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 "csputils.h"
  38
  39 void (GLAPIENTRY *Begin)(GLenum);
  40 void (GLAPIENTRY *End)(void);
  41 void (GLAPIENTRY *Viewport)(GLint, GLint, GLsizei, GLsizei);
  42 void (GLAPIENTRY *MatrixMode)(GLenum);
  43 void (GLAPIENTRY *LoadIdentity)(void);
  44 void (GLAPIENTRY *Translated)(double, double, double);
  45 void (GLAPIENTRY *Scaled)(double, double, double);
  46 void (GLAPIENTRY *Ortho)(double, double, double, double, double, double);
  47 void (GLAPIENTRY *Frustum)(double, double, double, double, double, double);
  48 void (GLAPIENTRY *PushMatrix)(void);
  49 void (GLAPIENTRY *PopMatrix)(void);
  50 void (GLAPIENTRY *Clear)(GLbitfield);
  51 GLuint (GLAPIENTRY *GenLists)(GLsizei);
  52 void (GLAPIENTRY *DeleteLists)(GLuint, GLsizei);
  53 void (GLAPIENTRY *NewList)(GLuint, GLenum);
  54 void (GLAPIENTRY *EndList)(void);
  55 void (GLAPIENTRY *CallList)(GLuint);
  56 void (GLAPIENTRY *CallLists)(GLsizei, GLenum, const GLvoid *);
  57 void (GLAPIENTRY *GenTextures)(GLsizei, GLuint *);
  58 void (GLAPIENTRY *DeleteTextures)(GLsizei, const GLuint *);
  59 void (GLAPIENTRY *TexEnvf)(GLenum, GLenum, GLfloat);
  60 void (GLAPIENTRY *TexEnvi)(GLenum, GLenum, GLint);
  61 void (GLAPIENTRY *Color4ub)(GLubyte, GLubyte, GLubyte, GLubyte);
  62 void (GLAPIENTRY *Color3f)(GLfloat, GLfloat, GLfloat);
  63 void (GLAPIENTRY *Color4f)(GLfloat, GLfloat, GLfloat, GLfloat);
  64 void (GLAPIENTRY *ClearColor)(GLclampf, GLclampf, GLclampf, GLclampf);
  65 void (GLAPIENTRY *ClearDepth)(GLclampd);
  66 void (GLAPIENTRY *DepthFunc)(GLenum);
  67 void (GLAPIENTRY *Enable)(GLenum);
  68 void (GLAPIENTRY *Disable)(GLenum);
  69 const GLubyte *(GLAPIENTRY *GetString)(GLenum);
  70 void (GLAPIENTRY *DrawBuffer)(GLenum);
  71 void (GLAPIENTRY *DepthMask)(GLboolean);
  72 void (GLAPIENTRY *BlendFunc)(GLenum, GLenum);
  73 void (GLAPIENTRY *Flush)(void);
  74 void (GLAPIENTRY *Finish)(void);
  75 void (GLAPIENTRY *PixelStorei)(GLenum, GLint);
  76 void (GLAPIENTRY *TexImage1D)(GLenum, GLint, GLint, GLsizei, GLint, GLenum, GLenum, const GLvoid *);
  77 void (GLAPIENTRY *TexImage2D)(GLenum, GLint, GLint, GLsizei, GLsizei, GLint, GLenum, GLenum, const GLvoid *);
  78 void (GLAPIENTRY *TexSubImage2D)(GLenum, GLint, GLint, GLint, GLsizei, GLsizei, GLenum, GLenum, const GLvoid *);
  79 void (GLAPIENTRY *TexParameteri)(GLenum, GLenum, GLint);
  80 void (GLAPIENTRY *TexParameterf)(GLenum, GLenum, GLfloat);
  81 void (GLAPIENTRY *TexParameterfv)(GLenum, GLenum, const GLfloat *);
  82 void (GLAPIENTRY *TexCoord2f)(GLfloat, GLfloat);
  83 void (GLAPIENTRY *Vertex2f)(GLfloat, GLfloat);
  84 void (GLAPIENTRY *Vertex3f)(GLfloat, GLfloat, GLfloat);
  85 void (GLAPIENTRY *Normal3f)(GLfloat, GLfloat, GLfloat);
  86 void (GLAPIENTRY *Lightfv)(GLenum, GLenum, const GLfloat *);
  87 void (GLAPIENTRY *ColorMaterial)(GLenum, GLenum);
  88 void (GLAPIENTRY *ShadeModel)(GLenum);
  89 void (GLAPIENTRY *GetIntegerv)(GLenum, GLint *);
  90
  91 /**
  92  * \defgroup glextfunctions OpenGL extension functions
  93  *
  94  * the pointers to these functions are acquired when the OpenGL
  95  * context is created
  96  * \{
  97  */
  98 void (GLAPIENTRY *GenBuffers)(GLsizei, GLuint *);
  99 void (GLAPIENTRY *DeleteBuffers)(GLsizei, const GLuint *);
 100 void (GLAPIENTRY *BindBuffer)(GLenum, GLuint);
 101 GLvoid* (GLAPIENTRY *MapBuffer)(GLenum, GLenum);
 102 GLboolean (GLAPIENTRY *UnmapBuffer)(GLenum);
 103 void (GLAPIENTRY *BufferData)(GLenum, intptr_t, const GLvoid *, GLenum);
 104 void (GLAPIENTRY *CombinerParameterfv)(GLenum, const GLfloat *);
 105 void (GLAPIENTRY *CombinerParameteri)(GLenum, GLint);
 106 void (GLAPIENTRY *CombinerInput)(GLenum, GLenum, GLenum, GLenum, GLenum,
 107                                GLenum);
 108 void (GLAPIENTRY *CombinerOutput)(GLenum, GLenum, GLenum, GLenum, GLenum,
 109                                 GLenum, GLenum, GLboolean, GLboolean,
 110                                 GLboolean);
 111 void (GLAPIENTRY *BeginFragmentShader)(void);
 112 void (GLAPIENTRY *EndFragmentShader)(void);
 113 void (GLAPIENTRY *SampleMap)(GLuint, GLuint, GLenum);
 114 void (GLAPIENTRY *ColorFragmentOp2)(GLenum, GLuint, GLuint, GLuint, GLuint,
 115                                   GLuint, GLuint, GLuint, GLuint, GLuint);
 116 void (GLAPIENTRY *ColorFragmentOp3)(GLenum, GLuint, GLuint, GLuint, GLuint,
 117                                   GLuint, GLuint, GLuint, GLuint, GLuint,
 118                                   GLuint, GLuint, GLuint);
 119 void (GLAPIENTRY *SetFragmentShaderConstant)(GLuint, const GLfloat *);
 120 void (GLAPIENTRY *ActiveTexture)(GLenum);
 121 void (GLAPIENTRY *BindTexture)(GLenum, GLuint);
 122 void (GLAPIENTRY *MultiTexCoord2f)(GLenum, GLfloat, GLfloat);
 123 void (GLAPIENTRY *GenPrograms)(GLsizei, GLuint *);
 124 void (GLAPIENTRY *DeletePrograms)(GLsizei, const GLuint *);
 125 void (GLAPIENTRY *BindProgram)(GLenum, GLuint);
 126 void (GLAPIENTRY *ProgramString)(GLenum, GLenum, GLsizei, const GLvoid *);
 127 void (GLAPIENTRY *GetProgramiv)(GLenum, GLenum, GLint *);
 128 void (GLAPIENTRY *ProgramEnvParameter4f)(GLenum, GLuint, GLfloat, GLfloat,
 129                                        GLfloat, GLfloat);
 130 int (GLAPIENTRY *SwapInterval)(int);
 131 void (GLAPIENTRY *TexImage3D)(GLenum, GLint, GLenum, GLsizei, GLsizei, GLsizei,
 132                             GLint, GLenum, GLenum, const GLvoid *);
 133 void* (GLAPIENTRY *AllocateMemoryMESA)(void *, int, size_t, float, float, float);
 134 void (GLAPIENTRY *FreeMemoryMESA)(void *, int, void *);
 135 /** \} */ // end of glextfunctions group
 136
 137 //! \defgroup glgeneral OpenGL general helper functions
 138
 139 //! \defgroup glcontext OpenGL context management helper functions
 140
 141 //! \defgroup gltexture OpenGL texture handling helper functions
 142
 143 //! \defgroup glconversion OpenGL conversion helper functions
 144
 145 static GLint hqtexfmt;
 146
 147 /**
 148  * \brief adjusts the GL_UNPACK_ALIGNMENT to fit the stride.
 149  * \param stride number of bytes per line for which alignment should fit.
 150  * \ingroup glgeneral
 151  */
 152 void glAdjustAlignment(int stride) {
 153   GLint gl_alignment;
 154   if (stride % 8 == 0)
 155     gl_alignment=8;
 156   else if (stride % 4 == 0)
 157     gl_alignment=4;
 158   else if (stride % 2 == 0)
 159     gl_alignment=2;
 160   else
 161     gl_alignment=1;
 162   PixelStorei(GL_UNPACK_ALIGNMENT, gl_alignment);
 163 }
 164
 165 struct gl_name_map_struct {
 166   GLint value;
 167   const char *name;
 168 };
 169
 170 #undef MAP
 171 #define MAP(a) {a, #a}
 172 //! mapping table for the glValName function
 173 static const struct gl_name_map_struct gl_name_map[] = {
 174   // internal format
 175   MAP(GL_R3_G3_B2), MAP(GL_RGB4), MAP(GL_RGB5), MAP(GL_RGB8),
 176   MAP(GL_RGB10), MAP(GL_RGB12), MAP(GL_RGB16), MAP(GL_RGBA2),
 177   MAP(GL_RGBA4), MAP(GL_RGB5_A1), MAP(GL_RGBA8), MAP(GL_RGB10_A2),
 178   MAP(GL_RGBA12), MAP(GL_RGBA16), MAP(GL_LUMINANCE8),
 179
 180   // format
 181   MAP(GL_RGB), MAP(GL_RGBA), MAP(GL_RED), MAP(GL_GREEN), MAP(GL_BLUE),
 182   MAP(GL_ALPHA), MAP(GL_LUMINANCE), MAP(GL_LUMINANCE_ALPHA),
 183   MAP(GL_COLOR_INDEX),
 184   // rest 1.2 only
 185   MAP(GL_BGR), MAP(GL_BGRA),
 186
 187   //type
 188   MAP(GL_BYTE), MAP(GL_UNSIGNED_BYTE), MAP(GL_SHORT), MAP(GL_UNSIGNED_SHORT),
 189   MAP(GL_INT), MAP(GL_UNSIGNED_INT), MAP(GL_FLOAT), MAP(GL_DOUBLE),
 190   MAP(GL_2_BYTES), MAP(GL_3_BYTES), MAP(GL_4_BYTES),
 191   // rest 1.2 only
 192   MAP(GL_UNSIGNED_BYTE_3_3_2), MAP(GL_UNSIGNED_BYTE_2_3_3_REV),
 193   MAP(GL_UNSIGNED_SHORT_5_6_5), MAP(GL_UNSIGNED_SHORT_5_6_5_REV),
 194   MAP(GL_UNSIGNED_SHORT_4_4_4_4), MAP(GL_UNSIGNED_SHORT_4_4_4_4_REV),
 195   MAP(GL_UNSIGNED_SHORT_5_5_5_1), MAP(GL_UNSIGNED_SHORT_1_5_5_5_REV),
 196   MAP(GL_UNSIGNED_INT_8_8_8_8), MAP(GL_UNSIGNED_INT_8_8_8_8_REV),
 197   MAP(GL_UNSIGNED_INT_10_10_10_2), MAP(GL_UNSIGNED_INT_2_10_10_10_REV),
 198   {0, 0}
 199 };
 200 #undef MAP
 201
 202 /**
 203  * \brief return the name of an OpenGL constant
 204  * \param value the constant
 205  * \return name of the constant or "Unknown format!"
