Mercurial > libavcodec.hg
annotate i386/fft_sse.c @ 6243:930ac362be5a libavcodec
const
author | michael |
---|---|
date | Fri, 01 Feb 2008 13:55:15 +0000 |
parents | 524faa5eabd1 |
children | 33896780c612 |
rev | line source |
---|---|
781 | 1 /* |
2 * FFT/MDCT transform with SSE optimizations | |
3 * Copyright (c) 2002 Fabrice Bellard. | |
4 * | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
5 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
6 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
7 * FFmpeg is free software; you can redistribute it and/or |
781 | 8 * modify it under the terms of the GNU Lesser General Public |
9 * 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:
3746
diff
changeset
|
10 * version 2.1 of the License, or (at your option) any later version. |
781 | 11 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3746
diff
changeset
|
12 * FFmpeg is distributed in the hope that it will be useful, |
781 | 13 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
15 * Lesser General Public License for more details. | |
16 * | |
17 * 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:
3746
diff
changeset
|
18 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2979
diff
changeset
|
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
781 | 20 */ |
5010
d5ba514e3f4a
Add libavcodec to compiler include flags in order to simplify header
diego
parents:
5000
diff
changeset
|
21 #include "dsputil.h" |
781 | 22 |
3166
ab1273ffe275
use xorps instead of mulps to toggle the sign of a float, as suggested by Software Optimization Guide for AMD64 Processors.
gpoirier
parents:
3036
diff
changeset
|
23 static const int p1p1p1m1[4] __attribute__((aligned(16))) = |
ab1273ffe275
use xorps instead of mulps to toggle the sign of a float, as suggested by Software Optimization Guide for AMD64 Processors.
gpoirier
parents:
3036
diff
changeset
|
24 { 0, 0, 0, 1 << 31 }; |
781 | 25 |
3166
ab1273ffe275
use xorps instead of mulps to toggle the sign of a float, as suggested by Software Optimization Guide for AMD64 Processors.
gpoirier
parents:
3036
diff
changeset
|
26 static const int p1p1m1p1[4] __attribute__((aligned(16))) = |
ab1273ffe275
use xorps instead of mulps to toggle the sign of a float, as suggested by Software Optimization Guide for AMD64 Processors.
gpoirier
parents:
3036
diff
changeset
|
27 { 0, 0, 1 << 31, 0 }; |
968
64f1a11b5f86
added define for builtins use - inverse fix by Romain Dolbeau
bellard
parents:
781
diff
changeset
|
28 |
3166
ab1273ffe275
use xorps instead of mulps to toggle the sign of a float, as suggested by Software Optimization Guide for AMD64 Processors.
gpoirier
parents:
3036
diff
changeset
|
29 static const int p1p1m1m1[4] __attribute__((aligned(16))) = |
ab1273ffe275
use xorps instead of mulps to toggle the sign of a float, as suggested by Software Optimization Guide for AMD64 Processors.
