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