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 #include "libavutil/attributes.h"
  32
  33 typedef void (pack_func_t)(unsigned char *dst, unsigned char *y,
  34     unsigned char *u, unsigned char *v, int w, int us, int vs);
  35
  36 struct vf_priv_s {
  37     int mode;
  38     pack_func_t *pack[2];
  39 };
  40
  41 static void pack_nn_C(unsigned char *dst, unsigned char *y,
  42     unsigned char *u, unsigned char *v, int w,
  43     int av_unused us, int av_unused vs)
  44 {
  45     int j;
  46     for (j = w/2; j; j--) {
  47         *dst++ = *y++;
  48         *dst++ = *u++;
  49         *dst++ = *y++;
  50         *dst++ = *v++;
  51     }
  52 }
  53
  54 static void pack_li_0_C(unsigned char *dst, unsigned char *y,
  55     unsigned char *u, unsigned char *v, int w, int us, int vs)
  56 {
  57     int j;
  58     for (j = w/2; j; j--) {
  59         *dst++ = *y++;
  60         *dst++ = (u[us+us] + 7*u[0])>>3;
  61         *dst++ = *y++;
  62         *dst++ = (v[vs+vs] + 7*v[0])>>3;
  63         u++; v++;
  64     }
  65 }
  66
  67 static void pack_li_1_C(unsigned char *dst, unsigned char *y,
  68     unsigned char *u, unsigned char *v, int w, int us, int vs)
  69 {
  70     int j;
  71     for (j = w/2; j; j--) {
  72         *dst++ = *y++;
  73         *dst++ = (3*u[us+us] + 5*u[0])>>3;
  74         *dst++ = *y++;
  75         *dst++ = (3*v[vs+vs] + 5*v[0])>>3;
  76         u++; v++;
  77     }
  78 }
  79
  80 #if HAVE_MMX
  81 static void pack_nn_MMX(unsigned char *dst, unsigned char *y,
  82     unsigned char *u, unsigned char *v, int w,
  83     int av_unused us, int av_unused vs)
  84 {
  85     __asm__ volatile (""
  86         ASMALIGN(4)
  87         "1: \n\t"
  88         "movq (%0), %%mm1 \n\t"
  89         "movq (%0), %%mm2 \n\t"
  90         "movq (%1), %%mm4 \n\t"
  91         "movq (%2), %%mm6 \n\t"
  92         "punpcklbw %%mm6, %%mm4 \n\t"
  93         "punpcklbw %%mm4, %%mm1 \n\t"
  94         "punpckhbw %%mm4, %%mm2 \n\t"
  95
  96         "add $8, %0 \n\t"
  97         "add $4, %1 \n\t"
  98         "add $4, %2 \n\t"
  99         "movq %%mm1, (%3) \n\t"
 100         "movq %%mm2, 8(%3) \n\t"
 101         "add $16, %3 \n\t"
 102         "decl %4 \n\t"
 103         "jnz 1b \n\t"
 104         "emms \n\t"
 105         :
 106         : "r" (y), "r" (u), "r" (v), "r" (dst), "r" (w/8)
 107         : "memory"
 108         );
 109     pack_nn_C(dst, y, u, v, (w&7), 0, 0);
 110 }
 111
 112 #if HAVE_EBX_AVAILABLE
 113 static void pack_li_0_MMX(unsigned char *dst, unsigned char *y,
 114     unsigned char *u, unsigned char *v, int w, int us, int vs)
 115 {
 116     __asm__ volatile (""
 117         "push %%"REG_BP" \n\t"
 118 #if ARCH_X86_64
 119         "mov %6, %%"REG_BP" \n\t"
 120 #else
 121         "movl 4(%%"REG_d"), %%"REG_BP" \n\t"
