Fix embarassing typo in last commit: Restore mistakenly removed ','.
[libav.git] / libavcodec / x86 / fft_sse.c
CommitLineData
bb6f5690
FB
1/*
2 * FFT/MDCT transform with SSE optimizations
46803f4f 3 * Copyright (c) 2008 Loren Merritt
bb6f5690 4 *
b78e7197
DB
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
bb6f5690
FB
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
b78e7197 10 * version 2.1 of the License, or (at your option) any later version.
bb6f5690 11 *
b78e7197 12 * FFmpeg is distributed in the hope that it will be useful,
bb6f5690
FB
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
b78e7197 18 * License along with FFmpeg; if not, write to the Free Software
5509bffa 19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
bb6f5690 20 */
245976da
DB
21
22#include "libavutil/x86_cpu.h"
23#include "libavcodec/dsputil.h"
f4863213 24#include "fft.h"
bb6f5690 25
74a841af 26DECLARE_ALIGNED(16, static const int, m1m1m1m1[4]) =
ebbafcb4
LM
27 { 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
28
5d0ddd1a
LM
29void ff_fft_dispatch_sse(FFTComplex *z, int nbits);
30void ff_fft_dispatch_interleave_sse(FFTComplex *z, int nbits);
bb6f5690 31
68951ecf 32void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
bb6f5690 33{
5d0ddd1a
LM
34 int n = 1 << s->nbits;
35
36 ff_fft_dispatch_interleave_sse(z, s->nbits);
37
38 if(n <= 16) {
39 x86_reg i = -8*n;
be449fca 40 __asm__ volatile(
5d0ddd1a
LM
41 "1: \n"
42 "movaps (%0,%1), %%xmm0 \n"
43 "movaps %%xmm0, %%xmm1 \n"
44 "unpcklps 16(%0,%1), %%xmm0 \n"
45 "unpckhps 16(%0,%1), %%xmm1 \n"
46 "movaps %%xmm0, (%0,%1) \n"
47 "movaps %%xmm1, 16(%0,%1) \n"
48 "add $32, %0 \n"
49 "jl 1b \n"
50 :"+r"(i)
51 :"r"(z+n)
52 :"memory"
53 );
54 }
55}
bb6f5690 56
5d0ddd1a
LM
57void ff_fft_permute_sse(FFTContext *s, FFTComplex *z)
58{
59 int n = 1 << s->nbits;
60 int i;
61 for(i=0; i<n; i+=2) {
be449fca 62 __asm__ volatile(
5d0ddd1a
LM
63 "movaps %2, %%xmm0 \n"
64 "movlps %%xmm0, %0 \n"
65 "movhps %%xmm0, %1 \n"
66 :"=m"(s->tmp_buf[s->revtab[i]]),
67 "=m"(s->tmp_buf[s->revtab[i+1]])
68 :"m"(z[i])
69 );
70 }
71 memcpy(z, s->tmp_buf, n*sizeof(FFTComplex));
bb6f5690 72}
e1958604 73
46803f4f 74void ff_imdct_half_sse(MDCTContext *s, FFTSample *output, const FFTSample *input)
ebbafcb4 75{
46803f4f
LM
76 av_unused x86_reg i, j, k, l;
77 long n = 1 << s->nbits;
78 long n2 = n >> 1;
79 long n4 = n >> 2;
80 long n8 = n >> 3;
81 const uint16_t *revtab = s->fft.revtab + n8;
ebbafcb4
LM
82 const FFTSample *tcos = s->tcos;
83 const FFTSample *tsin = s->tsin;
46803f4f 84 FFTComplex *z = (FFTComplex *)output;
ebbafcb4
LM
85
86 /* pre rotation */
46803f4f 87 for(k=n8-2; k>=0; k-=2) {
be449fca 88 __asm__ volatile(
46803f4f
LM
89 "movaps (%2,%1,2), %%xmm0 \n" // { z[k].re, z[k].im, z[k+1].re, z[k+1].im }
90 "movaps -16(%2,%0,2), %%xmm1 \n" // { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im }
91 "movaps %%xmm0, %%xmm2 \n"
92 "shufps $0x88, %%xmm1, %%xmm0 \n" // { z[k].re, z[k+1].re, z[-k-2].re, z[-k-1].re }
93 "shufps $0x77, %%xmm2, %%xmm1 \n" // { z[-k-1].im, z[-k-2].im, z[k+1].im, z[k].im }
94 "movlps (%3,%1), %%xmm4 \n"
95 "movlps (%4,%1), %%xmm5 \n"
96 "movhps -8(%3,%0), %%xmm4 \n" // { cos[k], cos[k+1], cos[-k-2], cos[-k-1] }
97 "movhps -8(%4,%0), %%xmm5 \n" // { sin[k], sin[k+1], sin[-k-2], sin[-k-1] }
98 "movaps %%xmm0, %%xmm2 \n"
