git.sesse.net Git - ffmpeg/blob - libavcodec/x86/mpegvideo_mmx_template.c

   1 /*
   2  * MPEG video MMX templates
   3  *
   4  * Copyright (c) 2002 Michael Niedermayer <michaelni@gmx.at>
   5  *
   6  * This file is part of Libav.
   7  *
   8  * Libav is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * Libav is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with Libav; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #undef MMREG_WIDTH
  24 #undef MM
  25 #undef MOVQ
  26 #undef SPREADW
  27 #undef PMAXW
  28 #undef PMAX
  29 #undef SAVE_SIGN
  30 #undef RESTORE_SIGN
  31
  32 #if HAVE_SSE2
  33 #define MMREG_WIDTH "16"
  34 #define MM "%%xmm"
  35 #define MOVQ "movdqa"
  36 #define SPREADW(a) \
  37             "pshuflw $0, "a", "a"       \n\t"\
  38             "punpcklwd "a", "a"         \n\t"
  39 #define PMAXW(a,b) "pmaxsw "a", "b"     \n\t"
  40 #define PMAX(a,b) \
  41             "movhlps "a", "b"           \n\t"\
  42             PMAXW(b, a)\
  43             "pshuflw $0x0E, "a", "b"    \n\t"\
  44             PMAXW(b, a)\
  45             "pshuflw $0x01, "a", "b"    \n\t"\
  46             PMAXW(b, a)
  47 #else
  48 #define MMREG_WIDTH "8"
  49 #define MM "%%mm"
  50 #define MOVQ "movq"
  51 #if HAVE_MMX2
  52 #define SPREADW(a) "pshufw $0, "a", "a" \n\t"
  53 #define PMAXW(a,b) "pmaxsw "a", "b"     \n\t"
  54 #define PMAX(a,b) \
  55             "pshufw $0x0E, "a", "b"     \n\t"\
  56             PMAXW(b, a)\
  57             "pshufw $0x01, "a", "b"     \n\t"\
  58             PMAXW(b, a)
  59 #else
  60 #define SPREADW(a) \
  61             "punpcklwd "a", "a"         \n\t"\
  62             "punpcklwd "a", "a"         \n\t"
  63 #define PMAXW(a,b) \
  64             "psubusw "a", "b"           \n\t"\
  65             "paddw "a", "b"             \n\t"
  66 #define PMAX(a,b)  \
  67             "movq "a", "b"              \n\t"\
  68             "psrlq $32, "a"             \n\t"\
  69             PMAXW(b, a)\
  70             "movq "a", "b"              \n\t"\
  71             "psrlq $16, "a"             \n\t"\
  72             PMAXW(b, a)
  73
  74 #endif
  75 #endif
  76
  77 #if HAVE_SSSE3
  78 #define SAVE_SIGN(a,b) \
  79             "movdqa "b", "a"            \n\t"\
  80             "pabsw  "b", "b"            \n\t"
  81 #define RESTORE_SIGN(a,b) \
  82             "psignw "a", "b"            \n\t"
  83 #else
  84 #define SAVE_SIGN(a,b) \
  85             "pxor "a", "a"              \n\t"\
  86             "pcmpgtw "b", "a"           \n\t" /* block[i] <= 0 ? 0xFF : 0x00 */\
  87             "pxor "a", "b"              \n\t"\
  88             "psubw "a", "b"             \n\t" /* ABS(block[i]) */
  89 #define RESTORE_SIGN(a,b) \
  90             "pxor "a", "b"              \n\t"\
  91             "psubw "a", "b"             \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
  92 #endif
  93
  94 static int RENAME(dct_quantize)(MpegEncContext *s,
  95                             DCTELEM *block, int n,
  96                             int qscale, int *overflow)
  97 {
  98     x86_reg last_non_zero_p1;
  99     int level=0, q; //=0 is because gcc says uninitialized ...
 100     const uint16_t *qmat, *bias;
 101     DECLARE_ALIGNED(16, int16_t, temp_block)[64];
 102
 103     assert((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
 104
 105     //s->fdct (block);
