libavcodec/i386/fft_sse.c

   1 /*
   2  * FFT/MDCT transform with SSE optimizations
   3  * Copyright (c) 2002 Fabrice Bellard.
   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
  25 static const int p1p1p1m1[4] __attribute__((aligned(16))) =
  26     { 0, 0, 0, 1 << 31 };
  27
  28 static const int p1p1m1p1[4] __attribute__((aligned(16))) =
  29     { 0, 0, 1 << 31, 0 };
  30
  31 static const int p1p1m1m1[4] __attribute__((aligned(16))) =
  32     { 0, 0, 1 << 31, 1 << 31 };
  33
  34 static const int p1m1p1m1[4] __attribute__((aligned(16))) =
  35     { 0, 1 << 31, 0, 1 << 31 };
  36
  37 static const int m1m1m1m1[4] __attribute__((aligned(16))) =
  38     { 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
  39
  40 #if 0
  41 static void print_v4sf(const char *str, __m128 a)
  42 {
  43     float *p = (float *)&a;
  44     printf("%s: %f %f %f %f\n",
  45            str, p[0], p[1], p[2], p[3]);
  46 }
  47 #endif
  48
  49 /* XXX: handle reverse case */
  50 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
  51 {
  52     int ln = s->nbits;
  53     x86_reg i;
  54     long j;
  55     long nblocks, nloops;
  56     FFTComplex *p, *cptr;
  57
  58     asm volatile(
  59         "movaps %0, %%xmm4 \n\t"
  60         "movaps %1, %%xmm5 \n\t"
  61         ::"m"(*p1p1m1m1),
  62           "m"(*(s->inverse ? p1p1m1p1 : p1p1p1m1))
  63     );
  64
  65     i = 8 << ln;
  66     asm volatile(
  67         "1: \n\t"
  68         "sub $32, %0 \n\t"
  69         /* do the pass 0 butterfly */
  70         "movaps   (%0,%1), %%xmm0 \n\t"
  71         "movaps    %%xmm0, %%xmm1 \n\t"
  72         "shufps     $0x4E, %%xmm0, %%xmm0 \n\t"
  73         "xorps     %%xmm4, %%xmm1 \n\t"
  74         "addps     %%xmm1, %%xmm0 \n\t"
  75         "movaps 16(%0,%1), %%xmm2 \n\t"
  76         "movaps    %%xmm2, %%xmm3 \n\t"
  77         "shufps     $0x4E, %%xmm2, %%xmm2 \n\t"
  78         "xorps     %%xmm4, %%xmm3 \n\t"
  79         "addps     %%xmm3, %%xmm2 \n\t"
  80         /* multiply third by -i */
  81         /* by toggling the sign bit */
  82         "shufps     $0xB4, %%xmm2, %%xmm2 \n\t"
  83         "xorps     %%xmm5, %%xmm2 \n\t"
  84         /* do the pass 1 butterfly */
  85         "movaps    %%xmm0, %%xmm1 \n\t"
  86         "addps     %%xmm2, %%xmm0 \n\t"
  87         "subps     %%xmm2, %%xmm1 \n\t"
  88         "movaps    %%xmm0,   (%0,%1) \n\t"
  89         "movaps    %%xmm1, 16(%0,%1) \n\t"
  90         "jg 1b \n\t"
  91         :"+r"(i)
  92         :"r"(z)
  93     );
  94     /* pass 2 .. ln-1 */
  95
  96     nblocks = 1 << (ln-3);
  97     nloops = 1 << 2;
  98     cptr = s->exptab1;
  99     do {
 100         p = z;
 101         j = nblocks;
 102         do {
 103             i = nloops*8;
 104             asm volatile(
 105                 "1: \n\t"
 106                 "sub $32, %0 \n\t"
 107                 "movaps    (%2,%0), %%xmm1 \n\t"
 108                 "movaps    (%1,%0), %%xmm0 \n\t"
 109                 "movaps  16(%2,%0), %%xmm5 \n\t"
 110                 "movaps  16(%1,%0), %%xmm4 \n\t"