 122         "movl (%%"REG_d"), %%"REG_d" \n\t"
 123 #endif
 124         "pxor %%mm0, %%mm0 \n\t"
 125
 126         ASMALIGN(4)
 127         ".Lli0: \n\t"
 128         "movq (%%"REG_S"), %%mm1 \n\t"
 129         "movq (%%"REG_S"), %%mm2 \n\t"
 130
 131         "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t"
 132         "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t"
 133         "punpcklbw %%mm0, %%mm4 \n\t"
 134         "punpcklbw %%mm0, %%mm6 \n\t"
 135         "movq (%%"REG_a"), %%mm3 \n\t"
 136         "movq (%%"REG_b"), %%mm5 \n\t"
 137         "punpcklbw %%mm0, %%mm3 \n\t"
 138         "punpcklbw %%mm0, %%mm5 \n\t"
 139         "paddw %%mm3, %%mm4 \n\t"
 140         "paddw %%mm5, %%mm6 \n\t"
 141         "paddw %%mm3, %%mm4 \n\t"
 142         "paddw %%mm5, %%mm6 \n\t"
 143         "paddw %%mm3, %%mm4 \n\t"
 144         "paddw %%mm5, %%mm6 \n\t"
 145         "paddw %%mm3, %%mm4 \n\t"
 146         "paddw %%mm5, %%mm6 \n\t"
 147         "paddw %%mm3, %%mm4 \n\t"
 148         "paddw %%mm5, %%mm6 \n\t"
 149         "paddw %%mm3, %%mm4 \n\t"
 150         "paddw %%mm5, %%mm6 \n\t"
 151         "paddw %%mm3, %%mm4 \n\t"
 152         "paddw %%mm5, %%mm6 \n\t"
 153         "psrlw $3, %%mm4 \n\t"
 154         "psrlw $3, %%mm6 \n\t"
 155         "packuswb %%mm4, %%mm4 \n\t"
 156         "packuswb %%mm6, %%mm6 \n\t"
 157         "punpcklbw %%mm6, %%mm4 \n\t"
 158         "punpcklbw %%mm4, %%mm1 \n\t"
 159         "punpckhbw %%mm4, %%mm2 \n\t"
 160
 161         "movq %%mm1, (%%"REG_D") \n\t"
 162         "movq %%mm2, 8(%%"REG_D") \n\t"
 163
 164         "movq 8(%%"REG_S"), %%mm1 \n\t"
 165         "movq 8(%%"REG_S"), %%mm2 \n\t"
 166
 167         "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t"
 168         "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t"
 169         "punpckhbw %%mm0, %%mm4 \n\t"
 170         "punpckhbw %%mm0, %%mm6 \n\t"
 171         "movq (%%"REG_a"), %%mm3 \n\t"
 172         "movq (%%"REG_b"), %%mm5 \n\t"
 173         "punpckhbw %%mm0, %%mm3 \n\t"
 174         "punpckhbw %%mm0, %%mm5 \n\t"
 175         "paddw %%mm3, %%mm4 \n\t"
 176         "paddw %%mm5, %%mm6 \n\t"
 177         "paddw %%mm3, %%mm4 \n\t"
 178         "paddw %%mm5, %%mm6 \n\t"
 179         "paddw %%mm3, %%mm4 \n\t"
 180         "paddw %%mm5, %%mm6 \n\t"
 181         "paddw %%mm3, %%mm4 \n\t"
 182         "paddw %%mm5, %%mm6 \n\t"
 183         "paddw %%mm3, %%mm4 \n\t"
 184         "paddw %%mm5, %%mm6 \n\t"
 185         "paddw %%mm3, %%mm4 \n\t"
 186         "paddw %%mm5, %%mm6 \n\t"
 187         "paddw %%mm3, %%mm4 \n\t"
 188         "paddw %%mm5, %%mm6 \n\t"
 189         "psrlw $3, %%mm4 \n\t"
 190         "psrlw $3, %%mm6 \n\t"
 191         "packuswb %%mm4, %%mm4 \n\t"
 192         "packuswb %%mm6, %%mm6 \n\t"
 193         "punpcklbw %%mm6, %%mm4 \n\t"
 194         "punpcklbw %%mm4, %%mm1 \n\t"
