libvo/gl_common.c

   1 /*
   2  * common OpenGL routines
   3  *
   4  * copyleft (C) 2005-2010 Reimar Döffinger <Reimar.Doeffinger@gmx.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  * You can alternatively redistribute this file and/or
  25  * modify it under the terms of the GNU Lesser General Public
  26  * License as published by the Free Software Foundation; either
  27  * version 2.1 of the License, or (at your option) any later version.
  28  */
  29
  30 /**
  31  * \file gl_common.c
  32  * \brief OpenGL helper functions used by vo_gl.c and vo_gl2.c
  33  */
  34
  35 #include <stdlib.h>
  36 #include <stdio.h>
  37 #include <string.h>
  38 #include <ctype.h>
  39 #include <math.h>
  40 #include "old_vo_defines.h"
  41 #include "gl_common.h"
  42 #include "csputils.h"
  43 #include "aspect.h"
  44
  45 void (GLAPIENTRY *mpglBegin)(GLenum);
  46 void (GLAPIENTRY *mpglEnd)(void);
  47 void (GLAPIENTRY *mpglViewport)(GLint, GLint, GLsizei, GLsizei);
  48 void (GLAPIENTRY *mpglMatrixMode)(GLenum);
  49 void (GLAPIENTRY *mpglLoadIdentity)(void);
  50 void (GLAPIENTRY *mpglTranslated)(double, double, double);
  51 void (GLAPIENTRY *mpglScaled)(double, double, double);
  52 void (GLAPIENTRY *mpglOrtho)(double, double, double, double, double, double);
  53 void (GLAPIENTRY *mpglFrustum)(double, double, double, double, double, double);
  54 void (GLAPIENTRY *mpglPushMatrix)(void);
  55 void (GLAPIENTRY *mpglPopMatrix)(void);
  56 void (GLAPIENTRY *mpglClear)(GLbitfield);
  57 GLuint (GLAPIENTRY *mpglGenLists)(GLsizei);
  58 void (GLAPIENTRY *mpglDeleteLists)(GLuint, GLsizei);
  59 void (GLAPIENTRY *mpglNewList)(GLuint, GLenum);
  60 void (GLAPIENTRY *mpglEndList)(void);
  61 void (GLAPIENTRY *mpglCallList)(GLuint);
  62 void (GLAPIENTRY *mpglCallLists)(GLsizei, GLenum, const GLvoid *);
  63 void (GLAPIENTRY *mpglGenTextures)(GLsizei, GLuint *);
  64 void (GLAPIENTRY *mpglDeleteTextures)(GLsizei, const GLuint *);
  65 void (GLAPIENTRY *mpglTexEnvf)(GLenum, GLenum, GLfloat);
  66 void (GLAPIENTRY *mpglTexEnvi)(GLenum, GLenum, GLint);
  67 void (GLAPIENTRY *mpglColor4ub)(GLubyte, GLubyte, GLubyte, GLubyte);
  68 void (GLAPIENTRY *mpglColor3f)(GLfloat, GLfloat, GLfloat);
  69 void (GLAPIENTRY *mpglColor4f)(GLfloat, GLfloat, GLfloat, GLfloat);
  70 void (GLAPIENTRY *mpglClearColor)(GLclampf, GLclampf, GLclampf, GLclampf);
  71 void (GLAPIENTRY *mpglClearDepth)(GLclampd);
  72 void (GLAPIENTRY *mpglDepthFunc)(GLenum);
  73 void (GLAPIENTRY *mpglEnable)(GLenum);
  74 void (GLAPIENTRY *mpglDisable)(GLenum);
  75 const GLubyte *(GLAPIENTRY *mpglGetString)(GLenum);
  76 void (GLAPIENTRY *mpglDrawBuffer)(GLenum);
  77 void (GLAPIENTRY *mpglDepthMask)(GLboolean);
  78 void (GLAPIENTRY *mpglBlendFunc)(GLenum, GLenum);
  79 void (GLAPIENTRY *mpglFlush)(void);
  80 void (GLAPIENTRY *mpglFinish)(void);
  81 void (GLAPIENTRY *mpglPixelStorei)(GLenum, GLint);
  82 void (GLAPIENTRY *mpglTexImage1D)(GLenum, GLint, GLint, GLsizei, GLint, GLenum, GLenum, const GLvoid *);
  83 void (GLAPIENTRY *mpglTexImage2D)(GLenum, GLint, GLint, GLsizei, GLsizei, GLint, GLenum, GLenum, const GLvoid *);
  84 void (GLAPIENTRY *mpglTexSubImage2D)(GLenum, GLint, GLint, GLint, GLsizei, GLsizei, GLenum, GLenum, const GLvoid *);
  85 void (GLAPIENTRY *mpglTexParameteri)(GLenum, GLenum, GLint);
  86 void (GLAPIENTRY *mpglTexParameterf)(GLenum, GLenum, GLfloat);
  87 void (GLAPIENTRY *mpglTexParameterfv)(GLenum, GLenum, const GLfloat *);
  88 void (GLAPIENTRY *mpglTexCoord2f)(GLfloat, GLfloat);
  89 void (GLAPIENTRY *mpglVertex2f)(GLfloat, GLfloat);
  90 void (GLAPIENTRY *mpglVertex3f)(GLfloat, GLfloat, GLfloat);
  91 void (GLAPIENTRY *mpglNormal3f)(GLfloat, GLfloat, GLfloat);
  92 void (GLAPIENTRY *mpglLightfv)(GLenum, GLenum, const GLfloat *);
  93 void (GLAPIENTRY *mpglColorMaterial)(GLenum, GLenum);
  94 void (GLAPIENTRY *mpglShadeModel)(GLenum);
  95 void (GLAPIENTRY *mpglGetIntegerv)(GLenum, GLint *);
  96 void (GLAPIENTRY *mpglColorMask)(GLboolean, GLboolean, GLboolean, GLboolean);
  97
  98 /**
  99  * \defgroup glextfunctions OpenGL extension functions
 100  *
 101  * the pointers to these functions are acquired when the OpenGL
 102  * context is created
 103  * \{
 104  */
 105 void (GLAPIENTRY *mpglGenBuffers)(GLsizei, GLuint *);
 106 void (GLAPIENTRY *mpglDeleteBuffers)(GLsizei, const GLuint *);
 107 void (GLAPIENTRY *mpglBindBuffer)(GLenum, GLuint);
 108 GLvoid* (GLAPIENTRY *mpglMapBuffer)(GLenum, GLenum);
 109 GLboolean (GLAPIENTRY *mpglUnmapBuffer)(GLenum);
 110 void (GLAPIENTRY *mpglBufferData)(GLenum, intptr_t, const GLvoid *, GLenum);
 111 void (GLAPIENTRY *mpglCombinerParameterfv)(GLenum, const GLfloat *);
 112 void (GLAPIENTRY *mpglCombinerParameteri)(GLenum, GLint);
 113 void (GLAPIENTRY *mpglCombinerInput)(GLenum, GLenum, GLenum, GLenum, GLenum,
 114                                GLenum);
 115 void (GLAPIENTRY *mpglCombinerOutput)(GLenum, GLenum, GLenum, GLenum, GLenum,
 116                                 GLenum, GLenum, GLboolean, GLboolean,
 117                                 GLboolean);
 118 void (GLAPIENTRY *mpglBeginFragmentShader)(void);
 119 void (GLAPIENTRY *mpglEndFragmentShader)(void);
 120 void (GLAPIENTRY *mpglSampleMap)(GLuint, GLuint, GLenum);
 121 void (GLAPIENTRY *mpglColorFragmentOp2)(GLenum, GLuint, GLuint, GLuint, GLuint,
 122                                   GLuint, GLuint, GLuint, GLuint, GLuint);
 123 void (GLAPIENTRY *mpglColorFragmentOp3)(GLenum, GLuint, GLuint, GLuint, GLuint,
 124                                   GLuint, GLuint, GLuint, GLuint, GLuint,
 125                                   GLuint, GLuint, GLuint);
 126 void (GLAPIENTRY *mpglSetFragmentShaderConstant)(GLuint, const GLfloat *);
 127 void (GLAPIENTRY *mpglActiveTexture)(GLenum);
 128 void (GLAPIENTRY *mpglBindTexture)(GLenum, GLuint);
 129 void (GLAPIENTRY *mpglMultiTexCoord2f)(GLenum, GLfloat, GLfloat);
 130 void (GLAPIENTRY *mpglGenPrograms)(GLsizei, GLuint *);
 131 void (GLAPIENTRY *mpglDeletePrograms)(GLsizei, const GLuint *);
 132 void (GLAPIENTRY *mpglBindProgram)(GLenum, GLuint);
 133 void (GLAPIENTRY *mpglProgramString)(GLenum, GLenum, GLsizei, const GLvoid *);
 134 void (GLAPIENTRY *mpglGetProgramiv)(GLenum, GLenum, GLint *);
 135 void (GLAPIENTRY *mpglProgramEnvParameter4f)(GLenum, GLuint, GLfloat, GLfloat,
 136                                        GLfloat, GLfloat);
 137 int (GLAPIENTRY *mpglSwapInterval)(int);
 138 void (GLAPIENTRY *mpglTexImage3D)(GLenum, GLint, GLenum, GLsizei, GLsizei, GLsizei,
 139                             GLint, GLenum, GLenum, const GLvoid *);
 140 void* (GLAPIENTRY *mpglAllocateMemoryMESA)(void *, int, size_t, float, float, float);
 141 void (GLAPIENTRY *mpglFreeMemoryMESA)(void *, int, void *);
 142 /** \} */ // end of glextfunctions group
 143
 144 //! \defgroup glgeneral OpenGL general helper functions
 145
 146 //! \defgroup glcontext OpenGL context management helper functions
 147
 148 //! \defgroup gltexture OpenGL texture handling helper functions
 149
 150 //! \defgroup glconversion OpenGL conversion helper functions
 151
 152 static GLint hqtexfmt;
 153
 154 /**
 155  * \brief adjusts the GL_UNPACK_ALIGNMENT to fit the stride.
 156  * \param stride number of bytes per line for which alignment should fit.
