1/*
2 * FFT/MDCT transform with SSE optimizations
3 * Copyright (c) 2008 Loren Merritt
4 *
5 * This file is part of FFmpeg.
6 *
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.
11 *
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.
16 *
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
20 */
21
22#include "libavutil/x86_cpu.h"
23#include "libavcodec/dsputil.h"
24
25static const int m1m1m1m1[4] __attribute__((aligned(16))) =
26    { 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
27
28void ff_fft_dispatch_sse(FFTComplex *z, int nbits);
29void ff_fft_dispatch_interleave_sse(FFTComplex *z, int nbits);
30
31void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
32{
33    int n = 1 << s->nbits;
34
35    ff_fft_dispatch_interleave_sse(z, s->nbits);
36
37    if(n <= 16) {
38        x86_reg i = -8*n;
39        __asm__ volatile(
40            "1: \n"
41            "movaps     (%0,%1), %%xmm0 \n"
42            "movaps      %%xmm0, %%xmm1 \n"
43            "unpcklps 16(%0,%1), %%xmm0 \n"
44            "unpckhps 16(%0,%1), %%xmm1 \n"
45            "movaps      %%xmm0,   (%0,%1) \n"
46            "movaps      %%xmm1, 16(%0,%1) \n"
47            "add $32, %0 \n"
48            "jl 1b \n"
49            :"+r"(i)
50            :"r"(z+n)
51            :"memory"
52        );
53    }
54}
55
56void ff_fft_permute_sse(FFTContext *s, FFTComplex *z)
57{
58    int n = 1 << s->nbits;
59    int i;
60    for(i=0; i<n; i+=2) {
61        __asm__ volatile(
62            "movaps %2, %%xmm0 \n"
63            "movlps %%xmm0, %0 \n"
64            "movhps %%xmm0, %1 \n"
65            :"=m"(s->tmp_buf[s->revtab[i]]),
66             "=m"(s->tmp_buf[s->revtab[i+1]])
67            :"m"(z[i])
68        );
69    }
70    memcpy(z, s->tmp_buf, n*sizeof(FFTComplex));
71}
72
73void ff_imdct_half_sse(MDCTContext *s, FFTSample *output, const FFTSample *input)
74{
75    av_unused x86_reg i, j, k, l;
76    long n = 1 << s->nbits;
77    long n2 = n >> 1;
78    long n4 = n >> 2;
79    long n8 = n >> 3;
80    const uint16_t *revtab = s->fft.revtab + n8;
81    const FFTSample *tcos = s->tcos;
82    const FFTSample *tsin = s->tsin;
83    FFTComplex *z = (FFTComplex *)output;
84
85    /* pre rotation */
86    for(k=n8-2; k>=0; k-=2) {
87        __asm__ volatile(
88            "movaps     (%2,%1,2), %%xmm0 \n" // { z[k].re,    z[k].im,    z[k+1].re,  z[k+1].im  }
89            "movaps  -16(%2,%0,2), %%xmm1 \n" // { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im }
90            "movaps        %%xmm0, %%xmm2 \n"
91            "shufps $0x88, %%xmm1, %%xmm0 \n" // { z[k].re,    z[k+1].re,  z[-k-2].re, z[-k-1].re }
92            "shufps $0x77, %%xmm2, %%xmm1 \n" // { z[-k-1].im, z[-k-2].im, z[k+1].im,  z[k].im    }
93            "movlps       (%3,%1), %%xmm4 \n"
94            "movlps       (%4,%1), %%xmm5 \n"
95            "movhps     -8(%3,%0), %%xmm4 \n" // { cos[k],     cos[k+1],   cos[-k-2],  cos[-k-1]  }
96            "movhps     -8(%4,%0), %%xmm5 \n" // { sin[k],     sin[k+1],   sin[-k-2],  sin[-k-1]  }
97            "movaps        %%xmm0, %%xmm2 \n"
98            "movaps        %%xmm1, %%xmm3 \n"
99            "mulps         %%xmm5, %%xmm0 \n" // re*sin
100            "mulps         %%xmm4, %%xmm1 \n" // im*cos
101            "mulps         %%xmm4, %%xmm2 \n" // re*cos
102            "mulps         %%xmm5, %%xmm3 \n" // im*sin
103            "subps         %%xmm0, %%xmm1 \n" // -> re
