annotate i386/fft_3dn2.c @ 7047:4b5b586a3dd0 libavcodec

Support decoding of sample_g726.asf.
author michael
date Mon, 16 Jun 2008 23:12:50 +0000
parents f7cbb7733146
children fc843d00867c
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
1 /*
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
2 * FFT/MDCT transform with Extended 3DNow! optimizations
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
3 * Copyright (c) 2006 Zuxy MENG Jie, Loren Merritt
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
4 * Based on fft_sse.c copyright (c) 2002 Fabrice Bellard.
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
5 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3748
diff changeset
6 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3748
diff changeset
7 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3748
diff changeset
8 * FFmpeg is free software; you can redistribute it and/or
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
9 * modify it under the terms of the GNU Lesser General Public
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
10 * License as published by the Free Software Foundation; either
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3748
diff changeset
11 * version 2.1 of the License, or (at your option) any later version.
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
12 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3748
diff changeset
13 * FFmpeg is distributed in the hope that it will be useful,
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
16 * Lesser General Public License for more details.
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
17 *
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
18 * You should have received a copy of the GNU Lesser General Public
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3748
diff changeset
19 * License along with FFmpeg; if not, write to the Free Software
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
21 */
6763
f7cbb7733146 Use full path for #includes from another directory.
diego
parents: 6755
diff changeset
22
f7cbb7733146 Use full path for #includes from another directory.
diego
parents: 6755
diff changeset
23 #include "libavutil/x86_cpu.h"
f7cbb7733146 Use full path for #includes from another directory.
diego
parents: 6755
diff changeset
24 #include "libavcodec/dsputil.h"
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
25
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
26 static const int p1m1[2] __attribute__((aligned(8))) =
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
27 { 0, 1 << 31 };
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
28
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
29 static const int m1p1[2] __attribute__((aligned(8))) =
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
30 { 1 << 31, 0 };
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
31
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
32 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z)
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
33 {
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
34 int ln = s->nbits;
6755
33896780c612 Do not misuse long as the size of a register in x86.
ramiro
parents: 5010
diff changeset
35 long j;
33896780c612 Do not misuse long as the size of a register in x86.
ramiro
parents: 5010
diff changeset
36 x86_reg i;
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
37 long nblocks, nloops;
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
38 FFTComplex *p, *cptr;
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
39
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
40 asm volatile(
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
41 /* FEMMS is not a must here but recommended by AMD */
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
42 "femms \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
43 "movq %0, %%mm7 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
44 ::"m"(*(s->inverse ? m1p1 : p1m1))
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
45 );
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
46
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
47 i = 8 << ln;
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
48 asm volatile(
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
49 "1: \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
50 "sub $32, %0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
51 "movq (%0,%1), %%mm0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
52 "movq 16(%0,%1), %%mm1 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
53 "movq 8(%0,%1), %%mm2 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
54 "movq 24(%0,%1), %%mm3 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
55 "movq %%mm0, %%mm4 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
56 "movq %%mm1, %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
57 "pfadd %%mm2, %%mm0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
58 "pfadd %%mm3, %%mm1 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
59 "pfsub %%mm2, %%mm4 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
60 "pfsub %%mm3, %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
61 "movq %%mm0, %%mm2 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
62 "pswapd %%mm5, %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
63 "movq %%mm4, %%mm3 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
64 "pxor %%mm7, %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
65 "pfadd %%mm1, %%mm0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
66 "pfadd %%mm5, %%mm4 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
67 "pfsub %%mm1, %%mm2 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
68 "pfsub %%mm5, %%mm3 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
69 "movq %%mm0, (%0,%1) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
70 "movq %%mm4, 8(%0,%1) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
71 "movq %%mm2, 16(%0,%1) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
72 "movq %%mm3, 24(%0,%1) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
73 "jg 1b \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
74 :"+r"(i)
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
75 :"r"(z)
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
76 );
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
77 /* pass 2 .. ln-1 */
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
78
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
79 nblocks = 1 << (ln-3);
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
80 nloops = 1 << 2;
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
81 cptr = s->exptab1;