 157  * \ingroup glgeneral
 158  */
 159 void glAdjustAlignment(int stride) {
 160   GLint gl_alignment;
 161   if (stride % 8 == 0)
 162     gl_alignment=8;
 163   else if (stride % 4 == 0)
 164     gl_alignment=4;
 165   else if (stride % 2 == 0)
 166     gl_alignment=2;
 167   else
 168     gl_alignment=1;
 169   mpglPixelStorei(GL_UNPACK_ALIGNMENT, gl_alignment);
 170 }
 171
 172 struct gl_name_map_struct {
 173   GLint value;
 174   const char *name;
 175 };
 176
 177 #undef MAP
 178 #define MAP(a) {a, #a}
 179 //! mapping table for the glValName function
 180 static const struct gl_name_map_struct gl_name_map[] = {
 181   // internal format
 182   MAP(GL_R3_G3_B2), MAP(GL_RGB4), MAP(GL_RGB5), MAP(GL_RGB8),
 183   MAP(GL_RGB10), MAP(GL_RGB12), MAP(GL_RGB16), MAP(GL_RGBA2),
 184   MAP(GL_RGBA4), MAP(GL_RGB5_A1), MAP(GL_RGBA8), MAP(GL_RGB10_A2),
 185   MAP(GL_RGBA12), MAP(GL_RGBA16), MAP(GL_LUMINANCE8),
 186
 187   // format
 188   MAP(GL_RGB), MAP(GL_RGBA), MAP(GL_RED), MAP(GL_GREEN), MAP(GL_BLUE),
 189   MAP(GL_ALPHA), MAP(GL_LUMINANCE), MAP(GL_LUMINANCE_ALPHA),
 190   MAP(GL_COLOR_INDEX),
 191   // rest 1.2 only
 192   MAP(GL_BGR), MAP(GL_BGRA),
 193
 194   //type
 195   MAP(GL_BYTE), MAP(GL_UNSIGNED_BYTE), MAP(GL_SHORT), MAP(GL_UNSIGNED_SHORT),
 196   MAP(GL_INT), MAP(GL_UNSIGNED_INT), MAP(GL_FLOAT), MAP(GL_DOUBLE),
 197   MAP(GL_2_BYTES), MAP(GL_3_BYTES), MAP(GL_4_BYTES),
 198   // rest 1.2 only
 199   MAP(GL_UNSIGNED_BYTE_3_3_2), MAP(GL_UNSIGNED_BYTE_2_3_3_REV),
 200   MAP(GL_UNSIGNED_SHORT_5_6_5), MAP(GL_UNSIGNED_SHORT_5_6_5_REV),
 201   MAP(GL_UNSIGNED_SHORT_4_4_4_4), MAP(GL_UNSIGNED_SHORT_4_4_4_4_REV),
 202   MAP(GL_UNSIGNED_SHORT_5_5_5_1), MAP(GL_UNSIGNED_SHORT_1_5_5_5_REV),
 203   MAP(GL_UNSIGNED_INT_8_8_8_8), MAP(GL_UNSIGNED_INT_8_8_8_8_REV),
 204   MAP(GL_UNSIGNED_INT_10_10_10_2), MAP(GL_UNSIGNED_INT_2_10_10_10_REV),
 205   {0, 0}
 206 };
 207 #undef MAP
 208
 209 /**
 210  * \brief return the name of an OpenGL constant
 211  * \param value the constant
 212  * \return name of the constant or "Unknown format!"
 213  * \ingroup glgeneral
 214  */
 215 const char *glValName(GLint value)
 216 {
 217   int i = 0;
 218
 219   while (gl_name_map[i].name) {
 220     if (gl_name_map[i].value == value)
 221       return gl_name_map[i].name;
 222     i++;
 223   }
 224   return "Unknown format!";
 225 }
 226
 227 //! always return this format as internal texture format in glFindFormat
 228 #define TEXTUREFORMAT_ALWAYS GL_RGB8
 229 #undef TEXTUREFORMAT_ALWAYS
 230
 231 /**
 232  * \brief find the OpenGL settings coresponding to format.
 233  *
 234  * All parameters may be NULL.
 235  * \param fmt MPlayer format to analyze.
 236  * \param bpp [OUT] bits per pixel of that format.
 237  * \param gl_texfmt [OUT] internal texture format that fits the
 238  * image format, not necessarily the best for performance.
 239  * \param gl_format [OUT] OpenGL format for this image format.
 240  * \param gl_type [OUT] OpenGL type for this image format.
 241  * \return 1 if format is supported by OpenGL, 0 if not.
 242  * \ingroup gltexture
 243  */
 244 int glFindFormat(uint32_t fmt, int *bpp, GLint *gl_texfmt,
 245                   GLenum *gl_format, GLenum *gl_type)
 246 {
 247   int supported = 1;
 248   int dummy1;
 249   GLenum dummy2;
 250   GLint dummy3;
 251   if (!bpp) bpp = &dummy1;
 252   if (!gl_texfmt) gl_texfmt = &dummy3;
 253   if (!gl_format) gl_format = &dummy2;
 254   if (!gl_type) gl_type = &dummy2;
 255
 256   if (mp_get_chroma_shift(fmt, NULL, NULL)) {
 257     // reduce the possible cases a bit
 258     if (IMGFMT_IS_YUVP16_LE(fmt))
 259       fmt = IMGFMT_420P16_LE;
 260     else if (IMGFMT_IS_YUVP16_BE(fmt))
 261       fmt = IMGFMT_420P16_BE;
 262     else
 263       fmt = IMGFMT_YV12;
 264   }
 265
 266   *bpp = IMGFMT_IS_BGR(fmt)?IMGFMT_BGR_DEPTH(fmt):IMGFMT_RGB_DEPTH(fmt);
 267   *gl_texfmt = 3;
 268   switch (fmt) {
 269     case IMGFMT_RGB48NE:
 270       *gl_format = GL_RGB;
 271       *gl_type = GL_UNSIGNED_SHORT;
 272       break;
 273     case IMGFMT_RGB24:
 274       *gl_format = GL_RGB;
 275       *gl_type = GL_UNSIGNED_BYTE;
 276       break;
 277     case IMGFMT_RGBA:
 278       *gl_texfmt = 4;
 279       *gl_format = GL_RGBA;
 280       *gl_type = GL_UNSIGNED_BYTE;
 281       break;
 282     case IMGFMT_420P16:
 283       supported = 0; // no native YUV support
 284       *gl_texfmt = 1;
 285       *bpp = 16;
 286       *gl_format = GL_LUMINANCE;
 287       *gl_type = GL_UNSIGNED_SHORT;
 288       break;
 289     case IMGFMT_YV12:
 290       supported = 0; // no native YV12 support
 291     case IMGFMT_Y800:
 292     case IMGFMT_Y8:
 293       *gl_texfmt = 1;
 294       *bpp = 8;
 295       *gl_format = GL_LUMINANCE;
 296       *gl_type = GL_UNSIGNED_BYTE;
 297       break;
 298     case IMGFMT_UYVY:
 299     case IMGFMT_YUY2:
 300       *gl_texfmt = GL_YCBCR_MESA;
 301       *bpp = 16;
 302       *gl_format = GL_YCBCR_MESA;
 303       *gl_type = fmt == IMGFMT_UYVY ? GL_UNSIGNED_SHORT_8_8 : GL_UNSIGNED_SHORT_8_8_REV;
 304       break;
 305 #if 0
 306     // we do not support palettized formats, although the format the
 307     // swscale produces works
 308     case IMGFMT_RGB8:
 309       gl_format = GL_RGB;
 310       gl_type = GL_UNSIGNED_BYTE_2_3_3_REV;
 311       break;
 312 #endif
 313     case IMGFMT_RGB15:
 314       *gl_format = GL_RGBA;
 315       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 316       break;
 317     case IMGFMT_RGB16:
 318       *gl_format = GL_RGB;
 319       *gl_type = GL_UNSIGNED_SHORT_5_6_5_REV;
 320       break;
 321 #if 0
 322     case IMGFMT_BGR8:
 323       // special case as red and blue have a differen number of bits.
 324       // GL_BGR and GL_UNSIGNED_BYTE_3_3_2 isn't supported at least
 325       // by nVidia drivers, and in addition would give more bits to
 326       // blue than to red, which isn't wanted
 327       gl_format = GL_RGB;
 328       gl_type = GL_UNSIGNED_BYTE_3_3_2;
 329       break;
 330 #endif
 331     case IMGFMT_BGR15:
 332       *gl_format = GL_BGRA;
 333       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 334       break;
 335     case IMGFMT_BGR16:
 336       *gl_format = GL_RGB;
 337       *gl_type = GL_UNSIGNED_SHORT_5_6_5;
 338       break;
 339     case IMGFMT_BGR24:
 340       *gl_format = GL_BGR;
 341       *gl_type = GL_UNSIGNED_BYTE;
 342       break;
 343     case IMGFMT_BGRA:
 344       *gl_texfmt = 4;
 345       *gl_format = GL_BGRA;
 346       *gl_type = GL_UNSIGNED_BYTE;
 347       break;
 348     default:
 349       *gl_texfmt = 4;
 350       *gl_format = GL_RGBA;
 351       *gl_type = GL_UNSIGNED_BYTE;
 352       supported = 0;
 353   }
 354 #ifdef TEXTUREFORMAT_ALWAYS
 355   *gl_texfmt = TEXTUREFORMAT_ALWAYS;
 356 #endif
 357   return supported;
 358 }
 359
 360 #ifdef HAVE_LIBDL
 361 #include <dlfcn.h>
 362 #endif
 363 /**
 364  * \brief find address of a linked function
 365  * \param s name of function to find
 366  * \return address of function or NULL if not found
 367  */
 368 static void *getdladdr(const char *s) {
 369   void *ret = NULL;
 370 #ifdef HAVE_LIBDL
 371   void *handle = dlopen(NULL, RTLD_LAZY);
 372   if (!handle)
 373     return NULL;
 374   ret = dlsym(handle, s);
 375   dlclose(handle);
 376 #endif
 377   return ret;
 378 }
 379
 380 typedef struct {
 381   void *funcptr;
 382   const char *extstr;
 383   const char *funcnames[7];
 384   void *fallback;
 385 } extfunc_desc_t;
 386
 387 #define DEF_FUNC_DESC(name) {&mpgl##name, NULL, {"gl"#name, NULL}, gl ##name}
 388 static const extfunc_desc_t extfuncs[] = {
 389   // these aren't extension functions but we query them anyway to allow