gpoirier
parents:
3036
diff
changeset
|
30 { 0, 0, 1 << 31, 1 << 31 }; |
781 | 31 |
3746 | 32 static const int p1m1p1m1[4] __attribute__((aligned(16))) = |
33 { 0, 1 << 31, 0, 1 << 31 }; | |
34 | |
35 static const int m1m1m1m1[4] __attribute__((aligned(16))) = | |
36 { 1 << 31, 1 << 31, 1 << 31, 1 << 31 }; | |
37 | |
781 | 38 #if 0 |
39 static void print_v4sf(const char *str, __m128 a) | |
40 { | |
41 float *p = (float *)&a; | |
42 printf("%s: %f %f %f %f\n", | |
43 str, p[0], p[1], p[2], p[3]); | |
44 } | |
45 #endif | |
46 | |
47 /* XXX: handle reverse case */ | |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
968
diff
changeset
|
48 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z) |
781 | 49 { |
50 int ln = s->nbits; | |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
51 long i, j; |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
52 long nblocks, nloops; |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
53 FFTComplex *p, *cptr; |
781 | 54 |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
55 asm volatile( |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
56 "movaps %0, %%xmm4 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
57 "movaps %1, %%xmm5 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
58 ::"m"(*p1p1m1m1), |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
59 "m"(*(s->inverse ? p1p1m1p1 : p1p1p1m1)) |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
60 ); |
968
64f1a11b5f86
added define for builtins use - inverse fix by Romain Dolbeau
bellard
parents:
781
diff
changeset
|
61 |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
62 i = 8 << ln; |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
63 asm volatile( |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
64 "1: \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
65 "sub $32, %0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
66 /* do the pass 0 butterfly */ |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
67 "movaps (%0,%1), %%xmm0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
68 "movaps %%xmm0, %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
69 "shufps $0x4E, %%xmm0, %%xmm0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
70 "xorps %%xmm4, %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
71 "addps %%xmm1, %%xmm0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
72 "movaps 16(%0,%1), %%xmm2 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
73 "movaps %%xmm2, %%xmm3 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
74 "shufps $0x4E, %%xmm2, %%xmm2 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
75 "xorps %%xmm4, %%xmm3 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
76 "addps %%xmm3, %%xmm2 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
77 /* multiply third by -i */ |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
78 /* by toggling the sign bit */ |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
79 "shufps $0xB4, %%xmm2, %%xmm2 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
80 "xorps %%xmm5, %%xmm2 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
81 /* do the pass 1 butterfly */ |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
82 "movaps %%xmm0, %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
83 "addps %%xmm2, %%xmm0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
84 "subps %%xmm2, %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
85 "movaps %%xmm0, (%0,%1) \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
86 "movaps %%xmm1, 16(%0,%1) \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
87 "jg 1b \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
88 :"+r"(i) |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
89 :"r"(z) |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
90 ); |
781 | 91 /* pass 2 .. ln-1 */ |
92 | |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
93 nblocks = 1 << (ln-3); |
781 | 94 nloops = 1 << 2; |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
95 cptr = s->exptab1; |
781 | 96 do { |
97 p = z; | |
98 j = nblocks; | |
99 do { | |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
100 i = nloops*8; |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
101 asm volatile( |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
102 "1: \n\t" |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
