Mercurial > libavcodec.hg
annotate i386/fft_3dn2.c @ 7544:ee1cb5ab9f99 libavcodec
optimize imdct_half:
remove tmp buffer.
skip fft reinterleave pass, leaving data in a format more convenient for simd.
merge post-rotate with post-reorder.
author | lorenm |
---|---|
date | Tue, 12 Aug 2008 00:33:34 +0000 |
parents | a8a8205a9081 |
children | 97383e012cb9 |
rev | line source |
---|---|
3175 | 1 /* |
2 * FFT/MDCT transform with Extended 3DNow! optimizations | |
7544 | 3 * Copyright (c) 2006-2008 Zuxy MENG Jie, Loren Merritt |
3175 | 4 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3748
diff
changeset
|
5 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3748
diff
changeset
|
6 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3748
diff
changeset
|
7 * FFmpeg is free software; you can redistribute it and/or |
3175 | 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:
3748
diff
changeset
|
10 * version 2.1 of the License, or (at your option) any later version. |
3175 | 11 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3748
diff
changeset
|
12 * FFmpeg is distributed in the hope that it will be useful, |
3175 | 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:
3748
diff
changeset
|
18 * License along with FFmpeg; if not, write to the Free Software |
3175 | 19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
20 */ | |
6763 | 21 |
22 #include "libavutil/x86_cpu.h" | |
23 #include "libavcodec/dsputil.h" | |
3175 | 24 |
7544 | 25 DECLARE_ALIGNED_8(static const int, m1m1[2]) = { 1<<31, 1<<31 }; |
26 | |
7542 | 27 #ifdef EMULATE_3DNOWEXT |
7544 | 28 #define PSWAPD(s,d)\ |
29 "movq "#s","#d"\n"\ | |
30 "psrlq $32,"#d"\n"\ | |
31 "punpckldq "#s","#d"\n" | |
7542 | 32 #define ff_fft_calc_3dn2 ff_fft_calc_3dn |
33 #define ff_fft_dispatch_3dn2 ff_fft_dispatch_3dn | |
34 #define ff_fft_dispatch_interleave_3dn2 ff_fft_dispatch_interleave_3dn | |
35 #define ff_imdct_calc_3dn2 ff_imdct_calc_3dn | |
36 #define ff_imdct_half_3dn2 ff_imdct_half_3dn | |
7544 | 37 #else |
38 #define PSWAPD(s,d) "pswapd "#s","#d"\n" | |
7542 | 39 #endif |
3175 | 40 |
7542 | 41 void ff_fft_dispatch_3dn2(FFTComplex *z, int nbits); |
3175 | 42 |
43 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z) | |
44 { | |
7542 | 45 int n = 1<<s->nbits; |
46 int i; | |
47 ff_fft_dispatch_interleave_3dn2(z, s->nbits); | |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3561
diff
changeset
|
48 asm volatile("femms"); |
7542 | 49 if(n <= 8) |
50 for(i=0; i<n; i+=2) | |
51 FFSWAP(FFTSample, z[i].im, z[i+1].re); | |
3175 | 52 } |
53 | |
7544 | 54 void ff_imdct_half_3dn2(MDCTContext *s, FFTSample *output, const FFTSample *input) |
3555 | 55 { |
7544 | 56 x86_reg j, k; |
57 long n = 1 << s->nbits; | |
58 long n2 = n >> 1; | |
59 long n4 = n >> 2; | |
60 long n8 = n >> 3; | |
3555 | 61 const uint16_t *revtab = s->fft.revtab; |
62 const FFTSample *tcos = s->tcos; | |
63 const FFTSample *tsin = s->tsin; | |
64 const FFTSample *in1, *in2; | |
7544 | 65 FFTComplex *z = (FFTComplex *)output; |
3555 | 66 |
67 /* pre rotation */ | |
68 in1 = input; | |
69 in2 = input + n2 - 1; | |
7544 | 70 #ifdef EMULATE_3DNOWEXT |
71 asm volatile("movd %0, %%mm7" ::"r"(1<<31)); | |
72 #endif | |