 390   // different "backends" with one binary
 391   DEF_FUNC_DESC(Begin),
 392   DEF_FUNC_DESC(End),
 393   DEF_FUNC_DESC(Viewport),
 394   DEF_FUNC_DESC(MatrixMode),
 395   DEF_FUNC_DESC(LoadIdentity),
 396   DEF_FUNC_DESC(Translated),
 397   DEF_FUNC_DESC(Scaled),
 398   DEF_FUNC_DESC(Ortho),
 399   DEF_FUNC_DESC(Frustum),
 400   DEF_FUNC_DESC(PushMatrix),
 401   DEF_FUNC_DESC(PopMatrix),
 402   DEF_FUNC_DESC(Clear),
 403   DEF_FUNC_DESC(GenLists),
 404   DEF_FUNC_DESC(DeleteLists),
 405   DEF_FUNC_DESC(NewList),
 406   DEF_FUNC_DESC(EndList),
 407   DEF_FUNC_DESC(CallList),
 408   DEF_FUNC_DESC(CallLists),
 409   DEF_FUNC_DESC(GenTextures),
 410   DEF_FUNC_DESC(DeleteTextures),
 411   DEF_FUNC_DESC(TexEnvf),
 412   DEF_FUNC_DESC(TexEnvi),
 413   DEF_FUNC_DESC(Color4ub),
 414   DEF_FUNC_DESC(Color3f),
 415   DEF_FUNC_DESC(Color4f),
 416   DEF_FUNC_DESC(ClearColor),
 417   DEF_FUNC_DESC(ClearDepth),
 418   DEF_FUNC_DESC(DepthFunc),
 419   DEF_FUNC_DESC(Enable),
 420   DEF_FUNC_DESC(Disable),
 421   DEF_FUNC_DESC(DrawBuffer),
 422   DEF_FUNC_DESC(DepthMask),
 423   DEF_FUNC_DESC(BlendFunc),
 424   DEF_FUNC_DESC(Flush),
 425   DEF_FUNC_DESC(Finish),
 426   DEF_FUNC_DESC(PixelStorei),
 427   DEF_FUNC_DESC(TexImage1D),
 428   DEF_FUNC_DESC(TexImage2D),
 429   DEF_FUNC_DESC(TexSubImage2D),
 430   DEF_FUNC_DESC(TexParameteri),
 431   DEF_FUNC_DESC(TexParameterf),
 432   DEF_FUNC_DESC(TexParameterfv),
 433   DEF_FUNC_DESC(TexCoord2f),
 434   DEF_FUNC_DESC(Vertex2f),
 435   DEF_FUNC_DESC(Vertex3f),
 436   DEF_FUNC_DESC(Normal3f),
 437   DEF_FUNC_DESC(Lightfv),
 438   DEF_FUNC_DESC(ColorMaterial),
 439   DEF_FUNC_DESC(ShadeModel),
 440   DEF_FUNC_DESC(GetIntegerv),
 441   DEF_FUNC_DESC(ColorMask),
 442
 443   // here start the real extensions
 444   {&mpglGenBuffers, NULL, {"glGenBuffers", "glGenBuffersARB", NULL}},
 445   {&mpglDeleteBuffers, NULL, {"glDeleteBuffers", "glDeleteBuffersARB", NULL}},
 446   {&mpglBindBuffer, NULL, {"glBindBuffer", "glBindBufferARB", NULL}},
 447   {&mpglMapBuffer, NULL, {"glMapBuffer", "glMapBufferARB", NULL}},
 448   {&mpglUnmapBuffer, NULL, {"glUnmapBuffer", "glUnmapBufferARB", NULL}},
 449   {&mpglBufferData, NULL, {"glBufferData", "glBufferDataARB", NULL}},
 450   {&mpglCombinerParameterfv, "NV_register_combiners", {"glCombinerParameterfv", "glCombinerParameterfvNV", NULL}},
 451   {&mpglCombinerParameteri, "NV_register_combiners", {"glCombinerParameteri", "glCombinerParameteriNV", NULL}},
 452   {&mpglCombinerInput, "NV_register_combiners", {"glCombinerInput", "glCombinerInputNV", NULL}},
 453   {&mpglCombinerOutput, "NV_register_combiners", {"glCombinerOutput", "glCombinerOutputNV", NULL}},
 454   {&mpglBeginFragmentShader, "ATI_fragment_shader", {"glBeginFragmentShaderATI", NULL}},
 455   {&mpglEndFragmentShader, "ATI_fragment_shader", {"glEndFragmentShaderATI", NULL}},
 456   {&mpglSampleMap, "ATI_fragment_shader", {"glSampleMapATI", NULL}},
 457   {&mpglColorFragmentOp2, "ATI_fragment_shader", {"glColorFragmentOp2ATI", NULL}},
 458   {&mpglColorFragmentOp3, "ATI_fragment_shader", {"glColorFragmentOp3ATI", NULL}},
 459   {&mpglSetFragmentShaderConstant, "ATI_fragment_shader", {"glSetFragmentShaderConstantATI", NULL}},
 460   {&mpglActiveTexture, NULL, {"glActiveTexture", "glActiveTextureARB", NULL}},
 461   {&mpglBindTexture, NULL, {"glBindTexture", "glBindTextureARB", "glBindTextureEXT", NULL}},
 462   {&mpglMultiTexCoord2f, NULL, {"glMultiTexCoord2f", "glMultiTexCoord2fARB", NULL}},
 463   {&mpglGenPrograms, "_program", {"glGenProgramsARB", NULL}},
 464   {&mpglDeletePrograms, "_program", {"glDeleteProgramsARB", NULL}},
 465   {&mpglBindProgram, "_program", {"glBindProgramARB", NULL}},
 466   {&mpglProgramString, "_program", {"glProgramStringARB", NULL}},
 467   {&mpglGetProgramiv, "_program", {"glGetProgramivARB", NULL}},
 468   {&mpglProgramEnvParameter4f, "_program", {"glProgramEnvParameter4fARB", NULL}},
 469   {&mpglSwapInterval, "_swap_control", {"glXSwapIntervalSGI", "glXSwapInterval", "wglSwapIntervalSGI", "wglSwapInterval", "wglSwapIntervalEXT", NULL}},
 470   {&mpglTexImage3D, NULL, {"glTexImage3D", NULL}},
 471   {&mpglAllocateMemoryMESA, "GLX_MESA_allocate_memory", {"glXAllocateMemoryMESA", NULL}},
 472   {&mpglFreeMemoryMESA, "GLX_MESA_allocate_memory", {"glXFreeMemoryMESA", NULL}},
 473   {NULL}
 474 };
 475
 476 /**
 477  * \brief find the function pointers of some useful OpenGL extensions
 478  * \param getProcAddress function to resolve function names, may be NULL
 479  * \param ext2 an extra extension string
 480  */
 481 static void getFunctions(void *(*getProcAddress)(const GLubyte *),
 482                          const char *ext2) {
 483   const extfunc_desc_t *dsc;
 484   const char *extensions;
 485   char *allexts;
 486
 487   if (!getProcAddress)
 488     getProcAddress = (void *)getdladdr;
 489
 490   // special case, we need glGetString before starting to find the other functions
 491   mpglGetString = getProcAddress("glGetString");
 492   if (!mpglGetString)
 493       mpglGetString = glGetString;
 494
 495   extensions = (const char *)mpglGetString(GL_EXTENSIONS);
 496   if (!extensions) extensions = "";
 497   if (!ext2) ext2 = "";
 498   allexts = malloc(strlen(extensions) + strlen(ext2) + 2);
 499   strcpy(allexts, extensions);
 500   strcat(allexts, " ");
 501   strcat(allexts, ext2);
 502   mp_msg(MSGT_VO, MSGL_DBG2, "OpenGL extensions string:\n%s\n", allexts);
 503   for (dsc = extfuncs; dsc->funcptr; dsc++) {
 504     void *ptr = NULL;
 505     int i;
 506     if (!dsc->extstr || strstr(allexts, dsc->extstr)) {
 507       for (i = 0; !ptr && dsc->funcnames[i]; i++)
 508         ptr = getProcAddress((const GLubyte *)dsc->funcnames[i]);
 509     }
 510     if (!ptr)
 511         ptr = dsc->fallback;
 512     *(void **)dsc->funcptr = ptr;
 513   }
 514   if (strstr(allexts, "_texture_float"))
 515     hqtexfmt = GL_RGB32F;
 516   else if (strstr(allexts, "NV_float_buffer"))
 517     hqtexfmt = GL_FLOAT_RGB32_NV;
 518   else
 519     hqtexfmt = GL_RGB16;
 520   free(allexts);
 521 }
 522
 523 /**
 524  * \brief create a texture and set some defaults
 525  * \param target texture taget, usually GL_TEXTURE_2D
 526  * \param fmt internal texture format
 527  * \param format texture host data format
 528  * \param type texture host data type
 529  * \param filter filter used for scaling, e.g. GL_LINEAR
 530  * \param w texture width
 531  * \param h texture height
 532  * \param val luminance value to fill texture with
 533  * \ingroup gltexture
 534  */
 535 void glCreateClearTex(GLenum target, GLenum fmt, GLenum format, GLenum type, GLint filter,
 536                       int w, int h, unsigned char val) {
 537   GLfloat fval = (GLfloat)val / 255.0;
 538   GLfloat border[4] = {fval, fval, fval, fval};
 539   int stride;
 540   char *init;
 541   if (w == 0) w = 1;
 542   if (h == 0) h = 1;
 543   stride = w * glFmt2bpp(format, type);
 544   if (!stride) return;
 545   init = malloc(stride * h);
 546   memset(init, val, stride * h);
 547   glAdjustAlignment(stride);
 548   mpglPixelStorei(GL_UNPACK_ROW_LENGTH, w);
 549   mpglTexImage2D(target, 0, fmt, w, h, 0, format, type, init);
 550   mpglTexParameterf(target, GL_TEXTURE_PRIORITY, 1.0);
 551   mpglTexParameteri(target, GL_TEXTURE_MIN_FILTER, filter);
 552   mpglTexParameteri(target, GL_TEXTURE_MAG_FILTER, filter);
 553   mpglTexParameteri(target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 554   mpglTexParameteri(target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 555   // Border texels should not be used with CLAMP_TO_EDGE
 556   // We set a sane default anyway.