103 "sub $32, %0 \n\t" |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
104 "movaps (%2,%0), %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
105 "movaps (%1,%0), %%xmm0 \n\t" |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
106 "movaps 16(%2,%0), %%xmm5 \n\t" |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
107 "movaps 16(%1,%0), %%xmm4 \n\t" |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
108 "movaps %%xmm1, %%xmm2 \n\t" |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
109 "movaps %%xmm5, %%xmm6 \n\t" |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
110 "shufps $0xA0, %%xmm1, %%xmm1 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
111 "shufps $0xF5, %%xmm2, %%xmm2 \n\t" |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
112 "shufps $0xA0, %%xmm5, %%xmm5 \n\t" |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
113 "shufps $0xF5, %%xmm6, %%xmm6 \n\t" |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
114 "mulps (%3,%0,2), %%xmm1 \n\t" // cre*re cim*re |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
115 "mulps 16(%3,%0,2), %%xmm2 \n\t" // -cim*im cre*im |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
116 "mulps 32(%3,%0,2), %%xmm5 \n\t" // cre*re cim*re |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
117 "mulps 48(%3,%0,2), %%xmm6 \n\t" // -cim*im cre*im |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
118 "addps %%xmm2, %%xmm1 \n\t" |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
119 "addps %%xmm6, %%xmm5 \n\t" |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
120 "movaps %%xmm0, %%xmm3 \n\t" |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
121 "movaps %%xmm4, %%xmm7 \n\t" |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
122 "addps %%xmm1, %%xmm0 \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
123 "subps %%xmm1, %%xmm3 \n\t" |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
124 "addps %%xmm5, %%xmm4 \n\t" |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
125 "subps %%xmm5, %%xmm7 \n\t" |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
126 "movaps %%xmm0, (%1,%0) \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
127 "movaps %%xmm3, (%2,%0) \n\t" |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
128 "movaps %%xmm4, 16(%1,%0) \n\t" |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
129 "movaps %%xmm7, 16(%2,%0) \n\t" |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
130 "jg 1b \n\t" |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
131 :"+r"(i) |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
132 :"r"(p), "r"(p + nloops), "r"(cptr) |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
133 ); |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
134 p += nloops*2; |
781 | 135 } while (--j); |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
136 cptr += nloops*2; |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
137 nblocks >>= 1; |
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3166
diff
changeset
|
138 nloops <<= 1; |
781 | 139 } while (nblocks != 0); |
140 } | |
968
64f1a11b5f86
added define for builtins use - inverse fix by Romain Dolbeau
bellard
parents:
781
diff
changeset
|
141 |
3746 | 142 void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output, |
143 const FFTSample *input, FFTSample *tmp) | |
144 { | |
145 long k, n8, n4, n2, n; | |
146 const uint16_t *revtab = s->fft.revtab; | |
147 const FFTSample *tcos = s->tcos; | |
148 const FFTSample *tsin = s->tsin; | |
149 const FFTSample *in1, *in2; | |
150 FFTComplex *z = (FFTComplex *)tmp; | |
151 | |
152 n = 1 << s->nbits; | |
153 n2 = n >> 1; | |
154 n4 = n >> 2; | |
155 n8 = n >> 3; | |
156 | |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
157 #ifdef ARCH_X86_64 |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
158 asm volatile ("movaps %0, %%xmm8\n\t"::"m"(*p1m1p1m1)); |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
159 #define P1M1P1M1 "%%xmm8" |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
160 #else |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
161 #define P1M1P1M1 "%4" |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
162 #endif |
3746 | 163 |
164 /* pre rotation */ | |
165 in1 = input; | |
166 in2 = input + n2 - 4; | |
167 | |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
