2 * FFT/MDCT transform with SSE optimizations
3 * Copyright (c) 2008 Loren Merritt
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
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.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 #include "libavutil/x86_cpu.h"
23 #include "libavcodec/dsputil.h"
26 DECLARE_ALIGNED(16, static const int, m1m1m1m1)[4] =
27 { 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
29 void ff_fft_dispatch_sse(FFTComplex *z, int nbits);
30 void ff_fft_dispatch_interleave_sse(FFTComplex *z, int nbits);
32 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
34 int n = 1 << s->nbits;
36 ff_fft_dispatch_interleave_sse(z, s->nbits);
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"
57 void ff_fft_permute_sse(FFTContext *s, FFTComplex *z)
59 int n = 1 << s->nbits;
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]])
71 memcpy(z, s->tmp_buf, n*sizeof(FFTComplex));
74 void ff_imdct_half_sse(FFTContext *s, FFTSample *output, const FFTSample *input)
76 av_unused x86_reg i, j, k, l;
77 long n = 1 << s->mdct_bits;
81 const uint16_t *revtab = s->revtab + n8;
82 const FFTSample *tcos = s->tcos;
83 const FFTSample *tsin = s->tsin;
84 FFTComplex *z = (FFTComplex *)output;
87 for(k=n8-2; k>=0; k-=2) {
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)
113 // if we have enough regs, don't let gcc make the luts latency-bound
114 // but if not, latency is faster than spilling
115 __asm__("movlps %%xmm0, %0 \n"
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]])
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]]));
132 ff_fft_dispatch_sse(z, s->nbits);
134 /* post rotation + reinterleave + reorder */
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"
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"
170 :"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8)
175 void ff_imdct_calc_sse(FFTContext *s, FFTSample *output, const FFTSample *input)
178 long n = 1 << s->mdct_bits;
181 ff_imdct_half_sse(s, output+n4, input);
186 "movaps %4, %%xmm7 \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"
199 :"r"(output+n4), "r"(output+n4*3),