FFmpeg: libavcodec/x86/fft_sse.c Source File

00001 /*
00002  * FFT/MDCT transform with SSE optimizations
00003  * Copyright (c) 2008 Loren Merritt
00004  *
00005  * This file is part of FFmpeg.
00006  *
00007  * FFmpeg is free software; you can redistribute it and/or
00008  * modify it under the terms of the GNU Lesser General Public
00009  * License as published by the Free Software Foundation; either
00010  * version 2.1 of the License, or (at your option) any later version.
00011  *
00012  * FFmpeg is distributed in the hope that it will be useful,
00013  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00014  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00015  * Lesser General Public License for more details.
00016  *
00017  * You should have received a copy of the GNU Lesser General Public
00018  * License along with FFmpeg; if not, write to the Free Software
00019  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00020  */
00021 
00022 #include "libavutil/x86_cpu.h"
00023 #include "libavcodec/dsputil.h"
00024 #include "fft.h"
00025 
00026 DECLARE_ASM_CONST(16, int, ff_m1m1m1m1)[4] =
00027     { 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
00028 
00029 void ff_fft_dispatch_sse(FFTComplex *z, int nbits);
00030 void ff_fft_dispatch_interleave_sse(FFTComplex *z, int nbits);
00031 
00032 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
00033 {
00034     int n = 1 << s->nbits;
00035 
00036     ff_fft_dispatch_interleave_sse(z, s->nbits);
00037 
00038     if(n <= 16) {
00039         x86_reg i = -8*n;
00040         __asm__ volatile(
00041             "1: \n"
00042             "movaps     (%0,%1), %%xmm0 \n"
00043             "movaps      %%xmm0, %%xmm1 \n"
00044             "unpcklps 16(%0,%1), %%xmm0 \n"
00045             "unpckhps 16(%0,%1), %%xmm1 \n"
00046             "movaps      %%xmm0,   (%0,%1) \n"
00047             "movaps      %%xmm1, 16(%0,%1) \n"
00048             "add $32, %0 \n"
00049             "jl 1b \n"
00050             :"+r"(i)
00051             :"r"(z+n)
00052             :"memory"
00053         );
00054     }
00055 }
00056 
00057 void ff_fft_permute_sse(FFTContext *s, FFTComplex *z)
00058 {
00059     int n = 1 << s->nbits;
00060     int i;
00061     for(i=0; i<n; i+=2) {
00062         __asm__ volatile(
00063             "movaps %2, %%xmm0 \n"
00064             "movlps %%xmm0, %0 \n"
00065             "movhps %%xmm0, %1 \n"
00066             :"=m"(s->tmp_buf[s->revtab[i]]),
00067              "=m"(s->tmp_buf[s->revtab[i+1]])
00068             :"m"(z[i])
00069         );
00070     }
00071     memcpy(z, s->tmp_buf, n*sizeof(FFTComplex));
00072 }
00073 
00074 void ff_imdct_half_sse(FFTContext *s, FFTSample *output, const FFTSample *input)
00075 {
00076     av_unused x86_reg i, j, k, l;
00077     long n = 1 << s->mdct_bits;
00078     long n2 = n >> 1;
00079     long n4 = n >> 2;
00080     long n8 = n >> 3;
00081     const uint16_t *revtab = s->revtab + n8;
00082     const FFTSample *tcos = s->tcos;
00083     const FFTSample *tsin = s->tsin;
00084     FFTComplex *z = (FFTComplex *)output;
00085 
00086     /* pre rotation */
00087     for(k=n8-2; k>=0; k-=2) {
00088         __asm__ volatile(
00089             "movaps     (%2,%1,2), %%xmm0 \n" // { z[k].re,    z[k].im,    z[k+1].re,  z[k+1].im  }
00090             "movaps  -16(%2,%0,2), %%xmm1 \n" // { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im }
00091             "movaps        %%xmm0, %%xmm2 \n"
00092             "shufps $0x88, %%xmm1, %%xmm0 \n" // { z[k].re,    z[k+1].re,  z[-k-2].re, z[-k-1].re }
00093             "shufps $0x77, %%xmm2, %%xmm1 \n" // { z[-k-1].im, z[-k-2].im, z[k+1].im,  z[k].im    }
00094             "movlps       (%3,%1), %%xmm4 \n"
00095             "movlps       (%4,%1), %%xmm5 \n"
00096             "movhps     -8(%3,%0), %%xmm4 \n" // { cos[k],     cos[k+1],   cos[-k-2],  cos[-k-1]  }
00097             "movhps     -8(%4,%0), %%xmm5 \n" // { sin[k],     sin[k+1],   sin[-k-2],  sin[-k-1]  }
00098             "movaps        %%xmm0, %%xmm2 \n"
00099             "movaps        %%xmm1, %%xmm3 \n"