 195         "punpckhbw %%mm4, %%mm2 \n\t"
 196
 197         "add $16, %%"REG_S" \n\t"
 198         "add $8, %%"REG_a" \n\t"
 199         "add $8, %%"REG_b" \n\t"
 200
 201         "movq %%mm1, 16(%%"REG_D") \n\t"
 202         "movq %%mm2, 24(%%"REG_D") \n\t"
 203         "add $32, %%"REG_D" \n\t"
 204
 205         "decl %%ecx \n\t"
 206         "jnz .Lli0 \n\t"
 207         "emms \n\t"
 208         "pop %%"REG_BP" \n\t"
 209         :
 210         : "S" (y), "D" (dst), "a" (u), "b" (v), "c" (w/16),
 211 #if ARCH_X86_64
 212         "d" ((x86_reg)us), "r" ((x86_reg)vs)
 213 #else
 214         "d" (&us)
 215 #endif
 216         : "memory"
 217         );
 218     pack_li_0_C(dst, y, u, v, (w&15), us, vs);
 219 }
 220
 221 static void pack_li_1_MMX(unsigned char *dst, unsigned char *y,
 222     unsigned char *u, unsigned char *v, int w, int us, int vs)
 223 {
 224     __asm__ volatile (""
 225         "push %%"REG_BP" \n\t"
 226 #if ARCH_X86_64
 227         "mov %6, %%"REG_BP" \n\t"
 228 #else
 229         "movl 4(%%"REG_d"), %%"REG_BP" \n\t"
 230         "movl (%%"REG_d"), %%"REG_d" \n\t"
 231 #endif
 232         "pxor %%mm0, %%mm0 \n\t"
 233
 234         ASMALIGN(4)
 235         ".Lli1: \n\t"
 236         "movq (%%"REG_S"), %%mm1 \n\t"
 237         "movq (%%"REG_S"), %%mm2 \n\t"
 238
 239         "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t"
 240         "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t"
 241         "punpcklbw %%mm0, %%mm4 \n\t"
 242         "punpcklbw %%mm0, %%mm6 \n\t"
 243         "movq (%%"REG_a"), %%mm3 \n\t"
 244         "movq (%%"REG_b"), %%mm5 \n\t"
 245         "punpcklbw %%mm0, %%mm3 \n\t"
 246         "punpcklbw %%mm0, %%mm5 \n\t"
 247         "movq %%mm4, %%mm7 \n\t"
 248         "paddw %%mm4, %%mm4 \n\t"
 249         "paddw %%mm7, %%mm4 \n\t"
 250         "movq %%mm6, %%mm7 \n\t"
 251         "paddw %%mm6, %%mm6 \n\t"
 252         "paddw %%mm7, %%mm6 \n\t"
 253         "paddw %%mm3, %%mm4 \n\t"
 254         "paddw %%mm5, %%mm6 \n\t"
 255         "paddw %%mm3, %%mm4 \n\t"
 256         "paddw %%mm5, %%mm6 \n\t"
 257         "paddw %%mm3, %%mm4 \n\t"
 258         "paddw %%mm5, %%mm6 \n\t"
 259         "paddw %%mm3, %%mm4 \n\t"
 260         "paddw %%mm5, %%mm6 \n\t"
 261         "paddw %%mm3, %%mm4 \n\t"
 262         "paddw %%mm5, %%mm6 \n\t"
 263         "psrlw $3, %%mm4 \n\t"
 264         "psrlw $3, %%mm6 \n\t"
 265         "packuswb %%mm4, %%mm4 \n\t"
 266         "packuswb %%mm6, %%mm6 \n\t"
 267         "punpcklbw %%mm6, %%mm4 \n\t"
 268         "punpcklbw %%mm4, %%mm1 \n\t"
 269         "punpckhbw %%mm4, %%mm2 \n\t"
 270
 271         "movq %%mm1, (%%"REG_D") \n\t"
 272         "movq %%mm2, 8(%%"REG_D") \n\t"
 273
 274         "movq 8(%%"REG_S"), %%mm1 \n\t"
 275         "movq 8(%%"REG_S"), %%mm2 \n\t"
 276