 111                 "movaps     %%xmm1, %%xmm2 \n\t"
 112                 "movaps     %%xmm5, %%xmm6 \n\t"
 113                 "shufps      $0xA0, %%xmm1, %%xmm1 \n\t"
 114                 "shufps      $0xF5, %%xmm2, %%xmm2 \n\t"
 115                 "shufps      $0xA0, %%xmm5, %%xmm5 \n\t"
 116                 "shufps      $0xF5, %%xmm6, %%xmm6 \n\t"
 117                 "mulps   (%3,%0,2), %%xmm1 \n\t" //  cre*re cim*re
 118                 "mulps 16(%3,%0,2), %%xmm2 \n\t" // -cim*im cre*im
 119                 "mulps 32(%3,%0,2), %%xmm5 \n\t" //  cre*re cim*re
 120                 "mulps 48(%3,%0,2), %%xmm6 \n\t" // -cim*im cre*im
 121                 "addps      %%xmm2, %%xmm1 \n\t"
 122                 "addps      %%xmm6, %%xmm5 \n\t"
 123                 "movaps     %%xmm0, %%xmm3 \n\t"
 124                 "movaps     %%xmm4, %%xmm7 \n\t"
 125                 "addps      %%xmm1, %%xmm0 \n\t"
 126                 "subps      %%xmm1, %%xmm3 \n\t"
 127                 "addps      %%xmm5, %%xmm4 \n\t"
 128                 "subps      %%xmm5, %%xmm7 \n\t"
 129                 "movaps     %%xmm0, (%1,%0) \n\t"
 130                 "movaps     %%xmm3, (%2,%0) \n\t"
 131                 "movaps     %%xmm4, 16(%1,%0) \n\t"
 132                 "movaps     %%xmm7, 16(%2,%0) \n\t"
 133                 "jg 1b \n\t"
 134                 :"+r"(i)
 135                 :"r"(p), "r"(p + nloops), "r"(cptr)
 136             );
 137             p += nloops*2;
 138         } while (--j);
 139         cptr += nloops*2;
 140         nblocks >>= 1;
 141         nloops <<= 1;
 142     } while (nblocks != 0);
 143 }
 144
 145 void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output,
 146                        const FFTSample *input, FFTSample *tmp)
 147 {
 148     x86_reg k;
 149     long n8, n4, n2, n;
 150     const uint16_t *revtab = s->fft.revtab;
 151     const FFTSample *tcos = s->tcos;
 152     const FFTSample *tsin = s->tsin;
 153     const FFTSample *in1, *in2;
 154     FFTComplex *z = (FFTComplex *)tmp;
 155
 156     n = 1 << s->nbits;
 157     n2 = n >> 1;
 158     n4 = n >> 2;
 159     n8 = n >> 3;
 160
 161 #ifdef ARCH_X86_64
 162     asm volatile ("movaps %0, %%xmm8\n\t"::"m"(*p1m1p1m1));
 163 #define P1M1P1M1 "%%xmm8"
 164 #else
 165 #define P1M1P1M1 "%4"
 166 #endif
 167
 168     /* pre rotation */
 169     in1 = input;
 170     in2 = input + n2 - 4;
 171
 172     /* Complex multiplication */
 173     for (k = 0; k < n4; k += 4) {
 174         asm volatile (
 175             "movaps          %0, %%xmm0 \n\t"   // xmm0 = r0 X  r1 X : in2
 176             "movaps          %1, %%xmm3 \n\t"   // xmm3 = X  i1 X  i0: in1
 177             "movaps    -16+1*%0, %%xmm4 \n\t"   // xmm4 = r0 X  r1 X : in2
 178             "movaps     16+1*%1, %%xmm7 \n\t"   // xmm7 = X  i1 X  i0: in1
 179             "movlps          %2, %%xmm1 \n\t"   // xmm1 = X  X  R1 R0: tcos
 180             "movlps          %3, %%xmm2 \n\t"   // xmm2 = X  X  I1 I0: tsin
 181             "movlps      8+1*%2, %%xmm5 \n\t"   // xmm5 = X  X  R1 R0: tcos
 182             "movlps      8+1*%3, %%xmm6 \n\t"   // xmm6 = X  X  I1 I0: tsin