00100             "mulps         %%xmm5, %%xmm0 \n" // re*sin
00101             "mulps         %%xmm4, %%xmm1 \n" // im*cos
00102             "mulps         %%xmm4, %%xmm2 \n" // re*cos
00103             "mulps         %%xmm5, %%xmm3 \n" // im*sin
00104             "subps         %%xmm0, %%xmm1 \n" // -> re
00105             "addps         %%xmm3, %%xmm2 \n" // -> im
00106             "movaps        %%xmm1, %%xmm0 \n"
00107             "unpcklps      %%xmm2, %%xmm1 \n" // { z[k],    z[k+1]  }
00108             "unpckhps      %%xmm2, %%xmm0 \n" // { z[-k-2], z[-k-1] }
00109             ::"r"(-4*k), "r"(4*k),
00110               "r"(input+n4), "r"(tcos+n8), "r"(tsin+n8)
00111         );
00112 #if ARCH_X86_64
00113         // if we have enough regs, don't let gcc make the luts latency-bound
00114         // but if not, latency is faster than spilling
00115         __asm__("movlps %%xmm0, %0 \n"
00116             "movhps %%xmm0, %1 \n"
00117             "movlps %%xmm1, %2 \n"
00118             "movhps %%xmm1, %3 \n"
00119             :"=m"(z[revtab[-k-2]]),
00120              "=m"(z[revtab[-k-1]]),
00121              "=m"(z[revtab[ k  ]]),
00122              "=m"(z[revtab[ k+1]])
00123         );
00124 #else
00125         __asm__("movlps %%xmm0, %0" :"=m"(z[revtab[-k-2]]));
00126         __asm__("movhps %%xmm0, %0" :"=m"(z[revtab[-k-1]]));
00127         __asm__("movlps %%xmm1, %0" :"=m"(z[revtab[ k  ]]));
00128         __asm__("movhps %%xmm1, %0" :"=m"(z[revtab[ k+1]]));
00129 #endif
00130     }
00131 
00132     ff_fft_dispatch_sse(z, s->nbits);
00133 
00134     /* post rotation + reinterleave + reorder */
00135 
00136 #define CMUL(j,xmm0,xmm1)\
00137         "movaps   (%2,"#j",2), %%xmm6 \n"\
00138         "movaps 16(%2,"#j",2), "#xmm0"\n"\
00139         "movaps        %%xmm6, "#xmm1"\n"\
00140         "movaps        "#xmm0",%%xmm7 \n"\
00141         "mulps      (%3,"#j"), %%xmm6 \n"\
00142         "mulps      (%4,"#j"), "#xmm0"\n"\
00143         "mulps      (%4,"#j"), "#xmm1"\n"\
00144         "mulps      (%3,"#j"), %%xmm7 \n"\
00145         "subps         %%xmm6, "#xmm0"\n"\
00146         "addps         %%xmm7, "#xmm1"\n"
00147 
00148     j = -n2;
00149     k = n2-16;
00150     __asm__ volatile(
00151         "1: \n"
00152         CMUL(%0, %%xmm0, %%xmm1)
00153         CMUL(%1, %%xmm4, %%xmm5)
00154         "shufps    $0x1b, %%xmm1, %%xmm1 \n"
00155         "shufps    $0x1b, %%xmm5, %%xmm5 \n"
00156         "movaps   %%xmm4, %%xmm6 \n"
00157         "unpckhps %%xmm1, %%xmm4 \n"
00158         "unpcklps %%xmm1, %%xmm6 \n"
00159         "movaps   %%xmm0, %%xmm2 \n"
00160         "unpcklps %%xmm5, %%xmm0 \n"
00161         "unpckhps %%xmm5, %%xmm2 \n"
00162         "movaps   %%xmm6,   (%2,%1,2) \n"
00163         "movaps   %%xmm4, 16(%2,%1,2) \n"
00164         "movaps   %%xmm0,   (%2,%0,2) \n"
00165         "movaps   %%xmm2, 16(%2,%0,2) \n"
00166         "sub $16, %1 \n"
00167         "add $16, %0 \n"
00168         "jl 1b \n"
00169         :"+&r"(j), "+&r"(k)
00170         :"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8)
00171         :"memory"
00172     );
00173 }
00174 
00175 void ff_imdct_calc_sse(FFTContext *s, FFTSample *output, const FFTSample *input)
00176 {
00177     x86_reg j, k;
00178     long n = 1 << s->mdct_bits;
00179     long n4 = n >> 2;
00180 
00181     ff_imdct_half_sse(s, output+n4, input);
00182 
00183     j = -n;
00184     k = n-16;
00185     __asm__ volatile(
00186         "movaps "MANGLE(ff_m1m1m1m1)", %%xmm7 \n"
00187         "1: \n"
00188         "movaps       (%2,%1), %%xmm0 \n"
00189         "movaps       (%3,%0), %%xmm1 \n"
00190         "shufps $0x1b, %%xmm0, %%xmm0 \n"
00191         "shufps $0x1b, %%xmm1, %%xmm1 \n"
00192         "xorps         %%xmm7, %%xmm0 \n"
00193         "movaps        %%xmm1, (%3,%1) \n"
00194         "movaps        %%xmm0, (%2,%0) \n"
00195         "sub $16, %1 \n"
00196         "add $16, %0 \n"
00197         "jl 1b \n"
00198         :"+r"(j), "+r"(k)
00199         :"r"(output+n4), "r"(output+n4*3)
00200     );
00201 }
00202