1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
|
/*
* FFT/MDCT transform with SSE optimizations
* Copyright (c) 2008 Loren Merritt
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/x86_cpu.h"
#include "libavcodec/dsputil.h"
static const int m1m1m1m1[4] __attribute__((aligned(16))) =
{ 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
void ff_fft_dispatch_sse(FFTComplex *z, int nbits);
void ff_fft_dispatch_interleave_sse(FFTComplex *z, int nbits);
void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
{
int n = 1 << s->nbits;
ff_fft_dispatch_interleave_sse(z, s->nbits);
if(n <= 16) {
x86_reg i = -8*n;
asm volatile(
"1: \n"
"movaps (%0,%1), %%xmm0 \n"
"movaps %%xmm0, %%xmm1 \n"
"unpcklps 16(%0,%1), %%xmm0 \n"
"unpckhps 16(%0,%1), %%xmm1 \n"
"movaps %%xmm0, (%0,%1) \n"
"movaps %%xmm1, 16(%0,%1) \n"
"add $32, %0 \n"
"jl 1b \n"
:"+r"(i)
:"r"(z+n)
:"memory"
);
}
}
void ff_fft_permute_sse(FFTContext *s, FFTComplex *z)
{
int n = 1 << s->nbits;
int i;
for(i=0; i<n; i+=2) {
asm volatile(
"movaps %2, %%xmm0 \n"
"movlps %%xmm0, %0 \n"
"movhps %%xmm0, %1 \n"
:"=m"(s->tmp_buf[s->revtab[i]]),
"=m"(s->tmp_buf[s->revtab[i+1]])
:"m"(z[i])
);
}
memcpy(z, s->tmp_buf, n*sizeof(FFTComplex));
}
void ff_imdct_half_sse(MDCTContext *s, FFTSample *output, const FFTSample *input)
{
av_unused x86_reg i, j, k, l;
long n = 1 << s->nbits;
long n2 = n >> 1;
long n4 = n >> 2;
long n8 = n >> 3;
const uint16_t *revtab = s->fft.revtab + n8;
const FFTSample *tcos = s->tcos;
const FFTSample *tsin = s->tsin;
FFTComplex *z = (FFTComplex *)output;
/* pre rotation */
for(k=n8-2; k>=0; k-=2) {
asm volatile(
"movaps (%2,%1,2), %%xmm0 \n" // { z[k].re, z[k].im, z[k+1].re, z[k+1].im }
"movaps -16(%2,%0,2), %%xmm1 \n" // { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im }
"movaps %%xmm0, %%xmm2 \n"
"shufps $0x88, %%xmm1, %%xmm0 \n" // { z[k].re, z[k+1].re, z[-k-2].re, z[-k-1].re }
"shufps $0x77, %%xmm2, %%xmm1 \n" // { z[-k-1].im, z[-k-2].im, z[k+1].im, z[k].im }
"movlps (%3,%1), %%xmm4 \n"
"movlps (%4,%1), %%xmm5 \n"
"movhps -8(%3,%0), %%xmm4 \n" // { cos[k], cos[k+1], cos[-k-2], cos[-k-1] }
"movhps -8(%4,%0), %%xmm5 \n" // { sin[k], sin[k+1], sin[-k-2], sin[-k-1] }
"movaps %%xmm0, %%xmm2 \n"
"movaps %%xmm1, %%xmm3 \n"
"mulps %%xmm5, %%xmm0 \n" // re*sin
"mulps %%xmm4, %%xmm1 \n" // im*cos
"mulps %%xmm4, %%xmm2 \n" // re*cos
"mulps %%xmm5, %%xmm3 \n" // im*sin
