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 "gl_common.h"
  36 #include "csputils.h"
  37
  38 void (GLAPIENTRY *Begin)(GLenum);
  39 void (GLAPIENTRY *End)(void);
  40 void (GLAPIENTRY *Viewport)(GLint, GLint, GLsizei, GLsizei);
  41 void (GLAPIENTRY *MatrixMode)(GLenum);
  42 void (GLAPIENTRY *LoadIdentity)(void);
  43 void (GLAPIENTRY *Translated)(double, double, double);
  44 void (GLAPIENTRY *Scaled)(double, double, double);
  45 void (GLAPIENTRY *Ortho)(double, double, double, double, double, double);
  46 void (GLAPIENTRY *Frustum)(double, double, double, double, double, double);
  47 void (GLAPIENTRY *PushMatrix)(void);
  48 void (GLAPIENTRY *PopMatrix)(void);
  49 void (GLAPIENTRY *Clear)(GLbitfield);
  50 GLuint (GLAPIENTRY *GenLists)(GLsizei);
  51 void (GLAPIENTRY *DeleteLists)(GLuint, GLsizei);
  52 void (GLAPIENTRY *NewList)(GLuint, GLenum);
  53 void (GLAPIENTRY *EndList)(void);
  54 void (GLAPIENTRY *CallList)(GLuint);
  55 void (GLAPIENTRY *CallLists)(GLsizei, GLenum, const GLvoid *);
  56 void (GLAPIENTRY *GenTextures)(GLsizei, GLuint *);
  57 void (GLAPIENTRY *DeleteTextures)(GLsizei, const GLuint *);
  58 void (GLAPIENTRY *TexEnvf)(GLenum, GLenum, GLfloat);
  59 void (GLAPIENTRY *TexEnvi)(GLenum, GLenum, GLint);
  60 void (GLAPIENTRY *Color4ub)(GLubyte, GLubyte, GLubyte, GLubyte);
  61 void (GLAPIENTRY *Color3f)(GLfloat, GLfloat, GLfloat);
  62 void (GLAPIENTRY *Color4f)(GLfloat, GLfloat, GLfloat, GLfloat);
  63 void (GLAPIENTRY *ClearColor)(GLclampf, GLclampf, GLclampf, GLclampf);
  64 void (GLAPIENTRY *ClearDepth)(GLclampd);
  65 void (GLAPIENTRY *DepthFunc)(GLenum);
  66 void (GLAPIENTRY *Enable)(GLenum);
  67 void (GLAPIENTRY *Disable)(GLenum);
  68 const GLubyte *(GLAPIENTRY *GetString)(GLenum);
  69 void (GLAPIENTRY *DrawBuffer)(GLenum);
  70 void (GLAPIENTRY *DepthMask)(GLboolean);
  71 void (GLAPIENTRY *BlendFunc)(GLenum, GLenum);
  72 void (GLAPIENTRY *Flush)(void);
  73 void (GLAPIENTRY *Finish)(void);
  74 void (GLAPIENTRY *PixelStorei)(GLenum, GLint);
  75 void (GLAPIENTRY *TexImage1D)(GLenum, GLint, GLint, GLsizei, GLint, GLenum, GLenum, const GLvoid *);
  76 void (GLAPIENTRY *TexImage2D)(GLenum, GLint, GLint, GLsizei, GLsizei, GLint, GLenum, GLenum, const GLvoid *);
  77 void (GLAPIENTRY *TexSubImage2D)(GLenum, GLint, GLint, GLint, GLsizei, GLsizei, GLenum, GLenum, const GLvoid *);
  78 void (GLAPIENTRY *TexParameteri)(GLenum, GLenum, GLint);
  79 void (GLAPIENTRY *TexParameterf)(GLenum, GLenum, GLfloat);
  80 void (GLAPIENTRY *TexParameterfv)(GLenum, GLenum, const GLfloat *);
  81 void (GLAPIENTRY *TexCoord2f)(GLfloat, GLfloat);
  82 void (GLAPIENTRY *Vertex2f)(GLfloat, GLfloat);
  83 void (GLAPIENTRY *Vertex3f)(GLfloat, GLfloat, GLfloat);
  84 void (GLAPIENTRY *Normal3f)(GLfloat, GLfloat, GLfloat);
  85 void (GLAPIENTRY *Lightfv)(GLenum, GLenum, const GLfloat *);
  86 void (GLAPIENTRY *ColorMaterial)(GLenum, GLenum);
  87 void (GLAPIENTRY *ShadeModel)(GLenum);
  88 void (GLAPIENTRY *GetIntegerv)(GLenum, GLint *);
  89
  90 /**
  91  * \defgroup glextfunctions OpenGL extension functions
  92  *
  93  * the pointers to these functions are acquired when the OpenGL
  94  * context is created
  95  * \{
  96  */
  97 void (GLAPIENTRY *GenBuffers)(GLsizei, GLuint *);
  98 void (GLAPIENTRY *DeleteBuffers)(GLsizei, const GLuint *);
  99 void (GLAPIENTRY *BindBuffer)(GLenum, GLuint);
 100 GLvoid* (GLAPIENTRY *MapBuffer)(GLenum, GLenum);
 101 GLboolean (GLAPIENTRY *UnmapBuffer)(GLenum);
 102 void (GLAPIENTRY *BufferData)(GLenum, intptr_t, const GLvoid *, GLenum);
 103 void (GLAPIENTRY *CombinerParameterfv)(GLenum, const GLfloat *);
 104 void (GLAPIENTRY *CombinerParameteri)(GLenum, GLint);
 105 void (GLAPIENTRY *CombinerInput)(GLenum, GLenum, GLenum, GLenum, GLenum,
 106                                GLenum);
 107 void (GLAPIENTRY *CombinerOutput)(GLenum, GLenum, GLenum, GLenum, GLenum,
 108                                 GLenum, GLenum, GLboolean, GLboolean,
 109                                 GLboolean);
 110 void (GLAPIENTRY *BeginFragmentShader)(void);
 111 void (GLAPIENTRY *EndFragmentShader)(void);
 112 void (GLAPIENTRY *SampleMap)(GLuint, GLuint, GLenum);
 113 void (GLAPIENTRY *ColorFragmentOp2)(GLenum, GLuint, GLuint, GLuint, GLuint,
 114                                   GLuint, GLuint, GLuint, GLuint, GLuint);
 115 void (GLAPIENTRY *ColorFragmentOp3)(GLenum, GLuint, GLuint, GLuint, GLuint,
 116                                   GLuint, GLuint, GLuint, GLuint, GLuint,
 117                                   GLuint, GLuint, GLuint);
 118 void (GLAPIENTRY *SetFragmentShaderConstant)(GLuint, const GLfloat *);
 119 void (GLAPIENTRY *ActiveTexture)(GLenum);
 120 void (GLAPIENTRY *BindTexture)(GLenum, GLuint);
 121 void (GLAPIENTRY *MultiTexCoord2f)(GLenum, GLfloat, GLfloat);
 122 void (GLAPIENTRY *GenPrograms)(GLsizei, GLuint *);
 123 void (GLAPIENTRY *DeletePrograms)(GLsizei, const GLuint *);
 124 void (GLAPIENTRY *BindProgram)(GLenum, GLuint);
 125 void (GLAPIENTRY *ProgramString)(GLenum, GLenum, GLsizei, const GLvoid *);
 126 void (GLAPIENTRY *GetProgramiv)(GLenum, GLenum, GLint *);
 127 void (GLAPIENTRY *ProgramEnvParameter4f)(GLenum, GLuint, GLfloat, GLfloat,
 128                                        GLfloat, GLfloat);
 129 int (GLAPIENTRY *SwapInterval)(int);
 130 void (GLAPIENTRY *TexImage3D)(GLenum, GLint, GLenum, GLsizei, GLsizei, GLsizei,
 131                             GLint, GLenum, GLenum, const GLvoid *);
 132 void* (GLAPIENTRY *AllocateMemoryMESA)(void *, int, size_t, float, float, float);
 133 void (GLAPIENTRY *FreeMemoryMESA)(void *, int, void *);
 134 /** \} */ // end of glextfunctions group
 135
 136 //! \defgroup glgeneral OpenGL general helper functions
 137
 138 //! \defgroup glcontext OpenGL context management helper functions
 139
 140 //! \defgroup gltexture OpenGL texture handling helper functions
 141
 142 //! \defgroup glconversion OpenGL conversion helper functions
 143
 144 static GLint hqtexfmt;
 145
 146 /**
 147  * \brief adjusts the GL_UNPACK_ALIGNMENT to fit the stride.
 148  * \param stride number of bytes per line for which alignment should fit.
 149  * \ingroup glgeneral
 150  */
 151 void glAdjustAlignment(int stride) {
 152   GLint gl_alignment;
 153   if (stride % 8 == 0)
 154     gl_alignment=8;
 155   else if (stride % 4 == 0)
 156     gl_alignment=4;
 157   else if (stride % 2 == 0)
 158     gl_alignment=2;
 159   else
 160     gl_alignment=1;
 161   PixelStorei(GL_UNPACK_ALIGNMENT, gl_alignment);
 162 }
 163
 164 struct gl_name_map_struct {
 165   GLint value;
 166   const char *name;
 167 };
 168
 169 #undef MAP
 170 #define MAP(a) {a, #a}
 171 //! mapping table for the glValName function
 172 static const struct gl_name_map_struct gl_name_map[] = {
 173   // internal format
 174   MAP(GL_R3_G3_B2), MAP(GL_RGB4), MAP(GL_RGB5), MAP(GL_RGB8),
 175   MAP(GL_RGB10), MAP(GL_RGB12), MAP(GL_RGB16), MAP(GL_RGBA2),
 176   MAP(GL_RGBA4), MAP(GL_RGB5_A1), MAP(GL_RGBA8), MAP(GL_RGB10_A2),
 177   MAP(GL_RGBA12), MAP(GL_RGBA16), MAP(GL_LUMINANCE8),
 178
 179   // format
 180   MAP(GL_RGB), MAP(GL_RGBA), MAP(GL_RED), MAP(GL_GREEN), MAP(GL_BLUE),
 181   MAP(GL_ALPHA), MAP(GL_LUMINANCE), MAP(GL_LUMINANCE_ALPHA),
 182   MAP(GL_COLOR_INDEX),
 183   // rest 1.2 only
 184   MAP(GL_BGR), MAP(GL_BGRA),
 185
 186   //type
 187   MAP(GL_BYTE), MAP(GL_UNSIGNED_BYTE), MAP(GL_SHORT), MAP(GL_UNSIGNED_SHORT),
 188   MAP(GL_INT), MAP(GL_UNSIGNED_INT), MAP(GL_FLOAT), MAP(GL_DOUBLE),
 189   MAP(GL_2_BYTES), MAP(GL_3_BYTES), MAP(GL_4_BYTES),
 190   // rest 1.2 only
 191   MAP(GL_UNSIGNED_BYTE_3_3_2), MAP(GL_UNSIGNED_BYTE_2_3_3_REV),
 192   MAP(GL_UNSIGNED_SHORT_5_6_5), MAP(GL_UNSIGNED_SHORT_5_6_5_REV),
 193   MAP(GL_UNSIGNED_SHORT_4_4_4_4), MAP(GL_UNSIGNED_SHORT_4_4_4_4_REV),
 194   MAP(GL_UNSIGNED_SHORT_5_5_5_1), MAP(GL_UNSIGNED_SHORT_1_5_5_5_REV),
 195   MAP(GL_UNSIGNED_INT_8_8_8_8), MAP(GL_UNSIGNED_INT_8_8_8_8_REV),
 196   MAP(GL_UNSIGNED_INT_10_10_10_2), MAP(GL_UNSIGNED_INT_2_10_10_10_REV),
 197   {0, 0}
 198 };
 199 #undef MAP
 200
 201 /**
 202  * \brief return the name of an OpenGL constant
 203  * \param value the constant
 204  * \return name of the constant or "Unknown format!"
