annotate x86/fft_mmx.asm @ 12498:c997f09d1e10 libavcodec

Move hadamard_diff{,16}_{mmx,mmx2,sse2,ssse3}() from inline asm to yasm, which will hopefully solve the Win64/FATE failures caused by these functions.
author rbultje
date Fri, 17 Sep 2010 01:56:06 +0000
parents f61e22f8cf28
children 67e7e49058c2
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1 ;******************************************************************************
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2 ;* FFT transform with SSE/3DNow optimizations
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3 ;* Copyright (c) 2008 Loren Merritt
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
4 ;*
12188
6f064ab48463 more credits to D. J. Bernstein for fft
lorenm
parents: 11931
diff changeset
5 ;* This algorithm (though not any of the implementation details) is
6f064ab48463 more credits to D. J. Bernstein for fft
lorenm
parents: 11931
diff changeset
6 ;* based on libdjbfft by D. J. Bernstein.
6f064ab48463 more credits to D. J. Bernstein for fft
lorenm
parents: 11931
diff changeset
7 ;*
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
8 ;* This file is part of FFmpeg.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
9 ;*
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
10 ;* FFmpeg is free software; you can redistribute it and/or
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
11 ;* modify it under the terms of the GNU Lesser General Public
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
12 ;* License as published by the Free Software Foundation; either
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
13 ;* version 2.1 of the License, or (at your option) any later version.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
14 ;*
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
15 ;* FFmpeg is distributed in the hope that it will be useful,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
16 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
17 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
18 ;* Lesser General Public License for more details.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
19 ;*
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
20 ;* You should have received a copy of the GNU Lesser General Public
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
21 ;* License along with FFmpeg; if not, write to the Free Software
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
22 ;* 51, Inc., Foundation Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
23 ;******************************************************************************
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
24
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
25 ; These functions are not individually interchangeable with the C versions.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
26 ; While C takes arrays of FFTComplex, SSE/3DNow leave intermediate results
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
27 ; in blocks as conventient to the vector size.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
28 ; i.e. {4x real, 4x imaginary, 4x real, ...} (or 2x respectively)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
29
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
30 %include "x86inc.asm"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
31
12399
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
32 %ifdef ARCH_X86_64
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
33 %define pointer resq
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
34 %else
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
35 %define pointer resd
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
36 %endif
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
37
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
38 struc FFTContext
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
39 .nbits: resd 1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
40 .reverse: resd 1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
41 .revtab: pointer 1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
42 .tmpbuf: pointer 1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
43 .mdctsize: resd 1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
44 .mdctbits: resd 1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
45 .tcos: pointer 1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
46 .tsin: pointer 1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
47 endstruc
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
48
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
49 SECTION_RODATA
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
50
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
51 %define M_SQRT1_2 0.70710678118654752440
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
52 ps_root2: times 4 dd M_SQRT1_2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
53 ps_root2mppm: dd -M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, -M_SQRT1_2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
54 ps_m1p1: dd 1<<31, 0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
55
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
56 %assign i 16
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
57 %rep 13
11931
980030a3e315 Update x264asm header files to latest versions.
darkshikari
parents: 11060
diff changeset
58 cextern cos_ %+ i
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
59 %assign i i<<1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
60 %endrep
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
61
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
62 %ifdef ARCH_X86_64
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
63 %define pointer dq
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
64 %else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
65 %define pointer dd
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
66 %endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
67
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
68 %macro IF0 1+
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
69 %endmacro
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
70 %macro IF1 1+
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
71 %1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
72 %endmacro
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
73
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
74 section .text align=16
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
75
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
76 %macro T2_3DN 4 ; z0, z1, mem0, mem1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
77 mova %1, %3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
78 mova %2, %1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
79 pfadd %1, %4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
80 pfsub %2, %4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
81 %endmacro
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
82
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
83 %macro T4_3DN 6 ; z0, z1, z2, z3, tmp0, tmp1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
84 mova %5, %3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
85 pfsub %3, %4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
86 pfadd %5, %4 ; {t6,t5}
11931
980030a3e315 Update x264asm header files to latest versions.
