annotate i386/fft_3dn2.c @ 3613:5cb45b2211a0 libavcodec

simplify
author michael
date Thu, 24 Aug 2006 08:39:07 +0000
parents a3d97c60ea07
children ba1f57431c85
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 *
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
6 * This library is free software; you can redistribute it and/or
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
7 * modify it under the terms of the GNU Lesser General Public
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
8 * License as published by the Free Software Foundation; either
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
9 * version 2 of the License, or (at your option) any later version.
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
10 *
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
11 * This library is distributed in the hope that it will be useful,
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
14 * Lesser General Public License for more details.
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
15 *
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
16 * You should have received a copy of the GNU Lesser General Public
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
17 * License along with this library; if not, write to the Free Software
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
19 */
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
20 #include "../dsputil.h"
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
21
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
22 static const int p1m1[2] __attribute__((aligned(8))) =
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
23 { 0, 1 << 31 };
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
24
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
25 static const int m1p1[2] __attribute__((aligned(8))) =
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
26 { 1 << 31, 0 };
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
27
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
28 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z)
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
29 {
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
30 int ln = s->nbits;
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
31 long i, j;
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
32 long nblocks, nloops;
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
33 FFTComplex *p, *cptr;
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
34
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
35 asm volatile(
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
36 /* 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
37 "femms \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
38 "movq %0, %%mm7 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
39 ::"m"(*(s->inverse ? m1p1 : p1m1))
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
40 );
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
41
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
42 i = 8 << ln;
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
43 asm volatile(
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
44 "1: \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
45 "sub $32, %0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
46 "movq (%0,%1), %%mm0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
47 "movq 16(%0,%1), %%mm1 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
48 "movq 8(%0,%1), %%mm2 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
49 "movq 24(%0,%1), %%mm3 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
50 "movq %%mm0, %%mm4 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
51 "movq %%mm1, %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
52 "pfadd %%mm2, %%mm0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
53 "pfadd %%mm3, %%mm1 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
54 "pfsub %%mm2, %%mm4 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
55 "pfsub %%mm3, %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
56 "movq %%mm0, %%mm2 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
57 "pswapd %%mm5, %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
58 "movq %%mm4, %%mm3 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
59 "pxor %%mm7, %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
60 "pfadd %%mm1, %%mm0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
61 "pfadd %%mm5, %%mm4 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
62 "pfsub %%mm1, %%mm2 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
63 "pfsub %%mm5, %%mm3 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
64 "movq %%mm0, (%0,%1) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
65 "movq %%mm4, 8(%0,%1) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
66 "movq %%mm2, 16(%0,%1) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
67 "movq %%mm3, 24(%0,%1) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
68 "jg 1b \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
69 :"+r"(i)
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
70 :"r"(z)
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
71 );
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
72 /* pass 2 .. ln-1 */
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
73
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
74 nblocks = 1 << (ln-3);
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
75 nloops = 1 << 2;
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
76 cptr = s->exptab1;
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
77 do {
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
78 p = z;
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
79 j = nblocks;
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
80 do {
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
81 i = nloops*8;
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
82 asm volatile(
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
83 "1: \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
84 "sub $16, %0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
85 "movq (%1,%0), %%mm0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
86 "movq 8(%1,%0), %%mm1 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
87 "movq (%2,%0), %%mm2 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
88 "movq 8(%2,%0), %%mm3 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
89 "movq (%3,%0,2), %%mm4 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
90 "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
91 "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
92 "pswapd %%mm5, %%mm7 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
93 "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
94 "pfmul %%mm3, %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
95 "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
96 "pfmul %%mm3, %%mm7 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
97 "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
98 "pfpnacc %%mm7, %%mm5 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
99 "movq %%mm0, %%mm2 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
100 "movq %%mm1, %%mm3 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
101 "pfadd %%mm4, %%mm0 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
102 "pfadd %%mm5, %%mm1 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
103 "pfsub %%mm4, %%mm2 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
104 "pfsub %%mm5, %%mm3 \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
105 "movq %%mm0, (%1,%0) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
106 "movq %%mm1, 8(%1,%0) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
107 "movq %%mm2, (%2,%0) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