 183             "shufps $95, %%xmm0, %%xmm0 \n\t"   // xmm0 = r1 r1 r0 r0
 184             "shufps $160,%%xmm3, %%xmm3 \n\t"   // xmm3 = i1 i1 i0 i0
 185             "shufps $95, %%xmm4, %%xmm4 \n\t"   // xmm4 = r1 r1 r0 r0
 186             "shufps $160,%%xmm7, %%xmm7 \n\t"   // xmm7 = i1 i1 i0 i0
 187             "unpcklps    %%xmm2, %%xmm1 \n\t"   // xmm1 = I1 R1 I0 R0
 188             "unpcklps    %%xmm6, %%xmm5 \n\t"   // xmm5 = I1 R1 I0 R0
 189             "movaps      %%xmm1, %%xmm2 \n\t"   // xmm2 = I1 R1 I0 R0
 190             "movaps      %%xmm5, %%xmm6 \n\t"   // xmm6 = I1 R1 I0 R0
 191             "xorps   "P1M1P1M1", %%xmm2 \n\t"   // xmm2 = -I1 R1 -I0 R0
 192             "xorps   "P1M1P1M1", %%xmm6 \n\t"   // xmm6 = -I1 R1 -I0 R0
 193             "mulps       %%xmm1, %%xmm0 \n\t"   // xmm0 = rI rR rI rR
 194             "mulps       %%xmm5, %%xmm4 \n\t"   // xmm4 = rI rR rI rR
 195             "shufps $177,%%xmm2, %%xmm2 \n\t"   // xmm2 = R1 -I1 R0 -I0
 196             "shufps $177,%%xmm6, %%xmm6 \n\t"   // xmm6 = R1 -I1 R0 -I0
 197             "mulps       %%xmm2, %%xmm3 \n\t"   // xmm3 = Ri -Ii Ri -Ii
 198             "mulps       %%xmm6, %%xmm7 \n\t"   // xmm7 = Ri -Ii Ri -Ii
 199             "addps       %%xmm3, %%xmm0 \n\t"   // xmm0 = result
 200             "addps       %%xmm7, %%xmm4 \n\t"   // xmm4 = result
 201             ::"m"(in2[-2*k]), "m"(in1[2*k]),
 202               "m"(tcos[k]), "m"(tsin[k])
 203 #ifndef ARCH_X86_64
 204               ,"m"(*p1m1p1m1)
 205 #endif
 206         );
 207         /* Should be in the same block, hack for gcc2.95 & gcc3 */
 208         asm (
 209             "movlps      %%xmm0, %0     \n\t"
 210             "movhps      %%xmm0, %1     \n\t"
 211             "movlps      %%xmm4, %2     \n\t"
 212             "movhps      %%xmm4, %3     \n\t"
 213             :"=m"(z[revtab[k]]), "=m"(z[revtab[k + 1]]),
 214              "=m"(z[revtab[k + 2]]), "=m"(z[revtab[k + 3]])
 215         );
 216     }
 217
 218     ff_fft_calc_sse(&s->fft, z);
 219
 220 #ifndef ARCH_X86_64
 221 #undef P1M1P1M1
 222 #define P1M1P1M1 "%3"
 223 #endif
 224
 225     /* post rotation + reordering */
 226     for (k = 0; k < n4; k += 4) {
 227         asm (
 228             "movaps          %0, %%xmm0 \n\t"   // xmm0 = i1 r1 i0 r0: z
 229             "movaps     16+1*%0, %%xmm4 \n\t"   // xmm4 = i1 r1 i0 r0: z
 230             "movlps          %1, %%xmm1 \n\t"   // xmm1 = X  X  R1 R0: tcos
 231             "movlps      8+1*%1, %%xmm5 \n\t"   // xmm5 = X  X  R1 R0: tcos
 232             "movaps      %%xmm0, %%xmm3 \n\t"   // xmm3 = i1 r1 i0 r0
 233             "movaps      %%xmm4, %%xmm7 \n\t"   // xmm7 = i1 r1 i0 r0
 234             "movlps          %2, %%xmm2 \n\t"   // xmm2 = X  X  I1 I0: tsin
 235             "movlps      8+1*%2, %%xmm6 \n\t"   // xmm6 = X  X  I1 I0: tsin
 236             "shufps $160,%%xmm0, %%xmm0 \n\t"   // xmm0 = r1 r1 r0 r0
 237             "shufps $245,%%xmm3, %%xmm3 \n\t"   // xmm3 = i1 i1 i0 i0
 238             "shufps $160,%%xmm4, %%xmm4 \n\t"   // xmm4 = r1 r1 r0 r0
