libmpcodecs/vf_tfields.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
  23 #include "config.h"
  24 #include "options.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 #include "libvo/fastmemcpy.h"
  33
  34 struct vf_priv_s {
  35         int mode;
  36         int parity;
  37         int buffered_i;
  38         mp_image_t *buffered_mpi;
  39         double buffered_pts;
  40 };
  41
  42 static void deint(unsigned char *dest, int ds, unsigned char *src, int ss, int w, int h, int field)
  43 {
  44         int x, y;
  45         src += ss;
  46         dest += ds;
  47         h--;
  48         if (field) {
  49                 fast_memcpy(dest - ds, src - ss, w);
  50                 src += ss;
  51                 dest += ds;
  52                 h--;
  53         }
  54         for (y=h/2; y > 0; y--) {
  55                 dest[0] = src[0];
  56                 for (x=1; x<w-1; x++) {
  57                         if (((src[x-ss] < src[x]) && (src[x+ss] < src[x])) ||
  58                                 ((src[x-ss] > src[x]) && (src[x+ss] > src[x]))) {
  59                                 //dest[x] = (src[x+ss] + src[x-ss])>>1;
  60                                 dest[x] = ((src[x+ss]<<1) + (src[x-ss]<<1)
  61                                         + src[x+ss+1] + src[x-ss+1]
  62                                         + src[x+ss-1] + src[x-ss-1])>>3;
  63                         }
  64                         else dest[x] = src[x];
  65                 }
  66                 dest[w-1] = src[w-1];
  67                 dest += ds<<1;
  68                 src += ss<<1;
  69         }
  70         if (h & 1)
  71                 fast_memcpy(dest, src, w);
  72 }
  73
  74 #if HAVE_AMD3DNOW
  75 static void qpel_li_3DNOW(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up)
  76 {
  77         int i, j, ssd=ss;
  78         long crap1, crap2;
  79         if (up) {
  80                 ssd = -ss;
  81                 fast_memcpy(d, s, w);
  82                 d += ds;
  83                 s += ss;
  84         }
  85         for (i=h-1; i; i--) {
  86                 __asm__ volatile(
  87                         "1: \n\t"
  88                         "movq (%%"REG_S"), %%mm0 \n\t"
  89                         "movq (%%"REG_S",%%"REG_a"), %%mm1 \n\t"
  90                         "pavgusb %%mm0, %%mm1 \n\t"
  91                         "add $8, %%"REG_S" \n\t"
  92                         "pavgusb %%mm0, %%mm1 \n\t"
  93                         "movq %%mm1, (%%"REG_D") \n\t"
  94                         "add $8, %%"REG_D" \n\t"
  95                         "decl %%ecx \n\t"
  96                         "jnz 1b \n\t"
  97                         : "=S"(crap1), "=D"(crap2)
  98                         : "c"(w>>3), "S"(s), "D"(d), "a"((long)ssd)
  99                 );
 100                 for (j=w-(w&7); j<w; j++)
 101                         d[j] = (s[j+ssd] + 3*s[j])>>2;
 102                 d += ds;
 103                 s += ss;
 104         }
 105         if (!up) fast_memcpy(d, s, w);
 106         __asm__ volatile("emms \n\t" : : : "memory");
 107 }
 108 #endif
 109
 110 #if HAVE_MMX2
 111 static void qpel_li_MMX2(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up)
 112 {
 113         int i, j, ssd=ss;
 114         long crap1, crap2;
 115         if (up) {
 116                 ssd = -ss;
 117                 fast_memcpy(d, s, w);
 118                 d += ds;
 119                 s += ss;
 120         }
 121         for (i=h-1; i; i--) {
 122                 __asm__ volatile(