 205  * \ingroup glgeneral
 206  */
 207 const char *glValName(GLint value)
 208 {
 209   int i = 0;
 210
 211   while (gl_name_map[i].name) {
 212     if (gl_name_map[i].value == value)
 213       return gl_name_map[i].name;
 214     i++;
 215   }
 216   return "Unknown format!";
 217 }
 218
 219 //! always return this format as internal texture format in glFindFormat
 220 #define TEXTUREFORMAT_ALWAYS GL_RGB8
 221 #undef TEXTUREFORMAT_ALWAYS
 222
 223 /**
 224  * \brief find the OpenGL settings coresponding to format.
 225  *
 226  * All parameters may be NULL.
 227  * \param fmt MPlayer format to analyze.
 228  * \param bpp [OUT] bits per pixel of that format.
 229  * \param gl_texfmt [OUT] internal texture format that fits the
 230  * image format, not necessarily the best for performance.
 231  * \param gl_format [OUT] OpenGL format for this image format.
 232  * \param gl_type [OUT] OpenGL type for this image format.
 233  * \return 1 if format is supported by OpenGL, 0 if not.
 234  * \ingroup gltexture
 235  */
 236 int glFindFormat(uint32_t fmt, int *bpp, GLint *gl_texfmt,
 237                   GLenum *gl_format, GLenum *gl_type)
 238 {
 239   int supported = 1;
 240   int dummy1;
 241   GLenum dummy2;
 242   GLint dummy3;
 243   if (!bpp) bpp = &dummy1;
 244   if (!gl_texfmt) gl_texfmt = &dummy3;
 245   if (!gl_format) gl_format = &dummy2;
 246   if (!gl_type) gl_type = &dummy2;
 247
 248   if (mp_get_chroma_shift(fmt, NULL, NULL)) {
 249     // reduce the possible cases a bit
 250     if (IMGFMT_IS_YUVP16_LE(fmt))
 251       fmt = IMGFMT_420P16_LE;
 252     else if (IMGFMT_IS_YUVP16_BE(fmt))
 253       fmt = IMGFMT_420P16_BE;
 254     else
 255       fmt = IMGFMT_YV12;
 256   }
 257
 258   *bpp = IMGFMT_IS_BGR(fmt)?IMGFMT_BGR_DEPTH(fmt):IMGFMT_RGB_DEPTH(fmt);
 259   *gl_texfmt = 3;
 260   switch (fmt) {
 261     case IMGFMT_RGB48NE:
 262       *gl_format = GL_RGB;
 263       *gl_type = GL_UNSIGNED_SHORT;
 264       break;
 265     case IMGFMT_RGB24:
 266       *gl_format = GL_RGB;
 267       *gl_type = GL_UNSIGNED_BYTE;
 268       break;
 269     case IMGFMT_RGBA:
 270       *gl_texfmt = 4;
 271       *gl_format = GL_RGBA;
 272       *gl_type = GL_UNSIGNED_BYTE;
 273       break;
 274     case IMGFMT_420P16:
 275       supported = 0; // no native YUV support
 276       *gl_texfmt = 1;
 277       *bpp = 16;
 278       *gl_format = GL_LUMINANCE;
 279       *gl_type = GL_UNSIGNED_SHORT;
 280       break;
 281     case IMGFMT_YV12:
 282       supported = 0; // no native YV12 support
 283     case IMGFMT_Y800:
 284     case IMGFMT_Y8:
 285       *gl_texfmt = 1;
 286       *bpp = 8;
 287       *gl_format = GL_LUMINANCE;
 288       *gl_type = GL_UNSIGNED_BYTE;
 289       break;
 290     case IMGFMT_UYVY:
 291     case IMGFMT_YUY2:
 292       *gl_texfmt = GL_YCBCR_MESA;
 293       *bpp = 16;
 294       *gl_format = GL_YCBCR_MESA;
 295       *gl_type = fmt == IMGFMT_UYVY ? GL_UNSIGNED_SHORT_8_8 : GL_UNSIGNED_SHORT_8_8_REV;
 296       break;
 297 #if 0
 298     // we do not support palettized formats, although the format the
 299     // swscale produces works
 300     case IMGFMT_RGB8:
 301       gl_format = GL_RGB;
 302       gl_type = GL_UNSIGNED_BYTE_2_3_3_REV;
 303       break;
 304 #endif
 305     case IMGFMT_RGB15:
 306       *gl_format = GL_RGBA;
 307       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 308       break;
 309     case IMGFMT_RGB16:
 310       *gl_format = GL_RGB;
 311       *gl_type = GL_UNSIGNED_SHORT_5_6_5_REV;
 312       break;
 313 #if 0
 314     case IMGFMT_BGR8:
 315       // special case as red and blue have a differen number of bits.
 316       // GL_BGR and GL_UNSIGNED_BYTE_3_3_2 isn't supported at least
 317       // by nVidia drivers, and in addition would give more bits to
 318       // blue than to red, which isn't wanted
 319       gl_format = GL_RGB;
 320       gl_type = GL_UNSIGNED_BYTE_3_3_2;
 321       break;
 322 #endif
 323     case IMGFMT_BGR15:
 324       *gl_format = GL_BGRA;
 325       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 326       break;
 327     case IMGFMT_BGR16:
 328       *gl_format = GL_RGB;
 329       *gl_type = GL_UNSIGNED_SHORT_5_6_5;
 330       break;
 331     case IMGFMT_BGR24:
 332       *gl_format = GL_BGR;
 333       *gl_type = GL_UNSIGNED_BYTE;
 334       break;
 335     case IMGFMT_BGRA:
 336       *gl_texfmt = 4;
 337       *gl_format = GL_BGRA;
 338       *gl_type = GL_UNSIGNED_BYTE;
 339       break;
 340     default:
 341       *gl_texfmt = 4;
 342       *gl_format = GL_RGBA;
 343       *gl_type = GL_UNSIGNED_BYTE;
 344       supported = 0;
 345   }
 346 #ifdef TEXTUREFORMAT_ALWAYS
 347   *gl_texfmt = TEXTUREFORMAT_ALWAYS;
 348 #endif
 349   return supported;
 350 }
 351
 352 static void *setNull(const GLubyte *s) {
 353   return NULL;
 354 }
 355
 356 typedef struct {
 357   void *funcptr;
 358   const char *extstr;
 359   const char *funcnames[7];
 360   void *fallback;
 361 } extfunc_desc_t;
 362
 363 #define DEF_FUNC_DESC(name) {&name, NULL, {"gl"#name, NULL}, gl ##name}
 364 static const extfunc_desc_t extfuncs[] = {
 365   // these aren't extension functions but we query them anyway to allow
 366   // different "backends" with one binary
 367   DEF_FUNC_DESC(Begin),
 368   DEF_FUNC_DESC(End),
 369   DEF_FUNC_DESC(Viewport),
 370   DEF_FUNC_DESC(MatrixMode),
 371   DEF_FUNC_DESC(LoadIdentity),
 372   DEF_FUNC_DESC(Translated),
 373   DEF_FUNC_DESC(Scaled),
 374   DEF_FUNC_DESC(Ortho),
 375   DEF_FUNC_DESC(Frustum),
 376   DEF_FUNC_DESC(PushMatrix),
 377   DEF_FUNC_DESC(PopMatrix),
 378   DEF_FUNC_DESC(Clear),
 379   DEF_FUNC_DESC(GenLists),
 380   DEF_FUNC_DESC(DeleteLists),
 381   DEF_FUNC_DESC(NewList),
 382   DEF_FUNC_DESC(EndList),
 383   DEF_FUNC_DESC(CallList),
 384   DEF_FUNC_DESC(CallLists),
 385   DEF_FUNC_DESC(GenTextures),
 386   DEF_FUNC_DESC(DeleteTextures),
 387   DEF_FUNC_DESC(TexEnvf),
 388   DEF_FUNC_DESC(TexEnvi),
 389   DEF_FUNC_DESC(Color4ub),
 390   DEF_FUNC_DESC(Color3f),
 391   DEF_FUNC_DESC(Color4f),
 392   DEF_FUNC_DESC(ClearColor),
 393   DEF_FUNC_DESC(ClearDepth),
 394   DEF_FUNC_DESC(DepthFunc),
 395   DEF_FUNC_DESC(Enable),
 396   DEF_FUNC_DESC(Disable),
 397   DEF_FUNC_DESC(DrawBuffer),
 398   DEF_FUNC_DESC(DepthMask),
 399   DEF_FUNC_DESC(BlendFunc),
 400   DEF_FUNC_DESC(Flush),
 401   DEF_FUNC_DESC(Finish),
 402   DEF_FUNC_DESC(PixelStorei),
 403   DEF_FUNC_DESC(TexImage1D),
 404   DEF_FUNC_DESC(TexImage2D),
 405   DEF_FUNC_DESC(TexSubImage2D),
 406   DEF_FUNC_DESC(TexParameteri),
 407   DEF_FUNC_DESC(TexParameterf),
 408   DEF_FUNC_DESC(TexParameterfv),
 409   DEF_FUNC_DESC(TexCoord2f),
 410   DEF_FUNC_DESC(Vertex2f),
 411   DEF_FUNC_DESC(Vertex3f),
 412   DEF_FUNC_DESC(Normal3f),
 413   DEF_FUNC_DESC(Lightfv),
 414   DEF_FUNC_DESC(ColorMaterial),
 415   DEF_FUNC_DESC(ShadeModel),
 416   DEF_FUNC_DESC(GetIntegerv),
 417
 418   // here start the real extensions
 419   {&GenBuffers, NULL, {"glGenBuffers", "glGenBuffersARB", NULL}},
 420   {&DeleteBuffers, NULL, {"glDeleteBuffers", "glDeleteBuffersARB", NULL}},
 421   {&BindBuffer, NULL, {"glBindBuffer", "glBindBufferARB", NULL}},
 422   {&MapBuffer, NULL, {"glMapBuffer", "glMapBufferARB", NULL}},
 423   {&UnmapBuffer, NULL, {"glUnmapBuffer", "glUnmapBufferARB", NULL}},
 424   {&BufferData, NULL, {"glBufferData", "glBufferDataARB", NULL}},
 425   {&CombinerParameterfv, "NV_register_combiners", {"glCombinerParameterfv", "glCombinerParameterfvNV", NULL}},
 426   {&CombinerParameteri, "NV_register_combiners", {"glCombinerParameteri", "glCombinerParameteriNV", NULL}},
 427   {&CombinerInput, "NV_register_combiners", {"glCombinerInput", "glCombinerInputNV", NULL}},
 428   {&CombinerOutput, "NV_register_combiners", {"glCombinerOutput", "glCombinerOutputNV", NULL}},
 429   {&BeginFragmentShader, "ATI_fragment_shader", {"glBeginFragmentShaderATI", NULL}},
 430   {&EndFragmentShader, "ATI_fragment_shader", {"glEndFragmentShaderATI", NULL}},
 431   {&SampleMap, "ATI_fragment_shader", {"glSampleMapATI", NULL}},
 432   {&ColorFragmentOp2, "ATI_fragment_shader", {"glColorFragmentOp2ATI", NULL}},
 433   {&ColorFragmentOp3, "ATI_fragment_shader", {"glColorFragmentOp3ATI", NULL}},
 434   {&SetFragmentShaderConstant, "ATI_fragment_shader", {"glSetFragmentShaderConstantATI", NULL}},
 435   {&ActiveTexture, NULL, {"glActiveTexture", "glActiveTextureARB", NULL}},
 436   {&BindTexture, NULL, {"glBindTexture", "glBindTextureARB", "glBindTextureEXT", NULL}},
 437   {&MultiTexCoord2f, NULL, {"glMultiTexCoord2f", "glMultiTexCoord2fARB", NULL}},
 438   {&GenPrograms, "_program", {"glGenProgramsARB", NULL}},
 439   {&DeletePrograms, "_program", {"glDeleteProgramsARB", NULL}},
 440   {&BindProgram, "_program", {"glBindProgramARB", NULL}},
 441   {&ProgramString, "_program", {"glProgramStringARB", NULL}},
 442   {&GetProgramiv, "_program", {"glGetProgramivARB", NULL}},
 443   {&ProgramEnvParameter4f, "_program", {"glProgramEnvParameter4fARB", NULL}},
 444   {&SwapInterval, "_swap_control", {"glXSwapIntervalSGI", "glXSwapInterval", "wglSwapIntervalSGI", "wglSwapInterval", "wglSwapIntervalEXT", NULL}},
