Mercurial > mplayer.hg
annotate mp3lib/dct36_3dnow.c @ 11251:731652167590
-vf fil
author | michael |
---|---|
date | Thu, 23 Oct 2003 21:52:44 +0000 |
parents | 9163bdb578a6 |
children | 07e7a572bd84 |
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 * |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
17 * Original disclaimer: |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
18 * 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
|
19 * 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
|
20 * 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
|
21 * this program. Use it at your own risk. |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
22 * |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
23 * 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
|
24 */ |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
25 |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
26 #define real float /* ugly - but only way */ |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
27 |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
28 #include "../mangle.h" |
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 #ifdef __DCT36_OPTIMIZE_FOR_K7 |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
31 void dct36_3dnowex(real *inbuf, real *o1, |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
32 real *o2, real *wintab, real *tsbuf) |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
33 #else |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
34 void dct36_3dnow(real *inbuf, real *o1, |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
35 real *o2, real *wintab, real *tsbuf) |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
36 #endif |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
37 { |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
38 __asm__ __volatile__( |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
39 "movq (%%eax),%%mm0\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
40 "movq 4(%%eax),%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
41 "pfadd %%mm1,%%mm0\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
42 "movq %%mm0,4(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
43 "psrlq $32,%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
44 "movq 12(%%eax),%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
45 "punpckldq %%mm2,%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
46 "pfadd %%mm2,%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
47 "movq %%mm1,12(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
48 "psrlq $32,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
49 "movq 20(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
50 "punpckldq %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
51 "pfadd %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
52 "movq %%mm2,20(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
53 "psrlq $32,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
54 "movq 28(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
55 "punpckldq %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
56 "pfadd %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
57 "movq %%mm3,28(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
58 "psrlq $32,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
59 "movq 36(%%eax),%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
60 "punpckldq %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
61 "pfadd %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
62 "movq %%mm4,36(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
63 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
64 "movq 44(%%eax),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
65 "punpckldq %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
66 "pfadd %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
67 "movq %%mm5,44(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
68 "psrlq $32,%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
69 "movq 52(%%eax),%%mm7\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
70 "punpckldq %%mm7,%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
71 "pfadd %%mm7,%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
72 "movq %%mm6,52(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
73 "psrlq $32,%%mm7\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
74 "movq 60(%%eax),%%mm0\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
75 "punpckldq %%mm0,%%mm7\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
76 "pfadd %%mm0,%%mm7\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
77 "movq %%mm7,60(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
78 "psrlq $32,%%mm0\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
79 "movd 68(%%eax),%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
80 "pfadd %%mm1,%%mm0\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
81 "movd %%mm0,68(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
82 "movd 4(%%eax),%%mm0\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
83 "movd 12(%%eax),%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
84 "punpckldq %%mm1,%%mm0\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
85 "punpckldq 20(%%eax),%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
86 "pfadd %%mm1,%%mm0\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
87 "movd %%mm0,12(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
88 "psrlq $32,%%mm0\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
89 "movd %%mm0,20(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
90 "psrlq $32,%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
91 "movd 28(%%eax),%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
92 "punpckldq %%mm2,%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
93 "punpckldq 36(%%eax),%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
94 "pfadd %%mm2,%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
95 "movd %%mm1,28(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
96 "psrlq $32,%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
97 "movd %%mm1,36(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
98 "psrlq $32,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
99 "movd 44(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
100 "punpckldq %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
101 "punpckldq 52(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
102 "pfadd %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
103 "movd %%mm2,44(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
104 "psrlq $32,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
105 "movd %%mm2,52(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
106 "psrlq $32,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
107 "movd 60(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
108 "punpckldq %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
109 "punpckldq 68(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
110 "pfadd %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
111 "movd %%mm3,60(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
112 "psrlq $32,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
113 "movd %%mm3,68(%%eax)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
114 |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
115 "movq 24(%%eax),%%mm0\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
116 "movq 48(%%eax),%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
117 "movd "MANGLE(COS9)"+12,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
118 "punpckldq %%mm2,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
119 "movd "MANGLE(COS9)"+24,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
120 "punpckldq %%mm3,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
121 "pfmul %%mm2,%%mm0\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
122 "pfmul %%mm3,%%mm1\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
123 "pushl %%eax\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
124 "movl $1,%%eax\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
125 "movd %%eax,%%mm7\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
126 "pi2fd %%mm7,%%mm7\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
127 "popl %%eax\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
128 "movq 8(%%eax),%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