 123                         "pxor %%mm7, %%mm7 \n\t"
 124                         "2: \n\t"
 125                         "movq (%%"REG_S"), %%mm0 \n\t"
 126                         "movq (%%"REG_S",%%"REG_a"), %%mm1 \n\t"
 127                         "pavgb %%mm0, %%mm1 \n\t"
 128                         "add $8, %%"REG_S" \n\t"
 129                         "pavgb %%mm0, %%mm1 \n\t"
 130                         "movq %%mm1, (%%"REG_D") \n\t"
 131                         "add $8, %%"REG_D" \n\t"
 132                         "decl %%ecx \n\t"
 133                         "jnz 2b \n\t"
 134                         : "=S"(crap1), "=D"(crap2)
 135                         : "c"(w>>3), "S"(s), "D"(d), "a"((long)ssd)
 136                 );
 137                 for (j=w-(w&7); j<w; j++)
 138                         d[j] = (s[j+ssd] + 3*s[j])>>2;
 139                 d += ds;
 140                 s += ss;
 141         }
 142         if (!up) fast_memcpy(d, s, w);
 143         __asm__ volatile("emms \n\t" : : : "memory");
 144 }
 145 #endif
 146
 147 #if HAVE_MMX
 148 static void qpel_li_MMX(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up)
 149 {
 150         int i, j, ssd=ss;
 151         int crap1, crap2;
 152         if (up) {
 153                 ssd = -ss;
 154                 fast_memcpy(d, s, w);
 155                 d += ds;
 156                 s += ss;
 157         }
 158         for (i=h-1; i; i--) {
 159                 __asm__ volatile(
 160                         "pxor %%mm7, %%mm7 \n\t"
 161                         "3: \n\t"
 162                         "movq (%%"REG_S"), %%mm0 \n\t"
 163                         "movq (%%"REG_S"), %%mm1 \n\t"
 164                         "movq (%%"REG_S",%%"REG_a"), %%mm2 \n\t"
 165                         "movq (%%"REG_S",%%"REG_a"), %%mm3 \n\t"
 166                         "add $8, %%"REG_S" \n\t"
 167                         "punpcklbw %%mm7, %%mm0 \n\t"
 168                         "punpckhbw %%mm7, %%mm1 \n\t"
 169                         "punpcklbw %%mm7, %%mm2 \n\t"
 170                         "punpckhbw %%mm7, %%mm3 \n\t"
 171                         "paddw %%mm0, %%mm2 \n\t"
 172                         "paddw %%mm1, %%mm3 \n\t"
 173                         "paddw %%mm0, %%mm2 \n\t"
 174                         "paddw %%mm1, %%mm3 \n\t"
 175                         "paddw %%mm0, %%mm2 \n\t"
 176                         "paddw %%mm1, %%mm3 \n\t"
 177                         "psrlw $2, %%mm2 \n\t"
 178                         "psrlw $2, %%mm3 \n\t"
 179                         "packsswb %%mm3, %%mm2 \n\t"
 180                         "movq %%mm2, (%%"REG_D") \n\t"
 181                         "add $8, %%"REG_D" \n\t"
 182                         "decl %%ecx \n\t"
 183                         "jnz 3b \n\t"
 184                         : "=S"(crap1), "=D"(crap2)
 185                         : "c"(w>>3), "S"(s), "D"(d), "a"((long)ssd)
 186                 );
 187                 for (j=w-(w&7); j<w; j++)
 188                         d[j] = (s[j+ssd] + 3*s[j])>>2;
 189                 d += ds;
 190                 s += ss;
 191         }
 192         if (!up) fast_memcpy(d, s, w);
 193         __asm__ volatile("emms \n\t" : : : "memory");
 194 }
 195
 196 #if HAVE_EBX_AVAILABLE
 197 static void qpel_4tap_MMX(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up)
 198 {
 199         int i, j, ssd=ss;
 200         static const short filter[] = {
 201                 29, 29, 29, 29, 110, 110, 110, 110,
 202                 9, 9, 9, 9, 3, 3, 3, 3,