 557   mpglTexParameterfv(target, GL_TEXTURE_BORDER_COLOR, border);
 558   free(init);
 559 }
 560
 561 /**
 562  * \brief skips whitespace and comments
 563  * \param f file to read from
 564  */
 565 static void ppm_skip(FILE *f) {
 566   int c, comment = 0;
 567   do {
 568     c = fgetc(f);
 569     if (c == '#')
 570       comment = 1;
 571     if (c == '\n')
 572       comment = 0;
 573   } while (c != EOF && (isspace(c) || comment));
 574   if (c != EOF)
 575     ungetc(c, f);
 576 }
 577
 578 #define MAXDIM (16 * 1024)
 579
 580 static uint8_t *read_pnm(FILE *f, int *width, int *height,
 581                          int *bytes_per_pixel, int *maxval) {
 582   uint8_t *data;
 583   int type;
 584   unsigned w, h, m, val, bpp;
 585   *width = *height = *bytes_per_pixel = *maxval = 0;
 586   ppm_skip(f);
 587   if (fgetc(f) != 'P')
 588     return NULL;
 589   type = fgetc(f);
 590   if (type != '5' && type != '6')
 591     return NULL;
 592   ppm_skip(f);
 593   if (fscanf(f, "%u", &w) != 1)
 594     return NULL;
 595   ppm_skip(f);
 596   if (fscanf(f, "%u", &h) != 1)
 597     return NULL;
 598   ppm_skip(f);
 599   if (fscanf(f, "%u", &m) != 1)
 600     return NULL;
 601   val = fgetc(f);
 602   if (!isspace(val))
 603     return NULL;
 604   if (w > MAXDIM || h > MAXDIM)
 605     return NULL;
 606   bpp = (m > 255) ? 2 : 1;
 607   if (type == '6')
 608     bpp *= 3;
 609   data = malloc(w * h * bpp);
 610   if (fread(data, w * bpp, h, f) != h) {
 611     free(data);
 612     return NULL;
 613   }
 614   *width  = w;
 615   *height = h;
 616   *bytes_per_pixel = bpp;
 617   *maxval = m;
 618   return data;
 619 }
 620
 621 /**
 622  * \brief creates a texture from a PPM file
 623  * \param target texture taget, usually GL_TEXTURE_2D
 624  * \param fmt internal texture format, 0 for default
 625  * \param filter filter used for scaling, e.g. GL_LINEAR
 626  * \param f file to read PPM from
 627  * \param width [out] width of texture
 628  * \param height [out] height of texture
 629  * \param maxval [out] maxval value from PPM file
 630  * \return 0 on error, 1 otherwise
 631  * \ingroup gltexture
 632  */
 633 int glCreatePPMTex(GLenum target, GLenum fmt, GLint filter,
 634                    FILE *f, int *width, int *height, int *maxval) {
 635   int w, h, m, bpp;
 636   GLenum type;
 637   uint8_t *data = read_pnm(f, &w, &h, &bpp, &m);
 638   if (!data || (bpp != 3 && bpp != 6)) {
 639     free(data);
 640     return 0;
 641   }
 642   if (!fmt) {
 643     fmt = bpp == 6 ? hqtexfmt : 3;
 644     if (fmt == GL_FLOAT_RGB32_NV && target != GL_TEXTURE_RECTANGLE)
 645       fmt = GL_RGB16;
 646   }
 647   type = bpp == 6 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 648   glCreateClearTex(target, fmt, GL_RGB, type, filter, w, h, 0);
 649   glUploadTex(target, GL_RGB, type,
 650               data, w * bpp, 0, 0, w, h, 0);
 651   free(data);
 652   if (width) *width = w;
 653   if (height) *height = h;
 654   if (maxval) *maxval = m;
 655   return 1;
 656 }
 657
 658 /**
 659  * \brief return the number of bytes per pixel for the given format
 660  * \param format OpenGL format
 661  * \param type OpenGL type
 662  * \return bytes per pixel
 663  * \ingroup glgeneral
 664  *
 665  * Does not handle all possible variants, just those used by MPlayer
 666  */
 667 int glFmt2bpp(GLenum format, GLenum type) {
 668   int component_size = 0;
 669   switch (type) {
 670     case GL_UNSIGNED_BYTE_3_3_2:
 671     case GL_UNSIGNED_BYTE_2_3_3_REV:
 672       return 1;
 673     case GL_UNSIGNED_SHORT_5_5_5_1:
 674     case GL_UNSIGNED_SHORT_1_5_5_5_REV:
 675     case GL_UNSIGNED_SHORT_5_6_5:
 676     case GL_UNSIGNED_SHORT_5_6_5_REV:
 677       return 2;
 678     case GL_UNSIGNED_BYTE:
 679       component_size = 1;
 680       break;
 681     case GL_UNSIGNED_SHORT:
 682       component_size = 2;
 683       break;
 684   }
 685   switch (format) {
 686     case GL_LUMINANCE:
 687     case GL_ALPHA:
 688       return component_size;
 689     case GL_YCBCR_MESA:
 690       return 2;
 691     case GL_RGB:
 692     case GL_BGR:
 693       return 3 * component_size;
 694     case GL_RGBA:
 695     case GL_BGRA:
 696       return 4 * component_size;
 697   }
 698   return 0; // unknown
 699 }
 700
 701 /**
 702  * \brief upload a texture, handling things like stride and slices
 703  * \param target texture target, usually GL_TEXTURE_2D
 704  * \param format OpenGL format of data
 705  * \param type OpenGL type of data
 706  * \param dataptr data to upload
 707  * \param stride data stride
 708  * \param x x offset in texture
 709  * \param y y offset in texture
 710  * \param w width of the texture part to upload
 711  * \param h height of the texture part to upload
 712  * \param slice height of an upload slice, 0 for all at once
 713  * \ingroup gltexture
 714  */
 715 void glUploadTex(GLenum target, GLenum format, GLenum type,
 716                  const void *dataptr, int stride,
 717                  int x, int y, int w, int h, int slice) {
 718   const uint8_t *data = dataptr;
 719   int y_max = y + h;
 720   if (w <= 0 || h <= 0) return;
 721   if (slice <= 0)
 722     slice = h;
 723   if (stride < 0) {
 724     data += (h - 1) * stride;
 725     stride = -stride;
 726   }
 727   // this is not always correct, but should work for MPlayer
 728   glAdjustAlignment(stride);
 729   mpglPixelStorei(GL_UNPACK_ROW_LENGTH, stride / glFmt2bpp(format, type));
 730   for (; y + slice <= y_max; y += slice) {
 731     mpglTexSubImage2D(target, 0, x, y, w, slice, format, type, data);
 732     data += stride * slice;
 733   }
 734   if (y < y_max)
 735     mpglTexSubImage2D(target, 0, x, y, w, y_max - y, format, type, data);
 736 }
 737
 738 static void fillUVcoeff(GLfloat *ucoef, GLfloat *vcoef,
 739                         float uvcos, float uvsin) {
 740   int i;
 741   ucoef[0] = 0 * uvcos + 1.403 * uvsin;
 742   vcoef[0] = 0 * uvsin + 1.403 * uvcos;
 743   ucoef[1] = -0.344 * uvcos + -0.714 * uvsin;
 744   vcoef[1] = -0.344 * uvsin + -0.714 * uvcos;
 745   ucoef[2] = 1.770 * uvcos + 0 * uvsin;
 746   vcoef[2] = 1.770 * uvsin + 0 * uvcos;
 747   ucoef[3] = 0;
 748   vcoef[3] = 0;
 749   // Coefficients (probably) must be in [0, 1] range, whereas they originally
 750   // are in [-2, 2] range, so here comes the trick:
 751   // First put them in the [-0.5, 0.5] range, then add 0.5.
 752   // This can be undone with the HALF_BIAS and SCALE_BY_FOUR arguments
 753   // for CombinerInput and CombinerOutput (or the respective ATI variants)
 754   for (i = 0; i < 4; i++) {
 755     ucoef[i] = ucoef[i] * 0.25 + 0.5;
 756     vcoef[i] = vcoef[i] * 0.25 + 0.5;
 757   }
 758 }
 759
 760 /**
 761  * \brief Setup register combiners for YUV to RGB conversion.
 762  * \param uvcos used for saturation and hue adjustment
 763  * \param uvsin used for saturation and hue adjustment
 764  */
 765 static void glSetupYUVCombiners(float uvcos, float uvsin) {
 766   GLfloat ucoef[4];
 767   GLfloat vcoef[4];
 768   GLint i;
 769   if (!mpglCombinerInput || !mpglCombinerOutput ||
 770       !mpglCombinerParameterfv || !mpglCombinerParameteri) {
 771     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner functions missing!\n");
 772     return;
 773   }
 774   mpglGetIntegerv(GL_MAX_GENERAL_COMBINERS_NV, &i);
 775   if (i < 2)
 776     mp_msg(MSGT_VO, MSGL_ERR,
 777            "[gl] 2 general combiners needed for YUV combiner support (found %i)\n", i);
 778   mpglGetIntegerv(GL_MAX_TEXTURE_UNITS, &i);
 779   if (i < 3)
 780     mp_msg(MSGT_VO, MSGL_ERR,
 781            "[gl] 3 texture units needed for YUV combiner support (found %i)\n", i);
 782   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 783   mpglCombinerParameterfv(GL_CONSTANT_COLOR0_NV, ucoef);
 784   mpglCombinerParameterfv(GL_CONSTANT_COLOR1_NV, vcoef);
 785
 786   // UV first, like this green component cannot overflow
 787   mpglCombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_A_NV,
 788                     GL_TEXTURE1, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 789   mpglCombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_B_NV,
 790                     GL_CONSTANT_COLOR0_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 791   mpglCombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_C_NV,
 792                     GL_TEXTURE2, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 793   mpglCombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_D_NV,
 794                     GL_CONSTANT_COLOR1_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 795   mpglCombinerOutput(GL_COMBINER0_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 796                      GL_SPARE0_NV, GL_SCALE_BY_FOUR_NV, GL_NONE, GL_FALSE,
 797                      GL_FALSE, GL_FALSE);
 798
 799   // stage 2
 800   mpglCombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_A_NV, GL_SPARE0_NV,
 801                     GL_SIGNED_IDENTITY_NV, GL_RGB);
 802   mpglCombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_B_NV, GL_ZERO,
 803                     GL_UNSIGNED_INVERT_NV, GL_RGB);
 804   mpglCombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_C_NV,
 805                     GL_TEXTURE0, GL_SIGNED_IDENTITY_NV, GL_RGB);
 806   mpglCombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_D_NV, GL_ZERO,
 807                     GL_UNSIGNED_INVERT_NV, GL_RGB);
 808   mpglCombinerOutput(GL_COMBINER1_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 809                      GL_SPARE0_NV, GL_NONE, GL_NONE, GL_FALSE,
 810                      GL_FALSE, GL_FALSE);
 811
 812   // leave final combiner stage in default mode
 813   mpglCombinerParameteri(GL_NUM_GENERAL_COMBINERS_NV, 2);
 814 }
 815
 816 /**
 817  * \brief Setup ATI version of register combiners for YUV to RGB conversion.
 818  * \param csp_params parameters used for colorspace conversion
 819  * \param text if set use the GL_ATI_text_fragment_shader API as
 820  *             used on OS X.
