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#include "fft.h" 25 26DECLARE_ALIGNED(16, static const int, m1m1m1m1)[4] = 27 { 1 << 31, 1 << 31, 1 << 31, 1 << 31 }; 28 29void ff_fft_dispatch_sse(FFTComplex *z, int nbits); 30void ff_fft_dispatch_interleave_sse(FFTComplex *z, int nbits); 31 32void ff_fft_calc_sse(FFTContext *s, FFTComplex *z) 33{ 34 int n = 1 << s->nbits; 35 36 ff_fft_dispatch_interleave_sse(z, s->nbits); 37 38 if(n <= 16) { 39 x86_reg i = -8*n; 40 __asm__ volatile( 41 "1: \n" 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" 48 "add $32, %0 \n" 49 "jl 1b \n" 50 :"+r"(i) 51 :"r"(z+n) 52 :"memory" 53 ); 54 } 55} 56 57void ff_fft_permute_sse(FFTContext *s, FFTComplex *z) 58{ 59 int n = 1 << s->nbits; 60 int i; 61 for(i=0; i<n; i+=2) { 62 __asm__ volatile( 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]]) 68 :"m"(z[i]) 69 ); 70 } 71 memcpy(z, s->tmp_buf, n*sizeof(FFTComplex)); 72} 73 74void ff_imdct_half_sse(FFTContext *s, FFTSample *output, const FFTSample *input) 75{ 76 av_unused x86_reg i, j, k, l; 77 long n = 1 << s->mdct_bits; 78 long n2 = n >> 1; 79 long n4 = n >> 2; 80 long n8 = n >> 3; 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; 85 86 /* pre rotation */ 87 for(k=n8-2; k>=0; k-=2) { 88 __asm__ volatile( 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) 111 ); 112#if ARCH_X86_64 113 // 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]]) 123 ); 124#else 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]])); 129#endif 130 } 131 132 ff_fft_dispatch_sse(z, s->nbits); 133 134 /* post rotation + reinterleave + reorder */ 135 136#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" 147 148 j = -n2; 149 k = n2-16; 150 __asm__ volatile( 151 "1: \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" 166 "sub $16, %1 \n" 167 "add $16, %0 \n" 168 "jl 1b \n" 169 :"+&r"(j), "+&r"(k) 170 :"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8) 171 :"memory" 172 ); 173} 174 175void ff_imdct_calc_sse(FFTContext *s, FFTSample *output, const FFTSample *input) 176{ 177 x86_reg j, k; 178 long n = 1 << s->mdct_bits; 179 long n4 = n >> 2; 180 181 ff_imdct_half_sse(s, output+n4, input); 182 183 j = -n; 184 k = n-16; 185 __asm__ volatile( 186 "movaps %4, %%xmm7 \n" 187 "1: \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" 195 "sub $16, %1 \n" 196 "add $16, %0 \n" 197 "jl 1b \n" 198 :"+r"(j), "+r"(k) 199 :"r"(output+n4), "r"(output+n4*3), 200 "m"(*m1m1m1m1) 201 ); 202} 203 204