 203                 64, 64, 64, 64 };
 204         int crap1, crap2;
 205         if (up) {
 206                 ssd = -ss;
 207                 fast_memcpy(d, s, w);
 208                 d += ds; s += ss;
 209         }
 210         for (j=0; j<w; j++)
 211                 d[j] = (s[j+ssd] + 3*s[j])>>2;
 212         d += ds; s += ss;
 213         for (i=h-3; i; i--) {
 214                 __asm__ volatile(
 215                         "pxor %%mm0, %%mm0 \n\t"
 216                         "movq (%%"REG_d"), %%mm4 \n\t"
 217                         "movq 8(%%"REG_d"), %%mm5 \n\t"
 218                         "movq 16(%%"REG_d"), %%mm6 \n\t"
 219                         "movq 24(%%"REG_d"), %%mm7 \n\t"
 220                         "4: \n\t"
 221
 222                         "movq (%%"REG_S",%%"REG_a"), %%mm1 \n\t"
 223                         "movq (%%"REG_S"), %%mm2 \n\t"
 224                         "movq (%%"REG_S",%%"REG_b"), %%mm3 \n\t"
 225                         "punpcklbw %%mm0, %%mm1 \n\t"
 226                         "punpcklbw %%mm0, %%mm2 \n\t"
 227                         "pmullw %%mm4, %%mm1 \n\t"
 228                         "punpcklbw %%mm0, %%mm3 \n\t"
 229                         "pmullw %%mm5, %%mm2 \n\t"
 230                         "paddusw %%mm2, %%mm1 \n\t"
 231                         "pmullw %%mm6, %%mm3 \n\t"
 232                         "movq (%%"REG_S",%%"REG_a",2), %%mm2 \n\t"
 233                         "psubusw %%mm3, %%mm1 \n\t"
 234                         "punpcklbw %%mm0, %%mm2 \n\t"
 235                         "pmullw %%mm7, %%mm2 \n\t"
 236                         "psubusw %%mm2, %%mm1 \n\t"
 237                         "psrlw $7, %%mm1 \n\t"
 238
 239                         "movq (%%"REG_S",%%"REG_a"), %%mm2 \n\t"
 240                         "movq (%%"REG_S"), %%mm3 \n\t"
 241                         "punpckhbw %%mm0, %%mm2 \n\t"
 242                         "punpckhbw %%mm0, %%mm3 \n\t"
 243                         "pmullw %%mm4, %%mm2 \n\t"
 244                         "pmullw %%mm5, %%mm3 \n\t"
 245                         "paddusw %%mm3, %%mm2 \n\t"
 246                         "movq (%%"REG_S",%%"REG_b"), %%mm3 \n\t"
 247                         "punpckhbw %%mm0, %%mm3 \n\t"
 248                         "pmullw %%mm6, %%mm3 \n\t"
 249                         "psubusw %%mm3, %%mm2 \n\t"
 250                         "movq (%%"REG_S",%%"REG_a",2), %%mm3 \n\t"
 251                         "punpckhbw %%mm0, %%mm3 \n\t"
 252                         "add $8, %%"REG_S" \n\t"
 253                         "pmullw %%mm7, %%mm3 \n\t"
 254                         "psubusw %%mm3, %%mm2 \n\t"
 255                         "psrlw $7, %%mm2 \n\t"
 256
 257                         "packuswb %%mm2, %%mm1 \n\t"
 258                         "movq %%mm1, (%%"REG_D") \n\t"
 259                         "add $8, %%"REG_D" \n\t"
 260                         "decl %%ecx \n\t"
 261                         "jnz 4b \n\t"
 262                         : "=S"(crap1), "=D"(crap2)
 263                         : "c"(w>>3), "S"(s), "D"(d), "a"((long)ssd), "b"((long)-ssd), "d"(filter)
 264                 );
 265                 for (j=w-(w&7); j<w; j++)
 266                         d[j] = (-9*s[j-ssd] + 111*s[j] + 29*s[j+ssd] - 3*s[j+ssd+ssd])>>7;
 267                 d += ds;
 268                 s += ss;
 269         }
 270         for (j=0; j<w; j++)
 271                 d[j] = (s[j+ssd] + 3*s[j])>>2;
 272         d += ds; s += ss;
 273         if (!up) fast_memcpy(d, s, w);