 821  */
 822 static void glSetupYUVFragmentATI(struct mp_csp_params *csp_params,
 823                                   int text) {
 824   GLint i;
 825   float yuv2rgb[3][4];
 826
 827   mpglGetIntegerv (GL_MAX_TEXTURE_UNITS, &i);
 828   if (i < 3)
 829     mp_msg(MSGT_VO, MSGL_ERR,
 830            "[gl] 3 texture units needed for YUV combiner (ATI) support (found %i)\n", i);
 831
 832   mp_get_yuv2rgb_coeffs(csp_params, yuv2rgb);
 833   for (i = 0; i < 3; i++) {
 834     int j;
 835     yuv2rgb[i][3] -= -0.5 * (yuv2rgb[i][1] + yuv2rgb[i][2]);
 836     for (j = 0; j < 4; j++) {
 837       yuv2rgb[i][j] *= 0.125;
 838       yuv2rgb[i][j] += 0.5;
 839       if (yuv2rgb[i][j] > 1)
 840         yuv2rgb[i][j] = 1;
 841       if (yuv2rgb[i][j] < 0)
 842         yuv2rgb[i][j] = 0;
 843     }
 844   }
 845   if (text == 0) {
 846     GLfloat c0[4] = {yuv2rgb[0][0], yuv2rgb[1][0], yuv2rgb[2][0]};
 847     GLfloat c1[4] = {yuv2rgb[0][1], yuv2rgb[1][1], yuv2rgb[2][1]};
 848     GLfloat c2[4] = {yuv2rgb[0][2], yuv2rgb[1][2], yuv2rgb[2][2]};
 849     GLfloat c3[4] = {yuv2rgb[0][3], yuv2rgb[1][3], yuv2rgb[2][3]};
 850     if (!mpglBeginFragmentShader || !mpglEndFragmentShader ||
 851         !mpglSetFragmentShaderConstant || !mpglSampleMap ||
 852         !mpglColorFragmentOp2 || !mpglColorFragmentOp3) {
 853       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner (ATI) functions missing!\n");
 854       return;
 855     }
 856     mpglGetIntegerv(GL_NUM_FRAGMENT_REGISTERS_ATI, &i);
 857     if (i < 3)
 858       mp_msg(MSGT_VO, MSGL_ERR,
 859              "[gl] 3 registers needed for YUV combiner (ATI) support (found %i)\n", i);
 860     mpglBeginFragmentShader();
 861     mpglSetFragmentShaderConstant(GL_CON_0_ATI, c0);
 862     mpglSetFragmentShaderConstant(GL_CON_1_ATI, c1);
 863     mpglSetFragmentShaderConstant(GL_CON_2_ATI, c2);
 864     mpglSetFragmentShaderConstant(GL_CON_3_ATI, c3);
 865     mpglSampleMap(GL_REG_0_ATI, GL_TEXTURE0, GL_SWIZZLE_STR_ATI);
 866     mpglSampleMap(GL_REG_1_ATI, GL_TEXTURE1, GL_SWIZZLE_STR_ATI);
 867     mpglSampleMap(GL_REG_2_ATI, GL_TEXTURE2, GL_SWIZZLE_STR_ATI);
 868     mpglColorFragmentOp2(GL_MUL_ATI, GL_REG_1_ATI, GL_NONE, GL_NONE,
 869                          GL_REG_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 870                          GL_CON_1_ATI, GL_NONE, GL_BIAS_BIT_ATI);
 871     mpglColorFragmentOp3(GL_MAD_ATI, GL_REG_2_ATI, GL_NONE, GL_NONE,
 872                          GL_REG_2_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 873                          GL_CON_2_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 874                          GL_REG_1_ATI, GL_NONE, GL_NONE);
 875     mpglColorFragmentOp3(GL_MAD_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 876                          GL_REG_0_ATI, GL_NONE, GL_NONE,
 877                          GL_CON_0_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 878                          GL_REG_2_ATI, GL_NONE, GL_NONE);
 879     mpglColorFragmentOp2(GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, GL_8X_BIT_ATI,
 880                          GL_REG_0_ATI, GL_NONE, GL_NONE,
 881                          GL_CON_3_ATI, GL_NONE, GL_BIAS_BIT_ATI);
 882     mpglEndFragmentShader();
 883   } else {
 884     static const char template[] =
 885       "!!ATIfs1.0\n"
 886       "StartConstants;\n"
 887       "  CONSTANT c0 = {%e, %e, %e};\n"
 888       "  CONSTANT c1 = {%e, %e, %e};\n"
 889       "  CONSTANT c2 = {%e, %e, %e};\n"
 890       "  CONSTANT c3 = {%e, %e, %e};\n"
 891       "EndConstants;\n"
 892       "StartOutputPass;\n"
 893       "  SampleMap r0, t0.str;\n"
 894       "  SampleMap r1, t1.str;\n"
 895       "  SampleMap r2, t2.str;\n"
 896       "  MUL r1.rgb, r1.bias, c1.bias;\n"
 897       "  MAD r2.rgb, r2.bias, c2.bias, r1;\n"
 898       "  MAD r0.rgb, r0, c0.bias, r2;\n"
 899       "  ADD r0.rgb.8x, r0, c3.bias;\n"
 900       "EndPass;\n";
 901     char buffer[512];
 902     snprintf(buffer, sizeof(buffer), template,
 903              yuv2rgb[0][0], yuv2rgb[1][0], yuv2rgb[2][0],
 904              yuv2rgb[0][1], yuv2rgb[1][1], yuv2rgb[2][1],
 905              yuv2rgb[0][2], yuv2rgb[1][2], yuv2rgb[2][2],
 906              yuv2rgb[0][3], yuv2rgb[1][3], yuv2rgb[2][3]);
 907     mp_msg(MSGT_VO, MSGL_DBG2, "[gl] generated fragment program:\n%s\n", buffer);
 908     loadGPUProgram(GL_TEXT_FRAGMENT_SHADER_ATI, buffer);
 909   }
 910 }
 911
 912 /**
 913  * \brief helper function for gen_spline_lookup_tex
 914  * \param x subpixel-position ((0,1) range) to calculate weights for
 915  * \param dst where to store transformed weights, must provide space for 4 GLfloats
 916  *
 917  * calculates the weights and stores them after appropriate transformation
 918  * for the scaler fragment program.
 919  */
 920 static void store_weights(float x, GLfloat *dst) {
 921   float w0 = (((-1 * x + 3) * x - 3) * x + 1) / 6;
 922   float w1 = ((( 3 * x - 6) * x + 0) * x + 4) / 6;
 923   float w2 = (((-3 * x + 3) * x + 3) * x + 1) / 6;
 924   float w3 = ((( 1 * x + 0) * x + 0) * x + 0) / 6;
 925   *dst++ = 1 + x - w1 / (w0 + w1);
 926   *dst++ = 1 - x + w3 / (w2 + w3);
 927   *dst++ = w0 + w1;
 928   *dst++ = 0;
 929 }
 930
 931 //! to avoid artefacts this should be rather large
 932 #define LOOKUP_BSPLINE_RES (2 * 1024)
 933 /**
 934  * \brief creates the 1D lookup texture needed for fast higher-order filtering
 935  * \param unit texture unit to attach texture to
 936  */
 937 static void gen_spline_lookup_tex(GLenum unit) {
 938   GLfloat *tex = calloc(4 * LOOKUP_BSPLINE_RES, sizeof(*tex));
 939   GLfloat *tp = tex;
 940   int i;
 941   for (i = 0; i < LOOKUP_BSPLINE_RES; i++) {
 942     float x = (float)(i + 0.5) / LOOKUP_BSPLINE_RES;
 943     store_weights(x, tp);
 944     tp += 4;
 945   }
 946   store_weights(0, tex);
 947   store_weights(1, &tex[4 * (LOOKUP_BSPLINE_RES - 1)]);
 948   mpglActiveTexture(unit);
 949   mpglTexImage1D(GL_TEXTURE_1D, 0, GL_RGBA16, LOOKUP_BSPLINE_RES, 0, GL_RGBA, GL_FLOAT, tex);
 950   mpglTexParameterf(GL_TEXTURE_1D, GL_TEXTURE_PRIORITY, 1.0);
 951   mpglTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 952   mpglTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 953   mpglTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_REPEAT);
 954   mpglActiveTexture(GL_TEXTURE0);
 955   free(tex);
 956 }
 957
 958 static const char *bilin_filt_template =
 959   "TEX yuv.%c, fragment.texcoord[%c], texture[%c], %s;\n";
 960
 961 #define BICUB_FILT_MAIN(textype) \
 962   /* first y-interpolation */ \
 963   "ADD coord, fragment.texcoord[%c].xyxy, cdelta.xyxw;\n" \
 964   "ADD coord2, fragment.texcoord[%c].xyxy, cdelta.zyzw;\n" \
 965   "TEX a.r, coord.xyxy, texture[%c], "textype";\n" \
 966   "TEX a.g, coord.zwzw, texture[%c], "textype";\n" \
 967   /* second y-interpolation */ \
 968   "TEX b.r, coord2.xyxy, texture[%c], "textype";\n" \
 969   "TEX b.g, coord2.zwzw, texture[%c], "textype";\n" \
 970   "LRP a.b, parmy.b, a.rrrr, a.gggg;\n" \
 971   "LRP a.a, parmy.b, b.rrrr, b.gggg;\n" \
 972   /* x-interpolation */ \
 973   "LRP yuv.%c, parmx.b, a.bbbb, a.aaaa;\n"
 974
 975 static const char *bicub_filt_template_2D =
 976   "MAD coord.xy, fragment.texcoord[%c], {%e, %e}, {0.5, 0.5};\n"
 977   "TEX parmx, coord.x, texture[%c], 1D;\n"
 978   "MUL cdelta.xz, parmx.rrgg, {-%e, 0, %e, 0};\n"