129 "movd "MANGLE(COS9)"+4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
130 "punpckldq %%mm3,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
131 "pfmul %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
132 "pfadd %%mm0,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
133 "movq 40(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
134 "movd "MANGLE(COS9)"+20,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
135 "punpckldq %%mm4,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
136 "pfmul %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
137 "pfadd %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
138 "movq 56(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
139 "movd "MANGLE(COS9)"+28,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
140 "punpckldq %%mm4,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
141 "pfmul %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
142 "pfadd %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
143 "movq (%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
144 "movq 16(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
145 "movd "MANGLE(COS9)"+8,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
146 "punpckldq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
147 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
148 "pfadd %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
149 "movq 32(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
150 "movd "MANGLE(COS9)"+16,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
151 "punpckldq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
152 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
153 "pfadd %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
154 "pfadd %%mm1,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
155 "movq 64(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
156 "movd "MANGLE(COS9)"+32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
157 "punpckldq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
158 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
159 "pfadd %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
160 "movq %%mm2,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
161 "pfadd %%mm3,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
162 "movq %%mm7,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
163 "punpckldq "MANGLE(tfcos36)"+0,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
164 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
165 "movq %%mm4,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
166 "pfacc %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
167 "movd 108(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
168 "punpckldq 104(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
169 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
170 #ifdef __DCT36_OPTIMIZE_FOR_K7 |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
171 "pswapd %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
172 "movq %%mm5,32(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
173 #else |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
174 "movd %%mm5,36(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
175 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
176 "movd %%mm5,32(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
177 #endif |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
178 "movq %%mm4,%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
179 "punpckldq %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
180 "pfsub %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
181 "punpckhdq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
182 "movd 32(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
183 "punpckldq 36(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
184 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
185 "movd 32(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
186 "punpckldq 36(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
187 "pfadd %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
188 "movd %%mm5,1024(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
189 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
190 "movd %%mm5,1152(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
191 "movq %%mm3,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
192 "pfsub %%mm2,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
193 "movq %%mm7,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
194 "punpckldq "MANGLE(tfcos36)"+32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
195 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
196 "movq %%mm4,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
197 "pfacc %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
198 "movd 140(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
199 "punpckldq 72(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
200 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
201 "movd %%mm5,68(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
202 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
203 "movd %%mm5,0(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
204 "movq %%mm4,%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
205 "punpckldq %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
206 "pfsub %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
207 "punpckhdq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
208 "movd 0(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
209 "punpckldq 68(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
210 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
211 "movd 0(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
212 "punpckldq 68(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
213 "pfadd %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
214 "movd %%mm5,0(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
215 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
216 "movd %%mm5,2176(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
217 "movq 8(%%eax),%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
218 "movq 40(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
219 "pfsub %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
220 "movq 56(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
221 "pfsub %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
222 "movd "MANGLE(COS9)"+12,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
223 "punpckldq %%mm3,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
224 "pfmul %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
225 "movq 16(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
226 "movq 32(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
227 "pfsub %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
228 "movq 64(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
229 "pfsub %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
230 "movd "MANGLE(COS9)"+24,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
231 "punpckldq %%mm4,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
232 "pfmul %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
233 "movq 48(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
234 "pfsub %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
235 "movq (%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
236 "pfadd %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
237 "movq %%mm2,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
238 "pfadd %%mm3,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
239 "movq %%mm7,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
240 "punpckldq "MANGLE(tfcos36)"+4,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
241 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
242 "movq %%mm4,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
243 "pfacc %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
244 "movd 112(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
245 "punpckldq 100(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
246 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
247 "movd %%mm5,40(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
248 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
249 "movd %%mm5,28(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
250 "movq %%mm4,%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
251 "punpckldq %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