 277         "movq (%%"REG_a",%%"REG_d",2), %%mm4 \n\t"
 278         "movq (%%"REG_b",%%"REG_BP",2), %%mm6 \n\t"
 279         "punpckhbw %%mm0, %%mm4 \n\t"
 280         "punpckhbw %%mm0, %%mm6 \n\t"
 281         "movq (%%"REG_a"), %%mm3 \n\t"
 282         "movq (%%"REG_b"), %%mm5 \n\t"
 283         "punpckhbw %%mm0, %%mm3 \n\t"
 284         "punpckhbw %%mm0, %%mm5 \n\t"
 285         "movq %%mm4, %%mm7 \n\t"
 286         "paddw %%mm4, %%mm4 \n\t"
 287         "paddw %%mm7, %%mm4 \n\t"
 288         "movq %%mm6, %%mm7 \n\t"
 289         "paddw %%mm6, %%mm6 \n\t"
 290         "paddw %%mm7, %%mm6 \n\t"
 291         "paddw %%mm3, %%mm4 \n\t"
 292         "paddw %%mm5, %%mm6 \n\t"
 293         "paddw %%mm3, %%mm4 \n\t"
 294         "paddw %%mm5, %%mm6 \n\t"
 295         "paddw %%mm3, %%mm4 \n\t"
 296         "paddw %%mm5, %%mm6 \n\t"
 297         "paddw %%mm3, %%mm4 \n\t"
 298         "paddw %%mm5, %%mm6 \n\t"
 299         "paddw %%mm3, %%mm4 \n\t"
 300         "paddw %%mm5, %%mm6 \n\t"
 301         "psrlw $3, %%mm4 \n\t"
 302         "psrlw $3, %%mm6 \n\t"
 303         "packuswb %%mm4, %%mm4 \n\t"
 304         "packuswb %%mm6, %%mm6 \n\t"
 305         "punpcklbw %%mm6, %%mm4 \n\t"
 306         "punpcklbw %%mm4, %%mm1 \n\t"
 307         "punpckhbw %%mm4, %%mm2 \n\t"
 308
 309         "add $16, %%"REG_S" \n\t"
 310         "add $8, %%"REG_a" \n\t"
 311         "add $8, %%"REG_b" \n\t"
 312
 313         "movq %%mm1, 16(%%"REG_D") \n\t"
 314         "movq %%mm2, 24(%%"REG_D") \n\t"
 315         "add $32, %%"REG_D" \n\t"
 316
 317         "decl %%ecx \n\t"
 318         "jnz .Lli1 \n\t"
 319         "emms \n\t"
 320         "pop %%"REG_BP" \n\t"
 321         :
 322         : "S" (y), "D" (dst), "a" (u), "b" (v), "c" (w/16),
 323 #if ARCH_X86_64
 324         "d" ((x86_reg)us), "r" ((x86_reg)vs)
 325 #else
 326         "d" (&us)
 327 #endif
 328         : "memory"
 329         );
 330     pack_li_1_C(dst, y, u, v, (w&15), us, vs);
 331 }
 332 #endif /* HAVE_EBX_AVAILABLE */
 333 #endif
 334
 335 static pack_func_t *pack_nn;
 336 static pack_func_t *pack_li_0;
 337 static pack_func_t *pack_li_1;
 338
 339 static void ilpack(unsigned char *dst, unsigned char *src[3],
 340     int dststride, int srcstride[3], int w, int h, pack_func_t *pack[2])
 341 {
 342     int i;
 343     unsigned char *y, *u, *v;
 344     int ys = srcstride[0], us = srcstride[1], vs = srcstride[2];
 345     int a, b;
 346
 347     y = src[0];
 348     u = src[1];
 349     v = src[2];
 350
 351     pack_nn(dst, y, u, v, w, 0, 0);
 352     y += ys; dst += dststride;
 353     pack_nn(dst, y, u+us, v+vs, w, 0, 0);
 354     y += ys; dst += dststride;
 355     for (i=2; i<h-2; i++) {
 356         a = (i&2) ? 1 : -1;
 357         b = (i&1) ^ ((i&2)>>1);
 358         pack[b](dst, y, u, v, w, us*a, vs*a);
 359         y += ys;
 360         if ((i&3) == 1) {