 445   {&TexImage3D, NULL, {"glTexImage3D", NULL}},
 446   {&AllocateMemoryMESA, "GLX_MESA_allocate_memory", {"glXAllocateMemoryMESA", NULL}},
 447   {&FreeMemoryMESA, "GLX_MESA_allocate_memory", {"glXFreeMemoryMESA", NULL}},
 448   {NULL}
 449 };
 450
 451 /**
 452  * \brief find the function pointers of some useful OpenGL extensions
 453  * \param getProcAddress function to resolve function names, may be NULL
 454  * \param ext2 an extra extension string
 455  */
 456 static void getFunctions(void *(*getProcAddress)(const GLubyte *),
 457                          const char *ext2) {
 458   const extfunc_desc_t *dsc;
 459   const char *extensions;
 460   char *allexts;
 461
 462   if (!getProcAddress)
 463     getProcAddress = setNull;
 464
 465   // special case, we need glGetString before starting to find the other functions
 466   GetString = getProcAddress("glGetString");
 467   if (!GetString)
 468       GetString = glGetString;
 469
 470   extensions = (const char *)GetString(GL_EXTENSIONS);
 471   if (!extensions) extensions = "";
 472   if (!ext2) ext2 = "";
 473   allexts = malloc(strlen(extensions) + strlen(ext2) + 2);
 474   strcpy(allexts, extensions);
 475   strcat(allexts, " ");
 476   strcat(allexts, ext2);
 477   mp_msg(MSGT_VO, MSGL_DBG2, "OpenGL extensions string:\n%s\n", allexts);
 478   for (dsc = extfuncs; dsc->funcptr; dsc++) {
 479     void *ptr = NULL;
 480     int i;
 481     if (!dsc->extstr || strstr(allexts, dsc->extstr)) {
 482       for (i = 0; !ptr && dsc->funcnames[i]; i++)
 483         ptr = getProcAddress((const GLubyte *)dsc->funcnames[i]);
 484     }
 485     if (!ptr)
 486         ptr = dsc->fallback;
 487     *(void **)dsc->funcptr = ptr;
 488   }
 489   if (strstr(allexts, "_texture_float"))
 490     hqtexfmt = GL_RGB32F;
 491   else if (strstr(allexts, "NV_float_buffer"))
 492     hqtexfmt = GL_FLOAT_RGB32_NV;
 493   else
 494     hqtexfmt = GL_RGB16;
 495   free(allexts);
 496 }
 497
 498 /**
 499  * \brief create a texture and set some defaults
 500  * \param target texture taget, usually GL_TEXTURE_2D
 501  * \param fmt internal texture format
 502  * \param format texture host data format
 503  * \param type texture host data type
 504  * \param filter filter used for scaling, e.g. GL_LINEAR
 505  * \param w texture width
 506  * \param h texture height
 507  * \param val luminance value to fill texture with
 508  * \ingroup gltexture
 509  */
 510 void glCreateClearTex(GLenum target, GLenum fmt, GLenum format, GLenum type, GLint filter,
 511                       int w, int h, unsigned char val) {
 512   GLfloat fval = (GLfloat)val / 255.0;
 513   GLfloat border[4] = {fval, fval, fval, fval};
 514   int stride = w * glFmt2bpp(format, type);
 515   char *init;
 516   if (!stride) return;
 517   init = malloc(stride * h);
 518   memset(init, val, stride * h);
 519   glAdjustAlignment(stride);
 520   PixelStorei(GL_UNPACK_ROW_LENGTH, w);
 521   TexImage2D(target, 0, fmt, w, h, 0, format, type, init);
 522   TexParameterf(target, GL_TEXTURE_PRIORITY, 1.0);
 523   TexParameteri(target, GL_TEXTURE_MIN_FILTER, filter);
 524   TexParameteri(target, GL_TEXTURE_MAG_FILTER, filter);
 525   TexParameteri(target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 526   TexParameteri(target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 527   // Border texels should not be used with CLAMP_TO_EDGE
 528   // We set a sane default anyway.
 529   TexParameterfv(target, GL_TEXTURE_BORDER_COLOR, border);
 530   free(init);
 531 }
 532
 533 /**
 534  * \brief skips whitespace and comments
 535  * \param f file to read from
 536  */
 537 static void ppm_skip(FILE *f) {
 538   int c, comment = 0;
 539   do {
 540     c = fgetc(f);
 541     if (c == '#')
 542       comment = 1;
 543     if (c == '\n')
 544       comment = 0;
 545   } while (c != EOF && (isspace(c) || comment));
 546   if (c != EOF)
 547     ungetc(c, f);
 548 }
 549
 550 #define MAXDIM (16 * 1024)
 551
 552 /**
 553  * \brief creates a texture from a PPM file
 554  * \param target texture taget, usually GL_TEXTURE_2D
 555  * \param fmt internal texture format, 0 for default
 556  * \param filter filter used for scaling, e.g. GL_LINEAR
 557  * \param f file to read PPM from
 558  * \param width [out] width of texture
 559  * \param height [out] height of texture
 560  * \param maxval [out] maxval value from PPM file
 561  * \return 0 on error, 1 otherwise
 562  * \ingroup gltexture
 563  */
 564 int glCreatePPMTex(GLenum target, GLenum fmt, GLint filter,
 565                    FILE *f, int *width, int *height, int *maxval) {
 566   unsigned w, h, m, val, bpp;
 567   char *data;
 568   GLenum type;
 569   ppm_skip(f);
 570   if (fgetc(f) != 'P' || fgetc(f) != '6')
 571     return 0;
 572   ppm_skip(f);
 573   if (fscanf(f, "%u", &w) != 1)
 574     return 0;
 575   ppm_skip(f);
 576   if (fscanf(f, "%u", &h) != 1)
 577     return 0;
 578   ppm_skip(f);
 579   if (fscanf(f, "%u", &m) != 1)
 580     return 0;
 581   val = fgetc(f);
 582   if (!isspace(val))
 583     return 0;
 584   if (w > MAXDIM || h > MAXDIM)
 585     return 0;
 586   bpp = (m > 255) ? 6 : 3;
 587   data = malloc(w * h * bpp);
 588   if (fread(data, w * bpp, h, f) != h)
 589     return 0;
 590   if (!fmt) {
 591     fmt = (m > 255) ? hqtexfmt : 3;
 592     if (fmt == GL_FLOAT_RGB32_NV && target != GL_TEXTURE_RECTANGLE)
 593       fmt = GL_RGB16;
 594   }
 595   type = m > 255 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 596   glCreateClearTex(target, fmt, GL_RGB, type, filter, w, h, 0);
 597   glUploadTex(target, GL_RGB, type,
 598               data, w * bpp, 0, 0, w, h, 0);
 599   free(data);
 600   if (width) *width = w;
 601   if (height) *height = h;
 602   if (maxval) *maxval = m;
 603   return 1;
 604 }
 605
 606 /**
 607  * \brief return the number of bytes per pixel for the given format
 608  * \param format OpenGL format
 609  * \param type OpenGL type
 610  * \return bytes per pixel
 611  * \ingroup glgeneral
 612  *
 613  * Does not handle all possible variants, just those used by MPlayer
 614  */
 615 int glFmt2bpp(GLenum format, GLenum type) {
 616   int component_size = 0;
 617   switch (type) {
 618     case GL_UNSIGNED_BYTE_3_3_2:
 619     case GL_UNSIGNED_BYTE_2_3_3_REV:
 620       return 1;
 621     case GL_UNSIGNED_SHORT_5_5_5_1:
 622     case GL_UNSIGNED_SHORT_1_5_5_5_REV:
 623     case GL_UNSIGNED_SHORT_5_6_5:
 624     case GL_UNSIGNED_SHORT_5_6_5_REV:
 625       return 2;
 626     case GL_UNSIGNED_BYTE:
 627       component_size = 1;
 628       break;
 629     case GL_UNSIGNED_SHORT:
 630       component_size = 2;
 631       break;
 632   }
 633   switch (format) {
 634     case GL_LUMINANCE:
 635     case GL_ALPHA:
 636       return component_size;
 637     case GL_YCBCR_MESA:
 638       return 2;
 639     case GL_RGB:
 640     case GL_BGR:
 641       return 3 * component_size;
 642     case GL_RGBA:
 643     case GL_BGRA:
 644       return 4 * component_size;
 645   }
 646   return 0; // unknown
 647 }
 648
 649 /**
 650  * \brief upload a texture, handling things like stride and slices
 651  * \param target texture target, usually GL_TEXTURE_2D
 652  * \param format OpenGL format of data
 653  * \param type OpenGL type of data
 654  * \param dataptr data to upload
 655  * \param stride data stride
 656  * \param x x offset in texture
 657  * \param y y offset in texture
 658  * \param w width of the texture part to upload
 659  * \param h height of the texture part to upload
 660  * \param slice height of an upload slice, 0 for all at once
 661  * \ingroup gltexture
 662  */
 663 void glUploadTex(GLenum target, GLenum format, GLenum type,
 664                  const void *dataptr, int stride,
 665                  int x, int y, int w, int h, int slice) {
 666   const uint8_t *data = dataptr;
 667   int y_max = y + h;
 668   if (w <= 0 || h <= 0) return;
 669   if (slice <= 0)
 670     slice = h;
 671   if (stride < 0) {
 672     data += (h - 1) * stride;
 673     stride = -stride;
 674   }
 675   // this is not always correct, but should work for MPlayer
 676   glAdjustAlignment(stride);
 677   PixelStorei(GL_UNPACK_ROW_LENGTH, stride / glFmt2bpp(format, type));
 678   for (; y + slice <= y_max; y += slice) {
 679     TexSubImage2D(target, 0, x, y, w, slice, format, type, data);
 680     data += stride * slice;
 681   }
 682   if (y < y_max)
 683     TexSubImage2D(target, 0, x, y, w, y_max - y, format, type, data);
 684 }
 685
 686 static void fillUVcoeff(GLfloat *ucoef, GLfloat *vcoef,
 687                         float uvcos, float uvsin) {
 688   int i;
 689   ucoef[0] = 0 * uvcos + 1.403 * uvsin;
 690   vcoef[0] = 0 * uvsin + 1.403 * uvcos;
 691   ucoef[1] = -0.344 * uvcos + -0.714 * uvsin;
 692   vcoef[1] = -0.344 * uvsin + -0.714 * uvcos;
 693   ucoef[2] = 1.770 * uvcos + 0 * uvsin;
 694   vcoef[2] = 1.770 * uvsin + 0 * uvcos;
 695   ucoef[3] = 0;
 696   vcoef[3] = 0;
 697   // Coefficients (probably) must be in [0, 1] range, whereas they originally
 698   // are in [-2, 2] range, so here comes the trick:
 699   // First put them in the [-0.5, 0.5] range, then add 0.5.
