libvo/gl_common.c

   1 /*
   2  * common OpenGL routines
   3  *
   4  * copyleft (C) 2005 Reimar Döffinger <Reimar.Doeffinger@stud.uni-karlsruhe.de>
   5  * Special thanks go to the xine team and Matthias Hopf, whose video_out_opengl.c
   6  * gave me lots of good ideas.
   7  *
   8  * This file is part of MPlayer.
   9  *
  10  * MPlayer is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published by
  12  * the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * MPlayer is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  * GNU General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public License along
  21  * with MPlayer; if not, write to the Free Software Foundation, Inc.,
  22  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  23  */
  24
  25 /**
  26  * \file gl_common.c
  27  * \brief OpenGL helper functions used by vo_gl.c and vo_gl2.c
  28  */
  29
  30 #include <stdlib.h>
  31 #include <stdio.h>
  32 #include <string.h>
  33 #include <ctype.h>
  34 #include <math.h>
  35 #include "old_vo_defines.h"
  36 #include "gl_common.h"
  37 #include "libavutil/common.h"
  38
  39 /**
  40  * \defgroup glextfunctions OpenGL extension functions
  41  *
  42  * the pointers to these functions are acquired when the OpenGL
  43  * context is created
  44  * \{
  45  */
  46 void (APIENTRY *GenBuffers)(GLsizei, GLuint *);
  47 void (APIENTRY *DeleteBuffers)(GLsizei, const GLuint *);
  48 void (APIENTRY *BindBuffer)(GLenum, GLuint);
  49 GLvoid* (APIENTRY *MapBuffer)(GLenum, GLenum);
  50 GLboolean (APIENTRY *UnmapBuffer)(GLenum);
  51 void (APIENTRY *BufferData)(GLenum, intptr_t, const GLvoid *, GLenum);
  52 void (APIENTRY *CombinerParameterfv)(GLenum, const GLfloat *);
  53 void (APIENTRY *CombinerParameteri)(GLenum, GLint);
  54 void (APIENTRY *CombinerInput)(GLenum, GLenum, GLenum, GLenum, GLenum,
  55                                GLenum);
  56 void (APIENTRY *CombinerOutput)(GLenum, GLenum, GLenum, GLenum, GLenum,
  57                                 GLenum, GLenum, GLboolean, GLboolean,
  58                                 GLboolean);
  59 void (APIENTRY *BeginFragmentShader)(void);
  60 void (APIENTRY *EndFragmentShader)(void);
  61 void (APIENTRY *SampleMap)(GLuint, GLuint, GLenum);
  62 void (APIENTRY *ColorFragmentOp2)(GLenum, GLuint, GLuint, GLuint, GLuint,
  63                                   GLuint, GLuint, GLuint, GLuint, GLuint);
  64 void (APIENTRY *ColorFragmentOp3)(GLenum, GLuint, GLuint, GLuint, GLuint,
  65                                   GLuint, GLuint, GLuint, GLuint, GLuint,
  66                                   GLuint, GLuint, GLuint);
  67 void (APIENTRY *SetFragmentShaderConstant)(GLuint, const GLfloat *);
  68 void (APIENTRY *ActiveTexture)(GLenum);
  69 void (APIENTRY *BindTexture)(GLenum, GLuint);
  70 void (APIENTRY *MultiTexCoord2f)(GLenum, GLfloat, GLfloat);
  71 void (APIENTRY *GenPrograms)(GLsizei, GLuint *);
  72 void (APIENTRY *DeletePrograms)(GLsizei, const GLuint *);
  73 void (APIENTRY *BindProgram)(GLenum, GLuint);
  74 void (APIENTRY *ProgramString)(GLenum, GLenum, GLsizei, const GLvoid *);
  75 void (APIENTRY *GetProgramiv)(GLenum, GLenum, GLint *);
  76 void (APIENTRY *ProgramEnvParameter4f)(GLenum, GLuint, GLfloat, GLfloat,
  77                                        GLfloat, GLfloat);
  78 int (APIENTRY *SwapInterval)(int);
  79 void (APIENTRY *TexImage3D)(GLenum, GLint, GLenum, GLsizei, GLsizei, GLsizei,
  80                             GLint, GLenum, GLenum, const GLvoid *);
  81 /** \} */ // end of glextfunctions group
  82
  83 //! \defgroup glgeneral OpenGL general helper functions
  84
  85 //! \defgroup glcontext OpenGL context management helper functions
  86
  87 //! \defgroup gltexture OpenGL texture handling helper functions
  88
  89 //! \defgroup glconversion OpenGL conversion helper functions
  90
  91 static GLint hqtexfmt;
  92
  93 /**
  94  * \brief adjusts the GL_UNPACK_ALIGNMENT to fit the stride.
  95  * \param stride number of bytes per line for which alignment should fit.
  96  * \ingroup glgeneral
  97  */
  98 void glAdjustAlignment(int stride) {
  99   GLint gl_alignment;
 100   if (stride % 8 == 0)
 101     gl_alignment=8;
 102   else if (stride % 4 == 0)
 103     gl_alignment=4;
 104   else if (stride % 2 == 0)
 105     gl_alignment=2;
 106   else
 107     gl_alignment=1;
 108   glPixelStorei (GL_UNPACK_ALIGNMENT, gl_alignment);
 109 }
 110
 111 struct gl_name_map_struct {
 112   GLint value;
 113   const char *name;
 114 };
 115
 116 #undef MAP
 117 #define MAP(a) {a, #a}
 118 //! mapping table for the glValName function
 119 static const struct gl_name_map_struct gl_name_map[] = {
 120   // internal format
 121   MAP(GL_R3_G3_B2), MAP(GL_RGB4), MAP(GL_RGB5), MAP(GL_RGB8),
 122   MAP(GL_RGB10), MAP(GL_RGB12), MAP(GL_RGB16), MAP(GL_RGBA2),
 123   MAP(GL_RGBA4), MAP(GL_RGB5_A1), MAP(GL_RGBA8), MAP(GL_RGB10_A2),
 124   MAP(GL_RGBA12), MAP(GL_RGBA16), MAP(GL_LUMINANCE8),
 125
 126   // format
 127   MAP(GL_RGB), MAP(GL_RGBA), MAP(GL_RED), MAP(GL_GREEN), MAP(GL_BLUE),
 128   MAP(GL_ALPHA), MAP(GL_LUMINANCE), MAP(GL_LUMINANCE_ALPHA),
 129   MAP(GL_COLOR_INDEX),
 130   // rest 1.2 only
 131   MAP(GL_BGR), MAP(GL_BGRA),
 132
 133   //type
 134   MAP(GL_BYTE), MAP(GL_UNSIGNED_BYTE), MAP(GL_SHORT), MAP(GL_UNSIGNED_SHORT),
 135   MAP(GL_INT), MAP(GL_UNSIGNED_INT), MAP(GL_FLOAT), MAP(GL_DOUBLE),
 136   MAP(GL_2_BYTES), MAP(GL_3_BYTES), MAP(GL_4_BYTES),
 137   // rest 1.2 only
 138   MAP(GL_UNSIGNED_BYTE_3_3_2), MAP(GL_UNSIGNED_BYTE_2_3_3_REV),
 139   MAP(GL_UNSIGNED_SHORT_5_6_5), MAP(GL_UNSIGNED_SHORT_5_6_5_REV),
 140   MAP(GL_UNSIGNED_SHORT_4_4_4_4), MAP(GL_UNSIGNED_SHORT_4_4_4_4_REV),
 141   MAP(GL_UNSIGNED_SHORT_5_5_5_1), MAP(GL_UNSIGNED_SHORT_1_5_5_5_REV),
 142   MAP(GL_UNSIGNED_INT_8_8_8_8), MAP(GL_UNSIGNED_INT_8_8_8_8_REV),
 143   MAP(GL_UNSIGNED_INT_10_10_10_2), MAP(GL_UNSIGNED_INT_2_10_10_10_REV),
 144   {0, 0}
 145 };
 146 #undef MAP
 147
 148 /**
 149  * \brief return the name of an OpenGL constant
 150  * \param value the constant
 151  * \return name of the constant or "Unknown format!"
 152  * \ingroup glgeneral
 153  */
 154 const char *glValName(GLint value)
 155 {
 156   int i = 0;
 157
 158   while (gl_name_map[i].name) {
 159     if (gl_name_map[i].value == value)
 160       return gl_name_map[i].name;
 161     i++;
 162   }
 163   return "Unknown format!";
 164 }
 165
 166 //! always return this format as internal texture format in glFindFormat
 167 #define TEXTUREFORMAT_ALWAYS GL_RGB8
 168 #undef TEXTUREFORMAT_ALWAYS
 169
 170 /**
 171  * \brief find the OpenGL settings coresponding to format.
 172  *
 173  * All parameters may be NULL.
 174  * \param fmt MPlayer format to analyze.
 175  * \param bpp [OUT] bits per pixel of that format.
 176  * \param gl_texfmt [OUT] internal texture format that fits the
 177  * image format, not necessarily the best for performance.
 178  * \param gl_format [OUT] OpenGL format for this image format.
 179  * \param gl_type [OUT] OpenGL type for this image format.
 180  * \return 1 if format is supported by OpenGL, 0 if not.
