libmpcodecs/vf_ilpack.c

   1 #include <stdio.h>
   2 #include <stdlib.h>
   3 #include <string.h>
   4 #include <inttypes.h>
   5
   6 #include "config.h"
   7 #include "mp_msg.h"
   8 #include "cpudetect.h"
   9 #include "asmalign.h"
  10
  11 #include "img_format.h"
  12 #include "mp_image.h"
  13 #include "vf.h"
  14
  15 #include "libvo/fastmemcpy.h"
  16 #include "postproc/rgb2rgb.h"
  17
  18 typedef void (pack_func_t)(unsigned char *dst, unsigned char *y,
  19         unsigned char *u, unsigned char *v, int w, int us, int vs);
  20
  21 struct vf_priv_s {
  22         int mode;
  23         pack_func_t *pack[2];
  24 };
  25
  26 static void pack_nn_C(unsigned char *dst, unsigned char *y,
  27         unsigned char *u, unsigned char *v, int w)
  28 {
  29         int j;
  30         for (j = w/2; j; j--) {
  31                 *dst++ = *y++;
  32                 *dst++ = *u++;
  33                 *dst++ = *y++;
  34                 *dst++ = *v++;
  35         }
  36 }
  37
  38 static void pack_li_0_C(unsigned char *dst, unsigned char *y,
  39         unsigned char *u, unsigned char *v, int w, int us, int vs)
  40 {
  41         int j;
  42         for (j = w/2; j; j--) {
  43                 *dst++ = *y++;
  44                 *dst++ = (u[us+us] + 7*u[0])>>3;
  45                 *dst++ = *y++;
  46                 *dst++ = (v[vs+vs] + 7*v[0])>>3;
  47                 u++; v++;
  48         }
  49 }
  50
  51 static void pack_li_1_C(unsigned char *dst, unsigned char *y,
  52         unsigned char *u, unsigned char *v, int w, int us, int vs)
  53 {
  54         int j;
  55         for (j = w/2; j; j--) {
  56                 *dst++ = *y++;
  57                 *dst++ = (3*u[us+us] + 5*u[0])>>3;
  58                 *dst++ = *y++;
  59                 *dst++ = (3*v[vs+vs] + 5*v[0])>>3;
  60                 u++; v++;
  61         }
  62 }
  63
  64 #ifdef HAVE_MMX
  65 static void pack_nn_MMX(unsigned char *dst, unsigned char *y,
  66         unsigned char *u, unsigned char *v, int w)
  67 {
  68         int j;
  69         asm volatile (""
  70                 ASMALIGN16
  71                 "1: \n\t"
  72                 "movq (%0), %%mm1 \n\t"
  73                 "movq (%0), %%mm2 \n\t"
  74                 "movq (%1), %%mm4 \n\t"
  75                 "movq (%2), %%mm6 \n\t"
  76                 "punpcklbw %%mm6, %%mm4 \n\t"
  77                 "punpcklbw %%mm4, %%mm1 \n\t"
  78                 "punpckhbw %%mm4, %%mm2 \n\t"
  79
  80                 "add $8, %0 \n\t"
  81                 "add $4, %1 \n\t"
  82                 "add $4, %2 \n\t"
  83                 "movq %%mm1, (%3) \n\t"
  84                 "movq %%mm2, 8(%3) \n\t"
  85                 "add $16, %3 \n\t"
  86                 "decl %4 \n\t"
  87                 "jnz 1b \n\t"
  88                 "emms \n\t"
  89                 :
  90                 : "r" (y), "r" (u), "r" (v), "r" (dst), "r" (w/8)
  91                 : "memory"
  92                 );
  93         pack_nn_C(dst, y, u, v, (w&7));
  94 }
  95
  96 static void pack_li_0_MMX(unsigned char *dst, unsigned char *y,
  97         unsigned char *u, unsigned char *v, int w, int us, int vs)
  98 {
  99         asm volatile (""
 100                 "push %%"REG_BP" \n\t"
 101 #ifdef ARCH_X86_64
 102                 "mov %6, %%"REG_BP" \n\t"
