libmpcodecs/vf_ilpack.c

   1 /*
   2  * This file is part of MPlayer.
   3  *
   4  * MPlayer is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * MPlayer is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along
  15  * with MPlayer; if not, write to the Free Software Foundation, Inc.,
  16  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  17  */
  18
  19 #include <stdio.h>
  20 #include <stdlib.h>
  21 #include <string.h>
  22 #include <inttypes.h>
  23
  24 #include "config.h"
  25 #include "mp_msg.h"
  26 #include "cpudetect.h"
  27
  28 #include "img_format.h"
  29 #include "mp_image.h"
  30 #include "vf.h"
  31
  32 typedef void (pack_func_t)(unsigned char *dst, unsigned char *y,
  33         unsigned char *u, unsigned char *v, int w, int us, int vs);
  34
  35 struct vf_priv_s {
  36         int mode;
  37         pack_func_t *pack[2];
  38 };
  39
  40 static void pack_nn_C(unsigned char *dst, unsigned char *y,
  41         unsigned char *u, unsigned char *v, int w)
  42 {
  43         int j;
  44         for (j = w/2; j; j--) {
  45                 *dst++ = *y++;
  46                 *dst++ = *u++;
  47                 *dst++ = *y++;
  48                 *dst++ = *v++;
  49         }
  50 }
  51
  52 static void pack_li_0_C(unsigned char *dst, unsigned char *y,
  53         unsigned char *u, unsigned char *v, int w, int us, int vs)
  54 {
  55         int j;
  56         for (j = w/2; j; j--) {
  57                 *dst++ = *y++;
  58                 *dst++ = (u[us+us] + 7*u[0])>>3;
  59                 *dst++ = *y++;
  60                 *dst++ = (v[vs+vs] + 7*v[0])>>3;
  61                 u++; v++;
  62         }
  63 }
  64
  65 static void pack_li_1_C(unsigned char *dst, unsigned char *y,
  66         unsigned char *u, unsigned char *v, int w, int us, int vs)
  67 {
  68         int j;
  69         for (j = w/2; j; j--) {
  70                 *dst++ = *y++;
  71                 *dst++ = (3*u[us+us] + 5*u[0])>>3;
  72                 *dst++ = *y++;
  73                 *dst++ = (3*v[vs+vs] + 5*v[0])>>3;
  74                 u++; v++;
  75         }
  76 }
  77
  78 #if HAVE_MMX
  79 static void pack_nn_MMX(unsigned char *dst, unsigned char *y,
  80         unsigned char *u, unsigned char *v, int w)
  81 {
  82         __asm__ volatile (""
  83                 ASMALIGN(4)
  84                 "1: \n\t"
  85                 "movq (%0), %%mm1 \n\t"
  86                 "movq (%0), %%mm2 \n\t"
  87                 "movq (%1), %%mm4 \n\t"
  88                 "movq (%2), %%mm6 \n\t"
  89                 "punpcklbw %%mm6, %%mm4 \n\t"
  90                 "punpcklbw %%mm4, %%mm1 \n\t"
  91                 "punpckhbw %%mm4, %%mm2 \n\t"
  92
  93                 "add $8, %0 \n\t"
  94                 "add $4, %1 \n\t"
  95                 "add $4, %2 \n\t"
  96                 "movq %%mm1, (%3) \n\t"
  97                 "movq %%mm2, 8(%3) \n\t"
  98                 "add $16, %3 \n\t"
  99                 "decl %4 \n\t"
 100                 "jnz 1b \n\t"
 101                 "emms \n\t"
 102                 :
 103                 : "r" (y), "r" (u), "r" (v), "r" (dst), "r" (w/8)
 104                 : "memory"
 105                 );
 106         pack_nn_C(dst, y, u, v, (w&7));
 107 }
 108
 109 #if HAVE_EBX_AVAILABLE
 110 static void pack_li_0_MMX(unsigned char *dst, unsigned char *y,