 979   "TEX parmy, coord.y, texture[%c], 1D;\n"
 980   "MUL cdelta.yw, parmy.rrgg, {0, -%e, 0, %e};\n"
 981   BICUB_FILT_MAIN("2D");
 982
 983 static const char *bicub_filt_template_RECT =
 984   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};\n"
 985   "TEX parmx, coord.x, texture[%c], 1D;\n"
 986   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};\n"
 987   "TEX parmy, coord.y, texture[%c], 1D;\n"
 988   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};\n"
 989   BICUB_FILT_MAIN("RECT");
 990
 991 #define CALCWEIGHTS(t, s) \
 992   "MAD "t", {-0.5, 0.1666, 0.3333, -0.3333}, "s", {1, 0, -0.5, 0.5};\n" \
 993   "MAD "t", "t", "s", {0, 0, -0.5, 0.5};\n" \
 994   "MAD "t", "t", "s", {-0.6666, 0, 0.8333, 0.1666};\n" \
 995   "RCP a.x, "t".z;\n" \
 996   "RCP a.y, "t".w;\n" \
 997   "MAD "t".xy, "t".xyxy, a.xyxy, {1, 1, 0, 0};\n" \
 998   "ADD "t".x, "t".xxxx, "s";\n" \
 999   "SUB "t".y, "t".yyyy, "s";\n"
1000
1001 static const char *bicub_notex_filt_template_2D =
1002   "MAD coord.xy, fragment.texcoord[%c], {%e, %e}, {0.5, 0.5};\n"
1003   "FRC coord.xy, coord.xyxy;\n"
1004   CALCWEIGHTS("parmx", "coord.xxxx")
1005   "MUL cdelta.xz, parmx.rrgg, {-%e, 0, %e, 0};\n"
1006   CALCWEIGHTS("parmy", "coord.yyyy")
1007   "MUL cdelta.yw, parmy.rrgg, {0, -%e, 0, %e};\n"
1008   BICUB_FILT_MAIN("2D");
1009
1010 static const char *bicub_notex_filt_template_RECT =
1011   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};\n"
1012   "FRC coord.xy, coord.xyxy;\n"
1013   CALCWEIGHTS("parmx", "coord.xxxx")
1014   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};\n"
1015   CALCWEIGHTS("parmy", "coord.yyyy")
1016   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};\n"
1017   BICUB_FILT_MAIN("RECT");
1018
1019 #define BICUB_X_FILT_MAIN(textype) \
1020   "ADD coord.xy, fragment.texcoord[%c].xyxy, cdelta.xyxy;\n" \
1021   "ADD coord2.xy, fragment.texcoord[%c].xyxy, cdelta.zyzy;\n" \
1022   "TEX a.r, coord, texture[%c], "textype";\n" \
1023   "TEX b.r, coord2, texture[%c], "textype";\n" \
1024   /* x-interpolation */ \
1025   "LRP yuv.%c, parmx.b, a.rrrr, b.rrrr;\n"
1026
1027 static const char *bicub_x_filt_template_2D =
1028   "MAD coord.x, fragment.texcoord[%c], {%e}, {0.5};\n"
1029   "TEX parmx, coord, texture[%c], 1D;\n"
1030   "MUL cdelta.xyz, parmx.rrgg, {-%e, 0, %e};\n"
1031   BICUB_X_FILT_MAIN("2D");
1032
1033 static const char *bicub_x_filt_template_RECT =
1034   "ADD coord.x, fragment.texcoord[%c], {0.5};\n"
1035   "TEX parmx, coord, texture[%c], 1D;\n"
1036   "MUL cdelta.xyz, parmx.rrgg, {-1, 0, 1};\n"
1037   BICUB_X_FILT_MAIN("RECT");
1038
1039 static const char *unsharp_filt_template =
1040   "PARAM dcoord%c = {%e, %e, %e, %e};\n"
1041   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;\n"
1042   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;\n"
1043   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;\n"
1044   "TEX b.r, coord.xyxy, texture[%c], %s;\n"
1045   "TEX b.g, coord.zwzw, texture[%c], %s;\n"
1046   "ADD b.r, b.r, b.g;\n"
1047   "TEX b.b, coord2.xyxy, texture[%c], %s;\n"
1048   "TEX b.g, coord2.zwzw, texture[%c], %s;\n"
1049   "DP3 b, b, {0.25, 0.25, 0.25};\n"
1050   "SUB b.r, a.r, b.r;\n"
1051   "MAD yuv.%c, b.r, {%e}, a.r;\n";
1052
1053 static const char *unsharp_filt_template2 =
1054   "PARAM dcoord%c = {%e, %e, %e, %e};\n"
1055   "PARAM dcoord2%c = {%e, 0, 0, %e};\n"
1056   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;\n"
1057   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;\n"
1058   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;\n"
1059   "TEX b.r, coord.xyxy, texture[%c], %s;\n"
1060   "TEX b.g, coord.zwzw, texture[%c], %s;\n"
1061   "ADD b.r, b.r, b.g;\n"
1062   "TEX b.b, coord2.xyxy, texture[%c], %s;\n"
1063   "TEX b.g, coord2.zwzw, texture[%c], %s;\n"
1064   "ADD b.r, b.r, b.b;\n"
1065   "ADD b.a, b.r, b.g;\n"
1066   "ADD coord, fragment.texcoord[%c].xyxy, dcoord2%c;\n"
1067   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord2%c;\n"
1068   "TEX b.r, coord.xyxy, texture[%c], %s;\n"
1069   "TEX b.g, coord.zwzw, texture[%c], %s;\n"
1070   "ADD b.r, b.r, b.g;\n"
1071   "TEX b.b, coord2.xyxy, texture[%c], %s;\n"
1072   "TEX b.g, coord2.zwzw, texture[%c], %s;\n"
1073   "DP4 b.r, b, {-0.1171875, -0.1171875, -0.1171875, -0.09765625};\n"
1074   "MAD b.r, a.r, {0.859375}, b.r;\n"
1075   "MAD yuv.%c, b.r, {%e}, a.r;\n";
1076
1077 static const char *yuv_prog_template =
1078   "PARAM ycoef = {%e, %e, %e};\n"
1079   "PARAM ucoef = {%e, %e, %e};\n"
1080   "PARAM vcoef = {%e, %e, %e};\n"
1081   "PARAM offsets = {%e, %e, %e};\n"
1082   "TEMP res;\n"
1083   "MAD res.rgb, yuv.rrrr, ycoef, offsets;\n"
1084   "MAD res.rgb, yuv.gggg, ucoef, res;\n"
1085   "MAD result.color.rgb, yuv.bbbb, vcoef, res;\n"
1086   "END";
1087
1088 static const char *yuv_pow_prog_template =
1089   "PARAM ycoef = {%e, %e, %e};\n"
1090   "PARAM ucoef = {%e, %e, %e};\n"
1091   "PARAM vcoef = {%e, %e, %e};\n"
1092   "PARAM offsets = {%e, %e, %e};\n"
1093   "PARAM gamma = {%e, %e, %e};\n"
1094   "TEMP res;\n"
1095   "MAD res.rgb, yuv.rrrr, ycoef, offsets;\n"
1096   "MAD res.rgb, yuv.gggg, ucoef, res;\n"
1097   "MAD_SAT res.rgb, yuv.bbbb, vcoef, res;\n"
1098   "POW result.color.r, res.r, gamma.r;\n"
1099   "POW result.color.g, res.g, gamma.g;\n"
1100   "POW result.color.b, res.b, gamma.b;\n"
1101   "END";
1102
1103 static const char *yuv_lookup_prog_template =
1104   "PARAM ycoef = {%e, %e, %e, 0};\n"
1105   "PARAM ucoef = {%e, %e, %e, 0};\n"
1106   "PARAM vcoef = {%e, %e, %e, 0};\n"
1107   "PARAM offsets = {%e, %e, %e, 0.125};\n"
1108   "TEMP res;\n"
1109   "MAD res, yuv.rrrr, ycoef, offsets;\n"
1110   "MAD res.rgb, yuv.gggg, ucoef, res;\n"
1111   "MAD res.rgb, yuv.bbbb, vcoef, res;\n"
1112   "TEX result.color.r, res.raaa, texture[%c], 2D;\n"
1113   "ADD res.a, res.a, 0.25;\n"
1114   "TEX result.color.g, res.gaaa, texture[%c], 2D;\n"
1115   "ADD res.a, res.a, 0.25;\n"
1116   "TEX result.color.b, res.baaa, texture[%c], 2D;\n"
1117   "END";
1118
1119 static const char *yuv_lookup3d_prog_template =
1120   "TEX result.color, yuv, texture[%c], 3D;\n"
1121   "END";
1122
1123 /**
1124  * \brief creates and initializes helper textures needed for scaling texture read
1125  * \param scaler scaler type to create texture for
1126  * \param texu contains next free texture unit number
1127  * \param texs texture unit ids for the scaler are stored in this array
1128  */
1129 static void create_scaler_textures(int scaler, int *texu, char *texs) {
1130   switch (scaler) {
1131     case YUV_SCALER_BILIN:
1132     case YUV_SCALER_BICUB_NOTEX:
1133     case YUV_SCALER_UNSHARP:
1134     case YUV_SCALER_UNSHARP2:
1135       break;
1136     case YUV_SCALER_BICUB:
1137     case YUV_SCALER_BICUB_X:
1138       texs[0] = (*texu)++;
1139       gen_spline_lookup_tex(GL_TEXTURE0 + texs[0]);
1140       texs[0] += '0';
1141       break;
1142     default:
1143       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown scaler type %i\n", scaler);
1144   }
1145 }
1146
1147 //! resolution of texture for gamma lookup table
1148 #define LOOKUP_RES 512
1149 //! resolution for 3D yuv->rgb conversion lookup table
1150 #define LOOKUP_3DRES 32
1151 /**
1152  * \brief creates and initializes helper textures needed for yuv conversion
1153  * \param params struct containing parameters like brightness, gamma, ...