darkshikari
parents: 11060
diff changeset
87 pxor %3, [ps_m1p1] ; {t8,t7}
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
88 mova %6, %1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
89 pswapd %3, %3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
90 pfadd %1, %5 ; {r0,i0}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
91 pfsub %6, %5 ; {r2,i2}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
92 mova %4, %2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
93 pfadd %2, %3 ; {r1,i1}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
94 pfsub %4, %3 ; {r3,i3}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
95 SWAP %3, %6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
96 %endmacro
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
97
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
98 ; in: %1={r0,i0,r1,i1} %2={r2,i2,r3,i3}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
99 ; out: %1={r0,r1,r2,r3} %2={i0,i1,i2,i3}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
100 %macro T4_SSE 3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
101 mova %3, %1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
102 shufps %1, %2, 0x64 ; {r0,i0,r3,i2}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
103 shufps %3, %2, 0xce ; {r1,i1,r2,i3}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
104 mova %2, %1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
105 addps %1, %3 ; {t1,t2,t6,t5}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
106 subps %2, %3 ; {t3,t4,t8,t7}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
107 mova %3, %1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
108 shufps %1, %2, 0x44 ; {t1,t2,t3,t4}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
109 shufps %3, %2, 0xbe ; {t6,t5,t7,t8}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
110 mova %2, %1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
111 addps %1, %3 ; {r0,i0,r1,i1}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
112 subps %2, %3 ; {r2,i2,r3,i3}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
113 mova %3, %1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
114 shufps %1, %2, 0x88 ; {r0,r1,r2,r3}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
115 shufps %3, %2, 0xdd ; {i0,i1,i2,i3}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
116 SWAP %2, %3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
117 %endmacro
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
118
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
119 %macro T8_SSE 6 ; r0,i0,r1,i1,t0,t1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
120 mova %5, %3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
121 shufps %3, %4, 0x44 ; {r4,i4,r6,i6}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
122 shufps %5, %4, 0xee ; {r5,i5,r7,i7}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
123 mova %6, %3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
124 subps %3, %5 ; {r5,i5,r7,i7}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
125 addps %6, %5 ; {t1,t2,t3,t4}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
126 mova %5, %3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
127 shufps %5, %5, 0xb1 ; {i5,r5,i7,r7}
11931
980030a3e315 Update x264asm header files to latest versions.
darkshikari
parents: 11060
diff changeset
128 mulps %3, [ps_root2mppm] ; {-r5,i5,r7,-i7}
980030a3e315 Update x264asm header files to latest versions.
darkshikari
parents: 11060
diff changeset
129 mulps %5, [ps_root2]
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
130 addps %3, %5 ; {t8,t7,ta,t9}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
131 mova %5, %6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
132 shufps %6, %3, 0x36 ; {t3,t2,t9,t8}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
133 shufps %5, %3, 0x9c ; {t1,t4,t7,ta}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
134 mova %3, %6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
135 addps %6, %5 ; {t1,t2,t9,ta}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
136 subps %3, %5 ; {t6,t5,tc,tb}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
137 mova %5, %6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
138 shufps %6, %3, 0xd8 ; {t1,t9,t5,tb}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
139 shufps %5, %3, 0x8d ; {t2,ta,t6,tc}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
140 mova %3, %1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
141 mova %4, %2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
142 addps %1, %6 ; {r0,r1,r2,r3}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
143 addps %2, %5 ; {i0,i1,i2,i3}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
144 subps %3, %6 ; {r4,r5,r6,r7}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
145 subps %4, %5 ; {i4,i5,i6,i7}
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
146 %endmacro
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
147
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
148 ; scheduled for cpu-bound sizes
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
149 %macro PASS_SMALL 3 ; (to load m4-m7), wre, wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
150 IF%1 mova m4, Z(4)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
151 IF%1 mova m5, Z(5)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
152 mova m0, %2 ; wre
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
153 mova m2, m4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
154 mova m1, %3 ; wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
155 mova m3, m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
156 mulps m2, m0 ; r2*wre
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
157 IF%1 mova m6, Z(6)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
158 mulps m3, m1 ; i2*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
159 IF%1 mova m7, Z(7)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
160 mulps m4, m1 ; r2*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
161 mulps m5, m0 ; i2*wre
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
162 addps m2, m3 ; r2*wre + i2*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
163 mova m3, m1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
164 mulps m1, m6 ; r3*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
165 subps m5, m4 ; i2*wre - r2*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
166 mova m4, m0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
167 mulps m3, m7 ; i3*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
168 mulps m4, m6 ; r3*wre
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
169 mulps m0, m7 ; i3*wre
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
170 subps m4, m3 ; r3*wre - i3*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
171 mova m3, Z(0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
172 addps m0, m1 ; i3*wre + r3*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
173 mova m1, m4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
174 addps m4, m2 ; t5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
175 subps m1, m2 ; t3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