 181  * \ingroup gltexture
 182  */
 183 int glFindFormat(uint32_t fmt, int *bpp, GLint *gl_texfmt,
 184                   GLenum *gl_format, GLenum *gl_type)
 185 {
 186   int supported = 1;
 187   int dummy1;
 188   GLenum dummy2;
 189   GLint dummy3;
 190   if (!bpp) bpp = &dummy1;
 191   if (!gl_texfmt) gl_texfmt = &dummy3;
 192   if (!gl_format) gl_format = &dummy2;
 193   if (!gl_type) gl_type = &dummy2;
 194
 195   *bpp = IMGFMT_IS_BGR(fmt)?IMGFMT_BGR_DEPTH(fmt):IMGFMT_RGB_DEPTH(fmt);
 196   *gl_texfmt = 3;
 197   switch (fmt) {
 198     case IMGFMT_RGB24:
 199       *gl_format = GL_RGB;
 200       *gl_type = GL_UNSIGNED_BYTE;
 201       break;
 202     case IMGFMT_RGBA:
 203       *gl_texfmt = 4;
 204       *gl_format = GL_RGBA;
 205       *gl_type = GL_UNSIGNED_BYTE;
 206       break;
 207     case IMGFMT_YV12:
 208       supported = 0; // no native YV12 support
 209     case IMGFMT_Y800:
 210     case IMGFMT_Y8:
 211       *gl_texfmt = 1;
 212       *bpp = 8;
 213       *gl_format = GL_LUMINANCE;
 214       *gl_type = GL_UNSIGNED_BYTE;
 215       break;
 216 #if 0
 217     // we do not support palettized formats, although the format the
 218     // swscale produces works
 219     case IMGFMT_RGB8:
 220       gl_format = GL_RGB;
 221       gl_type = GL_UNSIGNED_BYTE_2_3_3_REV;
 222       break;
 223 #endif
 224     case IMGFMT_RGB15:
 225       *gl_format = GL_RGBA;
 226       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 227       break;
 228     case IMGFMT_RGB16:
 229       *gl_format = GL_RGB;
 230       *gl_type = GL_UNSIGNED_SHORT_5_6_5_REV;
 231       break;
 232 #if 0
 233     case IMGFMT_BGR8:
 234       // special case as red and blue have a differen number of bits.
 235       // GL_BGR and GL_UNSIGNED_BYTE_3_3_2 isn't supported at least
 236       // by nVidia drivers, and in addition would give more bits to
 237       // blue than to red, which isn't wanted
 238       gl_format = GL_RGB;
 239       gl_type = GL_UNSIGNED_BYTE_3_3_2;
 240       break;
 241 #endif
 242     case IMGFMT_BGR15:
 243       *gl_format = GL_BGRA;
 244       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 245       break;
 246     case IMGFMT_BGR16:
 247       *gl_format = GL_RGB;
 248       *gl_type = GL_UNSIGNED_SHORT_5_6_5;
 249       break;
 250     case IMGFMT_BGR24:
 251       *gl_format = GL_BGR;
 252       *gl_type = GL_UNSIGNED_BYTE;
 253       break;
 254     case IMGFMT_BGRA:
 255       *gl_texfmt = 4;
 256       *gl_format = GL_BGRA;
 257       *gl_type = GL_UNSIGNED_BYTE;
 258       break;
 259     default:
 260       *gl_texfmt = 4;
 261       *gl_format = GL_RGBA;
 262       *gl_type = GL_UNSIGNED_BYTE;
 263       supported = 0;
 264   }
 265 #ifdef TEXTUREFORMAT_ALWAYS
 266   *gl_texfmt = TEXTUREFORMAT_ALWAYS;
 267 #endif
 268   return supported;
 269 }
 270
 271 static void *setNull(const GLubyte *s) {
 272   return NULL;
 273 }
 274
 275 typedef struct {
 276   void *funcptr;
 277   const char *extstr;
 278   const char *funcnames[7];
 279 } extfunc_desc_t;
 280
 281 static const extfunc_desc_t extfuncs[] = {
 282   {&GenBuffers, NULL, {"glGenBuffers", "glGenBuffersARB", NULL}},
 283   {&DeleteBuffers, NULL, {"glDeleteBuffers", "glDeleteBuffersARB", NULL}},
 284   {&BindBuffer, NULL, {"glBindBuffer", "glBindBufferARB", NULL}},
 285   {&MapBuffer, NULL, {"glMapBuffer", "glMapBufferARB", NULL}},
 286   {&UnmapBuffer, NULL, {"glUnmapBuffer", "glUnmapBufferARB", NULL}},
 287   {&BufferData, NULL, {"glBufferData", "glBufferDataARB", NULL}},
 288   {&CombinerParameterfv, "NV_register_combiners", {"glCombinerParameterfv", "glCombinerParameterfvNV", NULL}},
 289   {&CombinerParameteri, "NV_register_combiners", {"glCombinerParameteri", "glCombinerParameteriNV", NULL}},
 290   {&CombinerInput, "NV_register_combiners", {"glCombinerInput", "glCombinerInputNV", NULL}},
 291   {&CombinerOutput, "NV_register_combiners", {"glCombinerOutput", "glCombinerOutputNV", NULL}},
 292   {&BeginFragmentShader, "ATI_fragment_shader", {"glBeginFragmentShaderATI", NULL}},
 293   {&EndFragmentShader, "ATI_fragment_shader", {"glEndFragmentShaderATI", NULL}},
 294   {&SampleMap, "ATI_fragment_shader", {"glSampleMapATI", NULL}},
 295   {&ColorFragmentOp2, "ATI_fragment_shader", {"glColorFragmentOp2ATI", NULL}},
 296   {&ColorFragmentOp3, "ATI_fragment_shader", {"glColorFragmentOp3ATI", NULL}},
 297   {&SetFragmentShaderConstant, "ATI_fragment_shader", {"glSetFragmentShaderConstantATI", NULL}},
 298   {&ActiveTexture, NULL, {"glActiveTexture", "glActiveTextureARB", NULL}},
 299   {&BindTexture, NULL, {"glBindTexture", "glBindTextureARB", "glBindTextureEXT", NULL}},
 300   {&MultiTexCoord2f, NULL, {"glMultiTexCoord2f", "glMultiTexCoord2fARB", NULL}},
 301   {&GenPrograms, "_program", {"glGenProgramsARB", NULL}},
 302   {&DeletePrograms, "_program", {"glDeleteProgramsARB", NULL}},
 303   {&BindProgram, "_program", {"glBindProgramARB", NULL}},
 304   {&ProgramString, "_program", {"glProgramStringARB", NULL}},
 305   {&GetProgramiv, "_program", {"glGetProgramivARB", NULL}},
 306   {&ProgramEnvParameter4f, "_program", {"glProgramEnvParameter4fARB", NULL}},
 307   {&SwapInterval, "_swap_control", {"glXSwapInterval", "glXSwapIntervalEXT", "glXSwapIntervalSGI", "wglSwapInterval", "wglSwapIntervalEXT", "wglSwapIntervalSGI", NULL}},
 308   {&TexImage3D, NULL, {"glTexImage3D", NULL}},
 309   {NULL}
 310 };
 311
 312 /**
 313  * \brief find the function pointers of some useful OpenGL extensions
 314  * \param getProcAddress function to resolve function names, may be NULL
 315  * \param ext2 an extra extension string
 316  */
 317 static void getFunctions(void *(*getProcAddress)(const GLubyte *),
 318                          const char *ext2) {
 319   const extfunc_desc_t *dsc;
 320   const char *extensions = (const char *)glGetString(GL_EXTENSIONS);
 321   char *allexts;
 322   if (!extensions) extensions = "";
 323   if (!ext2) ext2 = "";
 324   allexts = malloc(strlen(extensions) + strlen(ext2) + 2);
 325   strcpy(allexts, extensions);
 326   strcat(allexts, " ");
 327   strcat(allexts, ext2);
 328   mp_msg(MSGT_VO, MSGL_V, "OpenGL extensions string:\n%s\n", allexts);
 329   if (!getProcAddress)
 330     getProcAddress = setNull;
 331   for (dsc = extfuncs; dsc->funcptr; dsc++) {
 332     void *ptr = NULL;
 333     int i;
 334     if (!dsc->extstr || strstr(allexts, dsc->extstr)) {
 335       for (i = 0; !ptr && dsc->funcnames[i]; i++)
 336         ptr = getProcAddress((const GLubyte *)dsc->funcnames[i]);
 337     }
 338     *(void **)dsc->funcptr = ptr;
 339   }
 340   if (strstr(allexts, "_texture_float"))
 341     hqtexfmt = GL_RGB32F;
 342   else if (strstr(allexts, "NV_float_buffer"))
 343     hqtexfmt = GL_FLOAT_RGB32_NV;
 344   else
 345     hqtexfmt = GL_RGB16;
 346   free(allexts);
 347 }
 348
 349 /**
 350  * \brief create a texture and set some defaults
 351  * \param target texture taget, usually GL_TEXTURE_2D
 352  * \param fmt internal texture format
 353  * \param format texture host data format
 354  * \param type texture host data type
 355  * \param filter filter used for scaling, e.g. GL_LINEAR
 356  * \param w texture width
 357  * \param h texture height
 358  * \param val luminance value to fill texture with
 359  * \ingroup gltexture
 360  */
 361 void glCreateClearTex(GLenum target, GLenum fmt, GLenum format, GLenum type, GLint filter,
 362                       int w, int h, unsigned char val) {
 363   GLfloat fval = (GLfloat)val / 255.0;
 364   GLfloat border[4] = {fval, fval, fval, fval};
 365   int stride = w * glFmt2bpp(format, type);
 366   char *init;
 367   if (!stride) return;
 368   init = malloc(stride * h);
 369   memset(init, val, stride * h);
 370   glAdjustAlignment(stride);
 371   glPixelStorei(GL_UNPACK_ROW_LENGTH, w);
 372   glTexImage2D(target, 0, fmt, w, h, 0, format, type, init);
 373   glTexParameterf(target, GL_TEXTURE_PRIORITY, 1.0);
 374   glTexParameteri(target, GL_TEXTURE_MIN_FILTER, filter);
 375   glTexParameteri(target, GL_TEXTURE_MAG_FILTER, filter);
 376   glTexParameteri(target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 377   glTexParameteri(target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 378   // Border texels should not be used with CLAMP_TO_EDGE
 379   // We set a sane default anyway.