 103 #else
 104                 "movl 4(%%"REG_d"), %%"REG_BP" \n\t"
 105                 "movl (%%"REG_d"), %%"REG_d" \n\t"
 106 #endif
 107                 "pxor %%mm0, %%mm0 \n\t"
 108
 109                 ASMALIGN16
 110                 ".Lli0: \n\t"
 111                 "movq (%%"REG_S"), %%mm1 \n\t"
 112                 "movq (%%"REG_S"), %%mm2 \n\t"
 113
 114                 "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t"
 115                 "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t"
 116                 "punpcklbw %%mm0, %%mm4 \n\t"
 117                 "punpcklbw %%mm0, %%mm6 \n\t"
 118                 "movq (%%"REG_a"), %%mm3 \n\t"
 119                 "movq (%%"REG_b"), %%mm5 \n\t"
 120                 "punpcklbw %%mm0, %%mm3 \n\t"
 121                 "punpcklbw %%mm0, %%mm5 \n\t"
 122                 "paddw %%mm3, %%mm4 \n\t"
 123                 "paddw %%mm5, %%mm6 \n\t"
 124                 "paddw %%mm3, %%mm4 \n\t"
 125                 "paddw %%mm5, %%mm6 \n\t"
 126                 "paddw %%mm3, %%mm4 \n\t"
 127                 "paddw %%mm5, %%mm6 \n\t"
 128                 "paddw %%mm3, %%mm4 \n\t"
 129                 "paddw %%mm5, %%mm6 \n\t"
 130                 "paddw %%mm3, %%mm4 \n\t"
 131                 "paddw %%mm5, %%mm6 \n\t"
 132                 "paddw %%mm3, %%mm4 \n\t"
 133                 "paddw %%mm5, %%mm6 \n\t"
 134                 "paddw %%mm3, %%mm4 \n\t"
 135                 "paddw %%mm5, %%mm6 \n\t"
 136                 "psrlw $3, %%mm4 \n\t"
 137                 "psrlw $3, %%mm6 \n\t"
 138                 "packuswb %%mm4, %%mm4 \n\t"
 139                 "packuswb %%mm6, %%mm6 \n\t"
 140                 "punpcklbw %%mm6, %%mm4 \n\t"
 141                 "punpcklbw %%mm4, %%mm1 \n\t"
 142                 "punpckhbw %%mm4, %%mm2 \n\t"
 143
 144                 "movq %%mm1, (%%"REG_D") \n\t"
 145                 "movq %%mm2, 8(%%"REG_D") \n\t"
 146
 147                 "movq 8(%%"REG_S"), %%mm1 \n\t"
 148                 "movq 8(%%"REG_S"), %%mm2 \n\t"
 149
 150                 "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t"
 151                 "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t"
 152                 "punpckhbw %%mm0, %%mm4 \n\t"
 153                 "punpckhbw %%mm0, %%mm6 \n\t"
 154                 "movq (%%"REG_a"), %%mm3 \n\t"
 155                 "movq (%%"REG_b"), %%mm5 \n\t"
 156                 "punpckhbw %%mm0, %%mm3 \n\t"
 157                 "punpckhbw %%mm0, %%mm5 \n\t"
 158                 "paddw %%mm3, %%mm4 \n\t"
 159                 "paddw %%mm5, %%mm6 \n\t"
 160                 "paddw %%mm3, %%mm4 \n\t"
 161                 "paddw %%mm5, %%mm6 \n\t"
 162                 "paddw %%mm3, %%mm4 \n\t"
 163                 "paddw %%mm5, %%mm6 \n\t"
 164                 "paddw %%mm3, %%mm4 \n\t"
 165                 "paddw %%mm5, %%mm6 \n\t"
 166                 "paddw %%mm3, %%mm4 \n\t"
 167                 "paddw %%mm5, %%mm6 \n\t"
 168                 "paddw %%mm3, %%mm4 \n\t"
 169                 "paddw %%mm5, %%mm6 \n\t"
 170                 "paddw %%mm3, %%mm4 \n\t"
 171                 "paddw %%mm5, %%mm6 \n\t"
 172                 "psrlw $3, %%mm4 \n\t"
 173                 "psrlw $3, %%mm6 \n\t"
 174                 "packuswb %%mm4, %%mm4 \n\t"