 206  * \ingroup glgeneral
 207  */
 208 const char *glValName(GLint value)
 209 {
 210   int i = 0;
 211
 212   while (gl_name_map[i].name) {
 213     if (gl_name_map[i].value == value)
 214       return gl_name_map[i].name;
 215     i++;
 216   }
 217   return "Unknown format!";
 218 }
 219
 220 //! always return this format as internal texture format in glFindFormat
 221 #define TEXTUREFORMAT_ALWAYS GL_RGB8
 222 #undef TEXTUREFORMAT_ALWAYS
 223
 224 /**
 225  * \brief find the OpenGL settings coresponding to format.
 226  *
 227  * All parameters may be NULL.
 228  * \param fmt MPlayer format to analyze.
 229  * \param bpp [OUT] bits per pixel of that format.
 230  * \param gl_texfmt [OUT] internal texture format that fits the
 231  * image format, not necessarily the best for performance.
 232  * \param gl_format [OUT] OpenGL format for this image format.
 233  * \param gl_type [OUT] OpenGL type for this image format.
 234  * \return 1 if format is supported by OpenGL, 0 if not.
 235  * \ingroup gltexture
 236  */
 237 int glFindFormat(uint32_t fmt, int *bpp, GLint *gl_texfmt,
 238                   GLenum *gl_format, GLenum *gl_type)
 239 {
 240   int supported = 1;
 241   int dummy1;
 242   GLenum dummy2;
 243   GLint dummy3;
 244   if (!bpp) bpp = &dummy1;
 245   if (!gl_texfmt) gl_texfmt = &dummy3;
 246   if (!gl_format) gl_format = &dummy2;
 247   if (!gl_type) gl_type = &dummy2;
 248
 249   if (mp_get_chroma_shift(fmt, NULL, NULL)) {
 250     // reduce the possible cases a bit
 251     if (IMGFMT_IS_YUVP16_LE(fmt))
 252       fmt = IMGFMT_420P16_LE;
 253     else if (IMGFMT_IS_YUVP16_BE(fmt))
 254       fmt = IMGFMT_420P16_BE;
 255     else
 256       fmt = IMGFMT_YV12;
 257   }
 258
 259   *bpp = IMGFMT_IS_BGR(fmt)?IMGFMT_BGR_DEPTH(fmt):IMGFMT_RGB_DEPTH(fmt);
 260   *gl_texfmt = 3;
 261   switch (fmt) {
 262     case IMGFMT_RGB48NE:
 263       *gl_format = GL_RGB;
 264       *gl_type = GL_UNSIGNED_SHORT;
 265       break;
 266     case IMGFMT_RGB24:
 267       *gl_format = GL_RGB;
 268       *gl_type = GL_UNSIGNED_BYTE;
 269       break;
 270     case IMGFMT_RGBA:
 271       *gl_texfmt = 4;
 272       *gl_format = GL_RGBA;
 273       *gl_type = GL_UNSIGNED_BYTE;
 274       break;
 275     case IMGFMT_420P16:
 276       supported = 0; // no native YUV support
 277       *gl_texfmt = 1;
 278       *bpp = 16;
 279       *gl_format = GL_LUMINANCE;
 280       *gl_type = GL_UNSIGNED_SHORT;
 281       break;
 282     case IMGFMT_YV12:
 283       supported = 0; // no native YV12 support
 284     case IMGFMT_Y800:
 285     case IMGFMT_Y8:
 286       *gl_texfmt = 1;
 287       *bpp = 8;
 288       *gl_format = GL_LUMINANCE;
 289       *gl_type = GL_UNSIGNED_BYTE;
 290       break;
 291     case IMGFMT_UYVY:
 292     case IMGFMT_YUY2:
 293       *gl_texfmt = GL_YCBCR_MESA;
 294       *bpp = 16;
 295       *gl_format = GL_YCBCR_MESA;
 296       *gl_type = fmt == IMGFMT_UYVY ? GL_UNSIGNED_SHORT_8_8 : GL_UNSIGNED_SHORT_8_8_REV;
 297       break;
 298 #if 0
 299     // we do not support palettized formats, although the format the
 300     // swscale produces works
 301     case IMGFMT_RGB8:
 302       gl_format = GL_RGB;
 303       gl_type = GL_UNSIGNED_BYTE_2_3_3_REV;
 304       break;
 305 #endif
 306     case IMGFMT_RGB15:
 307       *gl_format = GL_RGBA;
 308       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 309       break;
 310     case IMGFMT_RGB16:
 311       *gl_format = GL_RGB;
 312       *gl_type = GL_UNSIGNED_SHORT_5_6_5_REV;
 313       break;
 314 #if 0
 315     case IMGFMT_BGR8:
 316       // special case as red and blue have a differen number of bits.
 317       // GL_BGR and GL_UNSIGNED_BYTE_3_3_2 isn't supported at least
 318       // by nVidia drivers, and in addition would give more bits to
 319       // blue than to red, which isn't wanted
 320       gl_format = GL_RGB;
 321       gl_type = GL_UNSIGNED_BYTE_3_3_2;
 322       break;
 323 #endif
 324     case IMGFMT_BGR15:
 325       *gl_format = GL_BGRA;
 326       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 327       break;
 328     case IMGFMT_BGR16:
 329       *gl_format = GL_RGB;
 330       *gl_type = GL_UNSIGNED_SHORT_5_6_5;
 331       break;
 332     case IMGFMT_BGR24:
 333       *gl_format = GL_BGR;
 334       *gl_type = GL_UNSIGNED_BYTE;
 335       break;
 336     case IMGFMT_BGRA:
 337       *gl_texfmt = 4;
 338       *gl_format = GL_BGRA;
 339       *gl_type = GL_UNSIGNED_BYTE;
 340       break;
 341     default:
 342       *gl_texfmt = 4;
 343       *gl_format = GL_RGBA;
 344       *gl_type = GL_UNSIGNED_BYTE;
 345       supported = 0;
 346   }
 347 #ifdef TEXTUREFORMAT_ALWAYS
 348   *gl_texfmt = TEXTUREFORMAT_ALWAYS;
 349 #endif
 350   return supported;
 351 }
 352
 353 static void *setNull(const GLubyte *s) {
 354   return NULL;
 355 }
 356
 357 typedef struct {
 358   void *funcptr;
 359   const char *extstr;
 360   const char *funcnames[7];
 361   void *fallback;
 362 } extfunc_desc_t;
 363
 364 #define DEF_FUNC_DESC(name) {&name, NULL, {"gl"#name, NULL}, gl ##name}
 365 static const extfunc_desc_t extfuncs[] = {
 366   // these aren't extension functions but we query them anyway to allow
 367   // different "backends" with one binary
 368   DEF_FUNC_DESC(Begin),
 369   DEF_FUNC_DESC(End),
 370   DEF_FUNC_DESC(Viewport),
 371   DEF_FUNC_DESC(MatrixMode),
 372   DEF_FUNC_DESC(LoadIdentity),
 373   DEF_FUNC_DESC(Translated),
 374   DEF_FUNC_DESC(Scaled),
 375   DEF_FUNC_DESC(Ortho),
 376   DEF_FUNC_DESC(Frustum),
 377   DEF_FUNC_DESC(PushMatrix),
 378   DEF_FUNC_DESC(PopMatrix),
 379   DEF_FUNC_DESC(Clear),
 380   DEF_FUNC_DESC(GenLists),
 381   DEF_FUNC_DESC(DeleteLists),
 382   DEF_FUNC_DESC(NewList),
 383   DEF_FUNC_DESC(EndList),
 384   DEF_FUNC_DESC(CallList),
 385   DEF_FUNC_DESC(CallLists),
 386   DEF_FUNC_DESC(GenTextures),
 387   DEF_FUNC_DESC(DeleteTextures),
 388   DEF_FUNC_DESC(TexEnvf),
 389   DEF_FUNC_DESC(TexEnvi),
 390   DEF_FUNC_DESC(Color4ub),
 391   DEF_FUNC_DESC(Color3f),
 392   DEF_FUNC_DESC(Color4f),
 393   DEF_FUNC_DESC(ClearColor),
 394   DEF_FUNC_DESC(ClearDepth),
 395   DEF_FUNC_DESC(DepthFunc),
 396   DEF_FUNC_DESC(Enable),
 397   DEF_FUNC_DESC(Disable),
 398   DEF_FUNC_DESC(DrawBuffer),
 399   DEF_FUNC_DESC(DepthMask),
 400   DEF_FUNC_DESC(BlendFunc),
 401   DEF_FUNC_DESC(Flush),
 402   DEF_FUNC_DESC(Finish),
 403   DEF_FUNC_DESC(PixelStorei),
 404   DEF_FUNC_DESC(TexImage1D),
 405   DEF_FUNC_DESC(TexImage2D),
 406   DEF_FUNC_DESC(TexSubImage2D),
 407   DEF_FUNC_DESC(TexParameteri),
 408   DEF_FUNC_DESC(TexParameterf),
 409   DEF_FUNC_DESC(TexParameterfv),
 410   DEF_FUNC_DESC(TexCoord2f),
 411   DEF_FUNC_DESC(Vertex2f),
 412   DEF_FUNC_DESC(Vertex3f),
 413   DEF_FUNC_DESC(Normal3f),
 414   DEF_FUNC_DESC(Lightfv),
 415   DEF_FUNC_DESC(ColorMaterial),
 416   DEF_FUNC_DESC(ShadeModel),
 417   DEF_FUNC_DESC(GetIntegerv),
 418
 419   // here start the real extensions
 420   {&GenBuffers, NULL, {"glGenBuffers", "glGenBuffersARB", NULL}},
 421   {&DeleteBuffers, NULL, {"glDeleteBuffers", "glDeleteBuffersARB", NULL}},
 422   {&BindBuffer, NULL, {"glBindBuffer", "glBindBufferARB", NULL}},
 423   {&MapBuffer, NULL, {"glMapBuffer", "glMapBufferARB", NULL}},
 424   {&UnmapBuffer, NULL, {"glUnmapBuffer", "glUnmapBufferARB", NULL}},
 425   {&BufferData, NULL, {"glBufferData", "glBufferDataARB", NULL}},
 426   {&CombinerParameterfv, "NV_register_combiners", {"glCombinerParameterfv", "glCombinerParameterfvNV", NULL}},
 427   {&CombinerParameteri, "NV_register_combiners", {"glCombinerParameteri", "glCombinerParameteriNV", NULL}},
 428   {&CombinerInput, "NV_register_combiners", {"glCombinerInput", "glCombinerInputNV", NULL}},
 429   {&CombinerOutput, "NV_register_combiners", {"glCombinerOutput", "glCombinerOutputNV", NULL}},
 430   {&BeginFragmentShader, "ATI_fragment_shader", {"glBeginFragmentShaderATI", NULL}},
 431   {&EndFragmentShader, "ATI_fragment_shader", {"glEndFragmentShaderATI", NULL}},
 432   {&SampleMap, "ATI_fragment_shader", {"glSampleMapATI", NULL}},
 433   {&ColorFragmentOp2, "ATI_fragment_shader", {"glColorFragmentOp2ATI", NULL}},
 434   {&ColorFragmentOp3, "ATI_fragment_shader", {"glColorFragmentOp3ATI", NULL}},
 435   {&SetFragmentShaderConstant, "ATI_fragment_shader", {"glSetFragmentShaderConstantATI", NULL}},
 436   {&ActiveTexture, NULL, {"glActiveTexture", "glActiveTextureARB", NULL}},
 437   {&BindTexture, NULL, {"glBindTexture", "glBindTextureARB", "glBindTextureEXT", NULL}},
 438   {&MultiTexCoord2f, NULL, {"glMultiTexCoord2f", "glMultiTexCoord2fARB", NULL}},
 439   {&GenPrograms, "_program", {"glGenProgramsARB", NULL}},
 440   {&DeletePrograms, "_program", {"glDeleteProgramsARB", NULL}},
 441   {&BindProgram, "_program", {"glBindProgramARB", NULL}},
 442   {&ProgramString, "_program", {"glProgramStringARB", NULL}},
 443   {&GetProgramiv, "_program", {"glGetProgramivARB", NULL}},
 444   {&ProgramEnvParameter4f, "_program", {"glProgramEnvParameter4fARB", NULL}},
 445   {&SwapInterval, "_swap_control", {"glXSwapIntervalSGI", "glXSwapInterval", "wglSwapIntervalSGI", "wglSwapInterval", "wglSwapIntervalEXT", NULL}},