 380   glTexParameterfv(target, GL_TEXTURE_BORDER_COLOR, border);
 381   free(init);
 382 }
 383
 384 /**
 385  * \brief skips whitespace and comments
 386  * \param f file to read from
 387  */
 388 static void ppm_skip(FILE *f) {
 389   int c, comment = 0;
 390   do {
 391     c = fgetc(f);
 392     if (c == '#')
 393       comment = 1;
 394     if (c == '\n')
 395       comment = 0;
 396   } while (c != EOF && (isspace(c) || comment));
 397   if (c != EOF)
 398     ungetc(c, f);
 399 }
 400
 401 #define MAXDIM (16 * 1024)
 402
 403 /**
 404  * \brief creates a texture from a PPM file
 405  * \param target texture taget, usually GL_TEXTURE_2D
 406  * \param fmt internal texture format, 0 for default
 407  * \param filter filter used for scaling, e.g. GL_LINEAR
 408  * \param f file to read PPM from
 409  * \param width [out] width of texture
 410  * \param height [out] height of texture
 411  * \param maxval [out] maxval value from PPM file
 412  * \return 0 on error, 1 otherwise
 413  * \ingroup gltexture
 414  */
 415 int glCreatePPMTex(GLenum target, GLenum fmt, GLint filter,
 416                    FILE *f, int *width, int *height, int *maxval) {
 417   unsigned w, h, m, val, bpp;
 418   char *data;
 419   GLenum type;
 420   ppm_skip(f);
 421   if (fgetc(f) != 'P' || fgetc(f) != '6')
 422     return 0;
 423   ppm_skip(f);
 424   if (fscanf(f, "%u", &w) != 1)
 425     return 0;
 426   ppm_skip(f);
 427   if (fscanf(f, "%u", &h) != 1)
 428     return 0;
 429   ppm_skip(f);
 430   if (fscanf(f, "%u", &m) != 1)
 431     return 0;
 432   val = fgetc(f);
 433   if (!isspace(val))
 434     return 0;
 435   if (w > MAXDIM || h > MAXDIM)
 436     return 0;
 437   bpp = (m > 255) ? 6 : 3;
 438   data = malloc(w * h * bpp);
 439   if (fread(data, w * bpp, h, f) != h)
 440     return 0;
 441   if (!fmt) {
 442     fmt = (m > 255) ? hqtexfmt : 3;
 443     if (fmt == GL_FLOAT_RGB32_NV && target != GL_TEXTURE_RECTANGLE)
 444       fmt = GL_RGB16;
 445   }
 446   type = m > 255 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 447   glCreateClearTex(target, fmt, GL_RGB, type, filter, w, h, 0);
 448   glUploadTex(target, GL_RGB, type,
 449               data, w * bpp, 0, 0, w, h, 0);
 450   free(data);
 451   if (width) *width = w;
 452   if (height) *height = h;
 453   if (maxval) *maxval = m;
 454   return 1;
 455 }
 456
 457 /**
 458  * \brief return the number of bytes per pixel for the given format
 459  * \param format OpenGL format
 460  * \param type OpenGL type
 461  * \return bytes per pixel
 462  * \ingroup glgeneral
 463  *
 464  * Does not handle all possible variants, just those used by MPlayer
 465  */
 466 int glFmt2bpp(GLenum format, GLenum type) {
 467   int component_size = 0;
 468   switch (type) {
 469     case GL_UNSIGNED_BYTE_3_3_2:
 470     case GL_UNSIGNED_BYTE_2_3_3_REV:
 471       return 1;
 472     case GL_UNSIGNED_SHORT_5_5_5_1:
 473     case GL_UNSIGNED_SHORT_1_5_5_5_REV:
 474     case GL_UNSIGNED_SHORT_5_6_5:
 475     case GL_UNSIGNED_SHORT_5_6_5_REV:
 476       return 2;
 477     case GL_UNSIGNED_BYTE:
 478       component_size = 1;
 479       break;
 480     case GL_UNSIGNED_SHORT:
 481       component_size = 2;
 482       break;
 483   }
 484   switch (format) {
 485     case GL_LUMINANCE:
 486     case GL_ALPHA:
 487       return component_size;
 488     case GL_RGB:
 489     case GL_BGR:
 490       return 3 * component_size;
 491     case GL_RGBA:
 492     case GL_BGRA:
 493       return 4 * component_size;
 494   }
 495   return 0; // unknown
 496 }
 497
 498 /**
 499  * \brief upload a texture, handling things like stride and slices
 500  * \param target texture target, usually GL_TEXTURE_2D
 501  * \param format OpenGL format of data
 502  * \param type OpenGL type of data
 503  * \param dataptr data to upload
 504  * \param stride data stride
 505  * \param x x offset in texture
 506  * \param y y offset in texture
 507  * \param w width of the texture part to upload
 508  * \param h height of the texture part to upload
 509  * \param slice height of an upload slice, 0 for all at once
 510  * \ingroup gltexture
 511  */
 512 void glUploadTex(GLenum target, GLenum format, GLenum type,
 513                  const void *dataptr, int stride,
 514                  int x, int y, int w, int h, int slice) {
 515   const uint8_t *data = dataptr;
 516   int y_max = y + h;
 517   if (w <= 0 || h <= 0) return;
 518   if (slice <= 0)
 519     slice = h;
 520   if (stride < 0) {
 521     data += (h - 1) * stride;
 522     stride = -stride;
 523   }
 524   // this is not always correct, but should work for MPlayer
 525   glAdjustAlignment(stride);
 526   glPixelStorei(GL_UNPACK_ROW_LENGTH, stride / glFmt2bpp(format, type));
 527   for (; y + slice <= y_max; y += slice) {
 528     glTexSubImage2D(target, 0, x, y, w, slice, format, type, data);
 529     data += stride * slice;
 530   }
 531   if (y < y_max)
 532     glTexSubImage2D(target, 0, x, y, w, y_max - y, format, type, data);
 533 }
 534
 535 static void fillUVcoeff(GLfloat *ucoef, GLfloat *vcoef,
 536                         float uvcos, float uvsin) {
 537   int i;
 538   ucoef[0] = 0 * uvcos + 1.403 * uvsin;
 539   vcoef[0] = 0 * uvsin + 1.403 * uvcos;
 540   ucoef[1] = -0.344 * uvcos + -0.714 * uvsin;
 541   vcoef[1] = -0.344 * uvsin + -0.714 * uvcos;
 542   ucoef[2] = 1.770 * uvcos + 0 * uvsin;
 543   vcoef[2] = 1.770 * uvsin + 0 * uvcos;
 544   ucoef[3] = 0;
 545   vcoef[3] = 0;
 546   // Coefficients (probably) must be in [0, 1] range, whereas they originally
 547   // are in [-2, 2] range, so here comes the trick:
 548   // First put them in the [-0.5, 0.5] range, then add 0.5.
 549   // This can be undone with the HALF_BIAS and SCALE_BY_FOUR arguments
 550   // for CombinerInput and CombinerOutput (or the respective ATI variants)
 551   for (i = 0; i < 4; i++) {
 552     ucoef[i] = ucoef[i] * 0.25 + 0.5;
 553     vcoef[i] = vcoef[i] * 0.25 + 0.5;
 554   }
 555 }
 556
 557 /**
 558  * \brief Setup register combiners for YUV to RGB conversion.
 559  * \param uvcos used for saturation and hue adjustment
 560  * \param uvsin used for saturation and hue adjustment
 561  */
 562 static void glSetupYUVCombiners(float uvcos, float uvsin) {
 563   GLfloat ucoef[4];
 564   GLfloat vcoef[4];
 565   GLint i;
 566   if (!CombinerInput || !CombinerOutput ||
 567       !CombinerParameterfv || !CombinerParameteri) {
 568     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner functions missing!\n");
 569     return;
 570   }
 571   glGetIntegerv(GL_MAX_GENERAL_COMBINERS_NV, &i);
 572   if (i < 2)
 573     mp_msg(MSGT_VO, MSGL_ERR,
 574            "[gl] 2 general combiners needed for YUV combiner support (found %i)\n", i);
 575   glGetIntegerv(GL_MAX_TEXTURE_UNITS, &i);
 576   if (i < 3)
 577     mp_msg(MSGT_VO, MSGL_ERR,
 578            "[gl] 3 texture units needed for YUV combiner support (found %i)\n", i);
 579   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 580   CombinerParameterfv(GL_CONSTANT_COLOR0_NV, ucoef);
 581   CombinerParameterfv(GL_CONSTANT_COLOR1_NV, vcoef);
 582
 583   // UV first, like this green component cannot overflow
 584   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_A_NV,
 585                 GL_TEXTURE1, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 586   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_B_NV,
 587                 GL_CONSTANT_COLOR0_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 588   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_C_NV,
 589                 GL_TEXTURE2, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 590   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_D_NV,
 591                 GL_CONSTANT_COLOR1_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 592   CombinerOutput(GL_COMBINER0_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 593                  GL_SPARE0_NV, GL_SCALE_BY_FOUR_NV, GL_NONE, GL_FALSE,
 594                  GL_FALSE, GL_FALSE);
 595
 596   // stage 2
 597   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_A_NV, GL_SPARE0_NV,
 598                 GL_SIGNED_IDENTITY_NV, GL_RGB);
 599   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_B_NV, GL_ZERO,
 600                  GL_UNSIGNED_INVERT_NV, GL_RGB);
 601   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_C_NV,
 602                 GL_TEXTURE0, GL_SIGNED_IDENTITY_NV, GL_RGB);
 603   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_D_NV, GL_ZERO,
 604                 GL_UNSIGNED_INVERT_NV, GL_RGB);
 605   CombinerOutput(GL_COMBINER1_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 606                  GL_SPARE0_NV, GL_NONE, GL_NONE, GL_FALSE,
 607                  GL_FALSE, GL_FALSE);
 608
 609   // leave final combiner stage in default mode
 610   CombinerParameteri(GL_NUM_GENERAL_COMBINERS_NV, 2);
 611 }
 612
 613 /**
 614  * \brief Setup ATI version of register combiners for YUV to RGB conversion.