 700   // This can be undone with the HALF_BIAS and SCALE_BY_FOUR arguments
 701   // for CombinerInput and CombinerOutput (or the respective ATI variants)
 702   for (i = 0; i < 4; i++) {
 703     ucoef[i] = ucoef[i] * 0.25 + 0.5;
 704     vcoef[i] = vcoef[i] * 0.25 + 0.5;
 705   }
 706 }
 707
 708 /**
 709  * \brief Setup register combiners for YUV to RGB conversion.
 710  * \param uvcos used for saturation and hue adjustment
 711  * \param uvsin used for saturation and hue adjustment
 712  */
 713 static void glSetupYUVCombiners(float uvcos, float uvsin) {
 714   GLfloat ucoef[4];
 715   GLfloat vcoef[4];
 716   GLint i;
 717   if (!CombinerInput || !CombinerOutput ||
 718       !CombinerParameterfv || !CombinerParameteri) {
 719     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner functions missing!\n");
 720     return;
 721   }
 722   GetIntegerv(GL_MAX_GENERAL_COMBINERS_NV, &i);
 723   if (i < 2)
 724     mp_msg(MSGT_VO, MSGL_ERR,
 725            "[gl] 2 general combiners needed for YUV combiner support (found %i)\n", i);
 726   GetIntegerv(GL_MAX_TEXTURE_UNITS, &i);
 727   if (i < 3)
 728     mp_msg(MSGT_VO, MSGL_ERR,
 729            "[gl] 3 texture units needed for YUV combiner support (found %i)\n", i);
 730   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 731   CombinerParameterfv(GL_CONSTANT_COLOR0_NV, ucoef);
 732   CombinerParameterfv(GL_CONSTANT_COLOR1_NV, vcoef);
 733
 734   // UV first, like this green component cannot overflow
 735   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_A_NV,
 736                 GL_TEXTURE1, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 737   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_B_NV,
 738                 GL_CONSTANT_COLOR0_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 739   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_C_NV,
 740                 GL_TEXTURE2, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 741   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_D_NV,
 742                 GL_CONSTANT_COLOR1_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 743   CombinerOutput(GL_COMBINER0_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 744                  GL_SPARE0_NV, GL_SCALE_BY_FOUR_NV, GL_NONE, GL_FALSE,
 745                  GL_FALSE, GL_FALSE);
 746
 747   // stage 2
 748   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_A_NV, GL_SPARE0_NV,
 749                 GL_SIGNED_IDENTITY_NV, GL_RGB);
 750   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_B_NV, GL_ZERO,
 751                  GL_UNSIGNED_INVERT_NV, GL_RGB);
 752   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_C_NV,
 753                 GL_TEXTURE0, GL_SIGNED_IDENTITY_NV, GL_RGB);
 754   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_D_NV, GL_ZERO,
 755                 GL_UNSIGNED_INVERT_NV, GL_RGB);
 756   CombinerOutput(GL_COMBINER1_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 757                  GL_SPARE0_NV, GL_NONE, GL_NONE, GL_FALSE,
 758                  GL_FALSE, GL_FALSE);
 759
 760   // leave final combiner stage in default mode
 761   CombinerParameteri(GL_NUM_GENERAL_COMBINERS_NV, 2);
 762 }
 763
 764 /**
 765  * \brief Setup ATI version of register combiners for YUV to RGB conversion.
 766  * \param uvcos used for saturation and hue adjustment
 767  * \param uvsin used for saturation and hue adjustment
 768  *
 769  * ATI called this fragment shader, but the name is confusing in the
 770  * light of a very different OpenGL 2.0 extension with the same name
 771  */
 772 static void glSetupYUVCombinersATI(float uvcos, float uvsin) {
 773   GLfloat ucoef[4];
 774   GLfloat vcoef[4];
 775   GLint i;
 776   if (!BeginFragmentShader || !EndFragmentShader ||
 777       !SetFragmentShaderConstant || !SampleMap ||
 778       !ColorFragmentOp2 || !ColorFragmentOp3) {
 779     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner (ATI) functions missing!\n");
 780     return;
 781   }
 782   GetIntegerv(GL_NUM_FRAGMENT_REGISTERS_ATI, &i);
 783   if (i < 3)
 784     mp_msg(MSGT_VO, MSGL_ERR,
 785            "[gl] 3 registers needed for YUV combiner (ATI) support (found %i)\n", i);
 786   GetIntegerv (GL_MAX_TEXTURE_UNITS, &i);
 787   if (i < 3)
 788     mp_msg(MSGT_VO, MSGL_ERR,
 789            "[gl] 3 texture units needed for YUV combiner (ATI) support (found %i)\n", i);
 790   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 791   BeginFragmentShader();
 792   SetFragmentShaderConstant(GL_CON_0_ATI, ucoef);
 793   SetFragmentShaderConstant(GL_CON_1_ATI, vcoef);
 794   SampleMap(GL_REG_0_ATI, GL_TEXTURE0, GL_SWIZZLE_STR_ATI);
 795   SampleMap(GL_REG_1_ATI, GL_TEXTURE1, GL_SWIZZLE_STR_ATI);
 796   SampleMap(GL_REG_2_ATI, GL_TEXTURE2, GL_SWIZZLE_STR_ATI);
 797   // UV first, like this green component cannot overflow
 798   ColorFragmentOp2(GL_MUL_ATI, GL_REG_1_ATI, GL_NONE, GL_NONE,
 799                    GL_REG_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 800                    GL_CON_0_ATI, GL_NONE, GL_BIAS_BIT_ATI);
 801   ColorFragmentOp3(GL_MAD_ATI, GL_REG_2_ATI, GL_NONE, GL_4X_BIT_ATI,
 802                    GL_REG_2_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 803                    GL_CON_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 804                    GL_REG_1_ATI, GL_NONE, GL_NONE);
 805   ColorFragmentOp2(GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 806                    GL_REG_0_ATI, GL_NONE, GL_NONE,
 807                    GL_REG_2_ATI, GL_NONE, GL_NONE);
 808   EndFragmentShader();
 809 }
 810
 811 /**
 812  * \brief helper function for gen_spline_lookup_tex
 813  * \param x subpixel-position ((0,1) range) to calculate weights for
 814  * \param dst where to store transformed weights, must provide space for 4 GLfloats
 815  *
 816  * calculates the weights and stores them after appropriate transformation
 817  * for the scaler fragment program.