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
82 do {
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
83 p = z;
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
84 j = nblocks;
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
85 do {
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
86 i = nloops*8;
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
87 asm volatile(
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
88 "1: \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
89 "sub $16, %0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
90 "movq (%1,%0), %%mm0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
91 "movq 8(%1,%0), %%mm1 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
92 "movq (%2,%0), %%mm2 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
93 "movq 8(%2,%0), %%mm3 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
94 "movq (%3,%0,2), %%mm4 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
95 "movq 8(%3,%0,2), %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
96 "pswapd %%mm4, %%mm6 \n\t" // no need for cptr[2] & cptr[3]
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
97 "pswapd %%mm5, %%mm7 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
98 "pfmul %%mm2, %%mm4 \n\t" // cre*re cim*im
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
99 "pfmul %%mm3, %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
100 "pfmul %%mm2, %%mm6 \n\t" // cim*re cre*im
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
101 "pfmul %%mm3, %%mm7 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
102 "pfpnacc %%mm6, %%mm4 \n\t" // cre*re-cim*im cim*re+cre*im
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
103 "pfpnacc %%mm7, %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
104 "movq %%mm0, %%mm2 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
105 "movq %%mm1, %%mm3 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
106 "pfadd %%mm4, %%mm0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
107 "pfadd %%mm5, %%mm1 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
108 "pfsub %%mm4, %%mm2 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
109 "pfsub %%mm5, %%mm3 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
110 "movq %%mm0, (%1,%0) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
111 "movq %%mm1, 8(%1,%0) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
112 "movq %%mm2, (%2,%0) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
113 "movq %%mm3, 8(%2,%0) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
114 "jg 1b \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
115 :"+r"(i)
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
116 :"r"(p), "r"(p + nloops), "r"(cptr)
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
117 );
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
118 p += nloops*2;
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
119 } while (--j);
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
120 cptr += nloops*2;
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
121 nblocks >>= 1;
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
122 nloops <<= 1;
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
123 } while (nblocks != 0);
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
124 asm volatile("femms");
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
125 }
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
126
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
127 void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output,
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
128 const FFTSample *input, FFTSample *tmp)
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
129 {
6755
33896780c612 Do not misuse long as the size of a register in x86.
ramiro
parents: 5010
diff changeset
130 long n8, n4, n2, n;
33896780c612 Do not misuse long as the size of a register in x86.
ramiro
parents: 5010
diff changeset
131 x86_reg k;
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
132 const uint16_t *revtab = s->fft.revtab;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
133 const FFTSample *tcos = s->tcos;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
134 const FFTSample *tsin = s->tsin;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
135 const FFTSample *in1, *in2;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
136 FFTComplex *z = (FFTComplex *)tmp;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
137
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
138 n = 1 << s->nbits;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
139 n2 = n >> 1;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
140 n4 = n >> 2;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
141 n8 = n >> 3;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
142
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
143 /* pre rotation */
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
144 in1 = input;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
145 in2 = input + n2 - 1;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
146 for(k = 0; k < n4; k++) {
3560
f1a16d793fc5 gcc 2.95 and 3.4.x on x86 32bit without fomit-frame-pointer can't even find 5 registers for asm input.
lorenm
parents: 3559
diff changeset
147 // FIXME a single block is faster, but gcc 2.95 and 3.4.x on 32bit can't compile it
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
148 asm volatile(
3560
f1a16d793fc5 gcc 2.95 and 3.4.x on x86 32bit without fomit-frame-pointer can't even find 5 registers for asm input.
lorenm
parents: 3559
diff changeset
149 "movd %0, %%mm0 \n\t"
f1a16d793fc5 gcc 2.95 and 3.4.x on x86 32bit without fomit-frame-pointer can't even find 5 registers for asm input.
lorenm
parents: 3559
diff changeset
150 "movd %2, %%mm1 \n\t"
f1a16d793fc5 gcc 2.95 and 3.4.x on x86 32bit without fomit-frame-pointer can't even find 5 registers for asm input.
lorenm
parents: 3559
diff changeset
151 "punpckldq %1, %%mm0 \n\t"
f1a16d793fc5 gcc 2.95 and 3.4.x on x86 32bit without fomit-frame-pointer can't even find 5 registers for asm input.
lorenm
parents: 3559
diff changeset
152 "punpckldq %3, %%mm1 \n\t"
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
153 "movq %%mm0, %%mm2 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
154 "pfmul %%mm1, %%mm0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
155 "pswapd %%mm1, %%mm1 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
156 "pfmul %%mm1, %%mm2 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
157 "pfpnacc %%mm2, %%mm0 \n\t"
3560
f1a16d793fc5 gcc 2.95 and 3.4.x on x86 32bit without fomit-frame-pointer can't even find 5 registers for asm input.