"subps %%xmm0, %%xmm1 \n" // -> re
"addps %%xmm3, %%xmm2 \n" // -> im
"movaps %%xmm1, %%xmm0 \n"
"unpcklps %%xmm2, %%xmm1 \n" // { z[k], z[k+1] }
"unpckhps %%xmm2, %%xmm0 \n" // { z[-k-2], z[-k-1] }
::"r"(-4*k), "r"(4*k),
"r"(input+n4), "r"(tcos+n8), "r"(tsin+n8)
);
#ifdef ARCH_X86_64
// if we have enough regs, don't let gcc make the luts latency-bound
// but if not, latency is faster than spilling
asm("movlps %%xmm0, %0 \n"
"movhps %%xmm0, %1 \n"
"movlps %%xmm1, %2 \n"
"movhps %%xmm1, %3 \n"
:"=m"(z[revtab[-k-2]]),
"=m"(z[revtab[-k-1]]),
"=m"(z[revtab[ k ]]),
"=m"(z[revtab[ k+1]])
);
#else
asm("movlps %%xmm0, %0" :"=m"(z[revtab[-k-2]]));
asm("movhps %%xmm0, %0" :"=m"(z[revtab[-k-1]]));
asm("movlps %%xmm1, %0" :"=m"(z[revtab[ k ]]));
asm("movhps %%xmm1, %0" :"=m"(z[revtab[ k+1]]));
#endif
}
ff_fft_dispatch_sse(z, s->fft.nbits);
/* post rotation + reinterleave + reorder */
#define CMUL(j,xmm0,xmm1)\
"movaps (%2,"#j",2), %%xmm6 \n"\
"movaps 16(%2,"#j",2), "#xmm0"\n"\
"movaps %%xmm6, "#xmm1"\n"\
"movaps "#xmm0",%%xmm7 \n"\
"mulps (%3,"#j"), %%xmm6 \n"\
"mulps (%4,"#j"), "#xmm0"\n"\
"mulps (%4,"#j"), "#xmm1"\n"\
"mulps (%3,"#j"), %%xmm7 \n"\
"subps %%xmm6, "#xmm0"\n"\
"addps %%xmm7, "#xmm1"\n"
j = -n2;
k = n2-16;
asm volatile(
"1: \n"
CMUL(%0, %%xmm0, %%xmm1)
CMUL(%1, %%xmm4, %%xmm5)
"shufps $0x1b, %%xmm1, %%xmm1 \n"
"shufps $0x1b, %%xmm5, %%xmm5 \n"
"movaps %%xmm4, %%xmm6 \n"
"unpckhps %%xmm1, %%xmm4 \n"
"unpcklps %%xmm1, %%xmm6 \n"
"movaps %%xmm0, %%xmm2 \n"
"unpcklps %%xmm5, %%xmm0 \n"
"unpckhps %%xmm5, %%xmm2 \n"
"movaps %%xmm6, (%2,%1,2) \n"
"movaps %%xmm4, 16(%2,%1,2) \n"
"movaps %%xmm0, (%2,%0,2) \n"
"movaps %%xmm2, 16(%2,%0,2) \n"
"sub $16, %1 \n"
"add $16, %0 \n"
"jl 1b \n"
:"+&r"(j), "+&r"(k)
:"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8)
:"memory"
);
}
void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp)
{
x86_reg j, k;
long n = 1 << s->nbits;
long n4 = n >> 2;
ff_imdct_half_sse(s, output+n4, input);
j = -n;
k = n-16;
asm volatile(
"movaps %4, %%xmm7 \n"
"1: \n"
"movaps (%2,%1), %%xmm0 \n"
"movaps (%3,%0), %%xmm1 \n"
"shufps $0x1b, %%xmm0, %%xmm0 \n"
"shufps $0x1b, %%xmm1, %%xmm1 \n"
"xorps %%xmm7, %%xmm0 \n"
"movaps %%xmm1, (%3,%1) \n"
"movaps %%xmm0, (%2,%0) \n"
"sub $16, %1 \n"
"add $16, %0 \n"
"jl 1b \n"
:"+r"(j), "+r"(k)
:"r"(output+n4), "r"(output+n4*3),
"m"(*m1m1m1m1)
);
}
|
