |
| 332 "movq %%mm4,%%mm5\n\t" | |
| 333 "pfacc %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
334 "movd 116(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
335 "punpckldq 96(%%"REG_d"),%%mm6\n\t" |
| 30990 | 336 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
337 "movd %%mm5,44(%%"REG_c")\n\t" |
| 30990 | 338 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
339 "movd %%mm5,24(%%"REG_c")\n\t" |
| 30990 | 340 "movq %%mm4,%%mm6\n\t" |
| 341 "punpckldq %%mm6,%%mm5\n\t" | |
| 342 "pfsub %%mm6,%%mm5\n\t" | |
| 343 "punpckhdq %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
344 "movd 24(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
345 "punpckldq 44(%%"REG_d"),%%mm6\n\t" |
| 30990 | 346 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
347 "movd 24(%%"REG_S"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
348 "punpckldq 44(%%"REG_S"),%%mm6\n\t" |
| 30990 | 349 "pfadd %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
350 "movd %%mm5,768(%%"REG_D")\n\t" |
| 30990 | 351 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
352 "movd %%mm5,1408(%%"REG_D")\n\t" |
| 30990 | 353 "movq %%mm3,%%mm4\n\t" |
| 354 "pfsub %%mm2,%%mm4\n\t" | |
| 355 "movq %%mm7,%%mm5\n\t" | |
| 356 "punpckldq "MANGLE(tfcos36)"+24,%%mm5\n\t" | |
| 357 "pfmul %%mm5,%%mm4\n\t" | |
| 358 "movq %%mm4,%%mm5\n\t" | |
| 359 "pfacc %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
360 "movd 132(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
361 "punpckldq 80(%%"REG_d"),%%mm6\n\t" |
| 30990 | 362 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
363 "movd %%mm5,60(%%"REG_c")\n\t" |
| 30990 | 364 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
365 "movd %%mm5,8(%%"REG_c")\n\t" |
| 30990 | 366 "movq %%mm4,%%mm6\n\t" |
| 367 "punpckldq %%mm6,%%mm5\n\t" | |
| 368 "pfsub %%mm6,%%mm5\n\t" | |
| 369 "punpckhdq %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
370 "movd 8(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
371 "punpckldq 60(%%"REG_d"),%%mm6\n\t" |
| 30990 | 372 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
373 "movd 8(%%"REG_S"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
374 "punpckldq 60(%%"REG_S"),%%mm6\n\t" |
| 30990 | 375 "pfadd %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
376 "movd %%mm5,256(%%"REG_D")\n\t" |
| 30990 | 377 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
378 "movd %%mm5,1920(%%"REG_D")\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
379 "movq 8(%%"REG_a"),%%mm2\n\t" |
| 30990 | 380 "movd "MANGLE(COS9)"+28,%%mm3\n\t" |
| 381 "punpckldq %%mm3,%%mm3\n\t" | |
| 382 "pfmul %%mm3,%%mm2\n\t" | |
| 383 "pfsub %%mm0,%%mm2\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
384 "movq 40(%%"REG_a"),%%mm3\n\t" |
| 30990 | 385 "movd "MANGLE(COS9)"+4,%%mm4\n\t" |
| 386 "punpckldq %%mm4,%%mm4\n\t" | |
| 387 "pfmul %%mm4,%%mm3\n\t" | |
| 388 "pfadd %%mm3,%%mm2\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
389 "movq 56(%%"REG_a"),%%mm3\n\t" |
| 30990 | 390 "movd "MANGLE(COS9)"+20,%%mm4\n\t" |
| 391 "punpckldq %%mm4,%%mm4\n\t" | |
| 392 "pfmul %%mm4,%%mm3\n\t" | |
| 393 "pfsub %%mm3,%%mm2\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
394 "movq (%%"REG_a"),%%mm3\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
395 "movq 16(%%"REG_a"),%%mm4\n\t" |
| 30990 | 396 "movd "MANGLE(COS9)"+16,%%mm5\n\t" |
| 397 "punpckldq %%mm5,%%mm5\n\t" | |
| 398 "pfmul %%mm5,%%mm4\n\t" | |
| 399 "pfsub %%mm4,%%mm3\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
400 "movq 32(%%"REG_a"),%%mm4\n\t" |
| 30990 | 401 "movd "MANGLE(COS9)"+32,%%mm5\n\t" |
| 402 "punpckldq %%mm5,%%mm5\n\t" | |
| 403 "pfmul %%mm5,%%mm4\n\t" | |
| 404 "pfadd %%mm4,%%mm3\n\t" | |
| 405 "pfadd %%mm1,%%mm3\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
406 "movq 64(%%"REG_a"),%%mm4\n\t" |
| 30990 | 407 "movd "MANGLE(COS9)"+8,%%mm5\n\t" |
| 408 "punpckldq %%mm5,%%mm5\n\t" | |
| 409 "pfmul %%mm5,%%mm4\n\t" | |
| 410 "pfsub %%mm4,%%mm3\n\t" | |
| 411 "movq %%mm2,%%mm4\n\t" | |
| 412 "pfadd %%mm3,%%mm4\n\t" | |
| 413 "movq %%mm7,%%mm5\n\t" | |
| 414 "punpckldq "MANGLE(tfcos36)"+12,%%mm5\n\t" | |
| 415 "pfmul %%mm5,%%mm4\n\t" | |
| 416 "movq %%mm4,%%mm5\n\t" | |
| 417 "pfacc %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
418 "movd 120(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
419 "punpckldq 92(%%"REG_d"),%%mm6\n\t" |
| 30990 | 420 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
421 "movd %%mm5,48(%%"REG_c")\n\t" |