 274         __asm__ volatile("emms \n\t" : : : "memory");
 275 }
 276 #endif /* HAVE_EBX_AVAILABLE */
 277 #endif
 278
 279 static inline int clamp(int a)
 280 {
 281         // If a<512, this is equivalent to:
 282         // return (a<0) ? 0 : ( (a>255) ? 255 : a);
 283         return (~(a>>31)) & (a | ((a<<23)>>31));
 284 }
 285
 286 static void qpel_li_C(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up)
 287 {
 288         int i, j, ssd=ss;
 289         if (up) {
 290                 ssd = -ss;
 291                 fast_memcpy(d, s, w);
 292                 d += ds;
 293                 s += ss;
 294         }
 295         for (i=h-1; i; i--) {
 296                 for (j=0; j<w; j++)
 297                         d[j] = (s[j+ssd] + 3*s[j])>>2;
 298                 d += ds;
 299                 s += ss;
 300         }
 301         if (!up) fast_memcpy(d, s, w);
 302 }
 303
 304 static void qpel_4tap_C(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up)
 305 {
 306         int i, j, ssd=ss;
 307         if (up) {
 308                 ssd = -ss;
 309                 fast_memcpy(d, s, w);
 310                 d += ds; s += ss;
 311         }
 312         for (j=0; j<w; j++)
 313                 d[j] = (s[j+ssd] + 3*s[j] + 2)>>2;
 314         d += ds; s += ss;
 315         for (i=h-3; i; i--) {
 316                 for (j=0; j<w; j++)
 317                         d[j] = clamp((-9*s[j-ssd] + 111*s[j] + 29*s[j+ssd] - 3*s[j+ssd+ssd] + 64)>>7);
 318                 d += ds; s += ss;
 319         }
 320         for (j=0; j<w; j++)
 321                 d[j] = (s[j+ssd] + 3*s[j] + 2)>>2;
 322         d += ds; s += ss;
 323         if (!up) fast_memcpy(d, s, w);
 324 }
 325
 326 static void (*qpel_li)(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up);
 327 static void (*qpel_4tap)(unsigned char *d, unsigned char *s, int w, int h, int ds, int ss, int up);
 328
 329 static int continue_buffered_image(struct vf_instance *vf);
 330
 331 static int put_image(struct vf_instance *vf, mp_image_t *mpi, double pts)
 332 {
 333         vf->priv->buffered_mpi = mpi;
 334         vf->priv->buffered_pts = pts;
 335         vf->priv->buffered_i = 0;
 336         return continue_buffered_image(vf);
 337 }
 338
 339 extern const int under_mencoder;
 340
 341 static int continue_buffered_image(struct vf_instance *vf)
 342 {
 343         int i=vf->priv->buffered_i;
 344         double pts = vf->priv->buffered_pts;
 345         mp_image_t *mpi = vf->priv->buffered_mpi;
 346         int ret=0;
 347         mp_image_t *dmpi;
 348         void (*qpel)(unsigned char *, unsigned char *, int, int, int, int, int);
 349         int bpp=1;
 350         int tff;
 351
 352         if (i == 0)
 353                 vf_queue_frame(vf, continue_buffered_image);
 354         pts += i * .02;  // XXX not right
 355
 356         if (!(mpi->flags & MP_IMGFLAG_PLANAR)) bpp = mpi->bpp/8;
 357         if (vf->priv->parity < 0) {
 358                 if (mpi->fields & MP_IMGFIELD_ORDERED)
 359                         tff = mpi->fields & MP_IMGFIELD_TOP_FIRST;
 360                 else
 361                         tff = 1;
 362         }
 363         else tff = (vf->priv->parity&1)^1;
 364
 365         switch (vf->priv->mode) {
 366         case 2:
 367                 qpel = qpel_li;
 368                 break;
 369         case 3:
 370                 // TODO: add 3tap filter
 371                 qpel = qpel_4tap;
 372                 break;
 373         case 4:
 374                 qpel = qpel_4tap;
 375                 break;