 615  * \param uvcos used for saturation and hue adjustment
 616  * \param uvsin used for saturation and hue adjustment
 617  *
 618  * ATI called this fragment shader, but the name is confusing in the
 619  * light of a very different OpenGL 2.0 extension with the same name
 620  */
 621 static void glSetupYUVCombinersATI(float uvcos, float uvsin) {
 622   GLfloat ucoef[4];
 623   GLfloat vcoef[4];
 624   GLint i;
 625   if (!BeginFragmentShader || !EndFragmentShader ||
 626       !SetFragmentShaderConstant || !SampleMap ||
 627       !ColorFragmentOp2 || !ColorFragmentOp3) {
 628     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner (ATI) functions missing!\n");
 629     return;
 630   }
 631   glGetIntegerv(GL_NUM_FRAGMENT_REGISTERS_ATI, &i);
 632   if (i < 3)
 633     mp_msg(MSGT_VO, MSGL_ERR,
 634            "[gl] 3 registers needed for YUV combiner (ATI) support (found %i)\n", i);
 635   glGetIntegerv (GL_MAX_TEXTURE_UNITS, &i);
 636   if (i < 3)
 637     mp_msg(MSGT_VO, MSGL_ERR,
 638            "[gl] 3 texture units needed for YUV combiner (ATI) support (found %i)\n", i);
 639   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 640   BeginFragmentShader();
 641   SetFragmentShaderConstant(GL_CON_0_ATI, ucoef);
 642   SetFragmentShaderConstant(GL_CON_1_ATI, vcoef);
 643   SampleMap(GL_REG_0_ATI, GL_TEXTURE0, GL_SWIZZLE_STR_ATI);
 644   SampleMap(GL_REG_1_ATI, GL_TEXTURE1, GL_SWIZZLE_STR_ATI);
 645   SampleMap(GL_REG_2_ATI, GL_TEXTURE2, GL_SWIZZLE_STR_ATI);
 646   // UV first, like this green component cannot overflow
 647   ColorFragmentOp2(GL_MUL_ATI, GL_REG_1_ATI, GL_NONE, GL_NONE,
 648                    GL_REG_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 649                    GL_CON_0_ATI, GL_NONE, GL_BIAS_BIT_ATI);
 650   ColorFragmentOp3(GL_MAD_ATI, GL_REG_2_ATI, GL_NONE, GL_4X_BIT_ATI,
 651                    GL_REG_2_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 652                    GL_CON_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 653                    GL_REG_1_ATI, GL_NONE, GL_NONE);
 654   ColorFragmentOp2(GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 655                    GL_REG_0_ATI, GL_NONE, GL_NONE,
 656                    GL_REG_2_ATI, GL_NONE, GL_NONE);
 657   EndFragmentShader();
 658 }
 659
 660 /**
 661  * \brief helper function for gen_spline_lookup_tex
 662  * \param x subpixel-position ((0,1) range) to calculate weights for
 663  * \param dst where to store transformed weights, must provide space for 4 GLfloats
 664  *
 665  * calculates the weights and stores them after appropriate transformation
 666  * for the scaler fragment program.
 667  */
 668 static void store_weights(float x, GLfloat *dst) {
 669   float w0 = (((-1 * x + 3) * x - 3) * x + 1) / 6;
 670   float w1 = ((( 3 * x - 6) * x + 0) * x + 4) / 6;
 671   float w2 = (((-3 * x + 3) * x + 3) * x + 1) / 6;
 672   float w3 = ((( 1 * x + 0) * x + 0) * x + 0) / 6;
 673   *dst++ = 1 + x - w1 / (w0 + w1);
 674   *dst++ = 1 - x + w3 / (w2 + w3);
 675   *dst++ = w0 + w1;
 676   *dst++ = 0;
 677 }
 678
 679 //! to avoid artefacts this should be rather large
 680 #define LOOKUP_BSPLINE_RES (2 * 1024)
 681 /**
 682  * \brief creates the 1D lookup texture needed for fast higher-order filtering
 683  * \param unit texture unit to attach texture to
 684  */
 685 static void gen_spline_lookup_tex(GLenum unit) {
 686   GLfloat tex[4 * LOOKUP_BSPLINE_RES];
 687   GLfloat *tp = tex;
 688   int i;
 689   for (i = 0; i < LOOKUP_BSPLINE_RES; i++) {
 690     float x = (float)(i + 0.5) / LOOKUP_BSPLINE_RES;
 691     store_weights(x, tp);
 692     tp += 4;
 693   }
 694   store_weights(0, tex);
 695   store_weights(1, &tex[4 * (LOOKUP_BSPLINE_RES - 1)]);
 696   ActiveTexture(unit);
 697   glTexImage1D(GL_TEXTURE_1D, 0, GL_RGBA16, LOOKUP_BSPLINE_RES, 0, GL_RGBA, GL_FLOAT, tex);
 698   glTexParameterf(GL_TEXTURE_1D, GL_TEXTURE_PRIORITY, 1.0);
 699   glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 700   glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 701   glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_REPEAT);
 702   ActiveTexture(GL_TEXTURE0);
 703 }
 704
 705 static const char *bilin_filt_template =
 706   "TEX yuv.%c, fragment.texcoord[%c], texture[%c], %s;";
 707
 708 #define BICUB_FILT_MAIN(textype) \
 709   /* first y-interpolation */ \
 710   "ADD coord, fragment.texcoord[%c].xyxy, cdelta.xyxw;" \
 711   "ADD coord2, fragment.texcoord[%c].xyxy, cdelta.zyzw;" \
 712   "TEX a.r, coord.xyxy, texture[%c], "textype";" \
 713   "TEX a.g, coord.zwzw, texture[%c], "textype";" \
 714   /* second y-interpolation */ \
 715   "TEX b.r, coord2.xyxy, texture[%c], "textype";" \
 716   "TEX b.g, coord2.zwzw, texture[%c], "textype";" \
 717   "LRP a.b, parmy.b, a.rrrr, a.gggg;" \
 718   "LRP a.a, parmy.b, b.rrrr, b.gggg;" \
 719   /* x-interpolation */ \
 720   "LRP yuv.%c, parmx.b, a.bbbb, a.aaaa;"
 721
 722 static const char *bicub_filt_template_2D =
 723   "MAD coord.xy, fragment.texcoord[%c], {%f, %f}, {0.5, 0.5};"
 724   "TEX parmx, coord.x, texture[%c], 1D;"
 725   "MUL cdelta.xz, parmx.rrgg, {-%f, 0, %f, 0};"
 726   "TEX parmy, coord.y, texture[%c], 1D;"
 727   "MUL cdelta.yw, parmy.rrgg, {0, -%f, 0, %f};"
 728   BICUB_FILT_MAIN("2D");
 729
 730 static const char *bicub_filt_template_RECT =
 731   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};"
 732   "TEX parmx, coord.x, texture[%c], 1D;"
 733   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};"
 734   "TEX parmy, coord.y, texture[%c], 1D;"
 735   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};"
 736   BICUB_FILT_MAIN("RECT");
 737
 738 #define CALCWEIGHTS(t, s) \
 739   "MAD "t", {-0.5, 0.1666, 0.3333, -0.3333}, "s", {1, 0, -0.5, 0.5};" \
 740   "MAD "t", "t", "s", {0, 0, -0.5, 0.5};" \
 741   "MAD "t", "t", "s", {-0.6666, 0, 0.8333, 0.1666};" \
 742   "RCP a.x, "t".z;" \
 743   "RCP a.y, "t".w;" \
 744   "MAD "t".xy, "t".xyxy, a.xyxy, {1, 1, 0, 0};" \
 745   "ADD "t".x, "t".xxxx, "s";" \
 746   "SUB "t".y, "t".yyyy, "s";"
 747
 748 static const char *bicub_notex_filt_template_2D =
 749   "MAD coord.xy, fragment.texcoord[%c], {%f, %f}, {0.5, 0.5};"
 750   "FRC coord.xy, coord.xyxy;"
 751   CALCWEIGHTS("parmx", "coord.xxxx")
 752   "MUL cdelta.xz, parmx.rrgg, {-%f, 0, %f, 0};"
 753   CALCWEIGHTS("parmy", "coord.yyyy")
 754   "MUL cdelta.yw, parmy.rrgg, {0, -%f, 0, %f};"
 755   BICUB_FILT_MAIN("2D");
 756
 757 static const char *bicub_notex_filt_template_RECT =
 758   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};"
 759   "FRC coord.xy, coord.xyxy;"
 760   CALCWEIGHTS("parmx", "coord.xxxx")
 761   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};"
 762   CALCWEIGHTS("parmy", "coord.yyyy")
 763   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};"
 764   BICUB_FILT_MAIN("RECT");
 765
 766 #define BICUB_X_FILT_MAIN(textype) \
 767   "ADD coord.xy, fragment.texcoord[%c].xyxy, cdelta.xyxy;" \
 768   "ADD coord2.xy, fragment.texcoord[%c].xyxy, cdelta.zyzy;" \
 769   "TEX a.r, coord, texture[%c], "textype";" \
 770   "TEX b.r, coord2, texture[%c], "textype";" \
 771   /* x-interpolation */ \
 772   "LRP yuv.%c, parmx.b, a.rrrr, b.rrrr;"
 773
 774 static const char *bicub_x_filt_template_2D =
 775   "MAD coord.x, fragment.texcoord[%c], {%f}, {0.5};"
 776   "TEX parmx, coord, texture[%c], 1D;"
 777   "MUL cdelta.xyz, parmx.rrgg, {-%f, 0, %f};"
 778   BICUB_X_FILT_MAIN("2D");
 779
 780 static const char *bicub_x_filt_template_RECT =
 781   "ADD coord.x, fragment.texcoord[%c], {0.5};"
 782   "TEX parmx, coord, texture[%c], 1D;"
 783   "MUL cdelta.xyz, parmx.rrgg, {-1, 0, 1};"