lorenm
parents: 3559
diff changeset
158 ::"m"(in2[-2*k]), "m"(in1[2*k]),
f1a16d793fc5 gcc 2.95 and 3.4.x on x86 32bit without fomit-frame-pointer can't even find 5 registers for asm input.
lorenm
parents: 3559
diff changeset
159 "m"(tcos[k]), "m"(tsin[k])
f1a16d793fc5 gcc 2.95 and 3.4.x on x86 32bit without fomit-frame-pointer can't even find 5 registers for asm input.
lorenm
parents: 3559
diff changeset
160 );
f1a16d793fc5 gcc 2.95 and 3.4.x on x86 32bit without fomit-frame-pointer can't even find 5 registers for asm input.
lorenm
parents: 3559
diff changeset
161 asm volatile(
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
162 "movq %%mm0, %0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
163 :"=m"(z[revtab[k]])
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
164 );
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
165 }
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
166
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
167 ff_fft_calc(&s->fft, z);
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
168
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
169 /* post rotation + reordering */
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
170 for(k = 0; k < n4; k++) {
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
171 asm volatile(
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
172 "movq %0, %%mm0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
173 "movd %1, %%mm1 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
174 "punpckldq %2, %%mm1 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
175 "movq %%mm0, %%mm2 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
176 "pfmul %%mm1, %%mm0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
177 "pswapd %%mm1, %%mm1 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
178 "pfmul %%mm1, %%mm2 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
179 "pfpnacc %%mm2, %%mm0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
180 "movq %%mm0, %0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
181 :"+m"(z[k])
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
182 :"m"(tcos[k]), "m"(tsin[k])
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
183 );
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
184 }
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
185
3747
ba1f57431c85 tweak ff_imdct_calc_3dn2
lorenm
parents: 3590
diff changeset
186 k = n-8;
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
187 asm volatile("movd %0, %%mm7" ::"r"(1<<31));
3747
ba1f57431c85 tweak ff_imdct_calc_3dn2
lorenm
parents: 3590
diff changeset
188 asm volatile(
3748
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
189 "1: \n\t"
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
190 "movq (%4,%0), %%mm0 \n\t" // z[n8+k]
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
191 "neg %0 \n\t"
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
192 "pswapd -8(%4,%0), %%mm1 \n\t" // z[n8-1-k]
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
193 "movq %%mm0, %%mm2 \n\t"
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
194 "pxor %%mm7, %%mm2 \n\t"
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
195 "punpckldq %%mm1, %%mm2 \n\t"
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
196 "pswapd %%mm2, %%mm3 \n\t"
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
197 "punpckhdq %%mm1, %%mm0 \n\t"
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
198 "pswapd %%mm0, %%mm4 \n\t"
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
199 "pxor %%mm7, %%mm0 \n\t"
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
200 "pxor %%mm7, %%mm4 \n\t"
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
201 "movq %%mm3, -8(%3,%0) \n\t" // output[n-2-2*k] = { z[n8-1-k].im, -z[n8+k].re }
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
202 "movq %%mm4, -8(%2,%0) \n\t" // output[n2-2-2*k]= { -z[n8-1-k].re, z[n8+k].im }
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
203 "neg %0 \n\t"
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
204 "movq %%mm0, (%1,%0) \n\t" // output[2*k] = { -z[n8+k].im, z[n8-1-k].re }
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
205 "movq %%mm2, (%2,%0) \n\t" // output[n2+2*k] = { -z[n8+k].re, z[n8-1-k].im }
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
206 "sub $8, %0 \n\t"
82d22b7e80ff cosmetics (indentation)
lorenm
parents: 3747
diff changeset
207 "jge 1b \n\t"
3747
ba1f57431c85 tweak ff_imdct_calc_3dn2
lorenm
parents: 3590
diff changeset
208 :"+r"(k)
ba1f57431c85 tweak ff_imdct_calc_3dn2
lorenm
parents: 3590
diff changeset
209 :"r"(output), "r"(output+n2), "r"(output+n), "r"(z+n8)
ba1f57431c85 tweak ff_imdct_calc_3dn2
lorenm
parents: 3590
diff changeset
210 :"memory"
ba1f57431c85 tweak ff_imdct_calc_3dn2
lorenm
parents: 3590
diff changeset
211 );
3561
97325fecd35a emms -> femms
lorenm
parents: 3560
diff changeset
212 asm volatile("femms");
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
213 }
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
214