99 "movaps %%xmm1, %%xmm3 \n"
100 "mulps %%xmm5, %%xmm0 \n" // re*sin
101 "mulps %%xmm4, %%xmm1 \n" // im*cos
102 "mulps %%xmm4, %%xmm2 \n" // re*cos
103 "mulps %%xmm5, %%xmm3 \n" // im*sin
104 "subps %%xmm0, %%xmm1 \n" // -> re
105 "addps %%xmm3, %%xmm2 \n" // -> im
106 "movaps %%xmm1, %%xmm0 \n"
107 "unpcklps %%xmm2, %%xmm1 \n" // { z[k], z[k+1] }
108 "unpckhps %%xmm2, %%xmm0 \n" // { z[-k-2], z[-k-1] }
109 ::"r"(-4*k), "r"(4*k),
110 "r"(input+n4), "r"(tcos+n8), "r"(tsin+n8)
ebbafcb4 111 );
b250f9c6 112#if ARCH_X86_64
46803f4f
LM
113 // if we have enough regs, don't let gcc make the luts latency-bound
114 // but if not, latency is faster than spilling
be449fca 115 __asm__("movlps %%xmm0, %0 \n"
46803f4f
LM
116 "movhps %%xmm0, %1 \n"
117 "movlps %%xmm1, %2 \n"
118 "movhps %%xmm1, %3 \n"
119 :"=m"(z[revtab[-k-2]]),
120 "=m"(z[revtab[-k-1]]),
121 "=m"(z[revtab[ k ]]),
122 "=m"(z[revtab[ k+1]])
ebbafcb4 123 );
46803f4f 124#else
be449fca
DP
125 __asm__("movlps %%xmm0, %0" :"=m"(z[revtab[-k-2]]));
126 __asm__("movhps %%xmm0, %0" :"=m"(z[revtab[-k-1]]));
127 __asm__("movlps %%xmm1, %0" :"=m"(z[revtab[ k ]]));
128 __asm__("movhps %%xmm1, %0" :"=m"(z[revtab[ k+1]]));
25e4f8aa 129#endif
ebbafcb4 130 }
b9fa3208 131
46803f4f 132 ff_fft_dispatch_sse(z, s->fft.nbits);
b9fa3208 133
46803f4f 134 /* post rotation + reinterleave + reorder */
b9fa3208 135
46803f4f
LM
136#define CMUL(j,xmm0,xmm1)\
137 "movaps (%2,"#j",2), %%xmm6 \n"\
138 "movaps 16(%2,"#j",2), "#xmm0"\n"\
139 "movaps %%xmm6, "#xmm1"\n"\
140 "movaps "#xmm0",%%xmm7 \n"\
141 "mulps (%3,"#j"), %%xmm6 \n"\
142 "mulps (%4,"#j"), "#xmm0"\n"\
143 "mulps (%4,"#j"), "#xmm1"\n"\
144 "mulps (%3,"#j"), %%xmm7 \n"\
145 "subps %%xmm6, "#xmm0"\n"\
146 "addps %%xmm7, "#xmm1"\n"
ebbafcb4 147
46803f4f
LM
148 j = -n2;
149 k = n2-16;
be449fca 150 __asm__ volatile(
46803f4f
LM
151 "1: \n"
152 CMUL(%0, %%xmm0, %%xmm1)
153 CMUL(%1, %%xmm4, %%xmm5)
154 "shufps $0x1b, %%xmm1, %%xmm1 \n"
155 "shufps $0x1b, %%xmm5, %%xmm5 \n"
156 "movaps %%xmm4, %%xmm6 \n"
157 "unpckhps %%xmm1, %%xmm4 \n"
158 "unpcklps %%xmm1, %%xmm6 \n"
159 "movaps %%xmm0, %%xmm2 \n"
160 "unpcklps %%xmm5, %%xmm0 \n"
161 "unpckhps %%xmm5, %%xmm2 \n"
162 "movaps %%xmm6, (%2,%1,2) \n"
163 "movaps %%xmm4, 16(%2,%1,2) \n"
164 "movaps %%xmm0, (%2,%0,2) \n"
165 "movaps %%xmm2, 16(%2,%0,2) \n"
166 "sub $16, %1 \n"
167 "add $16, %0 \n"
168 "jl 1b \n"
169 :"+&r"(j), "+&r"(k)
170 :"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8)
ebbafcb4
LM
171 :"memory"
172 );
173}
174
0a570e82 175void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output, const FFTSample *input)
b9fa3208
LM
176{
177 x86_reg j, k;
46803f4f
LM
178 long n = 1 << s->nbits;
179 long n4 = n >> 2;
b9fa3208 180
46803f4f 181 ff_imdct_half_sse(s, output+n4, input);
b9fa3208
LM
182
183 j = -n;
184 k = n-16;
be449fca 185 __asm__ volatile(
46803f4f
LM
186 "movaps %4, %%xmm7 \n"
187 "1: \n"
188 "movaps (%2,%1), %%xmm0 \n"
189 "movaps (%3,%0), %%xmm1 \n"
190 "shufps $0x1b, %%xmm0, %%xmm0 \n"
191 "shufps $0x1b, %%xmm1, %%xmm1 \n"
192 "xorps %%xmm7, %%xmm0 \n"
193 "movaps %%xmm1, (%3,%1) \n"
194 "movaps %%xmm0, (%2,%0) \n"
195 "sub $16, %1 \n"
196 "add $16, %0 \n"
197 "jl 1b \n"
b9fa3208 198 :"+r"(j), "+r"(k)
46803f4f
LM
199 :"r"(output+n4), "r"(output+n4*3),
200 "m"(*m1m1m1m1)
b9fa3208
LM
201 );
202}
203