 376         }
 377
 378         switch (vf->priv->mode) {
 379         case 0:
 380                 for (; i<2; i++) {
 381                         dmpi = vf_get_image(vf->next, mpi->imgfmt,
 382                                 MP_IMGTYPE_EXPORT, MP_IMGFLAG_ACCEPT_STRIDE,
 383                                 mpi->width, mpi->height/2);
 384                         dmpi->planes[0] = mpi->planes[0] + (i^!tff)*mpi->stride[0];
 385                         dmpi->stride[0] = 2*mpi->stride[0];
 386                         if (mpi->flags & MP_IMGFLAG_PLANAR) {
 387                                 dmpi->planes[1] = mpi->planes[1] + (i^!tff)*mpi->stride[1];
 388                                 dmpi->planes[2] = mpi->planes[2] + (i^!tff)*mpi->stride[2];
 389                                 dmpi->stride[1] = 2*mpi->stride[1];
 390                                 dmpi->stride[2] = 2*mpi->stride[2];
 391                         }
 392                         ret |= vf_next_put_image(vf, dmpi, pts);
 393                         if (!under_mencoder)
 394                                 break;
 395                 }
 396                 break;
 397         case 1:
 398                 for (; i<2; i++) {
 399                         dmpi = vf_get_image(vf->next, mpi->imgfmt,
 400                                 MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE,
 401                                 mpi->width, mpi->height);
 402                         my_memcpy_pic(dmpi->planes[0] + (i^!tff)*dmpi->stride[0],
 403                                 mpi->planes[0] + (i^!tff)*mpi->stride[0],
 404                                 mpi->w*bpp, mpi->h/2, dmpi->stride[0]*2, mpi->stride[0]*2);
 405                         deint(dmpi->planes[0], dmpi->stride[0], mpi->planes[0], mpi->stride[0], mpi->w, mpi->h, (i^!tff));
 406                         if (mpi->flags & MP_IMGFLAG_PLANAR) {
 407                                 my_memcpy_pic(dmpi->planes[1] + (i^!tff)*dmpi->stride[1],
 408                                         mpi->planes[1] + (i^!tff)*mpi->stride[1],
 409                                         mpi->chroma_width, mpi->chroma_height/2,
 410                                         dmpi->stride[1]*2, mpi->stride[1]*2);
 411                                 my_memcpy_pic(dmpi->planes[2] + (i^!tff)*dmpi->stride[2],
 412                                         mpi->planes[2] + (i^!tff)*mpi->stride[2],
 413                                         mpi->chroma_width, mpi->chroma_height/2,
 414                                         dmpi->stride[2]*2, mpi->stride[2]*2);
 415                                 deint(dmpi->planes[1], dmpi->stride[1], mpi->planes[1], mpi->stride[1],
 416                                         mpi->chroma_width, mpi->chroma_height, (i^!tff));
 417                                 deint(dmpi->planes[2], dmpi->stride[2], mpi->planes[2], mpi->stride[2],
 418                                         mpi->chroma_width, mpi->chroma_height, (i^!tff));
 419                         }
 420                         ret |= vf_next_put_image(vf, dmpi, pts);
 421                         if (!under_mencoder)
 422                                 break;
 423                 }
 424                 break;
 425         case 2:
 426         case 3:
 427         case 4:
 428                 for (; i<2; i++) {
 429                         dmpi = vf_get_image(vf->next, mpi->imgfmt,
 430                                 MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE,
 431                                 mpi->width, mpi->height/2);