3555 | 73 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
|
74 // FIXME a single block is faster, but gcc 2.95 and 3.4.x on 32bit can't compile it |
3555 | 75 asm volatile( |
7544 | 76 "movd %0, %%mm0 \n" |
77 "movd %2, %%mm1 \n" | |
78 "punpckldq %1, %%mm0 \n" | |
79 "punpckldq %3, %%mm1 \n" | |
80 "movq %%mm0, %%mm2 \n" | |
81 PSWAPD( %%mm1, %%mm3 ) | |
82 "pfmul %%mm1, %%mm0 \n" | |
83 "pfmul %%mm3, %%mm2 \n" | |
84 #ifdef EMULATE_3DNOWEXT | |
85 "movq %%mm0, %%mm1 \n" | |
86 "punpckhdq %%mm2, %%mm0 \n" | |
87 "punpckldq %%mm2, %%mm1 \n" | |
88 "pxor %%mm7, %%mm0 \n" | |
89 "pfadd %%mm1, %%mm0 \n" | |
90 #else | |
91 "pfpnacc %%mm2, %%mm0 \n" | |
92 #endif | |
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
|
93 ::"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
|
94 "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
|
95 ); |
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
|
96 asm volatile( |
3555 | 97 "movq %%mm0, %0 \n\t" |
98 :"=m"(z[revtab[k]]) | |
99 ); | |
100 } | |
101 | |
7544 | 102 ff_fft_dispatch_3dn2(z, s->fft.nbits); |
103 | |
104 #define CMUL(j,mm0,mm1)\ | |
105 "movq (%2,"#j",2), %%mm6 \n"\ | |
106 "movq 8(%2,"#j",2), "#mm0"\n"\ | |
107 "movq %%mm6, "#mm1"\n"\ | |
108 "movq "#mm0",%%mm7 \n"\ | |
109 "pfmul (%3,"#j"), %%mm6 \n"\ | |
110 "pfmul (%4,"#j"), "#mm0"\n"\ | |
111 "pfmul (%4,"#j"), "#mm1"\n"\ | |
112 "pfmul (%3,"#j"), %%mm7 \n"\ | |
113 "pfsub %%mm6, "#mm0"\n"\ | |
114 "pfadd %%mm7, "#mm1"\n" | |
3555 | 115 |
7544 | 116 /* post rotation */ |
117 j = -n2; | |
118 k = n2-8; | |
119 asm volatile( | |
120 "1: \n" | |
121 CMUL(%0, %%mm0, %%mm1) | |
122 CMUL(%1, %%mm2, %%mm3) | |
123 "movd %%mm0, (%2,%0,2) \n" | |
124 "movd %%mm1,12(%2,%1,2) \n" | |
125 "movd %%mm2, (%2,%1,2) \n" | |
126 "movd %%mm3,12(%2,%0,2) \n" | |
127 "psrlq $32, %%mm0 \n" | |
128 "psrlq $32, %%mm1 \n" | |
129 "psrlq $32, %%mm2 \n" | |
130 "psrlq $32, %%mm3 \n" | |
131 "movd %%mm0, 8(%2,%0,2) \n" | |
132 "movd %%mm1, 4(%2,%1,2) \n" | |
133 "movd %%mm2, 8(%2,%1,2) \n" | |
134 "movd %%mm3, 4(%2,%0,2) \n" | |
135 "sub $8, %1 \n" | |
136 "add $8, %0 \n" | |
137 "jl 1b \n" | |
138 :"+r"(j), "+r"(k) | |
139 :"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8) | |
140 :"memory" | |
141 ); | |
142 asm volatile("femms"); | |
7263 | 143 } |
144 | |
145 void ff_imdct_calc_3dn2(MDCTContext *s, FFTSample *output, | |
146 const FFTSample *input, FFTSample *tmp) | |
147 { | |
7544 | 148 x86_reg j, k; |
149 long n = 1 << s->nbits; | |
150 long n4 = n >> 2; | |
7263 | 151 |
7544 | 152 ff_imdct_half_3dn2(s, output+n4, input); |
7263 | 153 |
7544 | 154 j = -n; |
3747 | 155 k = n-8; |
156 asm volatile( | |
7544 | 157 "movq %4, %%mm7 \n" |
158 "1: \n" | |
159 PSWAPD((%2,%1), %%mm0) | |
160 PSWAPD((%3,%0), %%mm1) | |
161 "pxor %%mm7, %%mm0 \n" | |
162 "movq %%mm1, (%3,%1) \n" | |
163 "movq %%mm0, (%2,%0) \n" | |
164 "sub $8, %1 \n" | |
165 "add $8, %0 \n" | |
166 "jl 1b \n" | |
167 :"+r"(j), "+r"(k) | |
168 :"r"(output+n4), "r"(output+n4*3), | |
169 "m"(*m1m1) | |
3747 | 170 ); |
3561 | 171 asm volatile("femms"); |
3555 | 172 } |
3590
a3d97c60ea07
ff_fft_calc_3dn/3dn2/sse: convert intrinsics to inline asm.
lorenm
parents:
3561
diff
changeset
|
173 |