 818  */
 819 static void store_weights(float x, GLfloat *dst) {
 820   float w0 = (((-1 * x + 3) * x - 3) * x + 1) / 6;
 821   float w1 = ((( 3 * x - 6) * x + 0) * x + 4) / 6;
 822   float w2 = (((-3 * x + 3) * x + 3) * x + 1) / 6;
 823   float w3 = ((( 1 * x + 0) * x + 0) * x + 0) / 6;
 824   *dst++ = 1 + x - w1 / (w0 + w1);
 825   *dst++ = 1 - x + w3 / (w2 + w3);
 826   *dst++ = w0 + w1;
 827   *dst++ = 0;
 828 }
 829
 830 //! to avoid artefacts this should be rather large
 831 #define LOOKUP_BSPLINE_RES (2 * 1024)
 832 /**
 833  * \brief creates the 1D lookup texture needed for fast higher-order filtering
 834  * \param unit texture unit to attach texture to
 835  */
 836 static void gen_spline_lookup_tex(GLenum unit) {
 837   GLfloat *tex = calloc(4 * LOOKUP_BSPLINE_RES, sizeof(*tex));
 838   GLfloat *tp = tex;
 839   int i;
 840   for (i = 0; i < LOOKUP_BSPLINE_RES; i++) {
 841     float x = (float)(i + 0.5) / LOOKUP_BSPLINE_RES;
 842     store_weights(x, tp);
 843     tp += 4;
 844   }
 845   store_weights(0, tex);
 846   store_weights(1, &tex[4 * (LOOKUP_BSPLINE_RES - 1)]);
 847   ActiveTexture(unit);
 848   TexImage1D(GL_TEXTURE_1D, 0, GL_RGBA16, LOOKUP_BSPLINE_RES, 0, GL_RGBA, GL_FLOAT, tex);
 849   TexParameterf(GL_TEXTURE_1D, GL_TEXTURE_PRIORITY, 1.0);
 850   TexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 851   TexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 852   TexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_REPEAT);
 853   ActiveTexture(GL_TEXTURE0);
 854   free(tex);
 855 }
 856
 857 static const char *bilin_filt_template =
 858   "TEX yuv.%c, fragment.texcoord[%c], texture[%c], %s;";
 859
 860 #define BICUB_FILT_MAIN(textype) \
 861   /* first y-interpolation */ \
 862   "ADD coord, fragment.texcoord[%c].xyxy, cdelta.xyxw;" \
 863   "ADD coord2, fragment.texcoord[%c].xyxy, cdelta.zyzw;" \
 864   "TEX a.r, coord.xyxy, texture[%c], "textype";" \
 865   "TEX a.g, coord.zwzw, texture[%c], "textype";" \
 866   /* second y-interpolation */ \
 867   "TEX b.r, coord2.xyxy, texture[%c], "textype";" \
 868   "TEX b.g, coord2.zwzw, texture[%c], "textype";" \
 869   "LRP a.b, parmy.b, a.rrrr, a.gggg;" \
 870   "LRP a.a, parmy.b, b.rrrr, b.gggg;" \
 871   /* x-interpolation */ \
 872   "LRP yuv.%c, parmx.b, a.bbbb, a.aaaa;"
 873
 874 static const char *bicub_filt_template_2D =
 875   "MAD coord.xy, fragment.texcoord[%c], {%e, %e}, {0.5, 0.5};"
 876   "TEX parmx, coord.x, texture[%c], 1D;"
 877   "MUL cdelta.xz, parmx.rrgg, {-%e, 0, %e, 0};"
 878   "TEX parmy, coord.y, texture[%c], 1D;"
 879   "MUL cdelta.yw, parmy.rrgg, {0, -%e, 0, %e};"
 880   BICUB_FILT_MAIN("2D");
 881
 882 static const char *bicub_filt_template_RECT =
 883   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};"
 884   "TEX parmx, coord.x, texture[%c], 1D;"
 885   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};"
 886   "TEX parmy, coord.y, texture[%c], 1D;"
 887   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};"
 888   BICUB_FILT_MAIN("RECT");
 889
 890 #define CALCWEIGHTS(t, s) \
 891   "MAD "t", {-0.5, 0.1666, 0.3333, -0.3333}, "s", {1, 0, -0.5, 0.5};" \
 892   "MAD "t", "t", "s", {0, 0, -0.5, 0.5};" \
 893   "MAD "t", "t", "s", {-0.6666, 0, 0.8333, 0.1666};" \
 894   "RCP a.x, "t".z;" \
 895   "RCP a.y, "t".w;" \
 896   "MAD "t".xy, "t".xyxy, a.xyxy, {1, 1, 0, 0};" \
 897   "ADD "t".x, "t".xxxx, "s";" \
 898   "SUB "t".y, "t".yyyy, "s";"
 899
 900 static const char *bicub_notex_filt_template_2D =
 901   "MAD coord.xy, fragment.texcoord[%c], {%e, %e}, {0.5, 0.5};"
 902   "FRC coord.xy, coord.xyxy;"
 903   CALCWEIGHTS("parmx", "coord.xxxx")
 904   "MUL cdelta.xz, parmx.rrgg, {-%e, 0, %e, 0};"
 905   CALCWEIGHTS("parmy", "coord.yyyy")
 906   "MUL cdelta.yw, parmy.rrgg, {0, -%e, 0, %e};"
 907   BICUB_FILT_MAIN("2D");
 908
 909 static const char *bicub_notex_filt_template_RECT =
 910   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};"
 911   "FRC coord.xy, coord.xyxy;"
 912   CALCWEIGHTS("parmx", "coord.xxxx")
 913   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};"
 914   CALCWEIGHTS("parmy", "coord.yyyy")
 915   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};"
 916   BICUB_FILT_MAIN("RECT");
 917
 918 #define BICUB_X_FILT_MAIN(textype) \
 919   "ADD coord.xy, fragment.texcoord[%c].xyxy, cdelta.xyxy;" \
 920   "ADD coord2.xy, fragment.texcoord[%c].xyxy, cdelta.zyzy;" \
 921   "TEX a.r, coord, texture[%c], "textype";" \
 922   "TEX b.r, coord2, texture[%c], "textype";" \
 923   /* x-interpolation */ \
 924   "LRP yuv.%c, parmx.b, a.rrrr, b.rrrr;"
 925
 926 static const char *bicub_x_filt_template_2D =
 927   "MAD coord.x, fragment.texcoord[%c], {%e}, {0.5};"
 928   "TEX parmx, coord, texture[%c], 1D;"
 929   "MUL cdelta.xyz, parmx.rrgg, {-%e, 0, %e};"
 930   BICUB_X_FILT_MAIN("2D");
 931
 932 static const char *bicub_x_filt_template_RECT =
 933   "ADD coord.x, fragment.texcoord[%c], {0.5};"
 934   "TEX parmx, coord, texture[%c], 1D;"
 935   "MUL cdelta.xyz, parmx.rrgg, {-1, 0, 1};"
 936   BICUB_X_FILT_MAIN("RECT");
 937
 938 static const char *unsharp_filt_template =
 939   "PARAM dcoord%c = {%e, %e, %e, %e};"
 940   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;"
 941   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;"
 942   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;"
 943   "TEX b.r, coord.xyxy, texture[%c], %s;"
 944   "TEX b.g, coord.zwzw, texture[%c], %s;"
 945   "ADD b.r, b.r, b.g;"
 946   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 947   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 948   "DP3 b, b, {0.25, 0.25, 0.25};"
 949   "SUB b.r, a.r, b.r;"
 950   "MAD yuv.%c, b.r, {%e}, a.r;";
 951
 952 static const char *unsharp_filt_template2 =
 953   "PARAM dcoord%c = {%e, %e, %e, %e};"
 954   "PARAM dcoord2%c = {%e, 0, 0, %e};"
 955   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;"
 956   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;"
 957   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;"
 958   "TEX b.r, coord.xyxy, texture[%c], %s;"
 959   "TEX b.g, coord.zwzw, texture[%c], %s;"
 960   "ADD b.r, b.r, b.g;"
 961   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 962   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 963   "ADD b.r, b.r, b.b;"
 964   "ADD b.a, b.r, b.g;"
 965   "ADD coord, fragment.texcoord[%c].xyxy, dcoord2%c;"
 966   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord2%c;"
 967   "TEX b.r, coord.xyxy, texture[%c], %s;"
 968   "TEX b.g, coord.zwzw, texture[%c], %s;"
 969   "ADD b.r, b.r, b.g;"
 970   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 971   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 972   "DP4 b.r, b, {-0.1171875, -0.1171875, -0.1171875, -0.09765625};"
 973   "MAD b.r, a.r, {0.859375}, b.r;"
 974   "MAD yuv.%c, b.r, {%e}, a.r;";
 975
 976 static const char *yuv_prog_template =
 977   "PARAM ycoef = {%e, %e, %e};"
 978   "PARAM ucoef = {%e, %e, %e};"
 979   "PARAM vcoef = {%e, %e, %e};"
 980   "PARAM offsets = {%e, %e, %e};"
 981   "TEMP res;"
 982   "MAD res.rgb, yuv.rrrr, ycoef, offsets;"
 983   "MAD res.rgb, yuv.gggg, ucoef, res;"
 984   "MAD result.color.rgb, yuv.bbbb, vcoef, res;"
 985   "END";
 986
 987 static const char *yuv_pow_prog_template =
 988   "PARAM ycoef = {%e, %e, %e};"
 989   "PARAM ucoef = {%e, %e, %e};"
 990   "PARAM vcoef = {%e, %e, %e};"
 991   "PARAM offsets = {%e, %e, %e};"
 992   "PARAM gamma = {%e, %e, %e};"
 993   "TEMP res;"
 994   "MAD res.rgb, yuv.rrrr, ycoef, offsets;"
 995   "MAD res.rgb, yuv.gggg, ucoef, res;"
 996   "MAD_SAT res.rgb, yuv.bbbb, vcoef, res;"
 997   "POW result.color.r, res.r, gamma.r;"
 998   "POW result.color.g, res.g, gamma.g;"
 999   "POW result.color.b, res.b, gamma.b;"
1000   "END";
1001
1002 static const char *yuv_lookup_prog_template =
1003   "PARAM ycoef = {%e, %e, %e, 0};"
1004   "PARAM ucoef = {%e, %e, %e, 0};"
1005   "PARAM vcoef = {%e, %e, %e, 0};"
1006   "PARAM offsets = {%e, %e, %e, 0.125};"
1007   "TEMP res;"
1008   "MAD res, yuv.rrrr, ycoef, offsets;"
1009   "MAD res.rgb, yuv.gggg, ucoef, res;"
1010   "MAD res.rgb, yuv.bbbb, vcoef, res;"
1011   "TEX result.color.r, res.raaa, texture[%c], 2D;"
1012   "ADD res.a, res.a, 0.25;"
1013   "TEX result.color.g, res.gaaa, texture[%c], 2D;"
1014   "ADD res.a, res.a, 0.25;"
1015   "TEX result.color.b, res.baaa, texture[%c], 2D;"
1016   "END";
1017
1018 static const char *yuv_lookup3d_prog_template =
1019   "TEX result.color, yuv, texture[%c], 3D;"
1020   "END";
1021
1022 /**
1023  * \brief creates and initializes helper textures needed for scaling texture read
1024  * \param scaler scaler type to create texture for
1025  * \param texu contains next free texture unit number
1026  * \param texs texture unit ids for the scaler are stored in this array
1027  */
1028 static void create_scaler_textures(int scaler, int *texu, char *texs) {
1029   switch (scaler) {
1030     case YUV_SCALER_BILIN:
1031     case YUV_SCALER_BICUB_NOTEX:
1032     case YUV_SCALER_UNSHARP:
1033     case YUV_SCALER_UNSHARP2:
1034       break;
1035     case YUV_SCALER_BICUB:
1036     case YUV_SCALER_BICUB_X:
1037       texs[0] = (*texu)++;
1038       gen_spline_lookup_tex(GL_TEXTURE0 + texs[0]);
1039       texs[0] += '0';
1040       break;
1041     default:
1042       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown scaler type %i\n", scaler);
1043   }
1044 }
1045
1046 //! resolution of texture for gamma lookup table
1047 #define LOOKUP_RES 512
1048 //! resolution for 3D yuv->rgb conversion lookup table
1049 #define LOOKUP_3DRES 32
1050 /**
1051  * \brief creates and initializes helper textures needed for yuv conversion
1052  * \param params struct containing parameters like brightness, gamma, ...
