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 
00025 DECLARE_ASM_CONST(16, int, m1m1m1m1)[4] =
00026     { 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
00027 
00028 void ff_fft_dispatch_sse(FFTComplex *z, int nbits);
00029 void ff_fft_dispatch_interleave_sse(FFTComplex *z, int nbits);
00030 
00031 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
00032 {
00033     int n = 1 << s->nbits;
00034 
00035     ff_fft_dispatch_interleave_sse(z, s->nbits);
00036 
00037     if(n <= 16) {
00038         x86_reg i = -8*n;
00039         __asm__ volatile(
00040             "1: \n"
00041             "movaps     (%0,%1), %%xmm0 \n"
00042             "movaps      %%xmm0, %%xmm1 \n"
00043             "unpcklps 16(%0,%1), %%xmm0 \n"
00044             "unpckhps 16(%0,%1), %%xmm1 \n"
00045             "movaps      %%xmm0,   (%0,%1) \n"
00046             "movaps      %%xmm1, 16(%0,%1) \n"
00047             "add $32, %0 \n"
00048             "jl 1b \n"
00049             :"+r"(i)
00050             :"r"(z+n)
00051             :"memory"
00052         );
00053     }
00054 }
00055 
00056 void ff_fft_permute_sse(FFTContext *s, FFTComplex *z)
00057 {
00058     int n = 1 << s->nbits;
00059     int i;
00060     for(i=0; i<n; i+=2) {
00061         __asm__ volatile(
00062             "movaps %2, %%xmm0 \n"
00063             "movlps %%xmm0, %0 \n"
00064             "movhps %%xmm0, %1 \n"
00065             :"=m"(s->tmp_buf[s->revtab[i]]),
00066              "=m"(s->tmp_buf[s->revtab[i+1]])
00067             :"m"(z[i])
00068         );
00069     }
00070     memcpy(z, s->tmp_buf, n*sizeof(FFTComplex));
00071 }
00072 
00073 void ff_imdct_half_sse(MDCTContext *s, FFTSample *output, const FFTSample *input)
00074 {
00075     av_unused x86_reg i, j, k, l;
00076     long n = 1 << s->nbits;
00077     long n2 = n >> 1;
00078     long n4 = n >> 2;
00079     long n8 = n >> 3;
00080     const uint16_t *revtab = s->fft.revtab + n8;
00081     const FFTSample *tcos = s->tcos;
00082     const FFTSample *tsin = s->tsin;
00083     FFTComplex *z = (FFTComplex *)output;
00084 
00085     /* pre rotation */
00086     for(k=n8-2; k>=0; k-=2) {
00087         __asm__ volatile(
00088             "movaps     (%2,%1,2), %%xmm0 \n" // { z[k].re,    z[k].im,    z[k+1].re,  z[k+1].im  }
00089             "movaps  -16(%2,%0,2), %%xmm1 \n" // { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im }
00090             "movaps        %%xmm0, %%xmm2 \n"
00091             "shufps $0x88, %%xmm1, %%xmm0 \n" // { z[k].re,    z[k+1].re,  z[-k-2].re, z[-k-1].re }
00092             "shufps $0x77, %%xmm2, %%xmm1 \n" // { z[-k-1].im, z[-k-2].im, z[k+1].im,  z[k].im    }
00093             "movlps       (%3,%1), %%xmm4 \n"
00094             "movlps       (%4,%1), %%xmm5 \n"
00095             "movhps     -8(%3,%0), %%xmm4 \n" // { cos[k],     cos[k+1],   cos[-k-2],  cos[-k-1]  }
00096             "movhps     -8(%4,%0), %%xmm5 \n" // { sin[k],     sin[k+1],   sin[-k-2],  sin[-k-1]  }
00097             "movaps        %%xmm0, %%xmm2 \n"
00098             "movaps        %%xmm1, %%xmm3 \n"
00099             "mulps         %%xmm5, %%xmm0 \n" // re*sin
00100             "mulps         %%xmm4, %%xmm1 \n" // im*cos
00101             "mulps         %%xmm4, %%xmm2 \n" // re*cos
00102             "mulps         %%xmm5, %%xmm3 \n" // im*sin
00103             "subps         %%xmm0, %%xmm1 \n" // -> re
00104             "addps         %%xmm3, %%xmm2 \n" // -> im
00105             "movaps        %%xmm1, %%xmm0 \n"
00106             "unpcklps      %%xmm2, %%xmm1 \n" // { z[k],    z[k+1]  }