1154  * \param texu contains next free texture unit number
1155  * \param texs texture unit ids for the conversion are stored in this array
1156  */
1157 static void create_conv_textures(gl_conversion_params_t *params, int *texu, char *texs) {
1158   unsigned char *lookup_data = NULL;
1159   int conv = YUV_CONVERSION(params->type);
1160   switch (conv) {
1161     case YUV_CONVERSION_FRAGMENT:
1162     case YUV_CONVERSION_FRAGMENT_POW:
1163       break;
1164     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1165       texs[0] = (*texu)++;
1166       mpglActiveTexture(GL_TEXTURE0 + texs[0]);
1167       lookup_data = malloc(4 * LOOKUP_RES);
1168       mp_gen_gamma_map(lookup_data, LOOKUP_RES, params->csp_params.rgamma);
1169       mp_gen_gamma_map(&lookup_data[LOOKUP_RES], LOOKUP_RES, params->csp_params.ggamma);
1170       mp_gen_gamma_map(&lookup_data[2 * LOOKUP_RES], LOOKUP_RES, params->csp_params.bgamma);
1171       glCreateClearTex(GL_TEXTURE_2D, GL_LUMINANCE8, GL_LUMINANCE, GL_UNSIGNED_BYTE, GL_LINEAR,
1172                        LOOKUP_RES, 4, 0);
1173       glUploadTex(GL_TEXTURE_2D, GL_LUMINANCE, GL_UNSIGNED_BYTE, lookup_data,
1174                   LOOKUP_RES, 0, 0, LOOKUP_RES, 4, 0);
1175       mpglActiveTexture(GL_TEXTURE0);
1176       texs[0] += '0';
1177       break;
1178     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1179       {
1180         int sz = LOOKUP_3DRES + 2; // texture size including borders
1181         if (!mpglTexImage3D) {
1182           mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing 3D texture function!\n");
1183           break;
1184         }
1185         texs[0] = (*texu)++;
1186         mpglActiveTexture(GL_TEXTURE0 + texs[0]);
1187         lookup_data = malloc(3 * sz * sz * sz);
1188         mp_gen_yuv2rgb_map(&params->csp_params, lookup_data, LOOKUP_3DRES);
1189         glAdjustAlignment(sz);
1190         mpglPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
1191         mpglTexImage3D(GL_TEXTURE_3D, 0, 3, sz, sz, sz, 1,
1192                        GL_RGB, GL_UNSIGNED_BYTE, lookup_data);
1193         mpglTexParameterf(GL_TEXTURE_3D, GL_TEXTURE_PRIORITY, 1.0);
1194         mpglTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1195         mpglTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1196         mpglTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_S, GL_CLAMP);
1197         mpglTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_T, GL_CLAMP);
1198         mpglTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_R, GL_CLAMP);
1199         mpglActiveTexture(GL_TEXTURE0);
1200         texs[0] += '0';
1201       }
1202       break;
1203     default:
1204       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", conv);
1205   }
1206   if (lookup_data)
1207     free(lookup_data);
1208 }
1209
1210 /**
1211  * \brief adds a scaling texture read at the current fragment program position
1212  * \param scaler type of scaler to insert
1213  * \param prog_pos current position in fragment program
1214  * \param remain how many bytes remain in the buffer given by prog_pos
1215  * \param texs array containing the texture unit identifiers for this scaler
1216  * \param in_tex texture unit the scaler should read from
1217  * \param out_comp component of the yuv variable the scaler stores the result in
1218  * \param rect if rectangular (pixel) adressing should be used for in_tex
1219  * \param texw width of the in_tex texture
1220  * \param texh height of the in_tex texture
1221  * \param strength strength of filter effect if the scaler does some kind of filtering
1222  */
1223 static void add_scaler(int scaler, char **prog_pos, int *remain, char *texs,
1224                     char in_tex, char out_comp, int rect, int texw, int texh,
1225                     double strength) {
1226   const char *ttype = rect ? "RECT" : "2D";
1227   const float ptw = rect ? 1.0 : 1.0 / texw;
1228   const float pth = rect ? 1.0 : 1.0 / texh;
1229   switch (scaler) {
1230     case YUV_SCALER_BILIN:
1231       snprintf(*prog_pos, *remain, bilin_filt_template, out_comp, in_tex,
1232                  in_tex, ttype);
1233       break;
1234     case YUV_SCALER_BICUB:
1235       if (rect)
1236         snprintf(*prog_pos, *remain, bicub_filt_template_RECT,
1237                  in_tex, texs[0], texs[0],
1238                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1239       else
1240         snprintf(*prog_pos, *remain, bicub_filt_template_2D,
1241                  in_tex, (float)texw, (float)texh,
1242                  texs[0], ptw, ptw, texs[0], pth, pth,
1243                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1244       break;
1245     case YUV_SCALER_BICUB_X:
1246       if (rect)
1247         snprintf(*prog_pos, *remain, bicub_x_filt_template_RECT,
1248                  in_tex, texs[0],
1249                  in_tex, in_tex, in_tex, in_tex, out_comp);
1250       else
1251         snprintf(*prog_pos, *remain, bicub_x_filt_template_2D,
1252                  in_tex, (float)texw,
1253                  texs[0], ptw, ptw,
1254                  in_tex, in_tex, in_tex, in_tex, out_comp);
1255       break;
1256     case YUV_SCALER_BICUB_NOTEX:
1257       if (rect)
1258         snprintf(*prog_pos, *remain, bicub_notex_filt_template_RECT,
1259                  in_tex,
1260                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1261       else
1262         snprintf(*prog_pos, *remain, bicub_notex_filt_template_2D,
1263                  in_tex, (float)texw, (float)texh, ptw, ptw, pth, pth,
1264                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1265       break;
1266     case YUV_SCALER_UNSHARP:
1267       snprintf(*prog_pos, *remain, unsharp_filt_template,
1268                out_comp, 0.5 * ptw, 0.5 * pth, 0.5 * ptw, -0.5 * pth,
1269                in_tex, out_comp, in_tex, out_comp, in_tex,
1270                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1271                in_tex, ttype, out_comp, strength);
1272       break;
1273     case YUV_SCALER_UNSHARP2:
1274       snprintf(*prog_pos, *remain, unsharp_filt_template2,
1275                out_comp, 1.2 * ptw, 1.2 * pth, 1.2 * ptw, -1.2 * pth,
1276                out_comp, 1.5 * ptw, 1.5 * pth,
1277                in_tex, out_comp, in_tex, out_comp, in_tex,
1278                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1279                in_tex, ttype, in_tex, out_comp, in_tex, out_comp,
1280                in_tex, ttype, in_tex, ttype, in_tex, ttype,
1281                in_tex, ttype, out_comp, strength);
1282       break;
1283   }
1284   *remain -= strlen(*prog_pos);
1285   *prog_pos += strlen(*prog_pos);
1286 }
1287
1288 static const struct {
1289   const char *name;
1290   GLenum cur;
1291   GLenum max;
1292 } progstats[] = {
1293   {"instructions", 0x88A0, 0x88A1},
1294   {"native instructions", 0x88A2, 0x88A3},
1295   {"temporaries", 0x88A4, 0x88A5},
1296   {"native temporaries", 0x88A6, 0x88A7},
1297   {"parameters", 0x88A8, 0x88A9},
1298   {"native parameters", 0x88AA, 0x88AB},
1299   {"attribs", 0x88AC, 0x88AD},
1300   {"native attribs", 0x88AE, 0x88AF},
1301   {"ALU instructions", 0x8805, 0x880B},
1302   {"TEX instructions", 0x8806, 0x880C},
1303   {"TEX indirections", 0x8807, 0x880D},
1304   {"native ALU instructions", 0x8808, 0x880E},
1305   {"native TEX instructions", 0x8809, 0x880F},
1306   {"native TEX indirections", 0x880A, 0x8810},
1307   {NULL, 0, 0}
1308 };
1309
1310 /**
1311  * \brief load the specified GPU Program
1312  * \param target program target to load into, only GL_FRAGMENT_PROGRAM is tested
1313  * \param prog program string
1314  * \return 1 on success, 0 otherwise
1315  */
1316 int loadGPUProgram(GLenum target, char *prog) {
1317   int i;
1318   GLint cur = 0, max = 0, err = 0;
1319   if (!mpglProgramString) {
1320     mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing GPU program function\n");
1321     return 0;
1322   }
1323   mpglProgramString(target, GL_PROGRAM_FORMAT_ASCII, strlen(prog), prog);
1324   mpglGetIntegerv(GL_PROGRAM_ERROR_POSITION, &err);
1325   if (err != -1) {
1326     mp_msg(MSGT_VO, MSGL_ERR,
1327       "[gl] Error compiling fragment program, make sure your card supports\n"
1328       "[gl]   GL_ARB_fragment_program (use glxinfo to check).\n"
1329       "[gl]   Error message:\n  %s at %.10s\n",
1330       mpglGetString(GL_PROGRAM_ERROR_STRING), &prog[err]);
1331     return 0;
1332   }
1333   if (!mpglGetProgramiv || !mp_msg_test(MSGT_VO, MSGL_DBG2))
1334     return 1;
1335   mp_msg(MSGT_VO, MSGL_V, "[gl] Program statistics:\n");
1336   for (i = 0; progstats[i].name; i++) {
1337     mpglGetProgramiv(target, progstats[i].cur, &cur);
1338     mpglGetProgramiv(target, progstats[i].max, &max);
1339     mp_msg(MSGT_VO, MSGL_V, "[gl]   %s: %i/%i\n", progstats[i].name, cur, max);
1340   }
1341   return 1;
1342 }
1343
1344 #define MAX_PROGSZ (1024*1024)
1345
1346 /**
1347  * \brief setup a fragment program that will do YUV->RGB conversion
1348  * \param parms struct containing parameters like conversion and scaler type,
1349  *              brightness, ...
1350  */
1351 static void glSetupYUVFragprog(gl_conversion_params_t *params) {
1352   int type = params->type;
1353   int texw = params->texw;
1354   int texh = params->texh;
1355   int rect = params->target == GL_TEXTURE_RECTANGLE;
1356   static const char prog_hdr[] =
1357     "!!ARBfp1.0\n"
1358     "OPTION ARB_precision_hint_fastest;\n"
1359     // all scaler variables must go here so they aren't defined
1360     // multiple times when the same scaler is used more than once
1361     "TEMP coord, coord2, cdelta, parmx, parmy, a, b, yuv;\n";
1362   int prog_remain;
1363   char *yuv_prog, *prog_pos;
1364   int cur_texu = 3;
1365   char lum_scale_texs[1];
1366   char chrom_scale_texs[1];
1367   char conv_texs[1];
1368   GLint i;
1369   // this is the conversion matrix, with y, u, v factors
1370   // for red, green, blue and the constant offsets
1371   float yuv2rgb[3][4];
1372   create_conv_textures(params, &cur_texu, conv_texs);
1373   create_scaler_textures(YUV_LUM_SCALER(type), &cur_texu, lum_scale_texs);
1374   if (YUV_CHROM_SCALER(type) == YUV_LUM_SCALER(type))
1375     memcpy(chrom_scale_texs, lum_scale_texs, sizeof(chrom_scale_texs));
1376   else
1377     create_scaler_textures(YUV_CHROM_SCALER(type), &cur_texu, chrom_scale_texs);
1378   mpglGetIntegerv(GL_MAX_TEXTURE_IMAGE_UNITS, &i);
1379   if (i < cur_texu)
1380     mp_msg(MSGT_VO, MSGL_ERR,
1381            "[gl] %i texture units needed for this type of YUV fragment support (found %i)\n",
1382            cur_texu, i);
1383   if (!mpglProgramString) {
1384     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] ProgramString function missing!\n");
1385     return;
1386   }
1387   yuv_prog = malloc(MAX_PROGSZ);
1388   strcpy(yuv_prog, prog_hdr);
1389   prog_pos    = yuv_prog + sizeof(prog_hdr) - 1;
1390   prog_remain = MAX_PROGSZ - sizeof(prog_hdr);
1391   add_scaler(YUV_LUM_SCALER(type), &prog_pos, &prog_remain, lum_scale_texs,
1392              '0', 'r', rect, texw, texh, params->filter_strength);
1393   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1394              '1', 'g', rect, params->chrom_texw, params->chrom_texh, params->filter_strength);
1395   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1396              '2', 'b', rect, params->chrom_texw, params->chrom_texh, params->filter_strength);
1397   mp_get_yuv2rgb_coeffs(&params->csp_params, yuv2rgb);
1398   switch (YUV_CONVERSION(type)) {
1399     case YUV_CONVERSION_FRAGMENT:
1400       snprintf(prog_pos, prog_remain, yuv_prog_template,
1401                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1402                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1403                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1404                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C]);
1405       break;
1406     case YUV_CONVERSION_FRAGMENT_POW:
1407       snprintf(prog_pos, prog_remain, yuv_pow_prog_template,
1408                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1409                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1410                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1411                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1412                (float)1.0 / params->csp_params.rgamma, (float)1.0 / params->csp_params.bgamma, (float)1.0 / params->csp_params.bgamma);
1413       break;
1414     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1415       snprintf(prog_pos, prog_remain, yuv_lookup_prog_template,
1416                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1417                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1418                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1419                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1420                conv_texs[0], conv_texs[0], conv_texs[0]);
1421       break;
1422     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1423       snprintf(prog_pos, prog_remain, yuv_lookup3d_prog_template, conv_texs[0]);
1424       break;
1425     default:
1426       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(type));
1427       break;
1428   }
1429   mp_msg(MSGT_VO, MSGL_DBG2, "[gl] generated fragment program:\n%s\n", yuv_prog);
1430   loadGPUProgram(GL_FRAGMENT_PROGRAM, yuv_prog);
1431   free(yuv_prog);
1432 }
1433
1434 /**
1435  * \brief detect the best YUV->RGB conversion method available
1436  */
1437 int glAutodetectYUVConversion(void) {
1438   const char *extensions = mpglGetString(GL_EXTENSIONS);
1439   if (!extensions || !mpglMultiTexCoord2f)
1440     return YUV_CONVERSION_NONE;
1441   if (strstr(extensions, "GL_ARB_fragment_program"))
1442     return YUV_CONVERSION_FRAGMENT;
1443   if (strstr(extensions, "GL_ATI_text_fragment_shader"))
1444     return YUV_CONVERSION_TEXT_FRAGMENT;
1445   if (strstr(extensions, "GL_ATI_fragment_shader"))
1446     return YUV_CONVERSION_COMBINERS_ATI;
1447   return YUV_CONVERSION_NONE;
1448 }
1449
1450 /**
1451  * \brief setup YUV->RGB conversion
1452  * \param parms struct containing parameters like conversion and scaler type,
1453  *              brightness, ...