1053  * \param texu contains next free texture unit number
1054  * \param texs texture unit ids for the conversion are stored in this array
1055  */
1056 static void create_conv_textures(gl_conversion_params_t *params, int *texu, char *texs) {
1057   unsigned char *lookup_data = NULL;
1058   int conv = YUV_CONVERSION(params->type);
1059   switch (conv) {
1060     case YUV_CONVERSION_FRAGMENT:
1061     case YUV_CONVERSION_FRAGMENT_POW:
1062       break;
1063     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1064       texs[0] = (*texu)++;
1065       ActiveTexture(GL_TEXTURE0 + texs[0]);
1066       lookup_data = malloc(4 * LOOKUP_RES);
1067       mp_gen_gamma_map(lookup_data, LOOKUP_RES, params->csp_params.rgamma);
1068       mp_gen_gamma_map(&lookup_data[LOOKUP_RES], LOOKUP_RES, params->csp_params.ggamma);
1069       mp_gen_gamma_map(&lookup_data[2 * LOOKUP_RES], LOOKUP_RES, params->csp_params.bgamma);
1070       glCreateClearTex(GL_TEXTURE_2D, GL_LUMINANCE8, GL_LUMINANCE, GL_UNSIGNED_BYTE, GL_LINEAR,
1071                        LOOKUP_RES, 4, 0);
1072       glUploadTex(GL_TEXTURE_2D, GL_LUMINANCE, GL_UNSIGNED_BYTE, lookup_data,
1073                   LOOKUP_RES, 0, 0, LOOKUP_RES, 4, 0);
1074       ActiveTexture(GL_TEXTURE0);
1075       texs[0] += '0';
1076       break;
1077     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1078       {
1079         int sz = LOOKUP_3DRES + 2; // texture size including borders
1080         if (!TexImage3D) {
1081           mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing 3D texture function!\n");
1082           break;
1083         }
1084         texs[0] = (*texu)++;
1085         ActiveTexture(GL_TEXTURE0 + texs[0]);
1086         lookup_data = malloc(3 * sz * sz * sz);
1087         mp_gen_yuv2rgb_map(&params->csp_params, lookup_data, LOOKUP_3DRES);
1088         glAdjustAlignment(sz);
1089         PixelStorei(GL_UNPACK_ROW_LENGTH, 0);
1090         TexImage3D(GL_TEXTURE_3D, 0, 3, sz, sz, sz, 1,
1091                    GL_RGB, GL_UNSIGNED_BYTE, lookup_data);
1092         TexParameterf(GL_TEXTURE_3D, GL_TEXTURE_PRIORITY, 1.0);
1093         TexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1094         TexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1095         TexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_S, GL_CLAMP);
1096         TexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_T, GL_CLAMP);
1097         TexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_R, GL_CLAMP);
1098         ActiveTexture(GL_TEXTURE0);
1099         texs[0] += '0';
1100       }
1101       break;
1102     default:
1103       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", conv);
1104   }
1105   if (lookup_data)
1106     free(lookup_data);
1107 }
1108
1109 /**
1110  * \brief adds a scaling texture read at the current fragment program position
1111  * \param scaler type of scaler to insert
1112  * \param prog_pos current position in fragment program
1113  * \param remain how many bytes remain in the buffer given by prog_pos
1114  * \param texs array containing the texture unit identifiers for this scaler
1115  * \param in_tex texture unit the scaler should read from
1116  * \param out_comp component of the yuv variable the scaler stores the result in
1117  * \param rect if rectangular (pixel) adressing should be used for in_tex
1118  * \param texw width of the in_tex texture
1119  * \param texh height of the in_tex texture
1120  * \param strength strength of filter effect if the scaler does some kind of filtering
1121  */
1122 static void add_scaler(int scaler, char **prog_pos, int *remain, char *texs,
1123                     char in_tex, char out_comp, int rect, int texw, int texh,
1124                     double strength) {
1125   const char *ttype = rect ? "RECT" : "2D";
1126   const float ptw = rect ? 1.0 : 1.0 / texw;
1127   const float pth = rect ? 1.0 : 1.0 / texh;
1128   switch (scaler) {
1129     case YUV_SCALER_BILIN:
1130       snprintf(*prog_pos, *remain, bilin_filt_template, out_comp, in_tex,
1131                  in_tex, ttype);
1132       break;
1133     case YUV_SCALER_BICUB:
1134       if (rect)
1135         snprintf(*prog_pos, *remain, bicub_filt_template_RECT,
1136                  in_tex, texs[0], texs[0],
1137                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1138       else
1139         snprintf(*prog_pos, *remain, bicub_filt_template_2D,
1140                  in_tex, (float)texw, (float)texh,
1141                  texs[0], ptw, ptw, texs[0], pth, pth,
1142                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1143       break;
1144     case YUV_SCALER_BICUB_X:
1145       if (rect)
1146         snprintf(*prog_pos, *remain, bicub_x_filt_template_RECT,
1147                  in_tex, texs[0],
1148                  in_tex, in_tex, in_tex, in_tex, out_comp);
1149       else
1150         snprintf(*prog_pos, *remain, bicub_x_filt_template_2D,
1151                  in_tex, (float)texw,
1152                  texs[0], ptw, ptw,
1153                  in_tex, in_tex, in_tex, in_tex, out_comp);
1154       break;
1155     case YUV_SCALER_BICUB_NOTEX:
1156       if (rect)
1157         snprintf(*prog_pos, *remain, bicub_notex_filt_template_RECT,
1158                  in_tex,
1159                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1160       else
1161         snprintf(*prog_pos, *remain, bicub_notex_filt_template_2D,
1162                  in_tex, (float)texw, (float)texh, ptw, ptw, pth, pth,
1163                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1164       break;
1165     case YUV_SCALER_UNSHARP:
1166       snprintf(*prog_pos, *remain, unsharp_filt_template,
1167                out_comp, 0.5 * ptw, 0.5 * pth, 0.5 * ptw, -0.5 * pth,
1168                in_tex, out_comp, in_tex, out_comp, in_tex,
1169                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1170                in_tex, ttype, out_comp, strength);
1171       break;
1172     case YUV_SCALER_UNSHARP2:
1173       snprintf(*prog_pos, *remain, unsharp_filt_template2,
1174                out_comp, 1.2 * ptw, 1.2 * pth, 1.2 * ptw, -1.2 * pth,
1175                out_comp, 1.5 * ptw, 1.5 * pth,
1176                in_tex, out_comp, in_tex, out_comp, in_tex,
1177                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1178                in_tex, ttype, in_tex, out_comp, in_tex, out_comp,
1179                in_tex, ttype, in_tex, ttype, in_tex, ttype,
1180                in_tex, ttype, out_comp, strength);
1181       break;
1182   }
1183   *remain -= strlen(*prog_pos);
1184   *prog_pos += strlen(*prog_pos);
1185 }
1186
1187 static const struct {
1188   const char *name;
1189   GLenum cur;
1190   GLenum max;
1191 } progstats[] = {
1192   {"instructions", 0x88A0, 0x88A1},
1193   {"native instructions", 0x88A2, 0x88A3},
1194   {"temporaries", 0x88A4, 0x88A5},
1195   {"native temporaries", 0x88A6, 0x88A7},
1196   {"parameters", 0x88A8, 0x88A9},
1197   {"native parameters", 0x88AA, 0x88AB},
1198   {"attribs", 0x88AC, 0x88AD},
1199   {"native attribs", 0x88AE, 0x88AF},
1200   {"ALU instructions", 0x8805, 0x880B},
1201   {"TEX instructions", 0x8806, 0x880C},
1202   {"TEX indirections", 0x8807, 0x880D},
1203   {"native ALU instructions", 0x8808, 0x880E},
1204   {"native TEX instructions", 0x8809, 0x880F},
1205   {"native TEX indirections", 0x880A, 0x8810},
1206   {NULL, 0, 0}
1207 };
1208
1209 /**
1210  * \brief load the specified GPU Program
1211  * \param target program target to load into, only GL_FRAGMENT_PROGRAM is tested
1212  * \param prog program string
1213  * \return 1 on success, 0 otherwise
1214  */
1215 int loadGPUProgram(GLenum target, char *prog) {
1216   int i;
1217   GLint cur = 0, max = 0, err = 0;
1218   if (!ProgramString) {
1219     mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing GPU program function\n");
1220     return 0;
1221   }
1222   ProgramString(target, GL_PROGRAM_FORMAT_ASCII, strlen(prog), prog);
1223   GetIntegerv(GL_PROGRAM_ERROR_POSITION, &err);
1224   if (err != -1) {
1225     mp_msg(MSGT_VO, MSGL_ERR,
1226       "[gl] Error compiling fragment program, make sure your card supports\n"
1227       "[gl]   GL_ARB_fragment_program (use glxinfo to check).\n"
1228       "[gl]   Error message:\n  %s at %.10s\n",
1229       GetString(GL_PROGRAM_ERROR_STRING), &prog[err]);
1230     return 0;
1231   }
1232   if (!GetProgramiv || !mp_msg_test(MSGT_VO, MSGL_DBG2))
1233     return 1;
1234   mp_msg(MSGT_VO, MSGL_V, "[gl] Program statistics:\n");
1235   for (i = 0; progstats[i].name; i++) {
1236     GetProgramiv(target, progstats[i].cur, &cur);
1237     GetProgramiv(target, progstats[i].max, &max);
1238     mp_msg(MSGT_VO, MSGL_V, "[gl]   %s: %i/%i\n", progstats[i].name, cur, max);
1239   }
1240   return 1;
1241 }
1242
1243 #define MAX_PROGSZ (1024*1024)
1244
1245 /**
1246  * \brief setup a fragment program that will do YUV->RGB conversion
1247  * \param parms struct containing parameters like conversion and scaler type,
1248  *              brightness, ...