 175                 "packuswb %%mm6, %%mm6 \n\t"
 176                 "punpcklbw %%mm6, %%mm4 \n\t"
 177                 "punpcklbw %%mm4, %%mm1 \n\t"
 178                 "punpckhbw %%mm4, %%mm2 \n\t"
 179
 180                 "add $16, %%"REG_S" \n\t"
 181                 "add $8, %%"REG_a" \n\t"
 182                 "add $8, %%"REG_b" \n\t"
 183
 184                 "movq %%mm1, 16(%%"REG_D") \n\t"
 185                 "movq %%mm2, 24(%%"REG_D") \n\t"
 186                 "add $32, %%"REG_D" \n\t"
 187
 188                 "decl %%ecx \n\t"
 189                 "jnz .Lli0 \n\t"
 190                 "emms \n\t"
 191                 "pop %%"REG_BP" \n\t"
 192                 :
 193                 : "S" (y), "D" (dst), "a" (u), "b" (v), "c" (w/16),
 194 #ifdef ARCH_X86_64
 195                 "d" ((long)us), "r" ((long)vs)
 196 #else
 197                 "d" (&us)
 198 #endif
 199                 : "memory"
 200                 );
 201         pack_li_0_C(dst, y, u, v, (w&15), us, vs);
 202 }
 203
 204 static void pack_li_1_MMX(unsigned char *dst, unsigned char *y,
 205         unsigned char *u, unsigned char *v, int w, int us, int vs)
 206 {
 207         asm volatile (""
 208                 "push %%"REG_BP" \n\t"
 209 #ifdef ARCH_X86_64
 210                 "mov %6, %%"REG_BP" \n\t"
 211 #else
 212                 "movl 4(%%"REG_d"), %%"REG_BP" \n\t"
 213                 "movl (%%"REG_d"), %%"REG_d" \n\t"
 214 #endif
 215                 "pxor %%mm0, %%mm0 \n\t"
 216
 217                 ASMALIGN16
 218                 ".Lli1: \n\t"
 219                 "movq (%%"REG_S"), %%mm1 \n\t"
 220                 "movq (%%"REG_S"), %%mm2 \n\t"
 221
 222                 "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t"
 223                 "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t"
 224                 "punpcklbw %%mm0, %%mm4 \n\t"
 225                 "punpcklbw %%mm0, %%mm6 \n\t"
 226                 "movq (%%"REG_a"), %%mm3 \n\t"
 227                 "movq (%%"REG_b"), %%mm5 \n\t"
 228                 "punpcklbw %%mm0, %%mm3 \n\t"
 229                 "punpcklbw %%mm0, %%mm5 \n\t"
 230                 "movq %%mm4, %%mm7 \n\t"
 231                 "paddw %%mm4, %%mm4 \n\t"
 232                 "paddw %%mm7, %%mm4 \n\t"
 233                 "movq %%mm6, %%mm7 \n\t"
 234                 "paddw %%mm6, %%mm6 \n\t"
 235                 "paddw %%mm7, %%mm6 \n\t"
 236                 "paddw %%mm3, %%mm4 \n\t"
 237                 "paddw %%mm5, %%mm6 \n\t"
 238                 "paddw %%mm3, %%mm4 \n\t"
 239                 "paddw %%mm5, %%mm6 \n\t"
 240                 "paddw %%mm3, %%mm4 \n\t"
 241                 "paddw %%mm5, %%mm6 \n\t"
 242                 "paddw %%mm3, %%mm4 \n\t"
 243                 "paddw %%mm5, %%mm6 \n\t"
 244                 "paddw %%mm3, %%mm4 \n\t"
 245                 "paddw %%mm5, %%mm6 \n\t"
 246                 "psrlw $3, %%mm4 \n\t"
 247                 "psrlw $3, %%mm6 \n\t"
 248                 "packuswb %%mm4, %%mm4 \n\t"
 249                 "packuswb %%mm6, %%mm6 \n\t"
 250                 "punpcklbw %%mm6, %%mm4 \n\t"
 251                 "punpcklbw %%mm4, %%mm1 \n\t"
 252                 "punpckhbw %%mm4, %%mm2 \n\t"
 253
 254                 "movq %%mm1, (%%"REG_D") \n\t"
 255                 "movq %%mm2, 8(%%"REG_D") \n\t"