 111         unsigned char *u, unsigned char *v, int w, int us, int vs)
 112 {
 113         __asm__ volatile (""
 114                 "push %%"REG_BP" \n\t"
 115 #if ARCH_X86_64
 116                 "mov %6, %%"REG_BP" \n\t"
 117 #else
 118                 "movl 4(%%"REG_d"), %%"REG_BP" \n\t"
 119                 "movl (%%"REG_d"), %%"REG_d" \n\t"
 120 #endif
 121                 "pxor %%mm0, %%mm0 \n\t"
 122
 123                 ASMALIGN(4)
 124                 ".Lli0: \n\t"
 125                 "movq (%%"REG_S"), %%mm1 \n\t"
 126                 "movq (%%"REG_S"), %%mm2 \n\t"
 127
 128                 "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t"
 129                 "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t"
 130                 "punpcklbw %%mm0, %%mm4 \n\t"
 131                 "punpcklbw %%mm0, %%mm6 \n\t"
 132                 "movq (%%"REG_a"), %%mm3 \n\t"
 133                 "movq (%%"REG_b"), %%mm5 \n\t"
 134                 "punpcklbw %%mm0, %%mm3 \n\t"
 135                 "punpcklbw %%mm0, %%mm5 \n\t"
 136                 "paddw %%mm3, %%mm4 \n\t"
 137                 "paddw %%mm5, %%mm6 \n\t"
 138                 "paddw %%mm3, %%mm4 \n\t"
 139                 "paddw %%mm5, %%mm6 \n\t"
 140                 "paddw %%mm3, %%mm4 \n\t"
 141                 "paddw %%mm5, %%mm6 \n\t"
 142                 "paddw %%mm3, %%mm4 \n\t"
 143                 "paddw %%mm5, %%mm6 \n\t"
 144                 "paddw %%mm3, %%mm4 \n\t"
 145                 "paddw %%mm5, %%mm6 \n\t"
 146                 "paddw %%mm3, %%mm4 \n\t"
 147                 "paddw %%mm5, %%mm6 \n\t"
 148                 "paddw %%mm3, %%mm4 \n\t"
 149                 "paddw %%mm5, %%mm6 \n\t"
 150                 "psrlw $3, %%mm4 \n\t"
 151                 "psrlw $3, %%mm6 \n\t"
 152                 "packuswb %%mm4, %%mm4 \n\t"
 153                 "packuswb %%mm6, %%mm6 \n\t"
 154                 "punpcklbw %%mm6, %%mm4 \n\t"
 155                 "punpcklbw %%mm4, %%mm1 \n\t"
 156                 "punpckhbw %%mm4, %%mm2 \n\t"
 157
 158                 "movq %%mm1, (%%"REG_D") \n\t"
 159                 "movq %%mm2, 8(%%"REG_D") \n\t"
 160
 161                 "movq 8(%%"REG_S"), %%mm1 \n\t"
 162                 "movq 8(%%"REG_S"), %%mm2 \n\t"
 163
 164                 "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t"
 165                 "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t"
 166                 "punpckhbw %%mm0, %%mm4 \n\t"
 167                 "punpckhbw %%mm0, %%mm6 \n\t"
 168                 "movq (%%"REG_a"), %%mm3 \n\t"
 169                 "movq (%%"REG_b"), %%mm5 \n\t"
 170                 "punpckhbw %%mm0, %%mm3 \n\t"
 171                 "punpckhbw %%mm0, %%mm5 \n\t"
 172                 "paddw %%mm3, %%mm4 \n\t"
 173                 "paddw %%mm5, %%mm6 \n\t"
 174                 "paddw %%mm3, %%mm4 \n\t"
 175                 "paddw %%mm5, %%mm6 \n\t"
 176                 "paddw %%mm3, %%mm4 \n\t"
 177                 "paddw %%mm5, %%mm6 \n\t"
 178                 "paddw %%mm3, %%mm4 \n\t"
 179                 "paddw %%mm5, %%mm6 \n\t"
 180                 "paddw %%mm3, %%mm4 \n\t"
 181                 "paddw %%mm5, %%mm6 \n\t"
 182                 "paddw %%mm3, %%mm4 \n\t"
 183                 "paddw %%mm5, %%mm6 \n\t"
 184                 "paddw %%mm3, %%mm4 \n\t"
 185                 "paddw %%mm5, %%mm6 \n\t"