176 subps m3, m4 ; r2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
177 addps m4, Z(0) ; r0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
178 mova m6, Z(2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
179 mova Z(4), m3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
180 mova Z(0), m4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
181 mova m3, m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
182 subps m5, m0 ; t4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
183 mova m4, m6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
184 subps m6, m5 ; r3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
185 addps m5, m4 ; r1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
186 mova Z(6), m6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
187 mova Z(2), m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
188 mova m2, Z(3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
189 addps m3, m0 ; t6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
190 subps m2, m1 ; i3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
191 mova m7, Z(1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
192 addps m1, Z(3) ; i1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
193 mova Z(7), m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
194 mova Z(3), m1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
195 mova m4, m7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
196 subps m7, m3 ; i2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
197 addps m3, m4 ; i0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
198 mova Z(5), m7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
199 mova Z(1), m3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
200 %endmacro
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
201
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
202 ; scheduled to avoid store->load aliasing
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
203 %macro PASS_BIG 1 ; (!interleave)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
204 mova m4, Z(4) ; r2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
205 mova m5, Z(5) ; i2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
206 mova m2, m4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
207 mova m0, [wq] ; wre
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
208 mova m3, m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
209 mova m1, [wq+o1q] ; wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
210 mulps m2, m0 ; r2*wre
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
211 mova m6, Z(6) ; r3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
212 mulps m3, m1 ; i2*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
213 mova m7, Z(7) ; i3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
214 mulps m4, m1 ; r2*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
215 mulps m5, m0 ; i2*wre
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
216 addps m2, m3 ; r2*wre + i2*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
217 mova m3, m1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
218 mulps m1, m6 ; r3*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
219 subps m5, m4 ; i2*wre - r2*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
220 mova m4, m0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
221 mulps m3, m7 ; i3*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
222 mulps m4, m6 ; r3*wre
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
223 mulps m0, m7 ; i3*wre
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
224 subps m4, m3 ; r3*wre - i3*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
225 mova m3, Z(0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
226 addps m0, m1 ; i3*wre + r3*wim
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
227 mova m1, m4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
228 addps m4, m2 ; t5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
229 subps m1, m2 ; t3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
230 subps m3, m4 ; r2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
231 addps m4, Z(0) ; r0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
232 mova m6, Z(2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
233 mova Z(4), m3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
234 mova Z(0), m4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
235 mova m3, m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
236 subps m5, m0 ; t4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
237 mova m4, m6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
238 subps m6, m5 ; r3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
239 addps m5, m4 ; r1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
240 IF%1 mova Z(6), m6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
241 IF%1 mova Z(2), m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
242 mova m2, Z(3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
243 addps m3, m0 ; t6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
244 subps m2, m1 ; i3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
245 mova m7, Z(1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
246 addps m1, Z(3) ; i1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
247 IF%1 mova Z(7), m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
248 IF%1 mova Z(3), m1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
249 mova m4, m7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
250 subps m7, m3 ; i2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
251 addps m3, m4 ; i0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
252 IF%1 mova Z(5), m7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
253 IF%1 mova Z(1), m3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
254 %if %1==0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
255 mova m4, m5 ; r1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
256 mova m0, m6 ; r3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
257 unpcklps m5, m1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
258 unpckhps m4, m1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
259 unpcklps m6, m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
260 unpckhps m0, m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
261 mova m1, Z(0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
262 mova m2, Z(4)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
263 mova Z(2), m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
264 mova Z(3), m4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
265 mova Z(6), m6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
266 mova Z(7), m0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
267 mova m5, m1 ; r0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
268 mova m4, m2 ; r2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
269 unpcklps m1, m3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
270 unpckhps m5, m3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
271 unpcklps m2, m7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