168 /* Complex multiplication */ |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
169 for (k = 0; k < n4; k += 4) { |
3746 | 170 asm volatile ( |
171 "movaps %0, %%xmm0 \n\t" // xmm0 = r0 X r1 X : in2 | |
172 "movaps %1, %%xmm3 \n\t" // xmm3 = X i1 X i0: in1 | |
5117
524faa5eabd1
work around issues with the old version of Gnu Assembler shipped on
gpoirier
parents:
5010
diff
changeset
|
173 "movaps -16+1*%0, %%xmm4 \n\t" // xmm4 = r0 X r1 X : in2 |
524faa5eabd1
work around issues with the old version of Gnu Assembler shipped on
gpoirier
parents:
5010
diff
changeset
|
174 "movaps 16+1*%1, %%xmm7 \n\t" // xmm7 = X i1 X i0: in1 |
3746 | 175 "movlps %2, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos |
176 "movlps %3, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin | |
5117
524faa5eabd1
work around issues with the old version of Gnu Assembler shipped on
gpoirier
parents:
5010
diff
changeset
|
177 "movlps 8+1*%2, %%xmm5 \n\t" // xmm5 = X X R1 R0: tcos |
524faa5eabd1
work around issues with the old version of Gnu Assembler shipped on
gpoirier
parents:
5010
diff
changeset
|
178 "movlps 8+1*%3, %%xmm6 \n\t" // xmm6 = X X I1 I0: tsin |
3746 | 179 "shufps $95, %%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0 |
180 "shufps $160,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0 | |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
181 "shufps $95, %%xmm4, %%xmm4 \n\t" // xmm4 = r1 r1 r0 r0 |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
182 "shufps $160,%%xmm7, %%xmm7 \n\t" // xmm7 = i1 i1 i0 i0 |
3746 | 183 "unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0 |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
184 "unpcklps %%xmm6, %%xmm5 \n\t" // xmm5 = I1 R1 I0 R0 |
3746 | 185 "movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0 |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
186 "movaps %%xmm5, %%xmm6 \n\t" // xmm6 = I1 R1 I0 R0 |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
187 "xorps "P1M1P1M1", %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0 |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
188 "xorps "P1M1P1M1", %%xmm6 \n\t" // xmm6 = -I1 R1 -I0 R0 |
3746 | 189 "mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
190 "mulps %%xmm5, %%xmm4 \n\t" // xmm4 = rI rR rI rR |
3746 | 191 "shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0 |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
192 "shufps $177,%%xmm6, %%xmm6 \n\t" // xmm6 = R1 -I1 R0 -I0 |
3746 | 193 "mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
194 "mulps %%xmm6, %%xmm7 \n\t" // xmm7 = Ri -Ii Ri -Ii |
3746 | 195 "addps %%xmm3, %%xmm0 \n\t" // xmm0 = result |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
196 "addps %%xmm7, %%xmm4 \n\t" // xmm4 = result |
3746 | 197 ::"m"(in2[-2*k]), "m"(in1[2*k]), |
198 "m"(tcos[k]), "m"(tsin[k]) | |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
199 #ifndef ARCH_X86_64 |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
200 ,"m"(*p1m1p1m1) |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
201 #endif |
3746 | 202 ); |
203 /* Should be in the same block, hack for gcc2.95 & gcc3 */ | |
204 asm ( | |
205 "movlps %%xmm0, %0 \n\t" | |
206 "movhps %%xmm0, %1 \n\t" | |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
207 "movlps %%xmm4, %2 \n\t" |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
208 "movhps %%xmm4, %3 \n\t" |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
209 :"=m"(z[revtab[k]]), "=m"(z[revtab[k + 1]]), |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
210 "=m"(z[revtab[k + 2]]), "=m"(z[revtab[k + 3]]) |
3746 | 211 ); |
212 } | |
213 | |
214 ff_fft_calc_sse(&s->fft, z); | |
215 | |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
216 #ifndef ARCH_X86_64 |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
217 #undef P1M1P1M1 |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
218 #define P1M1P1M1 "%3" |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
219 #endif |
3746 | 220 |
221 /* post rotation + reordering */ | |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
222 for (k = 0; k < n4; k += 4) { |
3746 | 223 asm ( |
224 "movaps %0, %%xmm0 \n\t" // xmm0 = i1 r1 i0 r0: z | |
5117
524faa5eabd1
work around issues with the old version of Gnu Assembler shipped on
gpoirier
parents:
5010
diff
changeset
|
225 "movaps 16+1*%0, %%xmm4 \n\t" // xmm4 = i1 r1 i0 r0: z |