104            "addps         %%xmm3, %%xmm2 \n" // -> im
105            "movaps        %%xmm1, %%xmm0 \n"
106            "unpcklps      %%xmm2, %%xmm1 \n" // { z[k],    z[k+1]  }
107            "unpckhps      %%xmm2, %%xmm0 \n" // { z[-k-2], z[-k-1] }
108            ::"r"(-4*k), "r"(4*k),
109              "r"(input+n4), "r"(tcos+n8), "r"(tsin+n8)
110        );
111#if ARCH_X86_64
112        // if we have enough regs, don't let gcc make the luts latency-bound
113        // but if not, latency is faster than spilling
114        __asm__("movlps %%xmm0, %0 \n"
115            "movhps %%xmm0, %1 \n"
116            "movlps %%xmm1, %2 \n"
117            "movhps %%xmm1, %3 \n"
118            :"=m"(z[revtab[-k-2]]),
119             "=m"(z[revtab[-k-1]]),
120             "=m"(z[revtab[ k  ]]),
121             "=m"(z[revtab[ k+1]])
122        );
123#else
124        __asm__("movlps %%xmm0, %0" :"=m"(z[revtab[-k-2]]));
125        __asm__("movhps %%xmm0, %0" :"=m"(z[revtab[-k-1]]));
126        __asm__("movlps %%xmm1, %0" :"=m"(z[revtab[ k  ]]));
127        __asm__("movhps %%xmm1, %0" :"=m"(z[revtab[ k+1]]));
128#endif
129    }
130
131    ff_fft_dispatch_sse(z, s->fft.nbits);
132
133    /* post rotation + reinterleave + reorder */
134
135#define CMUL(j,xmm0,xmm1)\
136        "movaps   (%2,"#j",2), %%xmm6 \n"\
137        "movaps 16(%2,"#j",2), "#xmm0"\n"\
138        "movaps        %%xmm6, "#xmm1"\n"\
139        "movaps        "#xmm0",%%xmm7 \n"\
140        "mulps      (%3,"#j"), %%xmm6 \n"\
141        "mulps      (%4,"#j"), "#xmm0"\n"\
142        "mulps      (%4,"#j"), "#xmm1"\n"\
143        "mulps      (%3,"#j"), %%xmm7 \n"\
144        "subps         %%xmm6, "#xmm0"\n"\
145        "addps         %%xmm7, "#xmm1"\n"
146
147    j = -n2;
148    k = n2-16;
149    __asm__ volatile(
150        "1: \n"
151        CMUL(%0, %%xmm0, %%xmm1)
152        CMUL(%1, %%xmm4, %%xmm5)
153        "shufps    $0x1b, %%xmm1, %%xmm1 \n"
154        "shufps    $0x1b, %%xmm5, %%xmm5 \n"
155        "movaps   %%xmm4, %%xmm6 \n"
156        "unpckhps %%xmm1, %%xmm4 \n"
157        "unpcklps %%xmm1, %%xmm6 \n"
158        "movaps   %%xmm0, %%xmm2 \n"
159        "unpcklps %%xmm5, %%xmm0 \n"
160        "unpckhps %%xmm5, %%xmm2 \n"
161        "movaps   %%xmm6,   (%2,%1,2) \n"
162        "movaps   %%xmm4, 16(%2,%1,2) \n"
163        "movaps   %%xmm0,   (%2,%0,2) \n"
164        "movaps   %%xmm2, 16(%2,%0,2) \n"
165        "sub $16, %1 \n"
166        "add $16, %0 \n"
167        "jl 1b \n"
168        :"+&r"(j), "+&r"(k)
169        :"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8)
170        :"memory"
171    );
172}
173
174void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output, const FFTSample *input)
175{
176    x86_reg j, k;
177    long n = 1 << s->nbits;
178    long n4 = n >> 2;
179
180    ff_imdct_half_sse(s, output+n4, input);
181
182    j = -n;
183    k = n-16;
184    __asm__ volatile(
185        "movaps %4, %%xmm7 \n"
186        "1: \n"
187        "movaps       (%2,%1), %%xmm0 \n"
188        "movaps       (%3,%0), %%xmm1 \n"
189        "shufps $0x1b, %%xmm0, %%xmm0 \n"
190        "shufps $0x1b, %%xmm1, %%xmm1 \n"
191        "xorps         %%xmm7, %%xmm0 \n"
192        "movaps        %%xmm1, (%3,%1) \n"
193        "movaps        %%xmm0, (%2,%0) \n"
194        "sub $16, %1 \n"
195        "add $16, %0 \n"
196        "jl 1b \n"
197        :"+r"(j), "+r"(k)
198        :"r"(output+n4), "r"(output+n4*3),
199         "m"(*m1m1m1m1)
200    );
201}
202
203