 446   {&TexImage3D, NULL, {"glTexImage3D", NULL}},
 447   {&AllocateMemoryMESA, "GLX_MESA_allocate_memory", {"glXAllocateMemoryMESA", NULL}},
 448   {&FreeMemoryMESA, "GLX_MESA_allocate_memory", {"glXFreeMemoryMESA", NULL}},
 449   {NULL}
 450 };
 451
 452 /**
 453  * \brief find the function pointers of some useful OpenGL extensions
 454  * \param getProcAddress function to resolve function names, may be NULL
 455  * \param ext2 an extra extension string
 456  */
 457 static void getFunctions(void *(*getProcAddress)(const GLubyte *),
 458                          const char *ext2) {
 459   const extfunc_desc_t *dsc;
 460   const char *extensions;
 461   char *allexts;
 462
 463   if (!getProcAddress)
 464     getProcAddress = setNull;
 465
 466   // special case, we need glGetString before starting to find the other functions
 467   GetString = getProcAddress("glGetString");
 468   if (!GetString)
 469       GetString = glGetString;
 470
 471   extensions = (const char *)GetString(GL_EXTENSIONS);
 472   if (!extensions) extensions = "";
 473   if (!ext2) ext2 = "";
 474   allexts = malloc(strlen(extensions) + strlen(ext2) + 2);
 475   strcpy(allexts, extensions);
 476   strcat(allexts, " ");
 477   strcat(allexts, ext2);
 478   mp_msg(MSGT_VO, MSGL_DBG2, "OpenGL extensions string:\n%s\n", allexts);
 479   for (dsc = extfuncs; dsc->funcptr; dsc++) {
 480     void *ptr = NULL;
 481     int i;
 482     if (!dsc->extstr || strstr(allexts, dsc->extstr)) {
 483       for (i = 0; !ptr && dsc->funcnames[i]; i++)
 484         ptr = getProcAddress((const GLubyte *)dsc->funcnames[i]);
 485     }
 486     if (!ptr)
 487         ptr = dsc->fallback;
 488     *(void **)dsc->funcptr = ptr;
 489   }
 490   if (strstr(allexts, "_texture_float"))
 491     hqtexfmt = GL_RGB32F;
 492   else if (strstr(allexts, "NV_float_buffer"))
 493     hqtexfmt = GL_FLOAT_RGB32_NV;
 494   else
 495     hqtexfmt = GL_RGB16;
 496   free(allexts);
 497 }
 498
 499 /**
 500  * \brief create a texture and set some defaults
 501  * \param target texture taget, usually GL_TEXTURE_2D
 502  * \param fmt internal texture format
 503  * \param format texture host data format
 504  * \param type texture host data type
 505  * \param filter filter used for scaling, e.g. GL_LINEAR
 506  * \param w texture width
 507  * \param h texture height
 508  * \param val luminance value to fill texture with
 509  * \ingroup gltexture
 510  */
 511 void glCreateClearTex(GLenum target, GLenum fmt, GLenum format, GLenum type, GLint filter,
 512                       int w, int h, unsigned char val) {
 513   GLfloat fval = (GLfloat)val / 255.0;
 514   GLfloat border[4] = {fval, fval, fval, fval};
 515   int stride = w * glFmt2bpp(format, type);
 516   char *init;
 517   if (!stride) return;
 518   init = malloc(stride * h);
 519   memset(init, val, stride * h);
 520   glAdjustAlignment(stride);
 521   PixelStorei(GL_UNPACK_ROW_LENGTH, w);
 522   TexImage2D(target, 0, fmt, w, h, 0, format, type, init);
 523   TexParameterf(target, GL_TEXTURE_PRIORITY, 1.0);
 524   TexParameteri(target, GL_TEXTURE_MIN_FILTER, filter);
 525   TexParameteri(target, GL_TEXTURE_MAG_FILTER, filter);
 526   TexParameteri(target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 527   TexParameteri(target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 528   // Border texels should not be used with CLAMP_TO_EDGE
 529   // We set a sane default anyway.
 530   TexParameterfv(target, GL_TEXTURE_BORDER_COLOR, border);
 531   free(init);
 532 }
 533
 534 /**
 535  * \brief skips whitespace and comments
 536  * \param f file to read from
 537  */
 538 static void ppm_skip(FILE *f) {
 539   int c, comment = 0;
 540   do {
 541     c = fgetc(f);
 542     if (c == '#')
 543       comment = 1;
 544     if (c == '\n')
 545       comment = 0;
 546   } while (c != EOF && (isspace(c) || comment));
 547   if (c != EOF)
 548     ungetc(c, f);
 549 }
 550
 551 #define MAXDIM (16 * 1024)
 552
 553 /**
 554  * \brief creates a texture from a PPM file
 555  * \param target texture taget, usually GL_TEXTURE_2D
 556  * \param fmt internal texture format, 0 for default
 557  * \param filter filter used for scaling, e.g. GL_LINEAR
 558  * \param f file to read PPM from
 559  * \param width [out] width of texture
 560  * \param height [out] height of texture
 561  * \param maxval [out] maxval value from PPM file
 562  * \return 0 on error, 1 otherwise
 563  * \ingroup gltexture
 564  */
 565 int glCreatePPMTex(GLenum target, GLenum fmt, GLint filter,
 566                    FILE *f, int *width, int *height, int *maxval) {
 567   unsigned w, h, m, val, bpp;
 568   char *data;
 569   GLenum type;
 570   ppm_skip(f);
 571   if (fgetc(f) != 'P' || fgetc(f) != '6')
 572     return 0;
 573   ppm_skip(f);
 574   if (fscanf(f, "%u", &w) != 1)
 575     return 0;
 576   ppm_skip(f);
 577   if (fscanf(f, "%u", &h) != 1)
 578     return 0;
 579   ppm_skip(f);
 580   if (fscanf(f, "%u", &m) != 1)
 581     return 0;
 582   val = fgetc(f);
 583   if (!isspace(val))
 584     return 0;
 585   if (w > MAXDIM || h > MAXDIM)
 586     return 0;
 587   bpp = (m > 255) ? 6 : 3;
 588   data = malloc(w * h * bpp);
 589   if (fread(data, w * bpp, h, f) != h)
 590     return 0;
 591   if (!fmt) {
 592     fmt = (m > 255) ? hqtexfmt : 3;
 593     if (fmt == GL_FLOAT_RGB32_NV && target != GL_TEXTURE_RECTANGLE)
 594       fmt = GL_RGB16;
 595   }
 596   type = m > 255 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 597   glCreateClearTex(target, fmt, GL_RGB, type, filter, w, h, 0);
 598   glUploadTex(target, GL_RGB, type,
 599               data, w * bpp, 0, 0, w, h, 0);
 600   free(data);
 601   if (width) *width = w;
 602   if (height) *height = h;
 603   if (maxval) *maxval = m;
 604   return 1;
 605 }
 606
 607 /**
 608  * \brief return the number of bytes per pixel for the given format
 609  * \param format OpenGL format
 610  * \param type OpenGL type
 611  * \return bytes per pixel
 612  * \ingroup glgeneral
 613  *
 614  * Does not handle all possible variants, just those used by MPlayer
 615  */
 616 int glFmt2bpp(GLenum format, GLenum type) {
 617   int component_size = 0;
 618   switch (type) {
 619     case GL_UNSIGNED_BYTE_3_3_2:
 620     case GL_UNSIGNED_BYTE_2_3_3_REV:
 621       return 1;
 622     case GL_UNSIGNED_SHORT_5_5_5_1:
 623     case GL_UNSIGNED_SHORT_1_5_5_5_REV:
 624     case GL_UNSIGNED_SHORT_5_6_5:
 625     case GL_UNSIGNED_SHORT_5_6_5_REV:
 626       return 2;
 627     case GL_UNSIGNED_BYTE:
 628       component_size = 1;
 629       break;
 630     case GL_UNSIGNED_SHORT:
 631       component_size = 2;
 632       break;
 633   }
 634   switch (format) {
 635     case GL_LUMINANCE:
 636     case GL_ALPHA:
 637       return component_size;
 638     case GL_YCBCR_MESA:
 639       return 2;
 640     case GL_RGB:
 641     case GL_BGR:
 642       return 3 * component_size;
 643     case GL_RGBA:
 644     case GL_BGRA:
 645       return 4 * component_size;
 646   }
 647   return 0; // unknown
 648 }
 649
 650 /**
 651  * \brief upload a texture, handling things like stride and slices
 652  * \param target texture target, usually GL_TEXTURE_2D
 653  * \param format OpenGL format of data
 654  * \param type OpenGL type of data
 655  * \param dataptr data to upload
 656  * \param stride data stride
 657  * \param x x offset in texture
 658  * \param y y offset in texture
 659  * \param w width of the texture part to upload
 660  * \param h height of the texture part to upload
 661  * \param slice height of an upload slice, 0 for all at once
 662  * \ingroup gltexture
 663  */
 664 void glUploadTex(GLenum target, GLenum format, GLenum type,
 665                  const void *dataptr, int stride,
 666                  int x, int y, int w, int h, int slice) {
 667   const uint8_t *data = dataptr;
 668   int y_max = y + h;
 669   if (w <= 0 || h <= 0) return;
 670   if (slice <= 0)
 671     slice = h;
 672   if (stride < 0) {
 673     data += (h - 1) * stride;
 674     stride = -stride;
 675   }
 676   // this is not always correct, but should work for MPlayer
 677   glAdjustAlignment(stride);
 678   PixelStorei(GL_UNPACK_ROW_LENGTH, stride / glFmt2bpp(format, type));
 679   for (; y + slice <= y_max; y += slice) {
 680     TexSubImage2D(target, 0, x, y, w, slice, format, type, data);
 681     data += stride * slice;
 682   }
 683   if (y < y_max)
 684     TexSubImage2D(target, 0, x, y, w, y_max - y, format, type, data);
 685 }
 686
 687 static void fillUVcoeff(GLfloat *ucoef, GLfloat *vcoef,
 688                         float uvcos, float uvsin) {
 689   int i;
 690   ucoef[0] = 0 * uvcos + 1.403 * uvsin;
 691   vcoef[0] = 0 * uvsin + 1.403 * uvcos;
 692   ucoef[1] = -0.344 * uvcos + -0.714 * uvsin;
 693   vcoef[1] = -0.344 * uvsin + -0.714 * uvcos;
 694   ucoef[2] = 1.770 * uvcos + 0 * uvsin;
 695   vcoef[2] = 1.770 * uvsin + 0 * uvcos;
 696   ucoef[3] = 0;
 697   vcoef[3] = 0;
 698   // Coefficients (probably) must be in [0, 1] range, whereas they originally
 699   // are in [-2, 2] range, so here comes the trick:
 700   // First put them in the [-0.5, 0.5] range, then add 0.5.