 186                 "psrlw $3, %%mm4 \n\t"
 187                 "psrlw $3, %%mm6 \n\t"
 188                 "packuswb %%mm4, %%mm4 \n\t"
 189                 "packuswb %%mm6, %%mm6 \n\t"
 190                 "punpcklbw %%mm6, %%mm4 \n\t"
 191                 "punpcklbw %%mm4, %%mm1 \n\t"
 192                 "punpckhbw %%mm4, %%mm2 \n\t"
 193
 194                 "add $16, %%"REG_S" \n\t"
 195                 "add $8, %%"REG_a" \n\t"
 196                 "add $8, %%"REG_b" \n\t"
 197
 198                 "movq %%mm1, 16(%%"REG_D") \n\t"
 199                 "movq %%mm2, 24(%%"REG_D") \n\t"
 200                 "add $32, %%"REG_D" \n\t"
 201
 202                 "decl %%ecx \n\t"
 203                 "jnz .Lli0 \n\t"
 204                 "emms \n\t"
 205                 "pop %%"REG_BP" \n\t"
 206                 :
 207                 : "S" (y), "D" (dst), "a" (u), "b" (v), "c" (w/16),
 208 #if ARCH_X86_64
 209                 "d" ((x86_reg)us), "r" ((x86_reg)vs)
 210 #else
 211                 "d" (&us)
 212 #endif
 213                 : "memory"
 214                 );
 215         pack_li_0_C(dst, y, u, v, (w&15), us, vs);
 216 }
 217
 218 static void pack_li_1_MMX(unsigned char *dst, unsigned char *y,
 219         unsigned char *u, unsigned char *v, int w, int us, int vs)
 220 {
 221         __asm__ volatile (""
 222                 "push %%"REG_BP" \n\t"
 223 #if ARCH_X86_64
 224                 "mov %6, %%"REG_BP" \n\t"
 225 #else
 226                 "movl 4(%%"REG_d"), %%"REG_BP" \n\t"
 227                 "movl (%%"REG_d"), %%"REG_d" \n\t"
 228 #endif
 229                 "pxor %%mm0, %%mm0 \n\t"
 230
 231                 ASMALIGN(4)
 232                 ".Lli1: \n\t"
 233                 "movq (%%"REG_S"), %%mm1 \n\t"
 234                 "movq (%%"REG_S"), %%mm2 \n\t"
 235
 236                 "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t"
 237                 "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t"
 238                 "punpcklbw %%mm0, %%mm4 \n\t"
 239                 "punpcklbw %%mm0, %%mm6 \n\t"
 240                 "movq (%%"REG_a"), %%mm3 \n\t"
 241                 "movq (%%"REG_b"), %%mm5 \n\t"
 242                 "punpcklbw %%mm0, %%mm3 \n\t"
 243                 "punpcklbw %%mm0, %%mm5 \n\t"
 244                 "movq %%mm4, %%mm7 \n\t"
 245                 "paddw %%mm4, %%mm4 \n\t"
 246                 "paddw %%mm7, %%mm4 \n\t"
 247                 "movq %%mm6, %%mm7 \n\t"
 248                 "paddw %%mm6, %%mm6 \n\t"
 249                 "paddw %%mm7, %%mm6 \n\t"
 250                 "paddw %%mm3, %%mm4 \n\t"
 251                 "paddw %%mm5, %%mm6 \n\t"
 252                 "paddw %%mm3, %%mm4 \n\t"
 253                 "paddw %%mm5, %%mm6 \n\t"
 254                 "paddw %%mm3, %%mm4 \n\t"
 255                 "paddw %%mm5, %%mm6 \n\t"
 256                 "paddw %%mm3, %%mm4 \n\t"
 257                 "paddw %%mm5, %%mm6 \n\t"
 258                 "paddw %%mm3, %%mm4 \n\t"
 259                 "paddw %%mm5, %%mm6 \n\t"
 260                 "psrlw $3, %%mm4 \n\t"
 261                 "psrlw $3, %%mm6 \n\t"
 262                 "packuswb %%mm4, %%mm4 \n\t"
 263                 "packuswb %%mm6, %%mm6 \n\t"
 264                 "punpcklbw %%mm6, %%mm4 \n\t"
 265                 "punpcklbw %%mm4, %%mm1 \n\t"
 266                 "punpckhbw %%mm4, %%mm2 \n\t"
 267
 268                 "movq %%mm1, (%%"REG_D") \n\t"