 784   BICUB_X_FILT_MAIN("RECT");
 785
 786 static const char *unsharp_filt_template =
 787   "PARAM dcoord%c = {%f, %f, %f, %f};"
 788   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;"
 789   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;"
 790   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;"
 791   "TEX b.r, coord.xyxy, texture[%c], %s;"
 792   "TEX b.g, coord.zwzw, texture[%c], %s;"
 793   "ADD b.r, b.r, b.g;"
 794   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 795   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 796   "DP3 b, b, {0.25, 0.25, 0.25};"
 797   "SUB b.r, a.r, b.r;"
 798   "MAD yuv.%c, b.r, {%f}, a.r;";
 799
 800 static const char *unsharp_filt_template2 =
 801   "PARAM dcoord%c = {%f, %f, %f, %f};"
 802   "PARAM dcoord2%c = {%f, 0, 0, %f};"
 803   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;"
 804   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;"
 805   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;"
 806   "TEX b.r, coord.xyxy, texture[%c], %s;"
 807   "TEX b.g, coord.zwzw, texture[%c], %s;"
 808   "ADD b.r, b.r, b.g;"
 809   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 810   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 811   "ADD b.r, b.r, b.b;"
 812   "ADD b.a, b.r, b.g;"
 813   "ADD coord, fragment.texcoord[%c].xyxy, dcoord2%c;"
 814   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord2%c;"
 815   "TEX b.r, coord.xyxy, texture[%c], %s;"
 816   "TEX b.g, coord.zwzw, texture[%c], %s;"
 817   "ADD b.r, b.r, b.g;"
 818   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 819   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 820   "DP4 b.r, b, {-0.1171875, -0.1171875, -0.1171875, -0.09765625};"
 821   "MAD b.r, a.r, {0.859375}, b.r;"
 822   "MAD yuv.%c, b.r, {%f}, a.r;";
 823
 824 static const char *yuv_prog_template =
 825   "PARAM ycoef = {%.4f, %.4f, %.4f};"
 826   "PARAM ucoef = {%.4f, %.4f, %.4f};"
 827   "PARAM vcoef = {%.4f, %.4f, %.4f};"
 828   "PARAM offsets = {%.4f, %.4f, %.4f};"
 829   "TEMP res;"
 830   "MAD res.rgb, yuv.rrrr, ycoef, offsets;"
 831   "MAD res.rgb, yuv.gggg, ucoef, res;"
 832   "MAD result.color.rgb, yuv.bbbb, vcoef, res;"
 833   "END";
 834
 835 static const char *yuv_pow_prog_template =
 836   "PARAM ycoef = {%.4f, %.4f, %.4f};"
 837   "PARAM ucoef = {%.4f, %.4f, %.4f};"
 838   "PARAM vcoef = {%.4f, %.4f, %.4f};"
 839   "PARAM offsets = {%.4f, %.4f, %.4f};"
 840   "PARAM gamma = {%.4f, %.4f, %.4f};"
 841   "TEMP res;"
 842   "MAD res.rgb, yuv.rrrr, ycoef, offsets;"
 843   "MAD res.rgb, yuv.gggg, ucoef, res;"
 844   "MAD_SAT res.rgb, yuv.bbbb, vcoef, res;"
 845   "POW result.color.r, res.r, gamma.r;"
 846   "POW result.color.g, res.g, gamma.g;"
 847   "POW result.color.b, res.b, gamma.b;"
 848   "END";
 849
 850 static const char *yuv_lookup_prog_template =
 851   "PARAM ycoef = {%.4f, %.4f, %.4f, 0};"
 852   "PARAM ucoef = {%.4f, %.4f, %.4f, 0};"
 853   "PARAM vcoef = {%.4f, %.4f, %.4f, 0};"
 854   "PARAM offsets = {%.4f, %.4f, %.4f, 0.125};"
 855   "TEMP res;"
 856   "MAD res, yuv.rrrr, ycoef, offsets;"
 857   "MAD res.rgb, yuv.gggg, ucoef, res;"
 858   "MAD res.rgb, yuv.bbbb, vcoef, res;"
 859   "TEX result.color.r, res.raaa, texture[%c], 2D;"
 860   "ADD res.a, res.a, 0.25;"
 861   "TEX result.color.g, res.gaaa, texture[%c], 2D;"
 862   "ADD res.a, res.a, 0.25;"
 863   "TEX result.color.b, res.baaa, texture[%c], 2D;"
 864   "END";
 865
 866 static const char *yuv_lookup3d_prog_template =
 867   "TEX result.color, yuv, texture[%c], 3D;"
 868   "END";
 869
 870 /**
 871  * \brief creates and initializes helper textures needed for scaling texture read
 872  * \param scaler scaler type to create texture for
 873  * \param texu contains next free texture unit number
 874  * \param texs texture unit ids for the scaler are stored in this array
 875  */
 876 static void create_scaler_textures(int scaler, int *texu, char *texs) {
 877   switch (scaler) {
 878     case YUV_SCALER_BILIN:
 879     case YUV_SCALER_BICUB_NOTEX:
 880     case YUV_SCALER_UNSHARP:
 881     case YUV_SCALER_UNSHARP2:
 882       break;
 883     case YUV_SCALER_BICUB:
 884     case YUV_SCALER_BICUB_X:
 885       texs[0] = (*texu)++;
 886       gen_spline_lookup_tex(GL_TEXTURE0 + texs[0]);
 887       texs[0] += '0';
 888       break;
 889     default:
 890       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown scaler type %i\n", scaler);
 891   }
 892 }
 893
 894 static void gen_gamma_map(unsigned char *map, int size, float gamma);
 895
 896 #define ROW_R 0
 897 #define ROW_G 1
 898 #define ROW_B 2
 899 #define COL_Y 0
 900 #define COL_U 1
 901 #define COL_V 2
 902 #define COL_C 3
 903
 904 static void get_yuv2rgb_coeffs(gl_conversion_params_t *params, float yuv2rgb[3][4]) {
 905   float uvcos = params->saturation * cos(params->hue);
 906   float uvsin = params->saturation * sin(params->hue);
 907   int i;
 908   float uv_coeffs[3][2] = {
 909     { 0.000,  1.596},
 910     {-0.391, -0.813},
 911     { 2.018,  0.000}
 912   };
 913   for (i = 0; i < 3; i++) {
 914     yuv2rgb[i][COL_C]  = params->brightness;
 915     yuv2rgb[i][COL_Y]  = 1.164 * params->contrast;
 916     yuv2rgb[i][COL_C] += (-16 / 255.0) * yuv2rgb[i][COL_Y];
 917     yuv2rgb[i][COL_U]  = uv_coeffs[i][0] * uvcos + uv_coeffs[i][1] * uvsin;
 918     yuv2rgb[i][COL_C] += (-128 / 255.0) * yuv2rgb[i][COL_U];
 919     yuv2rgb[i][COL_V]  = uv_coeffs[i][0] * uvsin + uv_coeffs[i][1] * uvcos;
 920     yuv2rgb[i][COL_C] += (-128 / 255.0) * yuv2rgb[i][COL_V];
 921     // this "centers" contrast control so that e.g. a contrast of 0
 922     // leads to a grey image, not a black one
 923     yuv2rgb[i][COL_C] += 0.5 - params->contrast / 2.0;
 924   }
 925 }
 926
 927 //! size of gamma map use to avoid slow exp function in gen_yuv2rgb_map
 928 #define GMAP_SIZE (1024)
 929 /**
 930  * \brief generate a 3D YUV -> RGB map
 931  * \param params struct containing parameters like brightness, gamma, ...
 932  * \param map where to store map. Must provide space for (size + 2)^3 elements
 933  * \param size size of the map, excluding border
 934  */
 935 static void gen_yuv2rgb_map(gl_conversion_params_t *params, unsigned char *map, int size) {
 936   int i, j, k, l;
 937   float step = 1.0 / size;
 938   float y, u, v;
 939   float yuv2rgb[3][4];
 940   unsigned char gmaps[3][GMAP_SIZE];
 941   gen_gamma_map(gmaps[0], GMAP_SIZE, params->rgamma);
 942   gen_gamma_map(gmaps[1], GMAP_SIZE, params->ggamma);
 943   gen_gamma_map(gmaps[2], GMAP_SIZE, params->bgamma);
 944   get_yuv2rgb_coeffs(params, yuv2rgb);
 945   for (i = 0; i < 3; i++)
 946     for (j = 0; j < 4; j++)
 947       yuv2rgb[i][j] *= GMAP_SIZE - 1;
 948   v = 0;
 949   for (i = -1; i <= size; i++) {
 950     u = 0;
 951     for (j = -1; j <= size; j++) {
 952       y = 0;
 953       for (k = -1; k <= size; k++) {
 954         for (l = 0; l < 3; l++) {
 955           float rgb = yuv2rgb[l][COL_Y] * y + yuv2rgb[l][COL_U] * u + yuv2rgb[l][COL_V] * v + yuv2rgb[l][COL_C];
 956           *map++ = gmaps[l][av_clip(rgb, 0, GMAP_SIZE - 1)];
 957         }
 958         y += (k == -1 || k == size - 1) ? step / 2 : step;
 959       }
 960       u += (j == -1 || j == size - 1) ? step / 2 : step;
 961     }
 962     v += (i == -1 || i == size - 1) ? step / 2 : step;
 963   }
 964 }
 965
 966 //! resolution of texture for gamma lookup table
 967 #define LOOKUP_RES 512
 968 //! resolution for 3D yuv->rgb conversion lookup table
 969 #define LOOKUP_3DRES 32
 970 /**
 971  * \brief creates and initializes helper textures needed for yuv conversion
 972  * \param params struct containing parameters like brightness, gamma, ...