 701   // This can be undone with the HALF_BIAS and SCALE_BY_FOUR arguments
 702   // for CombinerInput and CombinerOutput (or the respective ATI variants)
 703   for (i = 0; i < 4; i++) {
 704     ucoef[i] = ucoef[i] * 0.25 + 0.5;
 705     vcoef[i] = vcoef[i] * 0.25 + 0.5;
 706   }
 707 }
 708
 709 /**
 710  * \brief Setup register combiners for YUV to RGB conversion.
 711  * \param uvcos used for saturation and hue adjustment
 712  * \param uvsin used for saturation and hue adjustment
 713  */
 714 static void glSetupYUVCombiners(float uvcos, float uvsin) {
 715   GLfloat ucoef[4];
 716   GLfloat vcoef[4];
 717   GLint i;
 718   if (!CombinerInput || !CombinerOutput ||
 719       !CombinerParameterfv || !CombinerParameteri) {
 720     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner functions missing!\n");
 721     return;
 722   }
 723   GetIntegerv(GL_MAX_GENERAL_COMBINERS_NV, &i);
 724   if (i < 2)
 725     mp_msg(MSGT_VO, MSGL_ERR,
 726            "[gl] 2 general combiners needed for YUV combiner support (found %i)\n", i);
 727   GetIntegerv(GL_MAX_TEXTURE_UNITS, &i);
 728   if (i < 3)
 729     mp_msg(MSGT_VO, MSGL_ERR,
 730            "[gl] 3 texture units needed for YUV combiner support (found %i)\n", i);
 731   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 732   CombinerParameterfv(GL_CONSTANT_COLOR0_NV, ucoef);
 733   CombinerParameterfv(GL_CONSTANT_COLOR1_NV, vcoef);
 734
 735   // UV first, like this green component cannot overflow
 736   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_A_NV,
 737                 GL_TEXTURE1, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 738   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_B_NV,
 739                 GL_CONSTANT_COLOR0_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 740   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_C_NV,
 741                 GL_TEXTURE2, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 742   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_D_NV,
 743                 GL_CONSTANT_COLOR1_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 744   CombinerOutput(GL_COMBINER0_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 745                  GL_SPARE0_NV, GL_SCALE_BY_FOUR_NV, GL_NONE, GL_FALSE,
 746                  GL_FALSE, GL_FALSE);
 747
 748   // stage 2
 749   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_A_NV, GL_SPARE0_NV,
 750                 GL_SIGNED_IDENTITY_NV, GL_RGB);
 751   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_B_NV, GL_ZERO,
 752                  GL_UNSIGNED_INVERT_NV, GL_RGB);
 753   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_C_NV,
 754                 GL_TEXTURE0, GL_SIGNED_IDENTITY_NV, GL_RGB);
 755   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_D_NV, GL_ZERO,
 756                 GL_UNSIGNED_INVERT_NV, GL_RGB);
 757   CombinerOutput(GL_COMBINER1_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 758                  GL_SPARE0_NV, GL_NONE, GL_NONE, GL_FALSE,
 759                  GL_FALSE, GL_FALSE);
 760
 761   // leave final combiner stage in default mode
 762   CombinerParameteri(GL_NUM_GENERAL_COMBINERS_NV, 2);
 763 }
 764
 765 /**
 766  * \brief Setup ATI version of register combiners for YUV to RGB conversion.
 767  * \param uvcos used for saturation and hue adjustment
 768  * \param uvsin used for saturation and hue adjustment
 769  *
 770  * ATI called this fragment shader, but the name is confusing in the
 771  * light of a very different OpenGL 2.0 extension with the same name
 772  */
 773 static void glSetupYUVCombinersATI(float uvcos, float uvsin) {
 774   GLfloat ucoef[4];
 775   GLfloat vcoef[4];
 776   GLint i;
 777   if (!BeginFragmentShader || !EndFragmentShader ||
 778       !SetFragmentShaderConstant || !SampleMap ||
 779       !ColorFragmentOp2 || !ColorFragmentOp3) {
 780     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner (ATI) functions missing!\n");
 781     return;
 782   }
 783   GetIntegerv(GL_NUM_FRAGMENT_REGISTERS_ATI, &i);
 784   if (i < 3)
 785     mp_msg(MSGT_VO, MSGL_ERR,
 786            "[gl] 3 registers needed for YUV combiner (ATI) support (found %i)\n", i);
 787   GetIntegerv (GL_MAX_TEXTURE_UNITS, &i);
 788   if (i < 3)
 789     mp_msg(MSGT_VO, MSGL_ERR,
 790            "[gl] 3 texture units needed for YUV combiner (ATI) support (found %i)\n", i);
 791   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 792   BeginFragmentShader();
 793   SetFragmentShaderConstant(GL_CON_0_ATI, ucoef);
 794   SetFragmentShaderConstant(GL_CON_1_ATI, vcoef);
 795   SampleMap(GL_REG_0_ATI, GL_TEXTURE0, GL_SWIZZLE_STR_ATI);
 796   SampleMap(GL_REG_1_ATI, GL_TEXTURE1, GL_SWIZZLE_STR_ATI);
 797   SampleMap(GL_REG_2_ATI, GL_TEXTURE2, GL_SWIZZLE_STR_ATI);
 798   // UV first, like this green component cannot overflow
 799   ColorFragmentOp2(GL_MUL_ATI, GL_REG_1_ATI, GL_NONE, GL_NONE,
 800                    GL_REG_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 801                    GL_CON_0_ATI, GL_NONE, GL_BIAS_BIT_ATI);
 802   ColorFragmentOp3(GL_MAD_ATI, GL_REG_2_ATI, GL_NONE, GL_4X_BIT_ATI,
 803                    GL_REG_2_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 804                    GL_CON_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 805                    GL_REG_1_ATI, GL_NONE, GL_NONE);
 806   ColorFragmentOp2(GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 807                    GL_REG_0_ATI, GL_NONE, GL_NONE,
 808                    GL_REG_2_ATI, GL_NONE, GL_NONE);
 809   EndFragmentShader();
 810 }
 811
 812 /**
 813  * \brief helper function for gen_spline_lookup_tex
 814  * \param x subpixel-position ((0,1) range) to calculate weights for
 815  * \param dst where to store transformed weights, must provide space for 4 GLfloats
 816  *
 817  * calculates the weights and stores them after appropriate transformation
 818  * for the scaler fragment program.