108 "movq %%mm3, 8(%2,%0) \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
109 "jg 1b \n\t"
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
110 :"+r"(i)
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
111 :"r"(p), "r"(p + nloops), "r"(cptr)
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
112 );
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
113 p += nloops*2;
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
114 } while (--j);
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
115 cptr += nloops*2;
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
116 nblocks >>= 1;
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
117 nloops <<= 1;
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
118 } while (nblocks != 0);
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
119 asm volatile("femms");
3175
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
120 }
c20c181e0eca 3DNow! & Extended 3DNow! versions of FFT
corey
parents:
diff changeset
121
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
122 void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output,
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
123 const FFTSample *input, FFTSample *tmp)
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
124 {
3559
c02459cd0d31 slightly faster ff_imdct_calc_3dn2() on amd64. (gcc added a bunch of useless movsxd)
lorenm
parents: 3555
diff changeset
125 long k, n8, n4, n2, n;
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
126 const uint16_t *revtab = s->fft.revtab;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
127 const FFTSample *tcos = s->tcos;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
128 const FFTSample *tsin = s->tsin;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
129 const FFTSample *in1, *in2;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
130 FFTComplex *z = (FFTComplex *)tmp;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
131
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
132 n = 1 << s->nbits;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
133 n2 = n >> 1;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
134 n4 = n >> 2;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
135 n8 = n >> 3;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
136
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
137 /* pre rotation */
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
138 in1 = input;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
139 in2 = input + n2 - 1;
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
140 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
141 // 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
142 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
143 "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
144 "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
145 "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
146 "punpckldq %3, %%mm1 \n\t"
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
147 "movq %%mm0, %%mm2 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
148 "pfmul %%mm1, %%mm0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
149 "pswapd %%mm1, %%mm1 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
150 "pfmul %%mm1, %%mm2 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
151 "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
152 ::"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
153 "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
154 );
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
155 asm volatile(
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
156 "movq %%mm0, %0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
157 :"=m"(z[revtab[k]])
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
158 );
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
159 }
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
160
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
161 ff_fft_calc(&s->fft, z);
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
162
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
163 /* post rotation + reordering */
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
164 for(k = 0; k < n4; k++) {
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
165 asm volatile(
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
166 "movq %0, %%mm0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
167 "movd %1, %%mm1 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
168 "punpckldq %2, %%mm1 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
169 "movq %%mm0, %%mm2 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
170 "pfmul %%mm1, %%mm0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
171 "pswapd %%mm1, %%mm1 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
172 "pfmul %%mm1, %%mm2 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
173 "pfpnacc %%mm2, %%mm0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
174 "movq %%mm0, %0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
175 :"+m"(z[k])
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
176 :"m"(tcos[k]), "m"(tsin[k])
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
177 );
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
178 }
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
179
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
180 z += n8;
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
181 asm volatile("movd %0, %%mm7" ::"r"(1<<31));
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
182 for(k = 0; k < n8; k++) {
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
183 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
184 "movq %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
185 "pswapd %1, %%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
186 ::"m"(z[k]), "m"(z[-1-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
187 );
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
188 asm volatile(
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
189 "movq %%mm0, %%mm2 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
190 "pxor %%mm7, %%mm2 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
191 "punpckldq %%mm1, %%mm2 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
192 "pswapd %%mm2, %%mm3 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
193 "punpckhdq %%mm1, %%mm0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
194 "pswapd %%mm0, %%mm4 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
195 "pxor %%mm7, %%mm0 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
196 "pxor %%mm7, %%mm4 \n\t"
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
197 "movq %%mm0, %0 \n\t" // { -z[n8+k].im, z[n8-1-k].re }
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
198 "movq %%mm4, %1 \n\t" // { -z[n8-1-k].re, z[n8+k].im }
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
199 "movq %%mm2, %2 \n\t" // { -z[n8+k].re, z[n8-1-k].im }
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
200 "movq %%mm3, %3 \n\t" // { z[n8-1-k].im, -z[n8+k].re }
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
201 :"=m"(output[2*k]), "=m"(output[n2-2-2*k]),
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
202 "=m"(output[n2+2*k]), "=m"(output[n-2-2*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
203 ::"memory"
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
204 );
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
205 }
3561
97325fecd35a emms -> femms
lorenm
parents: 3560
diff changeset
206 asm volatile("femms");
3555
5ea82888103e 3dnow2 implementation of imdct.
lorenm
parents: 3175
diff changeset
207 }
3590
a3d97c60ea07 ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents: 3561
diff changeset
208