 973  * \param texu contains next free texture unit number
 974  * \param texs texture unit ids for the conversion are stored in this array
 975  */
 976 static void create_conv_textures(gl_conversion_params_t *params, int *texu, char *texs) {
 977   unsigned char *lookup_data = NULL;
 978   int conv = YUV_CONVERSION(params->type);
 979   switch (conv) {
 980     case YUV_CONVERSION_FRAGMENT:
 981     case YUV_CONVERSION_FRAGMENT_POW:
 982      break;
 983     case YUV_CONVERSION_FRAGMENT_LOOKUP:
 984       texs[0] = (*texu)++;
 985       ActiveTexture(GL_TEXTURE0 + texs[0]);
 986       lookup_data = malloc(4 * LOOKUP_RES);
 987       gen_gamma_map(lookup_data, LOOKUP_RES, params->rgamma);
 988       gen_gamma_map(&lookup_data[LOOKUP_RES], LOOKUP_RES, params->ggamma);
 989       gen_gamma_map(&lookup_data[2 * LOOKUP_RES], LOOKUP_RES, params->bgamma);
 990       glCreateClearTex(GL_TEXTURE_2D, GL_LUMINANCE8, GL_LUMINANCE, GL_UNSIGNED_BYTE, GL_LINEAR,
 991                        LOOKUP_RES, 4, 0);
 992       glUploadTex(GL_TEXTURE_2D, GL_LUMINANCE, GL_UNSIGNED_BYTE, lookup_data,
 993                   LOOKUP_RES, 0, 0, LOOKUP_RES, 4, 0);
 994       ActiveTexture(GL_TEXTURE0);
 995       texs[0] += '0';
 996       break;
 997     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
 998       {
 999         int sz = LOOKUP_3DRES + 2; // texture size including borders
1000         if (!TexImage3D) {
1001           mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing 3D texture function!\n");
1002           break;
1003         }
1004         texs[0] = (*texu)++;
1005         ActiveTexture(GL_TEXTURE0 + texs[0]);
1006         lookup_data = malloc(3 * sz * sz * sz);
1007         gen_yuv2rgb_map(params, lookup_data, LOOKUP_3DRES);
1008         glAdjustAlignment(sz);
1009         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
1010         TexImage3D(GL_TEXTURE_3D, 0, 3, sz, sz, sz, 1,
1011                    GL_RGB, GL_UNSIGNED_BYTE, lookup_data);
1012         glTexParameterf(GL_TEXTURE_3D, GL_TEXTURE_PRIORITY, 1.0);
1013         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1014         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1015         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_S, GL_CLAMP);
1016         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_T, GL_CLAMP);
1017         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_R, GL_CLAMP);
1018         ActiveTexture(GL_TEXTURE0);
1019         texs[0] += '0';
1020       }
1021       break;
1022     default:
1023       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", conv);
1024   }
1025   if (lookup_data)
1026     free(lookup_data);
1027 }
1028
1029 /**
1030  * \brief adds a scaling texture read at the current fragment program position
1031  * \param scaler type of scaler to insert
1032  * \param prog_pos current position in fragment program
1033  * \param remain how many bytes remain in the buffer given by prog_pos
1034  * \param texs array containing the texture unit identifiers for this scaler
1035  * \param in_tex texture unit the scaler should read from
1036  * \param out_comp component of the yuv variable the scaler stores the result in
1037  * \param rect if rectangular (pixel) adressing should be used for in_tex
1038  * \param texw width of the in_tex texture
1039  * \param texh height of the in_tex texture
1040  * \param strength strength of filter effect if the scaler does some kind of filtering
1041  */
1042 static void add_scaler(int scaler, char **prog_pos, int *remain, char *texs,
1043                     char in_tex, char out_comp, int rect, int texw, int texh,
1044                     double strength) {
1045   const char *ttype = rect ? "RECT" : "2D";
1046   const float ptw = rect ? 1.0 : 1.0 / texw;
1047   const float pth = rect ? 1.0 : 1.0 / texh;
1048   switch (scaler) {
1049     case YUV_SCALER_BILIN:
1050       snprintf(*prog_pos, *remain, bilin_filt_template, out_comp, in_tex,
1051                  in_tex, ttype);
1052       break;
1053     case YUV_SCALER_BICUB:
1054       if (rect)
1055         snprintf(*prog_pos, *remain, bicub_filt_template_RECT,
1056                  in_tex, texs[0], texs[0],
1057                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1058       else
1059         snprintf(*prog_pos, *remain, bicub_filt_template_2D,
1060                  in_tex, (float)texw, (float)texh,
1061                  texs[0], ptw, ptw, texs[0], pth, pth,
1062                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1063       break;
1064     case YUV_SCALER_BICUB_X:
1065       if (rect)
1066         snprintf(*prog_pos, *remain, bicub_x_filt_template_RECT,
1067                  in_tex, texs[0],
1068                  in_tex, in_tex, in_tex, in_tex, out_comp);
1069       else
1070         snprintf(*prog_pos, *remain, bicub_x_filt_template_2D,
1071                  in_tex, (float)texw,
1072                  texs[0], ptw, ptw,
1073                  in_tex, in_tex, in_tex, in_tex, out_comp);
1074       break;
1075     case YUV_SCALER_BICUB_NOTEX:
1076       if (rect)
1077         snprintf(*prog_pos, *remain, bicub_notex_filt_template_RECT,
1078                  in_tex,
1079                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1080       else
1081         snprintf(*prog_pos, *remain, bicub_notex_filt_template_2D,
1082                  in_tex, (float)texw, (float)texh, ptw, ptw, pth, pth,
1083                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1084       break;
1085     case YUV_SCALER_UNSHARP:
1086       snprintf(*prog_pos, *remain, unsharp_filt_template,
1087                out_comp, 0.5 * ptw, 0.5 * pth, 0.5 * ptw, -0.5 * pth,
1088                in_tex, out_comp, in_tex, out_comp, in_tex,
1089                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1090                in_tex, ttype, out_comp, strength);
1091       break;
1092     case YUV_SCALER_UNSHARP2:
1093       snprintf(*prog_pos, *remain, unsharp_filt_template2,
1094                out_comp, 1.2 * ptw, 1.2 * pth, 1.2 * ptw, -1.2 * pth,
1095                out_comp, 1.5 * ptw, 1.5 * pth,
1096                in_tex, out_comp, in_tex, out_comp, in_tex,
1097                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1098                in_tex, ttype, in_tex, out_comp, in_tex, out_comp,
1099                in_tex, ttype, in_tex, ttype, in_tex, ttype,
1100                in_tex, ttype, out_comp, strength);
1101       break;
1102   }
1103   *remain -= strlen(*prog_pos);
1104   *prog_pos += strlen(*prog_pos);
1105 }
1106
1107 static const struct {
1108   const char *name;
1109   GLenum cur;
1110   GLenum max;
1111 } progstats[] = {
1112   {"instructions", 0x88A0, 0x88A1},
1113   {"native instructions", 0x88A2, 0x88A3},
1114   {"temporaries", 0x88A4, 0x88A5},
1115   {"native temporaries", 0x88A6, 0x88A7},
1116   {"parameters", 0x88A8, 0x88A9},
1117   {"native parameters", 0x88AA, 0x88AB},
1118   {"attribs", 0x88AC, 0x88AD},
1119   {"native attribs", 0x88AE, 0x88AF},
1120   {"ALU instructions", 0x8805, 0x880B},
1121   {"TEX instructions", 0x8806, 0x880C},
1122   {"TEX indirections", 0x8807, 0x880D},
1123   {"native ALU instructions", 0x8808, 0x880E},
1124   {"native TEX instructions", 0x8809, 0x880F},
1125   {"native TEX indirections", 0x880A, 0x8810},
1126   {NULL, 0, 0}
1127 };
1128
1129 /**
1130  * \brief load the specified GPU Program
1131  * \param target program target to load into, only GL_FRAGMENT_PROGRAM is tested
1132  * \param prog program string
1133  * \return 1 on success, 0 otherwise
1134  */
1135 int loadGPUProgram(GLenum target, char *prog) {
1136   int i;
1137   GLint cur = 0, max = 0, err = 0;
1138   if (!ProgramString) {
1139     mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing GPU program function\n");
1140     return 0;
1141   }
1142   ProgramString(target, GL_PROGRAM_FORMAT_ASCII, strlen(prog), prog);
1143   glGetIntegerv(GL_PROGRAM_ERROR_POSITION, &err);
1144   if (err != -1) {
1145     mp_msg(MSGT_VO, MSGL_ERR,
1146       "[gl] Error compiling fragment program, make sure your card supports\n"
1147       "[gl]   GL_ARB_fragment_program (use glxinfo to check).\n"
1148       "[gl]   Error message:\n  %s at %.10s\n",
1149       glGetString(GL_PROGRAM_ERROR_STRING), &prog[err]);
1150     return 0;
1151   }
1152   if (!GetProgramiv || !mp_msg_test(MSGT_VO, MSGL_V))
1153     return 1;
1154   mp_msg(MSGT_VO, MSGL_V, "[gl] Program statistics:\n");
1155   for (i = 0; progstats[i].name; i++) {
1156     GetProgramiv(target, progstats[i].cur, &cur);
1157     GetProgramiv(target, progstats[i].max, &max);
1158     mp_msg(MSGT_VO, MSGL_V, "[gl]   %s: %i/%i\n", progstats[i].name, cur, max);
1159   }
1160   return 1;
1161 }
1162
1163 #define MAX_PROGSZ (1024*1024)
1164
1165 /**
1166  * \brief setup a fragment program that will do YUV->RGB conversion
1167  * \param parms struct containing parameters like conversion and scaler type,
1168  *              brightness, ...