 819  */
 820 static void store_weights(float x, GLfloat *dst) {
 821   float w0 = (((-1 * x + 3) * x - 3) * x + 1) / 6;
 822   float w1 = ((( 3 * x - 6) * x + 0) * x + 4) / 6;
 823   float w2 = (((-3 * x + 3) * x + 3) * x + 1) / 6;
 824   float w3 = ((( 1 * x + 0) * x + 0) * x + 0) / 6;
 825   *dst++ = 1 + x - w1 / (w0 + w1);
 826   *dst++ = 1 - x + w3 / (w2 + w3);
 827   *dst++ = w0 + w1;
 828   *dst++ = 0;
 829 }
 830
 831 //! to avoid artefacts this should be rather large
 832 #define LOOKUP_BSPLINE_RES (2 * 1024)
 833 /**
 834  * \brief creates the 1D lookup texture needed for fast higher-order filtering
 835  * \param unit texture unit to attach texture to
 836  */
 837 static void gen_spline_lookup_tex(GLenum unit) {
 838   GLfloat *tex = calloc(4 * LOOKUP_BSPLINE_RES, sizeof(*tex));
 839   GLfloat *tp = tex;
 840   int i;
 841   for (i = 0; i < LOOKUP_BSPLINE_RES; i++) {
 842     float x = (float)(i + 0.5) / LOOKUP_BSPLINE_RES;
 843     store_weights(x, tp);
 844     tp += 4;
 845   }
 846   store_weights(0, tex);
 847   store_weights(1, &tex[4 * (LOOKUP_BSPLINE_RES - 1)]);
 848   ActiveTexture(unit);
 849   TexImage1D(GL_TEXTURE_1D, 0, GL_RGBA16, LOOKUP_BSPLINE_RES, 0, GL_RGBA, GL_FLOAT, tex);
 850   TexParameterf(GL_TEXTURE_1D, GL_TEXTURE_PRIORITY, 1.0);
 851   TexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 852   TexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 853   TexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_REPEAT);
 854   ActiveTexture(GL_TEXTURE0);
 855   free(tex);
 856 }
 857
 858 static const char *bilin_filt_template =
 859   "TEX yuv.%c, fragment.texcoord[%c], texture[%c], %s;";
 860
 861 #define BICUB_FILT_MAIN(textype) \
 862   /* first y-interpolation */ \
 863   "ADD coord, fragment.texcoord[%c].xyxy, cdelta.xyxw;" \
 864   "ADD coord2, fragment.texcoord[%c].xyxy, cdelta.zyzw;" \
 865   "TEX a.r, coord.xyxy, texture[%c], "textype";" \
 866   "TEX a.g, coord.zwzw, texture[%c], "textype";" \
 867   /* second y-interpolation */ \
 868   "TEX b.r, coord2.xyxy, texture[%c], "textype";" \
 869   "TEX b.g, coord2.zwzw, texture[%c], "textype";" \
 870   "LRP a.b, parmy.b, a.rrrr, a.gggg;" \
 871   "LRP a.a, parmy.b, b.rrrr, b.gggg;" \
 872   /* x-interpolation */ \
 873   "LRP yuv.%c, parmx.b, a.bbbb, a.aaaa;"
 874
 875 static const char *bicub_filt_template_2D =
 876   "MAD coord.xy, fragment.texcoord[%c], {%e, %e}, {0.5, 0.5};"
 877   "TEX parmx, coord.x, texture[%c], 1D;"
 878   "MUL cdelta.xz, parmx.rrgg, {-%e, 0, %e, 0};"
 879   "TEX parmy, coord.y, texture[%c], 1D;"
 880   "MUL cdelta.yw, parmy.rrgg, {0, -%e, 0, %e};"
 881   BICUB_FILT_MAIN("2D");
 882
 883 static const char *bicub_filt_template_RECT =
 884   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};"
 885   "TEX parmx, coord.x, texture[%c], 1D;"
 886   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};"
 887   "TEX parmy, coord.y, texture[%c], 1D;"
 888   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};"
 889   BICUB_FILT_MAIN("RECT");
 890
 891 #define CALCWEIGHTS(t, s) \
 892   "MAD "t", {-0.5, 0.1666, 0.3333, -0.3333}, "s", {1, 0, -0.5, 0.5};" \
 893   "MAD "t", "t", "s", {0, 0, -0.5, 0.5};" \
 894   "MAD "t", "t", "s", {-0.6666, 0, 0.8333, 0.1666};" \
 895   "RCP a.x, "t".z;" \
 896   "RCP a.y, "t".w;" \
 897   "MAD "t".xy, "t".xyxy, a.xyxy, {1, 1, 0, 0};" \
 898   "ADD "t".x, "t".xxxx, "s";" \
 899   "SUB "t".y, "t".yyyy, "s";"
 900
 901 static const char *bicub_notex_filt_template_2D =
 902   "MAD coord.xy, fragment.texcoord[%c], {%e, %e}, {0.5, 0.5};"
 903   "FRC coord.xy, coord.xyxy;"
 904   CALCWEIGHTS("parmx", "coord.xxxx")
 905   "MUL cdelta.xz, parmx.rrgg, {-%e, 0, %e, 0};"
 906   CALCWEIGHTS("parmy", "coord.yyyy")
 907   "MUL cdelta.yw, parmy.rrgg, {0, -%e, 0, %e};"
 908   BICUB_FILT_MAIN("2D");
 909
 910 static const char *bicub_notex_filt_template_RECT =
 911   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};"
 912   "FRC coord.xy, coord.xyxy;"
 913   CALCWEIGHTS("parmx", "coord.xxxx")
 914   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};"
 915   CALCWEIGHTS("parmy", "coord.yyyy")
 916   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};"
 917   BICUB_FILT_MAIN("RECT");
 918
 919 #define BICUB_X_FILT_MAIN(textype) \
 920   "ADD coord.xy, fragment.texcoord[%c].xyxy, cdelta.xyxy;" \
 921   "ADD coord2.xy, fragment.texcoord[%c].xyxy, cdelta.zyzy;" \
 922   "TEX a.r, coord, texture[%c], "textype";" \
 923   "TEX b.r, coord2, texture[%c], "textype";" \
 924   /* x-interpolation */ \
 925   "LRP yuv.%c, parmx.b, a.rrrr, b.rrrr;"
 926
 927 static const char *bicub_x_filt_template_2D =
 928   "MAD coord.x, fragment.texcoord[%c], {%e}, {0.5};"
 929   "TEX parmx, coord, texture[%c], 1D;"
 930   "MUL cdelta.xyz, parmx.rrgg, {-%e, 0, %e};"
 931   BICUB_X_FILT_MAIN("2D");
 932
 933 static const char *bicub_x_filt_template_RECT =
 934   "ADD coord.x, fragment.texcoord[%c], {0.5};"
 935   "TEX parmx, coord, texture[%c], 1D;"
 936   "MUL cdelta.xyz, parmx.rrgg, {-1, 0, 1};"
 937   BICUB_X_FILT_MAIN("RECT");
 938
 939 static const char *unsharp_filt_template =
 940   "PARAM dcoord%c = {%e, %e, %e, %e};"
 941   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;"
 942   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;"
 943   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;"
 944   "TEX b.r, coord.xyxy, texture[%c], %s;"
 945   "TEX b.g, coord.zwzw, texture[%c], %s;"
 946   "ADD b.r, b.r, b.g;"
 947   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 948   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 949   "DP3 b, b, {0.25, 0.25, 0.25};"
 950   "SUB b.r, a.r, b.r;"
 951   "MAD yuv.%c, b.r, {%e}, a.r;";
 952
 953 static const char *unsharp_filt_template2 =
 954   "PARAM dcoord%c = {%e, %e, %e, %e};"
 955   "PARAM dcoord2%c = {%e, 0, 0, %e};"
 956   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;"
 957   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;"
 958   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;"
 959   "TEX b.r, coord.xyxy, texture[%c], %s;"
 960   "TEX b.g, coord.zwzw, texture[%c], %s;"
 961   "ADD b.r, b.r, b.g;"
 962   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 963   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 964   "ADD b.r, b.r, b.b;"
 965   "ADD b.a, b.r, b.g;"
 966   "ADD coord, fragment.texcoord[%c].xyxy, dcoord2%c;"
 967   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord2%c;"
 968   "TEX b.r, coord.xyxy, texture[%c], %s;"
 969   "TEX b.g, coord.zwzw, texture[%c], %s;"
 970   "ADD b.r, b.r, b.g;"
 971   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 972   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 973   "DP4 b.r, b, {-0.1171875, -0.1171875, -0.1171875, -0.09765625};"
 974   "MAD b.r, a.r, {0.859375}, b.r;"
 975   "MAD yuv.%c, b.r, {%e}, a.r;";
 976
 977 static const char *yuv_prog_template =
 978   "PARAM ycoef = {%e, %e, %e};"
 979   "PARAM ucoef = {%e, %e, %e};"
 980   "PARAM vcoef = {%e, %e, %e};"
 981   "PARAM offsets = {%e, %e, %e};"
 982   "TEMP res;"
 983   "MAD res.rgb, yuv.rrrr, ycoef, offsets;"
 984   "MAD res.rgb, yuv.gggg, ucoef, res;"
 985   "MAD result.color.rgb, yuv.bbbb, vcoef, res;"
 986   "END";
 987
 988 static const char *yuv_pow_prog_template =
 989   "PARAM ycoef = {%e, %e, %e};"
 990   "PARAM ucoef = {%e, %e, %e};"
 991   "PARAM vcoef = {%e, %e, %e};"
 992   "PARAM offsets = {%e, %e, %e};"
 993   "PARAM gamma = {%e, %e, %e};"
 994   "TEMP res;"
 995   "MAD res.rgb, yuv.rrrr, ycoef, offsets;"
 996   "MAD res.rgb, yuv.gggg, ucoef, res;"
 997   "MAD_SAT res.rgb, yuv.bbbb, vcoef, res;"
 998   "POW result.color.r, res.r, gamma.r;"
 999   "POW result.color.g, res.g, gamma.g;"
1000   "POW result.color.b, res.b, gamma.b;"
1001   "END";
1002
1003 static const char *yuv_lookup_prog_template =
1004   "PARAM ycoef = {%e, %e, %e, 0};"
1005   "PARAM ucoef = {%e, %e, %e, 0};"
1006   "PARAM vcoef = {%e, %e, %e, 0};"
1007   "PARAM offsets = {%e, %e, %e, 0.125};"
1008   "TEMP res;"
1009   "MAD res, yuv.rrrr, ycoef, offsets;"
1010   "MAD res.rgb, yuv.gggg, ucoef, res;"
1011   "MAD res.rgb, yuv.bbbb, vcoef, res;"
1012   "TEX result.color.r, res.raaa, texture[%c], 2D;"
1013   "ADD res.a, res.a, 0.25;"
1014   "TEX result.color.g, res.gaaa, texture[%c], 2D;"
1015   "ADD res.a, res.a, 0.25;"
1016   "TEX result.color.b, res.baaa, texture[%c], 2D;"
1017   "END";
1018
1019 static const char *yuv_lookup3d_prog_template =
1020   "TEX result.color, yuv, texture[%c], 3D;"
1021   "END";
1022
1023 /**
1024  * \brief creates and initializes helper textures needed for scaling texture read
1025  * \param scaler scaler type to create texture for
1026  * \param texu contains next free texture unit number
1027  * \param texs texture unit ids for the scaler are stored in this array
1028  */
1029 static void create_scaler_textures(int scaler, int *texu, char *texs) {
1030   switch (scaler) {
1031     case YUV_SCALER_BILIN:
1032     case YUV_SCALER_BICUB_NOTEX:
1033     case YUV_SCALER_UNSHARP:
1034     case YUV_SCALER_UNSHARP2:
1035       break;
1036     case YUV_SCALER_BICUB:
1037     case YUV_SCALER_BICUB_X:
1038       texs[0] = (*texu)++;
1039       gen_spline_lookup_tex(GL_TEXTURE0 + texs[0]);
1040       texs[0] += '0';
1041       break;
1042     default:
1043       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown scaler type %i\n", scaler);
1044   }
1045 }
1046
1047 //! resolution of texture for gamma lookup table
1048 #define LOOKUP_RES 512
1049 //! resolution for 3D yuv->rgb conversion lookup table
1050 #define LOOKUP_3DRES 32
1051 /**
1052  * \brief creates and initializes helper textures needed for yuv conversion
1053  * \param params struct containing parameters like brightness, gamma, ...