1249  */
1250 static void glSetupYUVFragprog(gl_conversion_params_t *params) {
1251   int type = params->type;
1252   int texw = params->texw;
1253   int texh = params->texh;
1254   int rect = params->target == GL_TEXTURE_RECTANGLE;
1255   static const char prog_hdr[] =
1256     "!!ARBfp1.0\n"
1257     "OPTION ARB_precision_hint_fastest;"
1258     // all scaler variables must go here so they aren't defined
1259     // multiple times when the same scaler is used more than once
1260     "TEMP coord, coord2, cdelta, parmx, parmy, a, b, yuv;";
1261   int prog_remain;
1262   char *yuv_prog, *prog_pos;
1263   int cur_texu = 3;
1264   char lum_scale_texs[1];
1265   char chrom_scale_texs[1];
1266   char conv_texs[1];
1267   GLint i;
1268   // this is the conversion matrix, with y, u, v factors
1269   // for red, green, blue and the constant offsets
1270   float yuv2rgb[3][4];
1271   create_conv_textures(params, &cur_texu, conv_texs);
1272   create_scaler_textures(YUV_LUM_SCALER(type), &cur_texu, lum_scale_texs);
1273   if (YUV_CHROM_SCALER(type) == YUV_LUM_SCALER(type))
1274     memcpy(chrom_scale_texs, lum_scale_texs, sizeof(chrom_scale_texs));
1275   else
1276     create_scaler_textures(YUV_CHROM_SCALER(type), &cur_texu, chrom_scale_texs);
1277   GetIntegerv(GL_MAX_TEXTURE_IMAGE_UNITS, &i);
1278   if (i < cur_texu)
1279     mp_msg(MSGT_VO, MSGL_ERR,
1280            "[gl] %i texture units needed for this type of YUV fragment support (found %i)\n",
1281            cur_texu, i);
1282   if (!ProgramString) {
1283     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] ProgramString function missing!\n");
1284     return;
1285   }
1286   yuv_prog = malloc(MAX_PROGSZ);
1287   strcpy(yuv_prog, prog_hdr);
1288   prog_pos    = yuv_prog + sizeof(prog_hdr) - 1;
1289   prog_remain = MAX_PROGSZ - sizeof(prog_hdr);
1290   add_scaler(YUV_LUM_SCALER(type), &prog_pos, &prog_remain, lum_scale_texs,
1291              '0', 'r', rect, texw, texh, params->filter_strength);
1292   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1293              '1', 'g', rect, params->chrom_texw, params->chrom_texh, params->filter_strength);
1294   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1295              '2', 'b', rect, params->chrom_texw, params->chrom_texh, params->filter_strength);
1296   mp_get_yuv2rgb_coeffs(&params->csp_params, yuv2rgb);
1297   switch (YUV_CONVERSION(type)) {
1298     case YUV_CONVERSION_FRAGMENT:
1299       snprintf(prog_pos, prog_remain, yuv_prog_template,
1300                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1301                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1302                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1303                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C]);
1304       break;
1305     case YUV_CONVERSION_FRAGMENT_POW:
1306       snprintf(prog_pos, prog_remain, yuv_pow_prog_template,
1307                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1308                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1309                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1310                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1311                (float)1.0 / params->csp_params.rgamma, (float)1.0 / params->csp_params.bgamma, (float)1.0 / params->csp_params.bgamma);
1312       break;
1313     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1314       snprintf(prog_pos, prog_remain, yuv_lookup_prog_template,
1315                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1316                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1317                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1318                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1319                conv_texs[0], conv_texs[0], conv_texs[0]);
1320       break;
1321     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1322       snprintf(prog_pos, prog_remain, yuv_lookup3d_prog_template, conv_texs[0]);
1323       break;
1324     default:
1325       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(type));
1326       break;
1327   }
1328   mp_msg(MSGT_VO, MSGL_DBG2, "[gl] generated fragment program:\n%s\n", yuv_prog);
1329   loadGPUProgram(GL_FRAGMENT_PROGRAM, yuv_prog);
1330   free(yuv_prog);
1331 }
1332
1333 /**
1334  * \brief setup YUV->RGB conversion
1335  * \param parms struct containing parameters like conversion and scaler type,
1336  *              brightness, ...
1337  * \ingroup glconversion
1338  */
1339 void glSetupYUVConversion(gl_conversion_params_t *params) {
1340   float uvcos = params->csp_params.saturation * cos(params->csp_params.hue);
1341   float uvsin = params->csp_params.saturation * sin(params->csp_params.hue);
1342   switch (YUV_CONVERSION(params->type)) {
1343     case YUV_CONVERSION_COMBINERS:
1344       glSetupYUVCombiners(uvcos, uvsin);
1345       break;
1346     case YUV_CONVERSION_COMBINERS_ATI:
1347       glSetupYUVCombinersATI(uvcos, uvsin);
1348       break;
1349     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1350     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1351     case YUV_CONVERSION_FRAGMENT:
1352     case YUV_CONVERSION_FRAGMENT_POW:
1353       glSetupYUVFragprog(params);
1354       break;
1355     case YUV_CONVERSION_NONE:
1356       break;
1357     default:
1358       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(params->type));
1359   }
1360 }
1361
1362 /**
1363  * \brief enable the specified YUV conversion
1364  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1365  * \param type type of YUV conversion
1366  * \ingroup glconversion
1367  */
1368 void glEnableYUVConversion(GLenum target, int type) {
1369   switch (YUV_CONVERSION(type)) {
1370     case YUV_CONVERSION_COMBINERS:
1371       ActiveTexture(GL_TEXTURE1);
1372       Enable(target);
1373       ActiveTexture(GL_TEXTURE2);
1374       Enable(target);
1375       ActiveTexture(GL_TEXTURE0);
1376       Enable(GL_REGISTER_COMBINERS_NV);
1377       break;
1378     case YUV_CONVERSION_COMBINERS_ATI:
1379       ActiveTexture(GL_TEXTURE1);
1380       Enable(target);
1381       ActiveTexture(GL_TEXTURE2);
1382       Enable(target);
1383       ActiveTexture(GL_TEXTURE0);
1384       Enable(GL_FRAGMENT_SHADER_ATI);
1385       break;
1386     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1387     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1388     case YUV_CONVERSION_FRAGMENT_POW:
1389     case YUV_CONVERSION_FRAGMENT:
1390     case YUV_CONVERSION_NONE:
1391       Enable(GL_FRAGMENT_PROGRAM);
1392       break;
1393   }
1394 }
1395
1396 /**
1397  * \brief disable the specified YUV conversion
1398  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1399  * \param type type of YUV conversion
1400  * \ingroup glconversion
1401  */
1402 void glDisableYUVConversion(GLenum target, int type) {
1403   switch (YUV_CONVERSION(type)) {
1404     case YUV_CONVERSION_COMBINERS:
1405       ActiveTexture(GL_TEXTURE1);
1406       Disable(target);
1407       ActiveTexture(GL_TEXTURE2);
1408       Disable(target);
1409       ActiveTexture(GL_TEXTURE0);
1410       Disable(GL_REGISTER_COMBINERS_NV);
1411       break;
1412     case YUV_CONVERSION_COMBINERS_ATI:
1413       ActiveTexture(GL_TEXTURE1);
1414       Disable(target);
1415       ActiveTexture(GL_TEXTURE2);
1416       Disable(target);
1417       ActiveTexture(GL_TEXTURE0);
1418       Disable(GL_FRAGMENT_SHADER_ATI);
1419       break;
1420     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1421     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1422     case YUV_CONVERSION_FRAGMENT_POW:
1423     case YUV_CONVERSION_FRAGMENT:
1424     case YUV_CONVERSION_NONE:
1425       Disable(GL_FRAGMENT_PROGRAM);
1426       break;
1427   }
1428 }
1429
1430 /**
1431  * \brief draw a texture part at given 2D coordinates
1432  * \param x screen top coordinate
1433  * \param y screen left coordinate
1434  * \param w screen width coordinate
1435  * \param h screen height coordinate
1436  * \param tx texture top coordinate in pixels
1437  * \param ty texture left coordinate in pixels
1438  * \param tw texture part width in pixels
1439  * \param th texture part height in pixels
1440  * \param sx width of texture in pixels
1441  * \param sy height of texture in pixels
1442  * \param rect_tex whether this texture uses texture_rectangle extension
1443  * \param is_yv12 if != 0, also draw the textures from units 1 and 2,
1444  *                bits 8 - 15 and 16 - 23 specify the x and y scaling of those textures
1445  * \param flip flip the texture upside down
1446  * \ingroup gltexture
1447  */
1448 void glDrawTex(GLfloat x, GLfloat y, GLfloat w, GLfloat h,
1449                GLfloat tx, GLfloat ty, GLfloat tw, GLfloat th,
1450                int sx, int sy, int rect_tex, int is_yv12, int flip) {
1451   int chroma_x_shift = (is_yv12 >>  8) & 31;
1452   int chroma_y_shift = (is_yv12 >> 16) & 31;
1453   GLfloat xscale = 1 << chroma_x_shift;
1454   GLfloat yscale = 1 << chroma_y_shift;
1455   GLfloat tx2 = tx / xscale, ty2 = ty / yscale, tw2 = tw / xscale, th2 = th / yscale;
1456   if (!rect_tex) {
1457     tx /= sx; ty /= sy; tw /= sx; th /= sy;
1458     tx2 = tx, ty2 = ty, tw2 = tw, th2 = th;
1459   }
1460   if (flip) {
1461     y += h;
1462     h = -h;
1463   }
1464   Begin(GL_QUADS);
1465   TexCoord2f(tx, ty);
1466   if (is_yv12) {
1467     MultiTexCoord2f(GL_TEXTURE1, tx2, ty2);
1468     MultiTexCoord2f(GL_TEXTURE2, tx2, ty2);
1469   }
1470   Vertex2f(x, y);
1471   TexCoord2f(tx, ty + th);
1472   if (is_yv12) {
1473     MultiTexCoord2f(GL_TEXTURE1, tx2, ty2 + th2);
1474     MultiTexCoord2f(GL_TEXTURE2, tx2, ty2 + th2);
1475   }
1476   Vertex2f(x, y + h);
1477   TexCoord2f(tx + tw, ty + th);
1478   if (is_yv12) {
1479     MultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2 + th2);
1480     MultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2 + th2);
1481   }
1482   Vertex2f(x + w, y + h);
1483   TexCoord2f(tx + tw, ty);
1484   if (is_yv12) {
1485     MultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2);
1486     MultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2);
1487   }
1488   Vertex2f(x + w, y);
1489   End();
1490 }
1491
1492 #ifdef CONFIG_GL_WIN32
1493 #include "w32_common.h"
1494 /**
1495  * \brief little helper since wglGetProcAddress definition does not fit our
1496  *        getProcAddress
1497  * \param procName name of function to look up
1498  * \return function pointer returned by wglGetProcAddress
1499  */
1500 static void *w32gpa(const GLubyte *procName) {
1501   HMODULE oglmod;
1502   void *res = wglGetProcAddress(procName);
1503   if (res) return res;
1504   oglmod = GetModuleHandle("opengl32.dll");
1505   return GetProcAddress(oglmod, procName);
1506 }
1507
1508 static int setGlWindow_w32(MPGLContext *ctx)
1509 {
1510   HWND win = vo_w32_window;
1511   int *vinfo = &ctx->vinfo.w32;
1512   HGLRC *context = &ctx->context.w32;
1513   int new_vinfo;
1514   HDC windc = vo_w32_get_dc(win);
1515   HGLRC new_context = 0;
1516   int keep_context = 0;
1517   int res = SET_WINDOW_FAILED;
1518
1519   // should only be needed when keeping context, but not doing glFinish
1520   // can cause flickering even when we do not keep it.