1169  */
1170 static void glSetupYUVFragprog(gl_conversion_params_t *params) {
1171   int type = params->type;
1172   int texw = params->texw;
1173   int texh = params->texh;
1174   int rect = params->target == GL_TEXTURE_RECTANGLE;
1175   static const char prog_hdr[] =
1176     "!!ARBfp1.0\n"
1177     "OPTION ARB_precision_hint_fastest;"
1178     // all scaler variables must go here so they aren't defined
1179     // multiple times when the same scaler is used more than once
1180     "TEMP coord, coord2, cdelta, parmx, parmy, a, b, yuv;";
1181   int prog_remain;
1182   char *yuv_prog, *prog_pos;
1183   int cur_texu = 3;
1184   char lum_scale_texs[1];
1185   char chrom_scale_texs[1];
1186   char conv_texs[1];
1187   GLint i;
1188   // this is the conversion matrix, with y, u, v factors
1189   // for red, green, blue and the constant offsets
1190   float yuv2rgb[3][4];
1191   create_conv_textures(params, &cur_texu, conv_texs);
1192   create_scaler_textures(YUV_LUM_SCALER(type), &cur_texu, lum_scale_texs);
1193   if (YUV_CHROM_SCALER(type) == YUV_LUM_SCALER(type))
1194     memcpy(chrom_scale_texs, lum_scale_texs, sizeof(chrom_scale_texs));
1195   else
1196     create_scaler_textures(YUV_CHROM_SCALER(type), &cur_texu, chrom_scale_texs);
1197   glGetIntegerv(GL_MAX_TEXTURE_IMAGE_UNITS, &i);
1198   if (i < cur_texu)
1199     mp_msg(MSGT_VO, MSGL_ERR,
1200            "[gl] %i texture units needed for this type of YUV fragment support (found %i)\n",
1201            cur_texu, i);
1202   if (!ProgramString) {
1203     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] ProgramString function missing!\n");
1204     return;
1205   }
1206   yuv_prog = malloc(MAX_PROGSZ);
1207   strcpy(yuv_prog, prog_hdr);
1208   prog_pos    = yuv_prog + sizeof(prog_hdr) - 1;
1209   prog_remain = MAX_PROGSZ - sizeof(prog_hdr);
1210   add_scaler(YUV_LUM_SCALER(type), &prog_pos, &prog_remain, lum_scale_texs,
1211              '0', 'r', rect, texw, texh, params->filter_strength);
1212   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1213              '1', 'g', rect, texw / 2, texh / 2, params->filter_strength);
1214   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1215              '2', 'b', rect, texw / 2, texh / 2, params->filter_strength);
1216   get_yuv2rgb_coeffs(params, yuv2rgb);
1217   switch (YUV_CONVERSION(type)) {
1218     case YUV_CONVERSION_FRAGMENT:
1219       snprintf(prog_pos, prog_remain, yuv_prog_template,
1220                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1221                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1222                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1223                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C]);
1224       break;
1225     case YUV_CONVERSION_FRAGMENT_POW:
1226       snprintf(prog_pos, prog_remain, yuv_pow_prog_template,
1227                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1228                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1229                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1230                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1231                (float)1.0 / params->rgamma, (float)1.0 / params->bgamma, (float)1.0 / params->bgamma);
1232       break;
1233     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1234       snprintf(prog_pos, prog_remain, yuv_lookup_prog_template,
1235                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1236                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1237                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1238                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1239                conv_texs[0], conv_texs[0], conv_texs[0]);
1240       break;
1241     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1242       snprintf(prog_pos, prog_remain, yuv_lookup3d_prog_template, conv_texs[0]);
1243       break;
1244     default:
1245       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(type));
1246       break;
1247   }
1248   mp_msg(MSGT_VO, MSGL_V, "[gl] generated fragment program:\n%s\n", yuv_prog);
1249   loadGPUProgram(GL_FRAGMENT_PROGRAM, yuv_prog);
1250   free(yuv_prog);
1251 }
1252
1253 /**
1254  * \brief little helper function to create a lookup table for gamma
1255  * \param map buffer to create map into
1256  * \param size size of buffer
1257  * \param gamma gamma value
1258  */
1259 static void gen_gamma_map(unsigned char *map, int size, float gamma) {
1260   int i;
1261   if (gamma == 1.0) {
1262     for (i = 0; i < size; i++)
1263       map[i] = 255 * i / (size - 1);
1264     return;
1265   }
1266   gamma = 1.0 / gamma;
1267   for (i = 0; i < size; i++) {
1268     float tmp = (float)i / (size - 1.0);
1269     tmp = pow(tmp, gamma);
1270     if (tmp > 1.0) tmp = 1.0;
1271     if (tmp < 0.0) tmp = 0.0;
1272     map[i] = 255 * tmp;
1273   }
1274 }
1275
1276 /**
1277  * \brief setup YUV->RGB conversion
1278  * \param parms struct containing parameters like conversion and scaler type,
1279  *              brightness, ...
1280  * \ingroup glconversion
1281  */
1282 void glSetupYUVConversion(gl_conversion_params_t *params) {
1283   float uvcos = params->saturation * cos(params->hue);
1284   float uvsin = params->saturation * sin(params->hue);
1285   switch (YUV_CONVERSION(params->type)) {
1286     case YUV_CONVERSION_COMBINERS:
1287       glSetupYUVCombiners(uvcos, uvsin);
1288       break;
1289     case YUV_CONVERSION_COMBINERS_ATI:
1290       glSetupYUVCombinersATI(uvcos, uvsin);
1291       break;
1292     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1293     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1294     case YUV_CONVERSION_FRAGMENT:
1295     case YUV_CONVERSION_FRAGMENT_POW:
1296       glSetupYUVFragprog(params);
1297       break;
1298     default:
1299       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(params->type));
1300   }
1301 }
1302
1303 /**
1304  * \brief enable the specified YUV conversion
1305  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1306  * \param type type of YUV conversion
1307  * \ingroup glconversion
1308  */
1309 void glEnableYUVConversion(GLenum target, int type) {
1310   if (type <= 0) return;
1311   switch (YUV_CONVERSION(type)) {
1312     case YUV_CONVERSION_COMBINERS:
1313       ActiveTexture(GL_TEXTURE1);
1314       glEnable(target);
1315       ActiveTexture(GL_TEXTURE2);
1316       glEnable(target);
1317       ActiveTexture(GL_TEXTURE0);
1318       glEnable(GL_REGISTER_COMBINERS_NV);
1319       break;
1320     case YUV_CONVERSION_COMBINERS_ATI:
1321       ActiveTexture(GL_TEXTURE1);
1322       glEnable(target);
1323       ActiveTexture(GL_TEXTURE2);
1324       glEnable(target);
1325       ActiveTexture(GL_TEXTURE0);
1326       glEnable(GL_FRAGMENT_SHADER_ATI);
1327       break;
1328     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1329     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1330     case YUV_CONVERSION_FRAGMENT_POW:
1331     case YUV_CONVERSION_FRAGMENT:
1332       glEnable(GL_FRAGMENT_PROGRAM);
1333       break;
1334   }
1335 }
1336
1337 /**
1338  * \brief disable the specified YUV conversion
1339  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1340  * \param type type of YUV conversion
1341  * \ingroup glconversion
1342  */
1343 void glDisableYUVConversion(GLenum target, int type) {
1344   if (type <= 0) return;
1345   switch (YUV_CONVERSION(type)) {
1346     case YUV_CONVERSION_COMBINERS:
1347       ActiveTexture(GL_TEXTURE1);
1348       glDisable(target);
1349       ActiveTexture(GL_TEXTURE2);
1350       glDisable(target);
1351       ActiveTexture(GL_TEXTURE0);
1352       glDisable(GL_REGISTER_COMBINERS_NV);
1353       break;
1354     case YUV_CONVERSION_COMBINERS_ATI:
1355       ActiveTexture(GL_TEXTURE1);
1356       glDisable(target);
1357       ActiveTexture(GL_TEXTURE2);
1358       glDisable(target);
1359       ActiveTexture(GL_TEXTURE0);
1360       glDisable(GL_FRAGMENT_SHADER_ATI);
1361       break;
1362     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1363     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1364     case YUV_CONVERSION_FRAGMENT_POW:
1365     case YUV_CONVERSION_FRAGMENT:
1366       glDisable(GL_FRAGMENT_PROGRAM);
1367       break;
1368   }
1369 }
1370
1371 /**
1372  * \brief draw a texture part at given 2D coordinates
1373  * \param x screen top coordinate
1374  * \param y screen left coordinate
1375  * \param w screen width coordinate
1376  * \param h screen height coordinate
1377  * \param tx texture top coordinate in pixels
1378  * \param ty texture left coordinate in pixels
1379  * \param tw texture part width in pixels
1380  * \param th texture part height in pixels
1381  * \param sx width of texture in pixels
1382  * \param sy height of texture in pixels
1383  * \param rect_tex whether this texture uses texture_rectangle extension
1384  * \param is_yv12 if set, also draw the textures from units 1 and 2
1385  * \param flip flip the texture upside down
1386  * \ingroup gltexture
1387  */
1388 void glDrawTex(GLfloat x, GLfloat y, GLfloat w, GLfloat h,
1389                GLfloat tx, GLfloat ty, GLfloat tw, GLfloat th,
1390                int sx, int sy, int rect_tex, int is_yv12, int flip) {
1391   GLfloat tx2 = tx / 2, ty2 = ty / 2, tw2 = tw / 2, th2 = th / 2;
1392   if (!rect_tex) {
1393     tx /= sx; ty /= sy; tw /= sx; th /= sy;
1394     tx2 = tx, ty2 = ty, tw2 = tw, th2 = th;
1395   }
1396   if (flip) {
1397     y += h;
1398     h = -h;
1399   }
1400   glBegin(GL_QUADS);
1401   glTexCoord2f(tx, ty);
1402   if (is_yv12) {
1403     MultiTexCoord2f(GL_TEXTURE1, tx2, ty2);
1404     MultiTexCoord2f(GL_TEXTURE2, tx2, ty2);
1405   }
1406   glVertex2f(x, y);
1407   glTexCoord2f(tx, ty + th);
1408   if (is_yv12) {
1409     MultiTexCoord2f(GL_TEXTURE1, tx2, ty2 + th2);
1410     MultiTexCoord2f(GL_TEXTURE2, tx2, ty2 + th2);
1411   }
1412   glVertex2f(x, y + h);
1413   glTexCoord2f(tx + tw, ty + th);
1414   if (is_yv12) {
1415     MultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2 + th2);
1416     MultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2 + th2);
1417   }
1418   glVertex2f(x + w, y + h);
1419   glTexCoord2f(tx + tw, ty);
1420   if (is_yv12) {
1421     MultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2);
1422     MultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2);
1423   }
1424   glVertex2f(x + w, y);
1425   glEnd();
1426 }
1427
1428 #ifdef GL_WIN32
1429 #include "w32_common.h"
1430 /**
1431  * \brief little helper since wglGetProcAddress definition does not fit our
1432  *        getProcAddress
1433  * \param procName name of function to look up
1434  * \return function pointer returned by wglGetProcAddress
1435  */
1436 static void *w32gpa(const GLubyte *procName) {
1437   HMODULE oglmod;
1438   void *res = wglGetProcAddress(procName);
1439   if (res) return res;
1440   oglmod = GetModuleHandle("opengl32.dll");
1441   return GetProcAddress(oglmod, procName);
1442 }
1443
1444 int setGlWindow(int *vinfo, HGLRC *context, HWND win)
1445 {
1446   int new_vinfo;
1447   HDC windc = GetDC(win);
1448   HGLRC new_context = 0;
1449   int keep_context = 0;
1450   int res = SET_WINDOW_FAILED;
1451
1452   // should only be needed when keeping context, but not doing glFinish
1453   // can cause flickering even when we do not keep it.