1054  * \param texu contains next free texture unit number
1055  * \param texs texture unit ids for the conversion are stored in this array
1056  */
1057 static void create_conv_textures(gl_conversion_params_t *params, int *texu, char *texs) {
1058   unsigned char *lookup_data = NULL;
1059   int conv = YUV_CONVERSION(params->type);
1060   switch (conv) {
1061     case YUV_CONVERSION_FRAGMENT:
1062     case YUV_CONVERSION_FRAGMENT_POW:
1063       break;
1064     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1065       texs[0] = (*texu)++;
1066       ActiveTexture(GL_TEXTURE0 + texs[0]);
1067       lookup_data = malloc(4 * LOOKUP_RES);
1068       mp_gen_gamma_map(lookup_data, LOOKUP_RES, params->csp_params.rgamma);
1069       mp_gen_gamma_map(&lookup_data[LOOKUP_RES], LOOKUP_RES, params->csp_params.ggamma);
1070       mp_gen_gamma_map(&lookup_data[2 * LOOKUP_RES], LOOKUP_RES, params->csp_params.bgamma);
1071       glCreateClearTex(GL_TEXTURE_2D, GL_LUMINANCE8, GL_LUMINANCE, GL_UNSIGNED_BYTE, GL_LINEAR,
1072                        LOOKUP_RES, 4, 0);
1073       glUploadTex(GL_TEXTURE_2D, GL_LUMINANCE, GL_UNSIGNED_BYTE, lookup_data,
1074                   LOOKUP_RES, 0, 0, LOOKUP_RES, 4, 0);
1075       ActiveTexture(GL_TEXTURE0);
1076       texs[0] += '0';
1077       break;
1078     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1079       {
1080         int sz = LOOKUP_3DRES + 2; // texture size including borders
1081         if (!TexImage3D) {
1082           mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing 3D texture function!\n");
1083           break;
1084         }
1085         texs[0] = (*texu)++;
1086         ActiveTexture(GL_TEXTURE0 + texs[0]);
1087         lookup_data = malloc(3 * sz * sz * sz);
1088         mp_gen_yuv2rgb_map(&params->csp_params, lookup_data, LOOKUP_3DRES);
1089         glAdjustAlignment(sz);
1090         PixelStorei(GL_UNPACK_ROW_LENGTH, 0);
1091         TexImage3D(GL_TEXTURE_3D, 0, 3, sz, sz, sz, 1,
1092                    GL_RGB, GL_UNSIGNED_BYTE, lookup_data);
1093         TexParameterf(GL_TEXTURE_3D, GL_TEXTURE_PRIORITY, 1.0);
1094         TexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1095         TexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1096         TexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_S, GL_CLAMP);
1097         TexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_T, GL_CLAMP);
1098         TexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_R, GL_CLAMP);
1099         ActiveTexture(GL_TEXTURE0);
1100         texs[0] += '0';
1101       }
1102       break;
1103     default:
1104       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", conv);
1105   }
1106   if (lookup_data)
1107     free(lookup_data);
1108 }
1109
1110 /**
1111  * \brief adds a scaling texture read at the current fragment program position
1112  * \param scaler type of scaler to insert
1113  * \param prog_pos current position in fragment program
1114  * \param remain how many bytes remain in the buffer given by prog_pos
1115  * \param texs array containing the texture unit identifiers for this scaler
1116  * \param in_tex texture unit the scaler should read from
1117  * \param out_comp component of the yuv variable the scaler stores the result in
1118  * \param rect if rectangular (pixel) adressing should be used for in_tex
1119  * \param texw width of the in_tex texture
1120  * \param texh height of the in_tex texture
1121  * \param strength strength of filter effect if the scaler does some kind of filtering
1122  */
1123 static void add_scaler(int scaler, char **prog_pos, int *remain, char *texs,
1124                     char in_tex, char out_comp, int rect, int texw, int texh,
1125                     double strength) {
1126   const char *ttype = rect ? "RECT" : "2D";
1127   const float ptw = rect ? 1.0 : 1.0 / texw;
1128   const float pth = rect ? 1.0 : 1.0 / texh;
1129   switch (scaler) {
1130     case YUV_SCALER_BILIN:
1131       snprintf(*prog_pos, *remain, bilin_filt_template, out_comp, in_tex,
1132                  in_tex, ttype);
1133       break;
1134     case YUV_SCALER_BICUB:
1135       if (rect)
1136         snprintf(*prog_pos, *remain, bicub_filt_template_RECT,
1137                  in_tex, texs[0], texs[0],
1138                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1139       else
1140         snprintf(*prog_pos, *remain, bicub_filt_template_2D,
1141                  in_tex, (float)texw, (float)texh,
1142                  texs[0], ptw, ptw, texs[0], pth, pth,
1143                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1144       break;
1145     case YUV_SCALER_BICUB_X:
1146       if (rect)
1147         snprintf(*prog_pos, *remain, bicub_x_filt_template_RECT,
1148                  in_tex, texs[0],
1149                  in_tex, in_tex, in_tex, in_tex, out_comp);
1150       else
1151         snprintf(*prog_pos, *remain, bicub_x_filt_template_2D,
1152                  in_tex, (float)texw,
1153                  texs[0], ptw, ptw,
1154                  in_tex, in_tex, in_tex, in_tex, out_comp);
1155       break;
1156     case YUV_SCALER_BICUB_NOTEX:
1157       if (rect)
1158         snprintf(*prog_pos, *remain, bicub_notex_filt_template_RECT,
1159                  in_tex,
1160                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1161       else
1162         snprintf(*prog_pos, *remain, bicub_notex_filt_template_2D,
1163                  in_tex, (float)texw, (float)texh, ptw, ptw, pth, pth,
1164                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1165       break;
1166     case YUV_SCALER_UNSHARP:
1167       snprintf(*prog_pos, *remain, unsharp_filt_template,
1168                out_comp, 0.5 * ptw, 0.5 * pth, 0.5 * ptw, -0.5 * pth,
1169                in_tex, out_comp, in_tex, out_comp, in_tex,
1170                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1171                in_tex, ttype, out_comp, strength);
1172       break;
1173     case YUV_SCALER_UNSHARP2:
1174       snprintf(*prog_pos, *remain, unsharp_filt_template2,
1175                out_comp, 1.2 * ptw, 1.2 * pth, 1.2 * ptw, -1.2 * pth,
1176                out_comp, 1.5 * ptw, 1.5 * pth,
1177                in_tex, out_comp, in_tex, out_comp, in_tex,
1178                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1179                in_tex, ttype, in_tex, out_comp, in_tex, out_comp,
1180                in_tex, ttype, in_tex, ttype, in_tex, ttype,
1181                in_tex, ttype, out_comp, strength);
1182       break;
1183   }
1184   *remain -= strlen(*prog_pos);
1185   *prog_pos += strlen(*prog_pos);
1186 }
1187
1188 static const struct {
1189   const char *name;
1190   GLenum cur;
1191   GLenum max;
1192 } progstats[] = {
1193   {"instructions", 0x88A0, 0x88A1},
1194   {"native instructions", 0x88A2, 0x88A3},
1195   {"temporaries", 0x88A4, 0x88A5},
1196   {"native temporaries", 0x88A6, 0x88A7},
1197   {"parameters", 0x88A8, 0x88A9},
1198   {"native parameters", 0x88AA, 0x88AB},
1199   {"attribs", 0x88AC, 0x88AD},
1200   {"native attribs", 0x88AE, 0x88AF},
1201   {"ALU instructions", 0x8805, 0x880B},
1202   {"TEX instructions", 0x8806, 0x880C},
1203   {"TEX indirections", 0x8807, 0x880D},
1204   {"native ALU instructions", 0x8808, 0x880E},
1205   {"native TEX instructions", 0x8809, 0x880F},
1206   {"native TEX indirections", 0x880A, 0x8810},
1207   {NULL, 0, 0}
1208 };
1209
1210 /**
1211  * \brief load the specified GPU Program
1212  * \param target program target to load into, only GL_FRAGMENT_PROGRAM is tested
1213  * \param prog program string
1214  * \return 1 on success, 0 otherwise
1215  */
1216 int loadGPUProgram(GLenum target, char *prog) {
1217   int i;
1218   GLint cur = 0, max = 0, err = 0;
1219   if (!ProgramString) {
1220     mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing GPU program function\n");
1221     return 0;
1222   }
1223   ProgramString(target, GL_PROGRAM_FORMAT_ASCII, strlen(prog), prog);
1224   GetIntegerv(GL_PROGRAM_ERROR_POSITION, &err);
1225   if (err != -1) {
1226     mp_msg(MSGT_VO, MSGL_ERR,
1227       "[gl] Error compiling fragment program, make sure your card supports\n"
1228       "[gl]   GL_ARB_fragment_program (use glxinfo to check).\n"
1229       "[gl]   Error message:\n  %s at %.10s\n",
1230       GetString(GL_PROGRAM_ERROR_STRING), &prog[err]);
1231     return 0;
1232   }
1233   if (!GetProgramiv || !mp_msg_test(MSGT_VO, MSGL_DBG2))
1234     return 1;
1235   mp_msg(MSGT_VO, MSGL_V, "[gl] Program statistics:\n");
1236   for (i = 0; progstats[i].name; i++) {
1237     GetProgramiv(target, progstats[i].cur, &cur);
1238     GetProgramiv(target, progstats[i].max, &max);
1239     mp_msg(MSGT_VO, MSGL_V, "[gl]   %s: %i/%i\n", progstats[i].name, cur, max);
1240   }
1241   return 1;
1242 }
1243
1244 #define MAX_PROGSZ (1024*1024)
1245
1246 /**
1247  * \brief setup a fragment program that will do YUV->RGB conversion
1248  * \param parms struct containing parameters like conversion and scaler type,
1249  *              brightness, ...