 106     RENAMEl(ff_fdct) (block); //cannot be anything else ...
 107
 108     if(s->dct_error_sum)
 109         s->denoise_dct(s, block);
 110
 111     if (s->mb_intra) {
 112         int dummy;
 113         if (n < 4)
 114             q = s->y_dc_scale;
 115         else
 116             q = s->c_dc_scale;
 117         /* note: block[0] is assumed to be positive */
 118         if (!s->h263_aic) {
 119 #if 1
 120         __asm__ volatile (
 121                 "mul %%ecx                \n\t"
 122                 : "=d" (level), "=a"(dummy)
 123                 : "a" ((block[0]>>2) + q), "c" (ff_inverse[q<<1])
 124         );
 125 #else
 126         __asm__ volatile (
 127                 "xorl %%edx, %%edx        \n\t"
 128                 "divw %%cx                \n\t"
 129                 "movzwl %%ax, %%eax       \n\t"
 130                 : "=a" (level)
 131                 : "a" ((block[0]>>2) + q), "c" (q<<1)
 132                 : "%edx"
 133         );
 134 #endif
 135         } else
 136             /* For AIC we skip quant/dequant of INTRADC */
 137             level = (block[0] + 4)>>3;
 138
 139         block[0]=0; //avoid fake overflow
 140 //        temp_block[0] = (block[0] + (q >> 1)) / q;
 141         last_non_zero_p1 = 1;
 142         bias = s->q_intra_matrix16[qscale][1];
 143         qmat = s->q_intra_matrix16[qscale][0];
 144     } else {
 145         last_non_zero_p1 = 0;
 146         bias = s->q_inter_matrix16[qscale][1];
 147         qmat = s->q_inter_matrix16[qscale][0];
 148     }
 149
 150     if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
 151
 152         __asm__ volatile(
 153             "movd %%"REG_a", "MM"3              \n\t" // last_non_zero_p1
 154             SPREADW(MM"3")
 155             "pxor "MM"7, "MM"7                  \n\t" // 0
 156             "pxor "MM"4, "MM"4                  \n\t" // 0
 157             MOVQ" (%2), "MM"5                   \n\t" // qmat[0]
 158             "pxor "MM"6, "MM"6                  \n\t"
 159             "psubw (%3), "MM"6                  \n\t" // -bias[0]
 160             "mov $-128, %%"REG_a"               \n\t"
 161             ".p2align 4                         \n\t"
 162             "1:                                 \n\t"
 163             MOVQ" (%1, %%"REG_a"), "MM"0        \n\t" // block[i]
 164             SAVE_SIGN(MM"1", MM"0")                   // ABS(block[i])
 165             "psubusw "MM"6, "MM"0               \n\t" // ABS(block[i]) + bias[0]
 166             "pmulhw "MM"5, "MM"0                \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
 167             "por "MM"0, "MM"4                   \n\t"
 168             RESTORE_SIGN(MM"1", MM"0")                // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
 169             MOVQ" "MM"0, (%5, %%"REG_a")        \n\t"
 170             "pcmpeqw "MM"7, "MM"0               \n\t" // out==0 ? 0xFF : 0x00
 171             MOVQ" (%4, %%"REG_a"), "MM"1        \n\t"
 172             MOVQ" "MM"7, (%1, %%"REG_a")        \n\t" // 0
 173             "pandn "MM"1, "MM"0                 \n\t"
 174             PMAXW(MM"0", MM"3")
 175             "add $"MMREG_WIDTH", %%"REG_a"      \n\t"
 176             " js 1b                             \n\t"
 177             PMAX(MM"3", MM"0")
 178             "movd "MM"3, %%"REG_a"              \n\t"
 179             "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
 180             : "+a" (last_non_zero_p1)
 181             : "r" (block+64), "r" (qmat), "r" (bias),
 182               "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
 183               XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
 184                                 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
 185         );
 186     }else{ // FMT_H263
 187         __asm__ volatile(
 188             "movd %%"REG_a", "MM"3              \n\t" // last_non_zero_p1
 189             SPREADW(MM"3")
 190             "pxor "MM"7, "MM"7                  \n\t" // 0
 191             "pxor "MM"4, "MM"4                  \n\t" // 0
 192             "mov $-128, %%"REG_a"               \n\t"
 193             ".p2align 4                         \n\t"
 194             "1:                                 \n\t"
 195             MOVQ" (%1, %%"REG_a"), "MM"0        \n\t" // block[i]
 196             SAVE_SIGN(MM"1", MM"0")                   // ABS(block[i])
 197             MOVQ" (%3, %%"REG_a"), "MM"6        \n\t" // bias[0]
 198             "paddusw "MM"6, "MM"0               \n\t" // ABS(block[i]) + bias[0]
 199             MOVQ" (%2, %%"REG_a"), "MM"5        \n\t" // qmat[i]
 200             "pmulhw "MM"5, "MM"0                \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
 201             "por "MM"0, "MM"4                   \n\t"
 202             RESTORE_SIGN(MM"1", MM"0")                // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
 203             MOVQ" "MM"0, (%5, %%"REG_a")        \n\t"
 204             "pcmpeqw "MM"7, "MM"0               \n\t" // out==0 ? 0xFF : 0x00
 205             MOVQ" (%4, %%"REG_a"), "MM"1        \n\t"
 206             MOVQ" "MM"7, (%1, %%"REG_a")        \n\t" // 0
 207             "pandn "MM"1, "MM"0                 \n\t"
 208             PMAXW(MM"0", MM"3")
 209             "add $"MMREG_WIDTH", %%"REG_a"      \n\t"
 210             " js 1b                             \n\t"
 211             PMAX(MM"3", MM"0")
 212             "movd "MM"3, %%"REG_a"              \n\t"
 213             "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
 214             : "+a" (last_non_zero_p1)
 215             : "r" (block+64), "r" (qmat+64), "r" (bias+64),
 216               "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
 217               XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
 218                                 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
 219         );
 220     }
 221     __asm__ volatile(
 222         "movd %1, "MM"1                     \n\t" // max_qcoeff
 223         SPREADW(MM"1")
 224         "psubusw "MM"1, "MM"4               \n\t"
 225         "packuswb "MM"4, "MM"4              \n\t"
 226 #if HAVE_SSE2
 227         "packuswb "MM"4, "MM"4              \n\t"
 228 #endif
 229         "movd "MM"4, %0                     \n\t" // *overflow
 230         : "=g" (*overflow)
 231         : "g" (s->max_qcoeff)
 232     );
 233
 234     if(s->mb_intra) block[0]= level;
 235     else            block[0]= temp_block[0];
 236
 237     if(s->dsp.idct_permutation_type == FF_SIMPLE_IDCT_PERM){
 238         if(last_non_zero_p1 <= 1) goto end;
 239         block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
 240         block[0x20] = temp_block[0x10];
 241         if(last_non_zero_p1 <= 4) goto end;
 242         block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
 243         block[0x09] = temp_block[0x03];
 244         if(last_non_zero_p1 <= 7) goto end;
 245         block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
 246         block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
 247         if(last_non_zero_p1 <= 11) goto end;
 248         block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
 249         block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
 250         block[0x0C] = temp_block[0x05];
 251         if(last_non_zero_p1 <= 16) goto end;
 252         block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
 253         block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
 254         block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
 255         block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