1521   if (*context)
1522   Finish();
1523   new_vinfo = GetPixelFormat(windc);
1524   if (*context && *vinfo && new_vinfo && *vinfo == new_vinfo) {
1525       // we can keep the wglContext
1526       new_context = *context;
1527       keep_context = 1;
1528   } else {
1529     // create a context
1530     new_context = wglCreateContext(windc);
1531     if (!new_context) {
1532       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GL context!\n");
1533       goto out;
1534     }
1535   }
1536
1537   // set context
1538   if (!wglMakeCurrent(windc, new_context)) {
1539     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GL context!\n");
1540     if (!keep_context) {
1541       wglDeleteContext(new_context);
1542     }
1543     goto out;
1544   }
1545
1546   // set new values
1547   vo_w32_window = win;
1548   {
1549     RECT rect;
1550     GetClientRect(win, &rect);
1551     vo_dwidth = rect.right;
1552     vo_dheight = rect.bottom;
1553   }
1554   if (!keep_context) {
1555     if (*context)
1556       wglDeleteContext(*context);
1557     *context = new_context;
1558     *vinfo = new_vinfo;
1559     getFunctions(w32gpa, NULL);
1560
1561     // and inform that reinit is neccessary
1562     res = SET_WINDOW_REINIT;
1563   } else
1564     res = SET_WINDOW_OK;
1565
1566 out:
1567   vo_w32_release_dc(win, windc);
1568   return res;
1569 }
1570
1571 static void releaseGlContext_w32(MPGLContext *ctx) {
1572   int *vinfo = &ctx->vinfo.w32;
1573   HGLRC *context = &ctx->context.w32;
1574   *vinfo = 0;
1575   if (*context) {
1576     wglMakeCurrent(0, 0);
1577     wglDeleteContext(*context);
1578   }
1579   *context = 0;
1580 }
1581
1582 static void swapGlBuffers_w32(MPGLContext *ctx) {
1583   HDC vo_hdc = vo_w32_get_dc(vo_w32_window);
1584   SwapBuffers(vo_hdc);
1585   vo_w32_release_dc(vo_w32_window, vo_hdc);
1586 }
1587 #endif
1588 #ifdef CONFIG_GL_X11
1589 #ifdef HAVE_LIBDL
1590 #include <dlfcn.h>
1591 #endif
1592 #include "x11_common.h"
1593 /**
1594  * \brief find address of a linked function
1595  * \param s name of function to find
1596  * \return address of function or NULL if not found
1597  */
1598 static void *getdladdr(const char *s) {
1599   void *ret = NULL;
1600 #ifdef HAVE_LIBDL
1601   void *handle = dlopen(NULL, RTLD_LAZY);
1602   if (!handle)
1603     return NULL;
1604   ret = dlsym(handle, s);
1605   dlclose(handle);
1606 #endif
1607   return ret;
1608 }
1609
1610 /**
1611  * \brief Returns the XVisualInfo associated with Window win.
1612  * \param win Window whose XVisualInfo is returne.
1613  * \return XVisualInfo of the window. Caller must use XFree to free it.
1614  */
1615 static XVisualInfo *getWindowVisualInfo(Window win) {
1616   XWindowAttributes xw_attr;
1617   XVisualInfo vinfo_template;
1618   int tmp;
1619   XGetWindowAttributes(mDisplay, win, &xw_attr);
1620   vinfo_template.visualid = XVisualIDFromVisual(xw_attr.visual);
1621   return XGetVisualInfo(mDisplay, VisualIDMask, &vinfo_template, &tmp);
1622 }
1623
1624 static void appendstr(char **dst, const char *str)
1625 {
1626     int newsize;
1627     char *newstr;
1628     if (!str)
1629         return;
1630     newsize = strlen(*dst) + 1 + strlen(str) + 1;
1631     newstr = realloc(*dst, newsize);
1632     if (!newstr)
1633         return;
1634     *dst = newstr;
1635     strcat(*dst, " ");
1636     strcat(*dst, str);
1637 }
1638
1639 /**
1640  * \brief Changes the window in which video is displayed.
1641  * If possible only transfers the context to the new window, otherwise
1642  * creates a new one, which must be initialized by the caller.
1643  * \param vinfo Currently used visual.
1644  * \param context Currently used context.
1645  * \param win window that should be used for drawing.
1646  * \return one of SET_WINDOW_FAILED, SET_WINDOW_OK or SET_WINDOW_REINIT.
1647  * In case of SET_WINDOW_REINIT the context could not be transfered
1648  * and the caller must initialize it correctly.
1649  * \ingroup glcontext
1650  */
1651 static int setGlWindow_x11(MPGLContext *ctx)
1652 {
1653   XVisualInfo **vinfo = &ctx->vinfo.x11;
1654   GLXContext *context = &ctx->context.x11;
1655   Window win = vo_window;
1656   XVisualInfo *new_vinfo;
1657   GLXContext new_context = NULL;
1658   int keep_context = 0;
1659
1660   // should only be needed when keeping context, but not doing glFinish
1661   // can cause flickering even when we do not keep it.
1662   if (*context)
1663   Finish();
1664   new_vinfo = getWindowVisualInfo(win);
1665   if (*context && *vinfo && new_vinfo &&
1666       (*vinfo)->visualid == new_vinfo->visualid) {
1667       // we can keep the GLXContext
1668       new_context = *context;
1669       XFree(new_vinfo);
1670       new_vinfo = *vinfo;
1671       keep_context = 1;
1672   } else {
1673     // create a context
1674     new_context = glXCreateContext(mDisplay, new_vinfo, NULL, True);
1675     if (!new_context) {
1676       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GLX context!\n");
1677       XFree(new_vinfo);
1678       return SET_WINDOW_FAILED;
1679     }
1680   }
1681
1682   // set context
1683   if (!glXMakeCurrent(mDisplay, vo_window, new_context)) {
1684     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GLX context!\n");
1685     if (!keep_context) {
1686       glXDestroyContext(mDisplay, new_context);
1687       XFree(new_vinfo);
1688     }
1689     return SET_WINDOW_FAILED;
1690   }
1691
1692   // set new values
1693   vo_window = win;
1694   vo_x11_update_geometry();
1695   if (!keep_context) {
1696     void *(*getProcAddress)(const GLubyte *);
1697     const char *(*glXExtStr)(Display *, int);
1698     char *glxstr = strdup("");
1699     if (*context)
1700       glXDestroyContext(mDisplay, *context);
1701     *context = new_context;
1702     if (*vinfo)
1703       XFree(*vinfo);
1704     *vinfo = new_vinfo;
1705       getProcAddress = getdladdr("glXGetProcAddress");
1706     if (!getProcAddress)
1707       getProcAddress = getdladdr("glXGetProcAddressARB");
1708     if (!getProcAddress)
1709       getProcAddress = (void *)getdladdr;
1710     glXExtStr = getdladdr("glXQueryExtensionsString");
1711     if (glXExtStr)
1712         appendstr(&glxstr, glXExtStr(mDisplay, DefaultScreen(mDisplay)));
1713     glXExtStr = getdladdr("glXGetClientString");
1714     if (glXExtStr)
1715         appendstr(&glxstr, glXExtStr(mDisplay, GLX_EXTENSIONS));
1716     glXExtStr = getdladdr("glXGetServerString");
1717     if (glXExtStr)
1718         appendstr(&glxstr, glXExtStr(mDisplay, GLX_EXTENSIONS));
1719
1720     getFunctions(getProcAddress, glxstr);
1721     free(glxstr);
1722
1723     // and inform that reinit is neccessary
1724     return SET_WINDOW_REINIT;
1725   }
1726   return SET_WINDOW_OK;
1727 }
1728
1729 /**
1730  * \brief free the VisualInfo and GLXContext of an OpenGL context.
1731  * \ingroup glcontext
1732  */
1733 static void releaseGlContext_x11(MPGLContext *ctx) {
1734   XVisualInfo **vinfo = &ctx->vinfo.x11;
1735   GLXContext *context = &ctx->context.x11;
1736   if (*vinfo)
1737     XFree(*vinfo);
1738   *vinfo = NULL;
1739   if (*context)
1740   {
1741     Finish();
1742     glXMakeCurrent(mDisplay, None, NULL);
1743     glXDestroyContext(mDisplay, *context);
1744   }
1745   *context = 0;
1746 }
1747
1748 static void swapGlBuffers_x11(MPGLContext *ctx) {
1749   glXSwapBuffers(mDisplay, vo_window);
1750 }
1751
1752 static int x11_check_events(void) {
1753   return vo_x11_check_events(mDisplay);
1754 }
1755 #endif
1756
1757 int init_mpglcontext(MPGLContext *ctx, enum MPGLType type) {
1758   memset(ctx, 0, sizeof(*ctx));
1759   ctx->type = type;
1760   switch (ctx->type) {
1761 #ifdef CONFIG_GL_WIN32
1762   case GLTYPE_W32:
1763     ctx->setGlWindow = setGlWindow_w32;
1764     ctx->releaseGlContext = releaseGlContext_w32;
1765     ctx->swapGlBuffers = swapGlBuffers_w32;
1766     ctx->update_xinerama_info = w32_update_xinerama_info;
1767     ctx->border = vo_w32_border;
1768     ctx->check_events = vo_w32_check_events;
1769     ctx->fullscreen = vo_w32_fullscreen;
1770     ctx->ontop = vo_w32_ontop;
1771     return vo_w32_init();
1772 #endif
1773 #ifdef CONFIG_GL_X11
1774   case GLTYPE_X11:
1775     ctx->setGlWindow = setGlWindow_x11;
1776     ctx->releaseGlContext = releaseGlContext_x11;
1777     ctx->swapGlBuffers = swapGlBuffers_x11;
1778     ctx->update_xinerama_info = update_xinerama_info;
1779     ctx->border = vo_x11_border;
1780     ctx->check_events = x11_check_events;
1781     ctx->fullscreen = vo_x11_fullscreen;
1782     ctx->ontop = vo_x11_ontop;
1783     return vo_init();
1784 #endif
1785   default:
1786     return 0;
1787   }
1788 }
1789
1790 void uninit_mpglcontext(MPGLContext *ctx) {
1791   ctx->releaseGlContext(ctx);
1792   switch (ctx->type) {
1793 #ifdef CONFIG_GL_WIN32
1794   case GLTYPE_W32:
1795     vo_w32_uninit();
1796     break;
1797 #endif
1798 #ifdef CONFIG_GL_X11
1799   case GLTYPE_X11:
1800     vo_x11_uninit();
1801     break;
1802 #endif
1803   }
1804   memset(ctx, 0, sizeof(*ctx));
1805 }