252 "pfsub %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
253 "punpckhdq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
254 "movd 28(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
255 "punpckldq 40(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
256 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
257 "movd 28(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
258 "punpckldq 40(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
259 "pfadd %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
260 "movd %%mm5,896(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
261 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
262 "movd %%mm5,1280(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
263 "movq %%mm3,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
264 "pfsub %%mm2,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
265 "movq %%mm7,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
266 "punpckldq "MANGLE(tfcos36)"+28,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
267 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
268 "movq %%mm4,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
269 "pfacc %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
270 "movd 136(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
271 "punpckldq 76(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
272 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
273 "movd %%mm5,64(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
274 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
275 "movd %%mm5,4(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
276 "movq %%mm4,%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
277 "punpckldq %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
278 "pfsub %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
279 "punpckhdq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
280 "movd 4(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
281 "punpckldq 64(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
282 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
283 "movd 4(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
284 "punpckldq 64(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
285 "pfadd %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
286 "movd %%mm5,128(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
287 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
288 "movd %%mm5,2048(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
289 |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
290 "movq 8(%%eax),%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
291 "movd "MANGLE(COS9)"+20,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
292 "punpckldq %%mm3,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
293 "pfmul %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
294 "pfsub %%mm0,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
295 "movq 40(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
296 "movd "MANGLE(COS9)"+28,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
297 "punpckldq %%mm4,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
298 "pfmul %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
299 "pfsub %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
300 "movq 56(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
301 "movd "MANGLE(COS9)"+4,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
302 "punpckldq %%mm4,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
303 "pfmul %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
304 "pfadd %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
305 "movq (%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
306 "movq 16(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
307 "movd "MANGLE(COS9)"+32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
308 "punpckldq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
309 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
310 "pfsub %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
311 "movq 32(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
312 "movd "MANGLE(COS9)"+8,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
313 "punpckldq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
314 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
315 "pfsub %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
316 "pfadd %%mm1,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
317 "movq 64(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
318 "movd "MANGLE(COS9)"+16,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
319 "punpckldq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
320 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
321 "pfadd %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
322 "movq %%mm2,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
323 "pfadd %%mm3,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
324 "movq %%mm7,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
325 "punpckldq "MANGLE(tfcos36)"+8,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
326 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
327 "movq %%mm4,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
328 "pfacc %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
329 "movd 116(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
330 "punpckldq 96(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
331 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
332 "movd %%mm5,44(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
333 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
334 "movd %%mm5,24(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
335 "movq %%mm4,%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
336 "punpckldq %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
337 "pfsub %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
338 "punpckhdq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
339 "movd 24(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
340 "punpckldq 44(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
341 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
342 "movd 24(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
343 "punpckldq 44(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
344 "pfadd %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
345 "movd %%mm5,768(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
346 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
347 "movd %%mm5,1408(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
348 "movq %%mm3,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
349 "pfsub %%mm2,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
350 "movq %%mm7,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
351 "punpckldq "MANGLE(tfcos36)"+24,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
352 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
353 "movq %%mm4,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
354 "pfacc %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
355 "movd 132(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
356 "punpckldq 80(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
357 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
358 "movd %%mm5,60(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
359 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
360 "movd %%mm5,8(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
361 "movq %%mm4,%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
362 "punpckldq %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
363 "pfsub %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
364 "punpckhdq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
365 "movd 8(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
366 "punpckldq 60(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
367 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
368 "movd 8(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
369 "punpckldq 60(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
370 "pfadd %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
371 "movd %%mm5,256(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
372 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
373 "movd %%mm5,1920(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