3746 | 226 "movlps %1, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos |
5117
524faa5eabd1
work around issues with the old version of Gnu Assembler shipped on
gpoirier
parents:
5010
diff
changeset
|
227 "movlps 8+1*%1, %%xmm5 \n\t" // xmm5 = X X R1 R0: tcos |
3746 | 228 "movaps %%xmm0, %%xmm3 \n\t" // xmm3 = i1 r1 i0 r0 |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
229 "movaps %%xmm4, %%xmm7 \n\t" // xmm7 = i1 r1 i0 r0 |
3746 | 230 "movlps %2, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin |
5117
524faa5eabd1
work around issues with the old version of Gnu Assembler shipped on
gpoirier
parents:
5010
diff
changeset
|
231 "movlps 8+1*%2, %%xmm6 \n\t" // xmm6 = X X I1 I0: tsin |
3746 | 232 "shufps $160,%%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0 |
233 "shufps $245,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0 | |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
234 "shufps $160,%%xmm4, %%xmm4 \n\t" // xmm4 = r1 r1 r0 r0 |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
235 "shufps $245,%%xmm7, %%xmm7 \n\t" // xmm7 = i1 i1 i0 i0 |
3746 | 236 "unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0 |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
237 "unpcklps %%xmm6, %%xmm5 \n\t" // xmm5 = I1 R1 I0 R0 |
3746 | 238 "movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0 |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
239 "movaps %%xmm5, %%xmm6 \n\t" // xmm6 = I1 R1 I0 R0 |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
240 "xorps "P1M1P1M1", %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0 |
3746 | 241 "mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
242 "xorps "P1M1P1M1", %%xmm6 \n\t" // xmm6 = -I1 R1 -I0 R0 |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
243 "mulps %%xmm5, %%xmm4 \n\t" // xmm4 = rI rR rI rR |
3746 | 244 "shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0 |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
245 "shufps $177,%%xmm6, %%xmm6 \n\t" // xmm6 = R1 -I1 R0 -I0 |
3746 | 246 "mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
247 "mulps %%xmm6, %%xmm7 \n\t" // xmm7 = Ri -Ii Ri -Ii |
3746 | 248 "addps %%xmm3, %%xmm0 \n\t" // xmm0 = result |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
249 "addps %%xmm7, %%xmm4 \n\t" // xmm4 = result |
3746 | 250 "movaps %%xmm0, %0 \n\t" |
5117
524faa5eabd1
work around issues with the old version of Gnu Assembler shipped on
gpoirier
parents:
5010
diff
changeset
|
251 "movaps %%xmm4, 16+1*%0\n\t" |
3746 | 252 :"+m"(z[k]) |
253 :"m"(tcos[k]), "m"(tsin[k]) | |
5000
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
254 #ifndef ARCH_X86_64 |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
255 ,"m"(*p1m1p1m1) |
743a8b12b7de
Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%
gpoirier
parents:
3947
diff
changeset
|
256 #endif |
3746 | 257 ); |
258 } | |
259 | |
260 /* | |
261 Mnemonics: | |
262 0 = z[k].re | |
263 1 = z[k].im | |
264 2 = z[k + 1].re | |
265 3 = z[k + 1].im | |
266 4 = z[-k - 2].re | |
267 5 = z[-k - 2].im | |
268 6 = z[-k - 1].re | |
269 7 = z[-k - 1].im | |
270 */ | |
271 k = 16-n; | |
272 asm volatile("movaps %0, %%xmm7 \n\t"::"m"(*m1m1m1m1)); | |
273 asm volatile( | |
274 "1: \n\t" | |
275 "movaps -16(%4,%0), %%xmm1 \n\t" // xmm1 = 4 5 6 7 = z[-2-k] | |
276 "neg %0 \n\t" | |
277 "movaps (%4,%0), %%xmm0 \n\t" // xmm0 = 0 1 2 3 = z[k] | |
278 "xorps %%xmm7, %%xmm0 \n\t" // xmm0 = -0 -1 -2 -3 | |
279 "movaps %%xmm0, %%xmm2 \n\t" // xmm2 = -0 -1 -2 -3 | |
280 "shufps $141,%%xmm1, %%xmm0 \n\t" // xmm0 = -1 -3 4 6 | |
281 "shufps $216,%%xmm1, %%xmm2 \n\t" // xmm2 = -0 -2 5 7 | |
282 "shufps $156,%%xmm0, %%xmm0 \n\t" // xmm0 = -1 6 -3 4 ! | |
283 "shufps $156,%%xmm2, %%xmm2 \n\t" // xmm2 = -0 7 -2 5 ! | |
284 "movaps %%xmm0, (%1,%0) \n\t" // output[2*k] | |
285 "movaps %%xmm2, (%2,%0) \n\t" // output[n2+2*k] | |
286 "neg %0 \n\t" | |
287 "shufps $27, %%xmm0, %%xmm0 \n\t" // xmm0 = 4 -3 6 -1 | |
288 "xorps %%xmm7, %%xmm0 \n\t" // xmm0 = -4 3 -6 1 ! | |
289 "shufps $27, %%xmm2, %%xmm2 \n\t" // xmm2 = 5 -2 7 -0 ! | |
290 "movaps %%xmm0, -16(%2,%0) \n\t" // output[n2-4-2*k] | |
291 "movaps %%xmm2, -16(%3,%0) \n\t" // output[n-4-2*k] | |
292 "add $16, %0 \n\t" | |
293 "jle 1b \n\t" | |
294 :"+r"(k) | |
295 :"r"(output), "r"(output+n2), "r"(output+n), "r"(z+n8) | |
296 :"memory" | |
297 ); | |
298 } | |
299 |