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