 269                 "movq %%mm2, 8(%%"REG_D") \n\t"
 270
 271                 "movq 8(%%"REG_S"), %%mm1 \n\t"
 272                 "movq 8(%%"REG_S"), %%mm2 \n\t"
 273
 274                 "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t"
 275                 "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t"
 276                 "punpckhbw %%mm0, %%mm4 \n\t"
 277                 "punpckhbw %%mm0, %%mm6 \n\t"
 278                 "movq (%%"REG_a"), %%mm3 \n\t"
 279                 "movq (%%"REG_b"), %%mm5 \n\t"
 280                 "punpckhbw %%mm0, %%mm3 \n\t"
 281                 "punpckhbw %%mm0, %%mm5 \n\t"
 282                 "movq %%mm4, %%mm7 \n\t"
 283                 "paddw %%mm4, %%mm4 \n\t"
 284                 "paddw %%mm7, %%mm4 \n\t"
 285                 "movq %%mm6, %%mm7 \n\t"
 286                 "paddw %%mm6, %%mm6 \n\t"
 287                 "paddw %%mm7, %%mm6 \n\t"
 288                 "paddw %%mm3, %%mm4 \n\t"
 289                 "paddw %%mm5, %%mm6 \n\t"
 290                 "paddw %%mm3, %%mm4 \n\t"
 291                 "paddw %%mm5, %%mm6 \n\t"
 292                 "paddw %%mm3, %%mm4 \n\t"
 293                 "paddw %%mm5, %%mm6 \n\t"
 294                 "paddw %%mm3, %%mm4 \n\t"
 295                 "paddw %%mm5, %%mm6 \n\t"
 296                 "paddw %%mm3, %%mm4 \n\t"
 297                 "paddw %%mm5, %%mm6 \n\t"
 298                 "psrlw $3, %%mm4 \n\t"
 299                 "psrlw $3, %%mm6 \n\t"
 300                 "packuswb %%mm4, %%mm4 \n\t"
 301                 "packuswb %%mm6, %%mm6 \n\t"
 302                 "punpcklbw %%mm6, %%mm4 \n\t"
 303                 "punpcklbw %%mm4, %%mm1 \n\t"
 304                 "punpckhbw %%mm4, %%mm2 \n\t"
 305
 306                 "add $16, %%"REG_S" \n\t"
 307                 "add $8, %%"REG_a" \n\t"
 308                 "add $8, %%"REG_b" \n\t"
 309
 310                 "movq %%mm1, 16(%%"REG_D") \n\t"
 311                 "movq %%mm2, 24(%%"REG_D") \n\t"
 312                 "add $32, %%"REG_D" \n\t"
 313
 314                 "decl %%ecx \n\t"
 315                 "jnz .Lli1 \n\t"
 316                 "emms \n\t"
 317                 "pop %%"REG_BP" \n\t"
 318                 :
 319                 : "S" (y), "D" (dst), "a" (u), "b" (v), "c" (w/16),
 320 #if ARCH_X86_64
 321                 "d" ((x86_reg)us), "r" ((x86_reg)vs)
 322 #else
 323                 "d" (&us)
 324 #endif
 325                 : "memory"
 326                 );
 327         pack_li_1_C(dst, y, u, v, (w&15), us, vs);
 328 }
 329 #endif /* HAVE_EBX_AVAILABLE */
 330 #endif
 331
 332 static pack_func_t *pack_nn;
 333 static pack_func_t *pack_li_0;
 334 static pack_func_t *pack_li_1;
 335
 336 static void ilpack(unsigned char *dst, unsigned char *src[3],
 337         int dststride, int srcstride[3], int w, int h, pack_func_t *pack[2])
 338 {
 339         int i;
 340         unsigned char *y, *u, *v;
 341         int ys = srcstride[0], us = srcstride[1], vs = srcstride[2];
 342         int a, b;
 343
 344         y = src[0];
 345         u = src[1];
 346         v = src[2];
 347
 348         pack_nn(dst, y, u, v, w, 0, 0);
 349         y += ys; dst += dststride;
 350         pack_nn(dst, y, u+us, v+vs, w, 0, 0);
 351         y += ys; dst += dststride;
 352         for (i=2; i<h-2; i++) {
 353                 a = (i&2) ? 1 : -1;