272 unpckhps m4, m7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
273 mova Z(0), m1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
274 mova Z(1), m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
275 mova Z(4), m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
276 mova Z(5), m4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
277 %endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
278 %endmacro
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
279
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
280 %macro PUNPCK 3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
281 mova %3, %1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
282 punpckldq %1, %2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
283 punpckhdq %3, %2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
284 %endmacro
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
285
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
286 INIT_XMM
10452
c6aa538c0bc3 s/movdqa/movaps/ in sse1 fft. (regression in r20293)
lorenm
parents: 10019
diff changeset
287 %define mova movaps
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
288
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
289 %define Z(x) [r0+mmsize*x]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
290
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
291 align 16
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
292 fft4_sse:
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
293 mova m0, Z(0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
294 mova m1, Z(1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
295 T4_SSE m0, m1, m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
296 mova Z(0), m0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
297 mova Z(1), m1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
298 ret
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
299
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
300 align 16
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
301 fft8_sse:
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
302 mova m0, Z(0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
303 mova m1, Z(1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
304 T4_SSE m0, m1, m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
305 mova m2, Z(2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
306 mova m3, Z(3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
307 T8_SSE m0, m1, m2, m3, m4, m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
308 mova Z(0), m0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
309 mova Z(1), m1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
310 mova Z(2), m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
311 mova Z(3), m3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
312 ret
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
313
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
314 align 16
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
315 fft16_sse:
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
316 mova m0, Z(0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
317 mova m1, Z(1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
318 T4_SSE m0, m1, m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
319 mova m2, Z(2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
320 mova m3, Z(3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
321 T8_SSE m0, m1, m2, m3, m4, m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
322 mova m4, Z(4)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
323 mova m5, Z(5)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
324 mova Z(0), m0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
325 mova Z(1), m1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
326 mova Z(2), m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
327 mova Z(3), m3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
328 T4_SSE m4, m5, m6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
329 mova m6, Z(6)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
330 mova m7, Z(7)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
331 T4_SSE m6, m7, m0
11931
980030a3e315 Update x264asm header files to latest versions.
darkshikari
parents: 11060
diff changeset
332 PASS_SMALL 0, [cos_16], [cos_16+16]
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
333 ret
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
334
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
335
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
336 INIT_MMX
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
337
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
338 %macro FFT48_3DN 1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
339 align 16
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
340 fft4%1:
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
341 T2_3DN m0, m1, Z(0), Z(1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
342 mova m2, Z(2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
343 mova m3, Z(3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
344 T4_3DN m0, m1, m2, m3, m4, m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
345 PUNPCK m0, m1, m4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
346 PUNPCK m2, m3, m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
347 mova Z(0), m0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
348 mova Z(1), m4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
349 mova Z(2), m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
350 mova Z(3), m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
351 ret
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
352
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
353 align 16
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
354 fft8%1:
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
355 T2_3DN m0, m1, Z(0), Z(1)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
356 mova m2, Z(2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
357 mova m3, Z(3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
358 T4_3DN m0, m1, m2, m3, m4, m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
359 mova Z(0), m0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
360 mova Z(2), m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
361 T2_3DN m4, m5, Z(4), Z(5)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
362 T2_3DN m6, m7, Z(6), Z(7)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
363 pswapd m0, m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
364 pswapd m2, m7
11931
980030a3e315 Update x264asm header files to latest versions.
darkshikari
parents: 11060
diff changeset
365 pxor m0, [ps_m1p1]
980030a3e315 Update x264asm header files to latest versions.
darkshikari
parents: 11060
diff changeset
366 pxor m2, [ps_m1p1]
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
367 pfsub m5, m0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
368 pfadd m7, m2
11931
980030a3e315 Update x264asm header files to latest versions.
darkshikari
parents: 11060
diff changeset
369 pfmul m5, [ps_root2]
980030a3e315 Update x264asm header files to latest versions.