 256
 257                 "movq 8(%%"REG_S"), %%mm1 \n\t"
 258                 "movq 8(%%"REG_S"), %%mm2 \n\t"
 259
 260                 "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t"
 261                 "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t"
 262                 "punpckhbw %%mm0, %%mm4 \n\t"
 263                 "punpckhbw %%mm0, %%mm6 \n\t"
 264                 "movq (%%"REG_a"), %%mm3 \n\t"
 265                 "movq (%%"REG_b"), %%mm5 \n\t"
 266                 "punpckhbw %%mm0, %%mm3 \n\t"
 267                 "punpckhbw %%mm0, %%mm5 \n\t"
 268                 "movq %%mm4, %%mm7 \n\t"
 269                 "paddw %%mm4, %%mm4 \n\t"
 270                 "paddw %%mm7, %%mm4 \n\t"
 271                 "movq %%mm6, %%mm7 \n\t"
 272                 "paddw %%mm6, %%mm6 \n\t"
 273                 "paddw %%mm7, %%mm6 \n\t"
 274                 "paddw %%mm3, %%mm4 \n\t"
 275                 "paddw %%mm5, %%mm6 \n\t"
 276                 "paddw %%mm3, %%mm4 \n\t"
 277                 "paddw %%mm5, %%mm6 \n\t"
 278                 "paddw %%mm3, %%mm4 \n\t"
 279                 "paddw %%mm5, %%mm6 \n\t"
 280                 "paddw %%mm3, %%mm4 \n\t"
 281                 "paddw %%mm5, %%mm6 \n\t"
 282                 "paddw %%mm3, %%mm4 \n\t"
 283                 "paddw %%mm5, %%mm6 \n\t"
 284                 "psrlw $3, %%mm4 \n\t"
 285                 "psrlw $3, %%mm6 \n\t"
 286                 "packuswb %%mm4, %%mm4 \n\t"
 287                 "packuswb %%mm6, %%mm6 \n\t"
 288                 "punpcklbw %%mm6, %%mm4 \n\t"
 289                 "punpcklbw %%mm4, %%mm1 \n\t"
 290                 "punpckhbw %%mm4, %%mm2 \n\t"
 291
 292                 "add $16, %%"REG_S" \n\t"
 293                 "add $8, %%"REG_a" \n\t"
 294                 "add $8, %%"REG_b" \n\t"
 295
 296                 "movq %%mm1, 16(%%"REG_D") \n\t"
 297                 "movq %%mm2, 24(%%"REG_D") \n\t"
 298                 "add $32, %%"REG_D" \n\t"
 299
 300                 "decl %%ecx \n\t"
 301                 "jnz .Lli1 \n\t"
 302                 "emms \n\t"
 303                 "pop %%"REG_BP" \n\t"
 304                 :
 305                 : "S" (y), "D" (dst), "a" (u), "b" (v), "c" (w/16),
 306 #ifdef ARCH_X86_64
 307                 "d" ((long)us), "r" ((long)vs)
 308 #else
 309                 "d" (&us)
 310 #endif
 311                 : "memory"
 312                 );
 313         pack_li_1_C(dst, y, u, v, (w&15), us, vs);
 314 }
 315 #endif
 316
 317 static pack_func_t *pack_nn;
 318 static pack_func_t *pack_li_0;
 319 static pack_func_t *pack_li_1;
 320
 321 static void ilpack(unsigned char *dst, unsigned char *src[3],
 322         int dststride, int srcstride[3], int w, int h, pack_func_t *pack[2])
 323 {
 324         int i;
 325         unsigned char *y, *u, *v;
 326         int ys = srcstride[0], us = srcstride[1], vs = srcstride[2];
 327         int a, b;
 328
 329         y = src[0];
 330         u = src[1];
 331         v = src[2];
 332
 333         pack_nn(dst, y, u, v, w, 0, 0);
 334         y += ys; dst += dststride;
 335         pack_nn(dst, y, u+us, v+vs, w, 0, 0);
 336         y += ys; dst += dststride;
 337         for (i=2; i<h-2; i++) {
 338                 a = (i&2) ? 1 : -1;
 339                 b = (i&1) ^ ((i&2)>>1);