 354                 b = (i&1) ^ ((i&2)>>1);
 355                 pack[b](dst, y, u, v, w, us*a, vs*a);
 356                 y += ys;
 357                 if ((i&3) == 1) {
 358                         u -= us;
 359                         v -= vs;
 360                 } else {
 361                         u += us;
 362                         v += vs;
 363                 }
 364                 dst += dststride;
 365         }
 366         pack_nn(dst, y, u, v, w, 0, 0);
 367         y += ys; dst += dststride; u += us; v += vs;
 368         pack_nn(dst, y, u, v, w, 0, 0);
 369 }
 370
 371
 372 static int put_image(struct vf_instance *vf, mp_image_t *mpi, double pts)
 373 {
 374         mp_image_t *dmpi;
 375
 376         // hope we'll get DR buffer:
 377         dmpi=vf_get_image(vf->next, IMGFMT_YUY2,
 378                           MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE,
 379                           mpi->w, mpi->h);
 380
 381         ilpack(dmpi->planes[0], mpi->planes, dmpi->stride[0], mpi->stride, mpi->w, mpi->h, vf->priv->pack);
 382
 383         return vf_next_put_image(vf,dmpi, pts);
 384 }
 385
 386 static int config(struct vf_instance *vf,
 387                   int width, int height, int d_width, int d_height,
 388                   unsigned int flags, unsigned int outfmt)
 389 {
 390         /* FIXME - also support UYVY output? */
 391         return vf_next_config(vf, width, height, d_width, d_height, flags, IMGFMT_YUY2);
 392 }
 393
 394
 395 static int query_format(struct vf_instance *vf, unsigned int fmt)
 396 {
 397         /* FIXME - really any YUV 4:2:0 input format should work */
 398         switch (fmt) {
 399         case IMGFMT_YV12:
 400         case IMGFMT_IYUV:
 401         case IMGFMT_I420:
 402                 return vf_next_query_format(vf,IMGFMT_YUY2);
 403         }
 404         return 0;
 405 }
 406
 407 static int vf_open(vf_instance_t *vf, char *args)
 408 {
 409         vf->config=config;
 410         vf->query_format=query_format;
 411         vf->put_image=put_image;
 412         vf->priv = calloc(1, sizeof(struct vf_priv_s));
 413         vf->priv->mode = 1;
 414         if (args) sscanf(args, "%d", &vf->priv->mode);
 415
 416         pack_nn = (pack_func_t *)pack_nn_C;
 417         pack_li_0 = pack_li_0_C;
 418         pack_li_1 = pack_li_1_C;
 419 #if HAVE_MMX
 420         if(gCpuCaps.hasMMX) {
 421                 pack_nn = (pack_func_t *)pack_nn_MMX;
 422 #if HAVE_EBX_AVAILABLE
 423                 pack_li_0 = pack_li_0_MMX;
 424                 pack_li_1 = pack_li_1_MMX;
 425 #endif
 426         }
 427 #endif
 428
 429         switch(vf->priv->mode) {
 430         case 0:
 431                 vf->priv->pack[0] = vf->priv->pack[1] = pack_nn;
 432                 break;
 433         default:
 434                 mp_msg(MSGT_VFILTER, MSGL_WARN,
 435                         "ilpack: unknown mode %d (fallback to linear)\n",
 436                         vf->priv->mode);
 437         case 1:
 438                 vf->priv->pack[0] = pack_li_0;
 439                 vf->priv->pack[1] = pack_li_1;
 440                 break;
 441         }
 442
 443         return 1;
 444 }
 445
 446 const vf_info_t vf_info_ilpack = {
 447         "4:2:0 planar -> 4:2:2 packed reinterlacer",
 448         "ilpack",
 449         "Richard Felker",
 450         "",
 451         vf_open,
 452         NULL
 453 };