1250  */
1251 static void glSetupYUVFragprog(gl_conversion_params_t *params) {
1252   int type = params->type;
1253   int texw = params->texw;
1254   int texh = params->texh;
1255   int rect = params->target == GL_TEXTURE_RECTANGLE;
1256   static const char prog_hdr[] =
1257     "!!ARBfp1.0\n"
1258     "OPTION ARB_precision_hint_fastest;"
1259     // all scaler variables must go here so they aren't defined
1260     // multiple times when the same scaler is used more than once
1261     "TEMP coord, coord2, cdelta, parmx, parmy, a, b, yuv;";
1262   int prog_remain;
1263   char *yuv_prog, *prog_pos;
1264   int cur_texu = 3;
1265   char lum_scale_texs[1];
1266   char chrom_scale_texs[1];
1267   char conv_texs[1];
1268   GLint i;
1269   // this is the conversion matrix, with y, u, v factors
1270   // for red, green, blue and the constant offsets
1271   float yuv2rgb[3][4];
1272   create_conv_textures(params, &cur_texu, conv_texs);
1273   create_scaler_textures(YUV_LUM_SCALER(type), &cur_texu, lum_scale_texs);
1274   if (YUV_CHROM_SCALER(type) == YUV_LUM_SCALER(type))
1275     memcpy(chrom_scale_texs, lum_scale_texs, sizeof(chrom_scale_texs));
1276   else
1277     create_scaler_textures(YUV_CHROM_SCALER(type), &cur_texu, chrom_scale_texs);
1278   GetIntegerv(GL_MAX_TEXTURE_IMAGE_UNITS, &i);
1279   if (i < cur_texu)
1280     mp_msg(MSGT_VO, MSGL_ERR,
1281            "[gl] %i texture units needed for this type of YUV fragment support (found %i)\n",
1282            cur_texu, i);
1283   if (!ProgramString) {
1284     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] ProgramString function missing!\n");
1285     return;
1286   }
1287   yuv_prog = malloc(MAX_PROGSZ);
1288   strcpy(yuv_prog, prog_hdr);
1289   prog_pos    = yuv_prog + sizeof(prog_hdr) - 1;
1290   prog_remain = MAX_PROGSZ - sizeof(prog_hdr);
1291   add_scaler(YUV_LUM_SCALER(type), &prog_pos, &prog_remain, lum_scale_texs,
1292              '0', 'r', rect, texw, texh, params->filter_strength);
1293   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1294              '1', 'g', rect, params->chrom_texw, params->chrom_texh, params->filter_strength);
1295   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1296              '2', 'b', rect, params->chrom_texw, params->chrom_texh, params->filter_strength);
1297   mp_get_yuv2rgb_coeffs(&params->csp_params, yuv2rgb);
1298   switch (YUV_CONVERSION(type)) {
1299     case YUV_CONVERSION_FRAGMENT:
1300       snprintf(prog_pos, prog_remain, yuv_prog_template,
1301                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1302                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1303                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1304                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C]);
1305       break;
1306     case YUV_CONVERSION_FRAGMENT_POW:
1307       snprintf(prog_pos, prog_remain, yuv_pow_prog_template,
1308                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1309                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1310                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1311                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1312                (float)1.0 / params->csp_params.rgamma, (float)1.0 / params->csp_params.bgamma, (float)1.0 / params->csp_params.bgamma);
1313       break;
1314     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1315       snprintf(prog_pos, prog_remain, yuv_lookup_prog_template,
1316                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1317                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1318                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1319                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1320                conv_texs[0], conv_texs[0], conv_texs[0]);
1321       break;
1322     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1323       snprintf(prog_pos, prog_remain, yuv_lookup3d_prog_template, conv_texs[0]);
1324       break;
1325     default:
1326       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(type));
1327       break;
1328   }
1329   mp_msg(MSGT_VO, MSGL_DBG2, "[gl] generated fragment program:\n%s\n", yuv_prog);
1330   loadGPUProgram(GL_FRAGMENT_PROGRAM, yuv_prog);
1331   free(yuv_prog);
1332 }
1333
1334 /**
1335  * \brief setup YUV->RGB conversion
1336  * \param parms struct containing parameters like conversion and scaler type,
1337  *              brightness, ...
1338  * \ingroup glconversion
1339  */
1340 void glSetupYUVConversion(gl_conversion_params_t *params) {
1341   float uvcos = params->csp_params.saturation * cos(params->csp_params.hue);
1342   float uvsin = params->csp_params.saturation * sin(params->csp_params.hue);
1343   switch (YUV_CONVERSION(params->type)) {
1344     case YUV_CONVERSION_COMBINERS:
1345       glSetupYUVCombiners(uvcos, uvsin);
1346       break;
1347     case YUV_CONVERSION_COMBINERS_ATI:
1348       glSetupYUVCombinersATI(uvcos, uvsin);
1349       break;
1350     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1351     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1352     case YUV_CONVERSION_FRAGMENT:
1353     case YUV_CONVERSION_FRAGMENT_POW:
1354       glSetupYUVFragprog(params);
1355       break;
1356     case YUV_CONVERSION_NONE:
1357       break;
1358     default:
1359       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(params->type));
1360   }
1361 }
1362
1363 /**
1364  * \brief enable the specified YUV conversion
1365  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1366  * \param type type of YUV conversion
1367  * \ingroup glconversion
1368  */
1369 void glEnableYUVConversion(GLenum target, int type) {
1370   switch (YUV_CONVERSION(type)) {
1371     case YUV_CONVERSION_COMBINERS:
1372       ActiveTexture(GL_TEXTURE1);
1373       Enable(target);
1374       ActiveTexture(GL_TEXTURE2);
1375       Enable(target);
1376       ActiveTexture(GL_TEXTURE0);
1377       Enable(GL_REGISTER_COMBINERS_NV);
1378       break;
1379     case YUV_CONVERSION_COMBINERS_ATI:
1380       ActiveTexture(GL_TEXTURE1);
1381       Enable(target);
1382       ActiveTexture(GL_TEXTURE2);
1383       Enable(target);
1384       ActiveTexture(GL_TEXTURE0);
1385       Enable(GL_FRAGMENT_SHADER_ATI);
1386       break;
1387     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1388     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1389     case YUV_CONVERSION_FRAGMENT_POW:
1390     case YUV_CONVERSION_FRAGMENT:
1391     case YUV_CONVERSION_NONE:
1392       Enable(GL_FRAGMENT_PROGRAM);
1393       break;
1394   }
1395 }
1396
1397 /**
1398  * \brief disable the specified YUV conversion
1399  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1400  * \param type type of YUV conversion
1401  * \ingroup glconversion
1402  */
1403 void glDisableYUVConversion(GLenum target, int type) {
1404   switch (YUV_CONVERSION(type)) {
1405     case YUV_CONVERSION_COMBINERS:
1406       ActiveTexture(GL_TEXTURE1);
1407       Disable(target);
1408       ActiveTexture(GL_TEXTURE2);
1409       Disable(target);
1410       ActiveTexture(GL_TEXTURE0);
1411       Disable(GL_REGISTER_COMBINERS_NV);
1412       break;
1413     case YUV_CONVERSION_COMBINERS_ATI:
1414       ActiveTexture(GL_TEXTURE1);
1415       Disable(target);
1416       ActiveTexture(GL_TEXTURE2);
1417       Disable(target);
1418       ActiveTexture(GL_TEXTURE0);
1419       Disable(GL_FRAGMENT_SHADER_ATI);
1420       break;
1421     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1422     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1423     case YUV_CONVERSION_FRAGMENT_POW:
1424     case YUV_CONVERSION_FRAGMENT:
1425     case YUV_CONVERSION_NONE:
1426       Disable(GL_FRAGMENT_PROGRAM);
1427       break;
1428   }
1429 }
1430
1431 /**
1432  * \brief draw a texture part at given 2D coordinates
1433  * \param x screen top coordinate
1434  * \param y screen left coordinate
1435  * \param w screen width coordinate
1436  * \param h screen height coordinate
1437  * \param tx texture top coordinate in pixels
1438  * \param ty texture left coordinate in pixels
1439  * \param tw texture part width in pixels
1440  * \param th texture part height in pixels
1441  * \param sx width of texture in pixels
1442  * \param sy height of texture in pixels
1443  * \param rect_tex whether this texture uses texture_rectangle extension
1444  * \param is_yv12 if != 0, also draw the textures from units 1 and 2,
1445  *                bits 8 - 15 and 16 - 23 specify the x and y scaling of those textures
1446  * \param flip flip the texture upside down
1447  * \ingroup gltexture
1448  */
1449 void glDrawTex(GLfloat x, GLfloat y, GLfloat w, GLfloat h,
1450                GLfloat tx, GLfloat ty, GLfloat tw, GLfloat th,
1451                int sx, int sy, int rect_tex, int is_yv12, int flip) {
1452   int chroma_x_shift = (is_yv12 >>  8) & 31;
1453   int chroma_y_shift = (is_yv12 >> 16) & 31;
1454   GLfloat xscale = 1 << chroma_x_shift;
1455   GLfloat yscale = 1 << chroma_y_shift;
1456   GLfloat tx2 = tx / xscale, ty2 = ty / yscale, tw2 = tw / xscale, th2 = th / yscale;
1457   if (!rect_tex) {
1458     tx /= sx; ty /= sy; tw /= sx; th /= sy;
1459     tx2 = tx, ty2 = ty, tw2 = tw, th2 = th;
1460   }
1461   if (flip) {
1462     y += h;
1463     h = -h;
1464   }
1465   Begin(GL_QUADS);
1466   TexCoord2f(tx, ty);
1467   if (is_yv12) {
1468     MultiTexCoord2f(GL_TEXTURE1, tx2, ty2);
1469     MultiTexCoord2f(GL_TEXTURE2, tx2, ty2);
1470   }
1471   Vertex2f(x, y);
1472   TexCoord2f(tx, ty + th);
1473   if (is_yv12) {
1474     MultiTexCoord2f(GL_TEXTURE1, tx2, ty2 + th2);
1475     MultiTexCoord2f(GL_TEXTURE2, tx2, ty2 + th2);
1476   }
1477   Vertex2f(x, y + h);
1478   TexCoord2f(tx + tw, ty + th);
1479   if (is_yv12) {
1480     MultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2 + th2);
1481     MultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2 + th2);
1482   }
1483   Vertex2f(x + w, y + h);
1484   TexCoord2f(tx + tw, ty);
1485   if (is_yv12) {
1486     MultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2);
1487     MultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2);
1488   }
1489   Vertex2f(x + w, y);
1490   End();
1491 }
1492
1493 #ifdef CONFIG_GL_WIN32
1494 #include "w32_common.h"
1495 /**
1496  * \brief little helper since wglGetProcAddress definition does not fit our
1497  *        getProcAddress
1498  * \param procName name of function to look up
1499  * \return function pointer returned by wglGetProcAddress
1500  */
1501 static void *w32gpa(const GLubyte *procName) {
1502   HMODULE oglmod;
1503   void *res = wglGetProcAddress(procName);
1504   if (res) return res;
1505   oglmod = GetModuleHandle("opengl32.dll");
1506   return GetProcAddress(oglmod, procName);
1507 }
1508
1509 static int setGlWindow_w32(MPGLContext *ctx)
1510 {
1511   HWND win = vo_w32_window;
1512   int *vinfo = &ctx->vinfo.w32;
1513   HGLRC *context = &ctx->context.w32;
1514   int new_vinfo;
1515   HDC windc = vo_w32_get_dc(win);
1516   HGLRC new_context = 0;
1517   int keep_context = 0;
1518   int res = SET_WINDOW_FAILED;
1519
1520   // should only be needed when keeping context, but not doing glFinish
1521   // can cause flickering even when we do not keep it.