374 "movq 8(%%eax),%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
375 "movd "MANGLE(COS9)"+28,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
376 "punpckldq %%mm3,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
377 "pfmul %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
378 "pfsub %%mm0,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
379 "movq 40(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
380 "movd "MANGLE(COS9)"+4,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
381 "punpckldq %%mm4,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
382 "pfmul %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
383 "pfadd %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
384 "movq 56(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
385 "movd "MANGLE(COS9)"+20,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
386 "punpckldq %%mm4,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
387 "pfmul %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
388 "pfsub %%mm3,%%mm2\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
389 "movq (%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
390 "movq 16(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
391 "movd "MANGLE(COS9)"+16,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
392 "punpckldq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
393 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
394 "pfsub %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
395 "movq 32(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
396 "movd "MANGLE(COS9)"+32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
397 "punpckldq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
398 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
399 "pfadd %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
400 "pfadd %%mm1,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
401 "movq 64(%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
402 "movd "MANGLE(COS9)"+8,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
403 "punpckldq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
404 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
405 "pfsub %%mm4,%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
406 "movq %%mm2,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
407 "pfadd %%mm3,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
408 "movq %%mm7,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
409 "punpckldq "MANGLE(tfcos36)"+12,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
410 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
411 "movq %%mm4,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
412 "pfacc %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
413 "movd 120(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
414 "punpckldq 92(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
415 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
416 "movd %%mm5,48(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
417 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
418 "movd %%mm5,20(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
419 "movq %%mm4,%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
420 "punpckldq %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
421 "pfsub %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
422 "punpckhdq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
423 "movd 20(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
424 "punpckldq 48(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
425 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
426 "movd 20(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
427 "punpckldq 48(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
428 "pfadd %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
429 "movd %%mm5,640(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
430 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
431 "movd %%mm5,1536(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
432 "movq %%mm3,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
433 "pfsub %%mm2,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
434 "movq %%mm7,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
435 "punpckldq "MANGLE(tfcos36)"+20,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
436 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
437 "movq %%mm4,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
438 "pfacc %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
439 "movd 128(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
440 "punpckldq 84(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
441 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
442 "movd %%mm5,56(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
443 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
444 "movd %%mm5,12(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
445 "movq %%mm4,%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
446 "punpckldq %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
447 "pfsub %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
448 "punpckhdq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
449 "movd 12(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
450 "punpckldq 56(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
451 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
452 "movd 12(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
453 "punpckldq 56(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
454 "pfadd %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
455 "movd %%mm5,384(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
456 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
457 "movd %%mm5,1792(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
458 |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
459 "movq (%%eax),%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
460 "movq 16(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
461 "pfsub %%mm3,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
462 "movq 32(%%eax),%%mm3\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
463 "pfadd %%mm3,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
464 "movq 48(%%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 64(%%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 %%mm7,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
469 "punpckldq "MANGLE(tfcos36)"+16,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
470 "pfmul %%mm5,%%mm4\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
471 "movq %%mm4,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
472 "pfacc %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
473 "movd 124(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
474 "punpckldq 88(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
475 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
476 "movd %%mm5,52(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
477 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
478 "movd %%mm5,16(%%ecx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
479 "movq %%mm4,%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
480 "punpckldq %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
481 "pfsub %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
482 "punpckhdq %%mm5,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
483 "movd 16(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
484 "punpckldq 52(%%edx),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
485 "pfmul %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
486 "movd 16(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
487 "punpckldq 52(%%esi),%%mm6\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
488 "pfadd %%mm6,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
489 "movd %%mm5,512(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
490 "psrlq $32,%%mm5\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
491 "movd %%mm5,1664(%%ebx)\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
492 |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
493 "femms\n\t" |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
494 : |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
495 : "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
|
496 : "memory"); |
9163bdb578a6
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
alex
parents:
diff
changeset
|
497 } |