| 30990 | 422 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
423 "movd %%mm5,20(%%"REG_c")\n\t" |
| 30990 | 424 "movq %%mm4,%%mm6\n\t" |
| 425 "punpckldq %%mm6,%%mm5\n\t" | |
| 426 "pfsub %%mm6,%%mm5\n\t" | |
| 427 "punpckhdq %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
428 "movd 20(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
429 "punpckldq 48(%%"REG_d"),%%mm6\n\t" |
| 30990 | 430 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
431 "movd 20(%%"REG_S"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
432 "punpckldq 48(%%"REG_S"),%%mm6\n\t" |
| 30990 | 433 "pfadd %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
434 "movd %%mm5,640(%%"REG_D")\n\t" |
| 30990 | 435 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
436 "movd %%mm5,1536(%%"REG_D")\n\t" |
| 30990 | 437 "movq %%mm3,%%mm4\n\t" |
| 438 "pfsub %%mm2,%%mm4\n\t" | |
| 439 "movq %%mm7,%%mm5\n\t" | |
| 440 "punpckldq "MANGLE(tfcos36)"+20,%%mm5\n\t" | |
| 441 "pfmul %%mm5,%%mm4\n\t" | |
| 442 "movq %%mm4,%%mm5\n\t" | |
| 443 "pfacc %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
444 "movd 128(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
445 "punpckldq 84(%%"REG_d"),%%mm6\n\t" |
| 30990 | 446 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
447 "movd %%mm5,56(%%"REG_c")\n\t" |
| 30990 | 448 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
449 "movd %%mm5,12(%%"REG_c")\n\t" |
| 30990 | 450 "movq %%mm4,%%mm6\n\t" |
| 451 "punpckldq %%mm6,%%mm5\n\t" | |
| 452 "pfsub %%mm6,%%mm5\n\t" | |
| 453 "punpckhdq %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
454 "movd 12(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
455 "punpckldq 56(%%"REG_d"),%%mm6\n\t" |
| 30990 | 456 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
457 "movd 12(%%"REG_S"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
458 "punpckldq 56(%%"REG_S"),%%mm6\n\t" |
| 30990 | 459 "pfadd %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
460 "movd %%mm5,384(%%"REG_D")\n\t" |
| 30990 | 461 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
462 "movd %%mm5,1792(%%"REG_D")\n\t" |
|
10322
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
463 |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
464 "movq (%%"REG_a"),%%mm4\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
465 "movq 16(%%"REG_a"),%%mm3\n\t" |
| 30990 | 466 "pfsub %%mm3,%%mm4\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
467 "movq 32(%%"REG_a"),%%mm3\n\t" |
| 30990 | 468 "pfadd %%mm3,%%mm4\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
469 "movq 48(%%"REG_a"),%%mm3\n\t" |
| 30990 | 470 "pfsub %%mm3,%%mm4\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
471 "movq 64(%%"REG_a"),%%mm3\n\t" |
| 30990 | 472 "pfadd %%mm3,%%mm4\n\t" |
| 473 "movq %%mm7,%%mm5\n\t" | |
| 474 "punpckldq "MANGLE(tfcos36)"+16,%%mm5\n\t" | |
| 475 "pfmul %%mm5,%%mm4\n\t" | |
| 476 "movq %%mm4,%%mm5\n\t" | |
| 477 "pfacc %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
478 "movd 124(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
479 "punpckldq 88(%%"REG_d"),%%mm6\n\t" |
| 30990 | 480 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
481 "movd %%mm5,52(%%"REG_c")\n\t" |
| 30990 | 482 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
483 "movd %%mm5,16(%%"REG_c")\n\t" |
| 30990 | 484 "movq %%mm4,%%mm6\n\t" |
| 485 "punpckldq %%mm6,%%mm5\n\t" | |
| 486 "pfsub %%mm6,%%mm5\n\t" | |
| 487 "punpckhdq %%mm5,%%mm5\n\t" | |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
488 "movd 16(%%"REG_d"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
489 "punpckldq 52(%%"REG_d"),%%mm6\n\t" |
| 30990 | 490 "pfmul %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
491 "movd 16(%%"REG_S"),%%mm6\n\t" |
|
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
492 "punpckldq 52(%%"REG_S"),%%mm6\n\t" |
| 30990 | 493 "pfadd %%mm6,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
494 "movd %%mm5,512(%%"REG_D")\n\t" |
| 30990 | 495 "psrlq $32,%%mm5\n\t" |
|
31215
d0f70692a140
Make 3dnow and MMX code compile for 64 bit (just stupid search-and-replace).
reimar
parents:
31206
diff
changeset
|
496 "movd %%mm5,1664(%%"REG_D")\n\t" |
|
10322
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
497 |
| 30990 | 498 "femms\n\t" |
| 499 : | |
| 31206 | 500 : "a" (inbuf), "S" (o1), "c" (o2), "d" (wintab), "D" (tsbuf) |
| 30990 | 501 : "memory"); |
|
10322
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
502 } |