1522   if (*context)
1523   Finish();
1524   new_vinfo = GetPixelFormat(windc);
1525   if (*context && *vinfo && new_vinfo && *vinfo == new_vinfo) {
1526       // we can keep the wglContext
1527       new_context = *context;
1528       keep_context = 1;
1529   } else {
1530     // create a context
1531     new_context = wglCreateContext(windc);
1532     if (!new_context) {
1533       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GL context!\n");
1534       goto out;
1535     }
1536   }
1537
1538   // set context
1539   if (!wglMakeCurrent(windc, new_context)) {
1540     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GL context!\n");
1541     if (!keep_context) {
1542       wglDeleteContext(new_context);
1543     }
1544     goto out;
1545   }
1546
1547   // set new values
1548   vo_w32_window = win;
1549   {
1550     RECT rect;
1551     GetClientRect(win, &rect);
1552     vo_dwidth = rect.right;
1553     vo_dheight = rect.bottom;
1554   }
1555   if (!keep_context) {
1556     if (*context)
1557       wglDeleteContext(*context);
1558     *context = new_context;
1559     *vinfo = new_vinfo;
1560     getFunctions(w32gpa, NULL);
1561
1562     // and inform that reinit is neccessary
1563     res = SET_WINDOW_REINIT;
1564   } else
1565     res = SET_WINDOW_OK;
1566
1567 out:
1568   vo_w32_release_dc(win, windc);
1569   return res;
1570 }
1571
1572 static void releaseGlContext_w32(MPGLContext *ctx) {
1573   int *vinfo = &ctx->vinfo.w32;
1574   HGLRC *context = &ctx->context.w32;
1575   *vinfo = 0;
1576   if (*context) {
1577     wglMakeCurrent(0, 0);
1578     wglDeleteContext(*context);
1579   }
1580   *context = 0;
1581 }
1582
1583 static void swapGlBuffers_w32(MPGLContext *ctx) {
1584   HDC vo_hdc = vo_w32_get_dc(vo_w32_window);
1585   SwapBuffers(vo_hdc);
1586   vo_w32_release_dc(vo_w32_window, vo_hdc);
1587 }
1588 #endif
1589 #ifdef CONFIG_GL_X11
1590 #ifdef HAVE_LIBDL
1591 #include <dlfcn.h>
1592 #endif
1593 #include "x11_common.h"
1594 /**
1595  * \brief find address of a linked function
1596  * \param s name of function to find
1597  * \return address of function or NULL if not found
1598  */
1599 static void *getdladdr(const char *s) {
1600   void *ret = NULL;
1601 #ifdef HAVE_LIBDL
1602   void *handle = dlopen(NULL, RTLD_LAZY);
1603   if (!handle)
1604     return NULL;
1605   ret = dlsym(handle, s);
1606   dlclose(handle);
1607 #endif
1608   return ret;
1609 }
1610
1611 /**
1612  * \brief Returns the XVisualInfo associated with Window win.
1613  * \param win Window whose XVisualInfo is returne.
1614  * \return XVisualInfo of the window. Caller must use XFree to free it.
1615  */
1616 static XVisualInfo *getWindowVisualInfo(Window win) {
1617   XWindowAttributes xw_attr;
1618   XVisualInfo vinfo_template;
1619   int tmp;
1620   XGetWindowAttributes(mDisplay, win, &xw_attr);
1621   vinfo_template.visualid = XVisualIDFromVisual(xw_attr.visual);
1622   return XGetVisualInfo(mDisplay, VisualIDMask, &vinfo_template, &tmp);
1623 }
1624
1625 static void appendstr(char **dst, const char *str)
1626 {
1627     int newsize;
1628     char *newstr;
1629     if (!str)
1630         return;
1631     newsize = strlen(*dst) + 1 + strlen(str) + 1;
1632     newstr = realloc(*dst, newsize);
1633     if (!newstr)
1634         return;
1635     *dst = newstr;
1636     strcat(*dst, " ");
1637     strcat(*dst, str);
1638 }
1639
1640 /**
1641  * \brief Changes the window in which video is displayed.
1642  * If possible only transfers the context to the new window, otherwise
1643  * creates a new one, which must be initialized by the caller.
1644  * \param vinfo Currently used visual.
1645  * \param context Currently used context.
1646  * \param win window that should be used for drawing.
1647  * \return one of SET_WINDOW_FAILED, SET_WINDOW_OK or SET_WINDOW_REINIT.
1648  * In case of SET_WINDOW_REINIT the context could not be transfered
1649  * and the caller must initialize it correctly.
1650  * \ingroup glcontext
1651  */
1652 static int setGlWindow_x11(MPGLContext *ctx)
1653 {
1654   XVisualInfo **vinfo = &ctx->vinfo.x11;
1655   GLXContext *context = &ctx->context.x11;
1656   Window win = vo_window;
1657   XVisualInfo *new_vinfo;
1658   GLXContext new_context = NULL;
1659   int keep_context = 0;
1660
1661   // should only be needed when keeping context, but not doing glFinish
1662   // can cause flickering even when we do not keep it.
1663   if (*context)
1664   Finish();
1665   new_vinfo = getWindowVisualInfo(win);
1666   if (*context && *vinfo && new_vinfo &&
1667       (*vinfo)->visualid == new_vinfo->visualid) {
1668       // we can keep the GLXContext
1669       new_context = *context;
1670       XFree(new_vinfo);
1671       new_vinfo = *vinfo;
1672       keep_context = 1;
1673   } else {
1674     // create a context
1675     new_context = glXCreateContext(mDisplay, new_vinfo, NULL, True);
1676     if (!new_context) {
1677       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GLX context!\n");
1678       XFree(new_vinfo);
1679       return SET_WINDOW_FAILED;
1680     }
1681   }
1682
1683   // set context
1684   if (!glXMakeCurrent(mDisplay, vo_window, new_context)) {
1685     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GLX context!\n");
1686     if (!keep_context) {
1687       glXDestroyContext(mDisplay, new_context);
1688       XFree(new_vinfo);
1689     }
1690     return SET_WINDOW_FAILED;
1691   }
1692
1693   // set new values
1694   vo_window = win;
1695   vo_x11_update_geometry();
1696   if (!keep_context) {
1697     void *(*getProcAddress)(const GLubyte *);
1698     const char *(*glXExtStr)(Display *, int);
1699     char *glxstr = strdup("");
1700     if (*context)
1701       glXDestroyContext(mDisplay, *context);
1702     *context = new_context;
1703     if (*vinfo)
1704       XFree(*vinfo);
1705     *vinfo = new_vinfo;
1706       getProcAddress = getdladdr("glXGetProcAddress");
1707     if (!getProcAddress)
1708       getProcAddress = getdladdr("glXGetProcAddressARB");
1709     if (!getProcAddress)
1710       getProcAddress = (void *)getdladdr;
1711     glXExtStr = getdladdr("glXQueryExtensionsString");
1712     if (glXExtStr)
1713         appendstr(&glxstr, glXExtStr(mDisplay, DefaultScreen(mDisplay)));
1714     glXExtStr = getdladdr("glXGetClientString");
1715     if (glXExtStr)
1716         appendstr(&glxstr, glXExtStr(mDisplay, GLX_EXTENSIONS));
1717     glXExtStr = getdladdr("glXGetServerString");
1718     if (glXExtStr)
1719         appendstr(&glxstr, glXExtStr(mDisplay, GLX_EXTENSIONS));
1720
1721     getFunctions(getProcAddress, glxstr);
1722     free(glxstr);
1723
1724     // and inform that reinit is neccessary
1725     return SET_WINDOW_REINIT;
1726   }
1727   return SET_WINDOW_OK;
1728 }
1729
1730 /**
1731  * \brief free the VisualInfo and GLXContext of an OpenGL context.
1732  * \ingroup glcontext
1733  */
1734 static void releaseGlContext_x11(MPGLContext *ctx) {
1735   XVisualInfo **vinfo = &ctx->vinfo.x11;
1736   GLXContext *context = &ctx->context.x11;
1737   if (*vinfo)
1738     XFree(*vinfo);
1739   *vinfo = NULL;
1740   if (*context)
1741   {
1742     Finish();
1743     glXMakeCurrent(mDisplay, None, NULL);
1744     glXDestroyContext(mDisplay, *context);
1745   }
1746   *context = 0;
1747 }
1748
1749 static void swapGlBuffers_x11(MPGLContext *ctx) {
1750   glXSwapBuffers(mDisplay, vo_window);
1751 }
1752
1753 static int x11_check_events(void) {
1754   return vo_x11_check_events(mDisplay);
1755 }
1756
1757 static void gl_update_xinerama_info(void)
1758 {
1759     update_xinerama_info(global_vo);
1760 }
1761
1762 static void gl_border(void)
1763 {
1764     vo_x11_border();
1765 }
1766
1767 static void gl_fullscreen(void)
1768 {
1769     vo_x11_fullscreen();
1770 }
1771
1772 static void gl_ontop(void)
1773 {
1774     vo_x11_ontop();
1775 }
1776 #endif
1777
1778 int init_mpglcontext(MPGLContext *ctx, enum MPGLType type) {
1779   memset(ctx, 0, sizeof(*ctx));
1780   ctx->type = type;
1781   switch (ctx->type) {
1782 #ifdef CONFIG_GL_WIN32
1783   case GLTYPE_W32:
1784     ctx->setGlWindow = setGlWindow_w32;
1785     ctx->releaseGlContext = releaseGlContext_w32;
1786     ctx->swapGlBuffers = swapGlBuffers_w32;
1787     ctx->update_xinerama_info = w32_update_xinerama_info;
1788     ctx->border = vo_w32_border;
1789     ctx->check_events = vo_w32_check_events;
1790     ctx->fullscreen = vo_w32_fullscreen;
1791     ctx->ontop = vo_w32_ontop;
1792     return vo_w32_init();
1793 #endif
1794 #ifdef CONFIG_GL_X11
1795   case GLTYPE_X11:
1796     ctx->setGlWindow = setGlWindow_x11;
1797     ctx->releaseGlContext = releaseGlContext_x11;
1798     ctx->swapGlBuffers = swapGlBuffers_x11;
1799     ctx->update_xinerama_info = gl_update_xinerama_info;
1800     ctx->border = gl_border;
1801     ctx->check_events = x11_check_events;
1802     ctx->fullscreen = gl_fullscreen;
1803     ctx->ontop = gl_ontop;
1804     return vo_init();
1805 #endif
1806   default:
1807     return 0;
1808   }
1809 }
1810
1811 void uninit_mpglcontext(MPGLContext *ctx) {
1812   ctx->releaseGlContext(ctx);
1813   switch (ctx->type) {
1814 #ifdef CONFIG_GL_WIN32
1815   case GLTYPE_W32:
1816     vo_w32_uninit();
1817     break;
1818 #endif
1819 #ifdef CONFIG_GL_X11
1820   case GLTYPE_X11:
1821     vo_x11_uninit();
1822     break;
1823 #endif
1824   }
1825   memset(ctx, 0, sizeof(*ctx));
1826 }