 361             u -= us;
 362             v -= vs;
 363         } else {
 364             u += us;
 365             v += vs;
 366         }
 367         dst += dststride;
 368     }
 369     pack_nn(dst, y, u, v, w, 0, 0);
 370     y += ys; dst += dststride; u += us; v += vs;
 371     pack_nn(dst, y, u, v, w, 0, 0);
 372 }
 373
 374
 375 static int put_image(struct vf_instance *vf, mp_image_t *mpi, double pts)
 376 {
 377     mp_image_t *dmpi;
 378
 379     // hope we'll get DR buffer:
 380     dmpi=vf_get_image(vf->next, IMGFMT_YUY2,
 381               MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE,
 382               mpi->w, mpi->h);
 383
 384     ilpack(dmpi->planes[0], mpi->planes, dmpi->stride[0], mpi->stride, mpi->w, mpi->h, vf->priv->pack);
 385
 386     return vf_next_put_image(vf,dmpi, pts);
 387 }
 388
 389 static int config(struct vf_instance *vf,
 390           int width, int height, int d_width, int d_height,
 391           unsigned int flags, unsigned int outfmt)
 392 {
 393     /* FIXME - also support UYVY output? */
 394     return vf_next_config(vf, width, height, d_width, d_height, flags, IMGFMT_YUY2);
 395 }
 396
 397
 398 static int query_format(struct vf_instance *vf, unsigned int fmt)
 399 {
 400     /* FIXME - really any YUV 4:2:0 input format should work */
 401     switch (fmt) {
 402     case IMGFMT_YV12:
 403     case IMGFMT_IYUV:
 404     case IMGFMT_I420:
 405         return vf_next_query_format(vf,IMGFMT_YUY2);
 406     }
 407     return 0;
 408 }
 409
 410 static int vf_open(vf_instance_t *vf, char *args)
 411 {
 412     vf->config=config;
 413     vf->query_format=query_format;
 414     vf->put_image=put_image;
 415     vf->priv = calloc(1, sizeof(struct vf_priv_s));
 416     vf->priv->mode = 1;
 417     if (args) sscanf(args, "%d", &vf->priv->mode);
 418
 419     pack_nn = pack_nn_C;
 420     pack_li_0 = pack_li_0_C;
 421     pack_li_1 = pack_li_1_C;
 422 #if HAVE_MMX
 423     if(gCpuCaps.hasMMX) {
 424         pack_nn = pack_nn_MMX;
 425 #if HAVE_EBX_AVAILABLE
 426         pack_li_0 = pack_li_0_MMX;
 427         pack_li_1 = pack_li_1_MMX;
 428 #endif
 429     }
 430 #endif
 431
 432     switch(vf->priv->mode) {
 433     case 0:
 434         vf->priv->pack[0] = vf->priv->pack[1] = pack_nn;
 435         break;
 436     default:
 437         mp_msg(MSGT_VFILTER, MSGL_WARN,
 438             "ilpack: unknown mode %d (fallback to linear)\n",
 439             vf->priv->mode);
 440     case 1:
 441         vf->priv->pack[0] = pack_li_0;
 442         vf->priv->pack[1] = pack_li_1;
 443         break;
 444     }
 445
 446     return 1;
 447 }
 448
 449 const vf_info_t vf_info_ilpack = {
 450     "4:2:0 planar -> 4:2:2 packed reinterlacer",
 451     "ilpack",
 452     "Richard Felker",
 453     "",
 454     vf_open,
 455     NULL
 456 };