 432                         qpel(dmpi->planes[0], mpi->planes[0] + (i^!tff)*mpi->stride[0],
 433                                 mpi->w*bpp, mpi->h/2, dmpi->stride[0], mpi->stride[0]*2, (i^!tff));
 434                         if (mpi->flags & MP_IMGFLAG_PLANAR) {
 435                                 qpel(dmpi->planes[1],
 436                                         mpi->planes[1] + (i^!tff)*mpi->stride[1],
 437                                         mpi->chroma_width, mpi->chroma_height/2,
 438                                         dmpi->stride[1], mpi->stride[1]*2, (i^!tff));
 439                                 qpel(dmpi->planes[2],
 440                                         mpi->planes[2] + (i^!tff)*mpi->stride[2],
 441                                         mpi->chroma_width, mpi->chroma_height/2,
 442                                         dmpi->stride[2], mpi->stride[2]*2, (i^!tff));
 443                         }
 444                         ret |= vf_next_put_image(vf, dmpi, pts);
 445                         if (!under_mencoder)
 446                                 break;
 447                 }
 448                 break;
 449         }
 450         vf->priv->buffered_i = 1;
 451         return ret;
 452 }
 453
 454 static int query_format(struct vf_instance *vf, unsigned int fmt)
 455 {
 456         /* FIXME - figure out which formats exactly work */
 457         switch (fmt) {
 458         default:
 459                 if (vf->priv->mode == 1)
 460                         return 0;
 461         case IMGFMT_YV12:
 462         case IMGFMT_IYUV:
 463         case IMGFMT_I420:
 464                 return vf_next_query_format(vf, fmt);
 465         }
 466         return 0;
 467 }
 468
 469 static int config(struct vf_instance *vf,
 470         int width, int height, int d_width, int d_height,
 471         unsigned int flags, unsigned int outfmt)
 472 {
 473         switch (vf->priv->mode) {
 474         case 0:
 475         case 2:
 476         case 3:
 477         case 4:
 478                 return vf_next_config(vf,width,height/2,d_width,d_height,flags,outfmt);
 479         case 1:
 480                 return vf_next_config(vf,width,height,d_width,d_height,flags,outfmt);
 481         }
 482         return 0;
 483 }
 484
 485 static void uninit(struct vf_instance *vf)
 486 {
 487         free(vf->priv);
 488 }
 489
 490 static int vf_open(vf_instance_t *vf, char *args)
 491 {
 492         struct vf_priv_s *p;
 493         vf->config = config;
 494         vf->put_image = put_image;
 495         vf->query_format = query_format;
 496         vf->uninit = uninit;
 497         vf->default_reqs = VFCAP_ACCEPT_STRIDE;
 498         vf->priv = p = calloc(1, sizeof(struct vf_priv_s));
 499         vf->priv->mode = 4;
 500         vf->priv->parity = -1;
 501         if (args) sscanf(args, "%d:%d", &vf->priv->mode, &vf->priv->parity);
 502         qpel_li = qpel_li_C;
 503         qpel_4tap = qpel_4tap_C;
 504 #if HAVE_MMX
 505         if(gCpuCaps.hasMMX) qpel_li = qpel_li_MMX;
 506 #if HAVE_EBX_AVAILABLE
 507         if(gCpuCaps.hasMMX) qpel_4tap = qpel_4tap_MMX;
 508 #endif
 509 #endif
 510 #if HAVE_MMX2
 511         if(gCpuCaps.hasMMX2) qpel_li = qpel_li_MMX2;
 512 #endif
 513 #if HAVE_AMD3DNOW
 514         if(gCpuCaps.has3DNow) qpel_li = qpel_li_3DNOW;
 515 #endif
 516         return 1;
 517 }
 518
 519 const vf_info_t vf_info_tfields = {
 520     "temporal field separation",
 521     "tfields",
 522     "Rich Felker",
 523     "",
 524     vf_open,
 525     NULL
 526 };