00107             "unpckhps      %%xmm2, %%xmm0 \n" // { z[-k-2], z[-k-1] }
00108             ::"r"(-4*k), "r"(4*k),
00109               "r"(input+n4), "r"(tcos+n8), "r"(tsin+n8)
00110         );
00111 #if ARCH_X86_64
00112         // if we have enough regs, don't let gcc make the luts latency-bound
00113         // but if not, latency is faster than spilling
00114         __asm__("movlps %%xmm0, %0 \n"
00115             "movhps %%xmm0, %1 \n"
00116             "movlps %%xmm1, %2 \n"
00117             "movhps %%xmm1, %3 \n"
00118             :"=m"(z[revtab[-k-2]]),
00119              "=m"(z[revtab[-k-1]]),
00120              "=m"(z[revtab[ k  ]]),
00121              "=m"(z[revtab[ k+1]])
00122         );
00123 #else
00124         __asm__("movlps %%xmm0, %0" :"=m"(z[revtab[-k-2]]));
00125         __asm__("movhps %%xmm0, %0" :"=m"(z[revtab[-k-1]]));
00126         __asm__("movlps %%xmm1, %0" :"=m"(z[revtab[ k  ]]));
00127         __asm__("movhps %%xmm1, %0" :"=m"(z[revtab[ k+1]]));
00128 #endif
00129     }
00130 
00131     ff_fft_dispatch_sse(z, s->fft.nbits);
00132 
00133     /* post rotation + reinterleave + reorder */
00134 
00135 #define CMUL(j,xmm0,xmm1)\
00136         "movaps   (%2,"#j",2), %%xmm6 \n"\
00137         "movaps 16(%2,"#j",2), "#xmm0"\n"\
00138         "movaps        %%xmm6, "#xmm1"\n"\
00139         "movaps        "#xmm0",%%xmm7 \n"\
00140         "mulps      (%3,"#j"), %%xmm6 \n"\
00141         "mulps      (%4,"#j"), "#xmm0"\n"\
00142         "mulps      (%4,"#j"), "#xmm1"\n"\
00143         "mulps      (%3,"#j"), %%xmm7 \n"\
00144         "subps         %%xmm6, "#xmm0"\n"\
00145         "addps         %%xmm7, "#xmm1"\n"
00146 
00147     j = -n2;
00148     k = n2-16;
00149     __asm__ volatile(
00150         "1: \n"
00151         CMUL(%0, %%xmm0, %%xmm1)
00152         CMUL(%1, %%xmm4, %%xmm5)
00153         "shufps    $0x1b, %%xmm1, %%xmm1 \n"
00154         "shufps    $0x1b, %%xmm5, %%xmm5 \n"
00155         "movaps   %%xmm4, %%xmm6 \n"
00156         "unpckhps %%xmm1, %%xmm4 \n"
00157         "unpcklps %%xmm1, %%xmm6 \n"
00158         "movaps   %%xmm0, %%xmm2 \n"
00159         "unpcklps %%xmm5, %%xmm0 \n"
00160         "unpckhps %%xmm5, %%xmm2 \n"
00161         "movaps   %%xmm6,   (%2,%1,2) \n"
00162         "movaps   %%xmm4, 16(%2,%1,2) \n"
00163         "movaps   %%xmm0,   (%2,%0,2) \n"
00164         "movaps   %%xmm2, 16(%2,%0,2) \n"
00165         "sub $16, %1 \n"
00166         "add $16, %0 \n"
00167         "jl 1b \n"
00168         :"+&r"(j), "+&r"(k)
00169         :"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8)
00170         :"memory"
00171     );
00172 }
00173 
00174 void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output, const FFTSample *input)
00175 {
00176     x86_reg j, k;
00177     long n = 1 << s->nbits;
00178     long n4 = n >> 2;
00179 
00180     ff_imdct_half_sse(s, output+n4, input);
00181 
00182     j = -n;
00183     k = n-16;
00184     __asm__ volatile(
00185         "movaps "MANGLE(m1m1m1m1)", %%xmm7 \n"
00186         "1: \n"
00187         "movaps       (%2,%1), %%xmm0 \n"
00188         "movaps       (%3,%0), %%xmm1 \n"
00189         "shufps $0x1b, %%xmm0, %%xmm0 \n"
00190         "shufps $0x1b, %%xmm1, %%xmm1 \n"
00191         "xorps         %%xmm7, %%xmm0 \n"
00192         "movaps        %%xmm1, (%3,%1) \n"
00193         "movaps        %%xmm0, (%2,%0) \n"
00194         "sub $16, %1 \n"
00195         "add $16, %0 \n"
00196         "jl 1b \n"
00197         :"+r"(j), "+r"(k)
00198         :"r"(output+n4), "r"(output+n4*3)
00199     );
00200 }
00201