darkshikari
parents: 11060
diff changeset
370 pfmul m7, [ps_root2]
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
371 T4_3DN m1, m3, m5, m7, m0, m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
372 mova Z(5), m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
373 mova Z(7), m7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
374 mova m0, Z(0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
375 mova m2, Z(2)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
376 T4_3DN m0, m2, m4, m6, m5, m7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
377 PUNPCK m0, m1, m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
378 PUNPCK m2, m3, m7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
379 mova Z(0), m0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
380 mova Z(1), m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
381 mova Z(2), m2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
382 mova Z(3), m7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
383 PUNPCK m4, Z(5), m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
384 PUNPCK m6, Z(7), m7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
385 mova Z(4), m4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
386 mova Z(5), m5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
387 mova Z(6), m6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
388 mova Z(7), m7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
389 ret
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
390 %endmacro
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
391
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
392 FFT48_3DN _3dn2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
393
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
394 %macro pswapd 2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
395 %ifidn %1, %2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
396 movd [r0+12], %1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
397 punpckhdq %1, [r0+8]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
398 %else
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
399 movq %1, %2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
400 psrlq %1, 32
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
401 punpckldq %1, %2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
402 %endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
403 %endmacro
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
404
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
405 FFT48_3DN _3dn
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
406
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
407
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
408 %define Z(x) [zq + o1q*(x&6)*((x/6)^1) + o3q*(x/6) + mmsize*(x&1)]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
409
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
410 %macro DECL_PASS 2+ ; name, payload
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
411 align 16
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
412 %1:
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
413 DEFINE_ARGS z, w, n, o1, o3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
414 lea o3q, [nq*3]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
415 lea o1q, [nq*8]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
416 shl o3q, 4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
417 .loop:
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
418 %2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
419 add zq, mmsize*2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
420 add wq, mmsize
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
421 sub nd, mmsize/8
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
422 jg .loop
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
423 rep ret
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
424 %endmacro
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
425
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
426 INIT_XMM
10452
c6aa538c0bc3 s/movdqa/movaps/ in sse1 fft. (regression in r20293)
lorenm
parents: 10019
diff changeset
427 %define mova movaps
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
428 DECL_PASS pass_sse, PASS_BIG 1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
429 DECL_PASS pass_interleave_sse, PASS_BIG 0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
431 INIT_MMX
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
432 %define mulps pfmul
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
433 %define addps pfadd
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
434 %define subps pfsub
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
435 %define unpcklps punpckldq
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
436 %define unpckhps punpckhdq
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
437 DECL_PASS pass_3dn, PASS_SMALL 1, [wq], [wq+o1q]
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
438 DECL_PASS pass_interleave_3dn, PASS_BIG 0
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
439 %define pass_3dn2 pass_3dn
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
440 %define pass_interleave_3dn2 pass_interleave_3dn
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
441
11060
daff45175333 Make the jump-table section-relative for x86_64 with PIC enabled.
reimar
parents: 10452
diff changeset
442 %ifdef PIC
daff45175333 Make the jump-table section-relative for x86_64 with PIC enabled.
reimar
parents: 10452
diff changeset
443 %define SECTION_REL - $$
daff45175333 Make the jump-table section-relative for x86_64 with PIC enabled.
reimar
parents: 10452
diff changeset
444 %else
daff45175333 Make the jump-table section-relative for x86_64 with PIC enabled.
reimar
parents: 10452
diff changeset
445 %define SECTION_REL
daff45175333 Make the jump-table section-relative for x86_64 with PIC enabled.
reimar
parents: 10452
diff changeset
446 %endif
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
447
12399
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
448 %macro FFT_DISPATCH 2; clobbers 5 GPRs, 8 XMMs
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
449 lea r2, [dispatch_tab%1]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
450 mov r2, [r2 + (%2q-2)*gprsize]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
451 %ifdef PIC
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
452 lea r3, [$$]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
453 add r2, r3
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
454 %endif
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
455 call r2
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
456 %endmacro ; FFT_DISPATCH
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
457
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
458 %macro DECL_FFT 2-3 ; nbits, cpu, suffix
11060
daff45175333 Make the jump-table section-relative for x86_64 with PIC enabled.
reimar
parents: 10452
diff changeset
459 %xdefine list_of_fft fft4%2 SECTION_REL, fft8%2 SECTION_REL
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
460 %if %1==5
11060
daff45175333 Make the jump-table section-relative for x86_64 with PIC enabled.
reimar
parents: 10452
diff changeset
461 %xdefine list_of_fft list_of_fft, fft16%2 SECTION_REL
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
462 %endif
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
463
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
464 %assign n 1<<%1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
465 %rep 17-%1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
466 %assign n2 n/2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
467 %assign n4 n/4
11060
daff45175333 Make the jump-table section-relative for x86_64 with PIC enabled.