1454   if (*context)
1455   glFinish();
1456   new_vinfo = GetPixelFormat(windc);
1457   if (*context && *vinfo && new_vinfo && *vinfo == new_vinfo) {
1458       // we can keep the wglContext
1459       new_context = *context;
1460       keep_context = 1;
1461   } else {
1462     // create a context
1463     new_context = wglCreateContext(windc);
1464     if (!new_context) {
1465       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GL context!\n");
1466       goto out;
1467     }
1468   }
1469
1470   // set context
1471   if (!wglMakeCurrent(windc, new_context)) {
1472     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GL context!\n");
1473     if (!keep_context) {
1474       wglDeleteContext(new_context);
1475     }
1476     goto out;
1477   }
1478
1479   // set new values
1480   vo_w32_window = win;
1481   {
1482     RECT rect;
1483     GetClientRect(win, &rect);
1484     vo_dwidth = rect.right;
1485     vo_dheight = rect.bottom;
1486   }
1487   if (!keep_context) {
1488     if (*context)
1489       wglDeleteContext(*context);
1490     *context = new_context;
1491     *vinfo = new_vinfo;
1492     getFunctions(w32gpa, NULL);
1493
1494     // and inform that reinit is neccessary
1495     res = SET_WINDOW_REINIT;
1496   } else
1497     res = SET_WINDOW_OK;
1498
1499 out:
1500   ReleaseDC(win, windc);
1501   return res;
1502 }
1503
1504 void releaseGlContext(int *vinfo, HGLRC *context) {
1505   *vinfo = 0;
1506   if (*context) {
1507     wglMakeCurrent(0, 0);
1508     wglDeleteContext(*context);
1509   }
1510   *context = 0;
1511 }
1512
1513 void swapGlBuffers() {
1514   HDC vo_hdc = GetDC(vo_w32_window);
1515   SwapBuffers(vo_hdc);
1516   ReleaseDC(vo_w32_window, vo_hdc);
1517 }
1518 #else
1519 #ifdef HAVE_LIBDL
1520 #include <dlfcn.h>
1521 #endif
1522 #include "x11_common.h"
1523 /**
1524  * \brief find address of a linked function
1525  * \param s name of function to find
1526  * \return address of function or NULL if not found
1527  *
1528  * Copied from xine
1529  */
1530 static void *getdladdr(const char *s) {
1531 #ifdef HAVE_LIBDL
1532 #if defined(__sun) || defined(__sgi)
1533   static void *handle = NULL;
1534   if (!handle)
1535     handle = dlopen(NULL, RTLD_LAZY);
1536   return dlsym(handle, s);
1537 #else
1538   return dlsym(0, s);
1539 #endif
1540 #else
1541   return NULL;
1542 #endif
1543 }
1544
1545 /**
1546  * \brief Returns the XVisualInfo associated with Window win.
1547  * \param win Window whose XVisualInfo is returne.
1548  * \return XVisualInfo of the window. Caller must use XFree to free it.
1549  */
1550 static XVisualInfo *getWindowVisualInfo(Window win) {
1551   XWindowAttributes xw_attr;
1552   XVisualInfo vinfo_template;
1553   int tmp;
1554   XGetWindowAttributes(mDisplay, win, &xw_attr);
1555   vinfo_template.visualid = XVisualIDFromVisual(xw_attr.visual);
1556   return XGetVisualInfo(mDisplay, VisualIDMask, &vinfo_template, &tmp);
1557 }
1558
1559 /**
1560  * \brief Changes the window in which video is displayed.
1561  * If possible only transfers the context to the new window, otherwise
1562  * creates a new one, which must be initialized by the caller.
1563  * \param vinfo Currently used visual.
1564  * \param context Currently used context.
1565  * \param win window that should be used for drawing.
1566  * \return one of SET_WINDOW_FAILED, SET_WINDOW_OK or SET_WINDOW_REINIT.
1567  * In case of SET_WINDOW_REINIT the context could not be transfered
1568  * and the caller must initialize it correctly.
1569  * \ingroup glcontext
1570  */
1571 int setGlWindow(XVisualInfo **vinfo, GLXContext *context, Window win)
1572 {
1573   XVisualInfo *new_vinfo;
1574   GLXContext new_context = NULL;
1575   int keep_context = 0;
1576
1577   // should only be needed when keeping context, but not doing glFinish
1578   // can cause flickering even when we do not keep it.
1579   if (*context)
1580   glFinish();
1581   new_vinfo = getWindowVisualInfo(win);
1582   if (*context && *vinfo && new_vinfo &&
1583       (*vinfo)->visualid == new_vinfo->visualid) {
1584       // we can keep the GLXContext
1585       new_context = *context;
1586       XFree(new_vinfo);
1587       new_vinfo = *vinfo;
1588       keep_context = 1;
1589   } else {
1590     // create a context
1591     new_context = glXCreateContext(mDisplay, new_vinfo, NULL, True);
1592     if (!new_context) {
1593       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GLX context!\n");
1594       XFree(new_vinfo);
1595       return SET_WINDOW_FAILED;
1596     }
1597   }
1598
1599   // set context
1600   if (!glXMakeCurrent(mDisplay, vo_window, new_context)) {
1601     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GLX context!\n");
1602     if (!keep_context) {
1603       glXDestroyContext(mDisplay, new_context);
1604       XFree(new_vinfo);
1605     }
1606     return SET_WINDOW_FAILED;
1607   }
1608
1609   // set new values
1610   vo_window = win;
1611   vo_x11_update_geometry();
1612   if (!keep_context) {
1613     void *(*getProcAddress)(const GLubyte *);
1614     const char *(*glXExtStr)(Display *, int);
1615     if (*context)
1616       glXDestroyContext(mDisplay, *context);
1617     *context = new_context;
1618     if (*vinfo)
1619       XFree(*vinfo);
1620     *vinfo = new_vinfo;
1621       getProcAddress = getdladdr("glXGetProcAddress");
1622     if (!getProcAddress)
1623       getProcAddress = getdladdr("glXGetProcAddressARB");
1624     if (!getProcAddress)
1625       getProcAddress = (void *)getdladdr;
1626     glXExtStr = getdladdr("glXQueryExtensionsString");
1627     getFunctions(getProcAddress, !glXExtStr ? NULL :
1628                  glXExtStr(mDisplay, DefaultScreen(mDisplay)));
1629
1630     // and inform that reinit is neccessary
1631     return SET_WINDOW_REINIT;
1632   }
1633   return SET_WINDOW_OK;
1634 }
1635
1636 /**
1637  * \brief free the VisualInfo and GLXContext of an OpenGL context.
1638  * \ingroup glcontext
1639  */
1640 void releaseGlContext(XVisualInfo **vinfo, GLXContext *context) {
1641   if (*vinfo)
1642     XFree(*vinfo);
1643   *vinfo = NULL;
1644   if (*context)
1645   {
1646     glFinish();
1647     glXMakeCurrent(mDisplay, None, NULL);
1648     glXDestroyContext(mDisplay, *context);
1649   }
1650   *context = 0;
1651 }
1652
1653 void swapGlBuffers(void) {
1654   glXSwapBuffers(mDisplay, vo_window);
1655 }
1656 #endif
1657