 256         if(last_non_zero_p1 <= 24) goto end;
 257         block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
 258         block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
 259         block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
 260         block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
 261         if(last_non_zero_p1 <= 32) goto end;
 262         block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
 263         block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
 264         block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
 265         block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
 266         if(last_non_zero_p1 <= 40) goto end;
 267         block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
 268         block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
 269         block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
 270         block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
 271         if(last_non_zero_p1 <= 48) goto end;
 272         block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
 273         block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
 274         block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
 275         block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
 276         if(last_non_zero_p1 <= 56) goto end;
 277         block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
 278         block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
 279         block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
 280         block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
 281     }else if(s->dsp.idct_permutation_type == FF_LIBMPEG2_IDCT_PERM){
 282         if(last_non_zero_p1 <= 1) goto end;
 283         block[0x04] = temp_block[0x01];
 284         block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
 285         if(last_non_zero_p1 <= 4) goto end;
 286         block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
 287         block[0x05] = temp_block[0x03];
 288         if(last_non_zero_p1 <= 7) goto end;
 289         block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
 290         block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
 291         if(last_non_zero_p1 <= 11) goto end;
 292         block[0x1C] = temp_block[0x19];
 293         block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
 294         block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
 295         if(last_non_zero_p1 <= 16) goto end;
 296         block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
 297         block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
 298         block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
 299         block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
 300         if(last_non_zero_p1 <= 24) goto end;
 301         block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
 302         block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
 303         block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
 304         block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
 305         if(last_non_zero_p1 <= 32) goto end;
 306         block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
 307         block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
 308         block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
 309         block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
 310         if(last_non_zero_p1 <= 40) goto end;
 311         block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
 312         block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
 313         block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
 314         block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
 315         if(last_non_zero_p1 <= 48) goto end;
 316         block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
 317         block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
 318             block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
 319         block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
 320         if(last_non_zero_p1 <= 56) goto end;
 321         block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
 322         block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
 323         block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
 324         block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
 325     }else{
 326         if(last_non_zero_p1 <= 1) goto end;
 327         block[0x01] = temp_block[0x01];
 328         block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
 329         if(last_non_zero_p1 <= 4) goto end;
 330         block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
 331         block[0x03] = temp_block[0x03];
 332         if(last_non_zero_p1 <= 7) goto end;
 333         block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
 334         block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
 335         if(last_non_zero_p1 <= 11) goto end;
 336         block[0x19] = temp_block[0x19];
 337         block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
 338         block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
 339         if(last_non_zero_p1 <= 16) goto end;
 340         block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
 341         block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
 342         block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
 343         block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
 344         if(last_non_zero_p1 <= 24) goto end;
 345         block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
 346         block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
 347         block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
 348         block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
 349         if(last_non_zero_p1 <= 32) goto end;
 350         block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
 351         block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
 352         block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
 353         block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
 354         if(last_non_zero_p1 <= 40) goto end;
 355         block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
 356         block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
 357         block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
 358         block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
 359         if(last_non_zero_p1 <= 48) goto end;
 360         block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
 361         block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
 362         block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
 363         block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
 364         if(last_non_zero_p1 <= 56) goto end;
 365         block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
 366         block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
 367         block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
 368         block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
 369     }
 370     end:
 371 /*
 372     for(i=0; i<last_non_zero_p1; i++)
 373     {
 374        int j= zigzag_direct_noperm[i];
 375        block[block_permute_op(j)]= temp_block[j];
 376     }
 377 */
 378
 379     return last_non_zero_p1 - 1;
 380 }