 340                 pack[b](dst, y, u, v, w, us*a, vs*a);
 341                 y += ys;
 342                 if ((i&3) == 1) {
 343                         u -= us;
 344                         v -= vs;
 345                 } else {
 346                         u += us;
 347                         v += vs;
 348                 }
 349                 dst += dststride;
 350         }
 351         pack_nn(dst, y, u, v, w, 0, 0);
 352         y += ys; dst += dststride; u += us; v += vs;
 353         pack_nn(dst, y, u, v, w, 0, 0);
 354 }
 355
 356
 357 static int put_image(struct vf_instance_s* vf, mp_image_t *mpi, double pts)
 358 {
 359         mp_image_t *dmpi;
 360
 361         // hope we'll get DR buffer:
 362         dmpi=vf_get_image(vf->next, IMGFMT_YUY2,
 363                           MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE,
 364                           mpi->w, mpi->h);
 365
 366         ilpack(dmpi->planes[0], mpi->planes, dmpi->stride[0], mpi->stride, mpi->w, mpi->h, vf->priv->pack);
 367
 368         return vf_next_put_image(vf,dmpi, pts);
 369 }
 370
 371 static int config(struct vf_instance_s* vf,
 372                   int width, int height, int d_width, int d_height,
 373                   unsigned int flags, unsigned int outfmt)
 374 {
 375         /* FIXME - also support UYVY output? */
 376         return vf_next_config(vf, width, height, d_width, d_height, flags, IMGFMT_YUY2);
 377 }
 378
 379
 380 static int query_format(struct vf_instance_s* vf, unsigned int fmt)
 381 {
 382         /* FIXME - really any YUV 4:2:0 input format should work */
 383         switch (fmt) {
 384         case IMGFMT_YV12:
 385         case IMGFMT_IYUV:
 386         case IMGFMT_I420:
 387                 return vf_next_query_format(vf,IMGFMT_YUY2);
 388         }
 389         return 0;
 390 }
 391
 392 static int open(vf_instance_t *vf, char* args)
 393 {
 394         vf->config=config;
 395         vf->query_format=query_format;
 396         vf->put_image=put_image;
 397         vf->priv = calloc(1, sizeof(struct vf_priv_s));
 398         vf->priv->mode = 1;
 399         if (args) sscanf(args, "%d", &vf->priv->mode);
 400
 401         pack_nn = (pack_func_t *)pack_nn_C;
 402         pack_li_0 = pack_li_0_C;
 403         pack_li_1 = pack_li_1_C;
 404 #ifdef HAVE_MMX
 405         if(gCpuCaps.hasMMX) {
 406                 pack_nn = (pack_func_t *)pack_nn_MMX;
 407                 pack_li_0 = pack_li_0_MMX;
 408                 pack_li_1 = pack_li_1_MMX;
 409         }
 410 #endif
 411
 412         switch(vf->priv->mode) {
 413         case 0:
 414                 vf->priv->pack[0] = vf->priv->pack[1] = pack_nn;
 415                 break;
 416         default:
 417                 mp_msg(MSGT_VFILTER, MSGL_WARN,
 418                         "ilpack: unknown mode %d (fallback to linear)\n",
 419                         vf->priv->mode);
 420         case 1:
 421                 vf->priv->pack[0] = pack_li_0;
 422                 vf->priv->pack[1] = pack_li_1;
 423                 break;
 424         }
 425
 426         return 1;
 427 }
 428
 429 vf_info_t vf_info_ilpack = {
 430         "4:2:0 planar -> 4:2:2 packed reinterlacer",
 431         "ilpack",
 432         "Richard Felker",
 433         "",
 434         open,
 435         NULL
 436 };
 437