reimar
parents: 10452
diff changeset
468 %xdefine list_of_fft list_of_fft, fft %+ n %+ %3%2 SECTION_REL
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
469
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
470 align 16
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
471 fft %+ n %+ %3%2:
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
472 call fft %+ n2 %+ %2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
473 add r0, n*4 - (n&(-2<<%1))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
474 call fft %+ n4 %+ %2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
475 add r0, n*2 - (n2&(-2<<%1))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
476 call fft %+ n4 %+ %2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
477 sub r0, n*6 + (n2&(-2<<%1))
11931
980030a3e315 Update x264asm header files to latest versions.
darkshikari
parents: 11060
diff changeset
478 lea r1, [cos_ %+ n]
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
479 mov r2d, n4/2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
480 jmp pass%3%2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
481
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
482 %assign n n*2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
483 %endrep
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
484 %undef n
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
485
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
486 align 8
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
487 dispatch_tab%3%2: pointer list_of_fft
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
488
8820
7400956a815d Put dispatch_tab in the rodata section for macho64.
astrange
parents: 8430
diff changeset
489 section .text
7400956a815d Put dispatch_tab in the rodata section for macho64.
astrange
parents: 8430
diff changeset
490
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
491 ; On x86_32, this function does the register saving and restoring for all of fft.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
492 ; The others pass args in registers and don't spill anything.
10019
c08ca946c80a Update x264 asm code to latest to add support for 64-bit Windows.
darkshikari
parents: 8820
diff changeset
493 cglobal fft_dispatch%3%2, 2,5,8, z, nbits
12399
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
494 FFT_DISPATCH %3%2, nbits
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
495 RET
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
496 %endmacro ; DECL_FFT
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
497
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
498 DECL_FFT 5, _sse
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
499 DECL_FFT 5, _sse, _interleave
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
500 DECL_FFT 4, _3dn
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
501 DECL_FFT 4, _3dn, _interleave
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
502 DECL_FFT 4, _3dn2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
503 DECL_FFT 4, _3dn2, _interleave
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
504
12399
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
505 INIT_XMM
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
506 %undef mulps
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
507 %undef addps
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
508 %undef subps
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
509 %undef unpcklps
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
510 %undef unpckhps
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
511
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
512 %macro PREROTATER 5 ;-2*k, 2*k, input+n4, tcos+n8, tsin+n8
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
513 movaps xmm0, [%3+%2*4]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
514 movaps xmm1, [%3+%1*4-0x10]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
515 movaps xmm2, xmm0
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
516 shufps xmm0, xmm1, 0x88
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
517 shufps xmm1, xmm2, 0x77
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
518 movlps xmm4, [%4+%2*2]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
519 movlps xmm5, [%5+%2*2+0x0]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
520 movhps xmm4, [%4+%1*2-0x8]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
521 movhps xmm5, [%5+%1*2-0x8]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
522 movaps xmm2, xmm0
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
523 movaps xmm3, xmm1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
524 mulps xmm0, xmm5
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
525 mulps xmm1, xmm4
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
526 mulps xmm2, xmm4
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
527 mulps xmm3, xmm5
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
528 subps xmm1, xmm0
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
529 addps xmm2, xmm3
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
530 movaps xmm0, xmm1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
531 unpcklps xmm1, xmm2
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
532 unpckhps xmm0, xmm2
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
533 %endmacro
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
534
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
535 %macro CMUL 6 ;j, xmm0, xmm1, 3, 4, 5
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
536 movaps xmm6, [%4+%1*2]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
537 movaps %2, [%4+%1*2+0x10]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
538 movaps %3, xmm6
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
539 movaps xmm7, %2
12432
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
540 mulps xmm6, [%5+%1]
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
541 mulps %2, [%6+%1]
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
542 mulps %3, [%6+%1]
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
543 mulps xmm7, [%5+%1]
12399
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
544 subps %2, xmm6
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
545 addps %3, xmm7
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
546 %endmacro
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
547
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
548 %macro POSROTATESHUF 5 ;j, k, z+n8, tcos+n8, tsin+n8
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
549 .post:
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
550 CMUL %1, xmm0, xmm1, %3, %4, %5
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
551 CMUL %2, xmm4, xmm5, %3, %4, %5
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
552 shufps xmm1, xmm1, 0x1b
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
553 shufps xmm5, xmm5, 0x1b
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
554 movaps xmm6, xmm4
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
555 unpckhps xmm4, xmm1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
556 unpcklps xmm6, xmm1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
557 movaps xmm2, xmm0
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
558 unpcklps xmm0, xmm5
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
559 unpckhps xmm2, xmm5
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
560 movaps [%3+%2*2], xmm6
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
561 movaps [%3+%2*2+0x10], xmm4
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
562 movaps [%3+%1*2], xmm0
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