 239             "shufps $245,%%xmm7, %%xmm7 \n\t"   // xmm7 = i1 i1 i0 i0
 240             "unpcklps    %%xmm2, %%xmm1 \n\t"   // xmm1 = I1 R1 I0 R0
 241             "unpcklps    %%xmm6, %%xmm5 \n\t"   // xmm5 = I1 R1 I0 R0
 242             "movaps      %%xmm1, %%xmm2 \n\t"   // xmm2 = I1 R1 I0 R0
 243             "movaps      %%xmm5, %%xmm6 \n\t"   // xmm6 = I1 R1 I0 R0
 244             "xorps   "P1M1P1M1", %%xmm2 \n\t"   // xmm2 = -I1 R1 -I0 R0
 245             "mulps       %%xmm1, %%xmm0 \n\t"   // xmm0 = rI rR rI rR
 246             "xorps   "P1M1P1M1", %%xmm6 \n\t"   // xmm6 = -I1 R1 -I0 R0
 247             "mulps       %%xmm5, %%xmm4 \n\t"   // xmm4 = rI rR rI rR
 248             "shufps $177,%%xmm2, %%xmm2 \n\t"   // xmm2 = R1 -I1 R0 -I0
 249             "shufps $177,%%xmm6, %%xmm6 \n\t"   // xmm6 = R1 -I1 R0 -I0
 250             "mulps       %%xmm2, %%xmm3 \n\t"   // xmm3 = Ri -Ii Ri -Ii
 251             "mulps       %%xmm6, %%xmm7 \n\t"   // xmm7 = Ri -Ii Ri -Ii
 252             "addps       %%xmm3, %%xmm0 \n\t"   // xmm0 = result
 253             "addps       %%xmm7, %%xmm4 \n\t"   // xmm4 = result
 254             "movaps      %%xmm0, %0     \n\t"
 255             "movaps      %%xmm4, 16+1*%0\n\t"
 256             :"+m"(z[k])
 257             :"m"(tcos[k]), "m"(tsin[k])
 258 #ifndef ARCH_X86_64
 259              ,"m"(*p1m1p1m1)
 260 #endif
 261         );
 262     }
 263
 264     /*
 265        Mnemonics:
 266        0 = z[k].re
 267        1 = z[k].im
 268        2 = z[k + 1].re
 269        3 = z[k + 1].im
 270        4 = z[-k - 2].re
 271        5 = z[-k - 2].im
 272        6 = z[-k - 1].re
 273        7 = z[-k - 1].im
 274     */
 275     k = 16-n;
 276     asm volatile("movaps %0, %%xmm7 \n\t"::"m"(*m1m1m1m1));
 277     asm volatile(
 278         "1: \n\t"
 279         "movaps  -16(%4,%0), %%xmm1 \n\t"   // xmm1 = 4 5 6 7 = z[-2-k]
 280         "neg %0 \n\t"
 281         "movaps     (%4,%0), %%xmm0 \n\t"   // xmm0 = 0 1 2 3 = z[k]
 282         "xorps       %%xmm7, %%xmm0 \n\t"   // xmm0 = -0 -1 -2 -3
 283         "movaps      %%xmm0, %%xmm2 \n\t"   // xmm2 = -0 -1 -2 -3
 284         "shufps $141,%%xmm1, %%xmm0 \n\t"   // xmm0 = -1 -3 4 6
 285         "shufps $216,%%xmm1, %%xmm2 \n\t"   // xmm2 = -0 -2 5 7
 286         "shufps $156,%%xmm0, %%xmm0 \n\t"   // xmm0 = -1 6 -3 4 !
 287         "shufps $156,%%xmm2, %%xmm2 \n\t"   // xmm2 = -0 7 -2 5 !
 288         "movaps      %%xmm0, (%1,%0) \n\t"  // output[2*k]
 289         "movaps      %%xmm2, (%2,%0) \n\t"  // output[n2+2*k]
 290         "neg %0 \n\t"
 291         "shufps $27, %%xmm0, %%xmm0 \n\t"   // xmm0 = 4 -3 6 -1
 292         "xorps       %%xmm7, %%xmm0 \n\t"   // xmm0 = -4 3 -6 1 !
 293         "shufps $27, %%xmm2, %%xmm2 \n\t"   // xmm2 = 5 -2 7 -0 !
 294         "movaps      %%xmm0, -16(%2,%0) \n\t" // output[n2-4-2*k]
 295         "movaps      %%xmm2, -16(%3,%0) \n\t" // output[n-4-2*k]
 296         "add $16, %0 \n\t"
 297         "jle 1b \n\t"
 298         :"+r"(k)
 299         :"r"(output), "r"(output+n2), "r"(output+n), "r"(z+n8)
 300         :"memory"
 301     );
 302 }
 303