563 movaps [%3+%1*2+0x10], xmm2
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
564 sub %2, 0x10
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
565 add %1, 0x10
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
566 jl .post
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
567 %endmacro
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
568
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
569 cglobal imdct_half_sse, 3,7,8; FFTContext *s, FFTSample *output, const FFTSample *input
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
570 %ifdef ARCH_X86_64
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
571 %define rrevtab r10
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
572 %define rtcos r11
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
573 %define rtsin r12
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
574 push r12
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
575 push r13
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
576 push r14
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
577 %else
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
578 %define rrevtab r6
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
579 %define rtsin r6
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
580 %define rtcos r5
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
581 %endif
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
582 mov r3d, [r0+FFTContext.mdctsize]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
583 add r2, r3
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
584 shr r3, 1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
585 mov rtcos, [r0+FFTContext.tcos]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
586 mov rtsin, [r0+FFTContext.tsin]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
587 add rtcos, r3
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
588 add rtsin, r3
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
589 %ifndef ARCH_X86_64
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
590 push rtcos
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
591 push rtsin
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
592 %endif
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
593 shr r3, 1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
594 mov rrevtab, [r0+FFTContext.revtab]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
595 add rrevtab, r3
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
596 %ifndef ARCH_X86_64
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
597 push rrevtab
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
598 %endif
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
599
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
600 sub r3, 4
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
601 %ifdef ARCH_X86_64
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
602 xor r4, r4
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
603 sub r4, r3
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
604 %endif
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
605 .pre:
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
606 %ifndef ARCH_X86_64
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
607 ;unspill
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
608 xor r4, r4
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
609 sub r4, r3
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
610 mov rtsin, [esp+4]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
611 mov rtcos, [esp+8]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
612 %endif
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
613
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
614 PREROTATER r4, r3, r2, rtcos, rtsin
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
615 %ifdef ARCH_X86_64
12432
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
616 movzx r5, word [rrevtab+r4-4]
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
617 movzx r6, word [rrevtab+r4-2]
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
618 movzx r13, word [rrevtab+r3]
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
619 movzx r14, word [rrevtab+r3+2]
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
620 movlps [r1+r5 *8], xmm0
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
621 movhps [r1+r6 *8], xmm0
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
622 movlps [r1+r13*8], xmm1
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
623 movhps [r1+r14*8], xmm1
12399
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
624 add r4, 4
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
625 %else
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
626 mov r6, [esp]
12432
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
627 movzx r5, word [r6+r4-4]
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
628 movzx r4, word [r6+r4-2]
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
629 movlps [r1+r5*8], xmm0
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
630 movhps [r1+r4*8], xmm0
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
631 movzx r5, word [r6+r3]
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
632 movzx r4, word [r6+r3+2]
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
633 movlps [r1+r5*8], xmm1
f61e22f8cf28 cosmetics in imdct_sse
lorenm
parents: 12399
diff changeset
634 movhps [r1+r4*8], xmm1
12399
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
635 %endif
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
636 sub r3, 4
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
637 jns .pre
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
638
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
639 mov r5, r0
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
640 mov r6, r1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
641 mov r0, r1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
642 mov r1d, [r5+FFTContext.nbits]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
643
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
644 FFT_DISPATCH _sse, r1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
645
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
646 mov r0d, [r5+FFTContext.mdctsize]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
647 add r6, r0
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
648 shr r0, 1
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
649 %ifndef ARCH_X86_64
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
650 %define rtcos r2
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
651 %define rtsin r3
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
652 mov rtcos, [esp+8]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
653 mov rtsin, [esp+4]
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
654 %endif
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
655 neg r0
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
656 mov r1, -16
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
657 sub r1, r0
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
658 POSROTATESHUF r0, r1, r6, rtcos, rtsin
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
659 %ifdef ARCH_X86_64
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
660 pop r14
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
661 pop r13
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
662 pop r12
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
663 %else
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
664 add esp, 12
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
665 %endif
020540442072 Convert ff_imdct_half_sse() to yasm.
alexc
parents: 12188
diff changeset
666 RET