1454  * \ingroup glconversion
1455  */
1456 void glSetupYUVConversion(gl_conversion_params_t *params) {
1457   float uvcos = params->csp_params.saturation * cos(params->csp_params.hue);
1458   float uvsin = params->csp_params.saturation * sin(params->csp_params.hue);
1459   if (params->chrom_texw == 0) params->chrom_texw = 1;
1460   if (params->chrom_texh == 0) params->chrom_texh = 1;
1461   switch (YUV_CONVERSION(params->type)) {
1462     case YUV_CONVERSION_COMBINERS:
1463       glSetupYUVCombiners(uvcos, uvsin);
1464       break;
1465     case YUV_CONVERSION_COMBINERS_ATI:
1466       glSetupYUVFragmentATI(&params->csp_params, 0);
1467       break;
1468     case YUV_CONVERSION_TEXT_FRAGMENT:
1469       glSetupYUVFragmentATI(&params->csp_params, 1);
1470       break;
1471     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1472     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1473     case YUV_CONVERSION_FRAGMENT:
1474     case YUV_CONVERSION_FRAGMENT_POW:
1475       glSetupYUVFragprog(params);
1476       break;
1477     case YUV_CONVERSION_NONE:
1478       break;
1479     default:
1480       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(params->type));
1481   }
1482 }
1483
1484 /**
1485  * \brief enable the specified YUV conversion
1486  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1487  * \param type type of YUV conversion
1488  * \ingroup glconversion
1489  */
1490 void glEnableYUVConversion(GLenum target, int type) {
1491   switch (YUV_CONVERSION(type)) {
1492     case YUV_CONVERSION_COMBINERS:
1493       mpglActiveTexture(GL_TEXTURE1);
1494       mpglEnable(target);
1495       mpglActiveTexture(GL_TEXTURE2);
1496       mpglEnable(target);
1497       mpglActiveTexture(GL_TEXTURE0);
1498       mpglEnable(GL_REGISTER_COMBINERS_NV);
1499       break;
1500     case YUV_CONVERSION_COMBINERS_ATI:
1501       mpglActiveTexture(GL_TEXTURE1);
1502       mpglEnable(target);
1503       mpglActiveTexture(GL_TEXTURE2);
1504       mpglEnable(target);
1505       mpglActiveTexture(GL_TEXTURE0);
1506       mpglEnable(GL_FRAGMENT_SHADER_ATI);
1507       break;
1508     case YUV_CONVERSION_TEXT_FRAGMENT:
1509       mpglActiveTexture(GL_TEXTURE1);
1510       mpglEnable(target);
1511       mpglActiveTexture(GL_TEXTURE2);
1512       mpglEnable(target);
1513       mpglActiveTexture(GL_TEXTURE0);
1514       mpglEnable(GL_TEXT_FRAGMENT_SHADER_ATI);
1515       break;
1516     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1517     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1518     case YUV_CONVERSION_FRAGMENT_POW:
1519     case YUV_CONVERSION_FRAGMENT:
1520     case YUV_CONVERSION_NONE:
1521       mpglEnable(GL_FRAGMENT_PROGRAM);
1522       break;
1523   }
1524 }
1525
1526 /**
1527  * \brief disable the specified YUV conversion
1528  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1529  * \param type type of YUV conversion
1530  * \ingroup glconversion
1531  */
1532 void glDisableYUVConversion(GLenum target, int type) {
1533   switch (YUV_CONVERSION(type)) {
1534     case YUV_CONVERSION_COMBINERS:
1535       mpglActiveTexture(GL_TEXTURE1);
1536       mpglDisable(target);
1537       mpglActiveTexture(GL_TEXTURE2);
1538       mpglDisable(target);
1539       mpglActiveTexture(GL_TEXTURE0);
1540       mpglDisable(GL_REGISTER_COMBINERS_NV);
1541       break;
1542     case YUV_CONVERSION_COMBINERS_ATI:
1543       mpglActiveTexture(GL_TEXTURE1);
1544       mpglDisable(target);
1545       mpglActiveTexture(GL_TEXTURE2);
1546       mpglDisable(target);
1547       mpglActiveTexture(GL_TEXTURE0);
1548       mpglDisable(GL_FRAGMENT_SHADER_ATI);
1549       break;
1550     case YUV_CONVERSION_TEXT_FRAGMENT:
1551       mpglDisable(GL_TEXT_FRAGMENT_SHADER_ATI);
1552       // HACK: at least the Mac OS X 10.5 PPC Radeon drivers are broken and
1553       // without this disable the texture units while the program is still
1554       // running (10.4 PPC seems to work without this though).
1555       mpglFlush();
1556       mpglActiveTexture(GL_TEXTURE1);
1557       mpglDisable(target);
1558       mpglActiveTexture(GL_TEXTURE2);
1559       mpglDisable(target);
1560       mpglActiveTexture(GL_TEXTURE0);
1561       break;
1562     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1563     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1564     case YUV_CONVERSION_FRAGMENT_POW:
1565     case YUV_CONVERSION_FRAGMENT:
1566     case YUV_CONVERSION_NONE:
1567       mpglDisable(GL_FRAGMENT_PROGRAM);
1568       break;
1569   }
1570 }
1571
1572 void glEnable3DLeft(int type) {
1573   GLint buffer;
1574   switch (type) {
1575     case GL_3D_RED_CYAN:
1576       mpglColorMask(GL_TRUE,  GL_FALSE, GL_FALSE, GL_FALSE);
1577       break;
1578     case GL_3D_GREEN_MAGENTA:
1579       mpglColorMask(GL_FALSE, GL_TRUE,  GL_FALSE, GL_FALSE);
1580       break;
1581     case GL_3D_QUADBUFFER:
1582       mpglGetIntegerv(GL_DRAW_BUFFER, &buffer);
1583       switch (buffer) {
1584         case GL_FRONT:
1585         case GL_FRONT_LEFT:
1586         case GL_FRONT_RIGHT:
1587           buffer = GL_FRONT_LEFT;
1588           break;
1589         case GL_BACK:
1590         case GL_BACK_LEFT:
1591         case GL_BACK_RIGHT:
1592           buffer = GL_BACK_LEFT;
1593           break;
1594       }
1595       mpglDrawBuffer(buffer);
1596       break;
1597   }
1598 }
1599
1600 void glEnable3DRight(int type) {
1601   GLint buffer;
1602   switch (type) {
1603     case GL_3D_RED_CYAN:
1604       mpglColorMask(GL_FALSE, GL_TRUE,  GL_TRUE,  GL_FALSE);
1605       break;
1606     case GL_3D_GREEN_MAGENTA:
1607       mpglColorMask(GL_TRUE,  GL_FALSE, GL_TRUE,  GL_FALSE);
1608       break;
1609     case GL_3D_QUADBUFFER:
1610       mpglGetIntegerv(GL_DRAW_BUFFER, &buffer);
1611       switch (buffer) {
1612         case GL_FRONT:
1613         case GL_FRONT_LEFT:
1614         case GL_FRONT_RIGHT:
1615           buffer = GL_FRONT_RIGHT;
1616           break;
1617         case GL_BACK:
1618         case GL_BACK_LEFT:
1619         case GL_BACK_RIGHT:
1620           buffer = GL_BACK_RIGHT;
1621           break;
1622       }
1623       mpglDrawBuffer(buffer);
1624       break;
1625   }
1626 }
1627
1628 void glDisable3D(int type) {
1629   GLint buffer;
1630   switch (type) {
1631     case GL_3D_RED_CYAN:
1632     case GL_3D_GREEN_MAGENTA:
1633       mpglColorMask(GL_TRUE,  GL_TRUE,  GL_TRUE,  GL_TRUE);
1634       break;
1635     case GL_3D_QUADBUFFER:
1636       mpglDrawBuffer(vo_doublebuffering ? GL_BACK : GL_FRONT);
1637       mpglGetIntegerv(GL_DRAW_BUFFER, &buffer);
1638       switch (buffer) {
1639         case GL_FRONT:
1640         case GL_FRONT_LEFT:
1641         case GL_FRONT_RIGHT:
1642           buffer = GL_FRONT;
1643           break;
1644         case GL_BACK:
1645         case GL_BACK_LEFT:
1646         case GL_BACK_RIGHT:
1647           buffer = GL_BACK;
1648           break;
1649       }
1650       mpglDrawBuffer(buffer);
1651       break;
1652   }
1653 }
1654
1655 /**
1656  * \brief draw a texture part at given 2D coordinates
1657  * \param x screen top coordinate
1658  * \param y screen left coordinate
1659  * \param w screen width coordinate
1660  * \param h screen height coordinate
1661  * \param tx texture top coordinate in pixels
1662  * \param ty texture left coordinate in pixels
1663  * \param tw texture part width in pixels
1664  * \param th texture part height in pixels
1665  * \param sx width of texture in pixels
1666  * \param sy height of texture in pixels
1667  * \param rect_tex whether this texture uses texture_rectangle extension
1668  * \param is_yv12 if != 0, also draw the textures from units 1 and 2,
1669  *                bits 8 - 15 and 16 - 23 specify the x and y scaling of those textures
1670  * \param flip flip the texture upside down
1671  * \ingroup gltexture
1672  */
1673 void glDrawTex(GLfloat x, GLfloat y, GLfloat w, GLfloat h,
1674                GLfloat tx, GLfloat ty, GLfloat tw, GLfloat th,
1675                int sx, int sy, int rect_tex, int is_yv12, int flip) {
1676   int chroma_x_shift = (is_yv12 >>  8) & 31;
1677   int chroma_y_shift = (is_yv12 >> 16) & 31;
1678   GLfloat xscale = 1 << chroma_x_shift;
1679   GLfloat yscale = 1 << chroma_y_shift;
1680   GLfloat tx2 = tx / xscale, ty2 = ty / yscale, tw2 = tw / xscale, th2 = th / yscale;
1681   if (!rect_tex) {
1682     tx /= sx; ty /= sy; tw /= sx; th /= sy;
1683     tx2 = tx, ty2 = ty, tw2 = tw, th2 = th;
1684   }
1685   if (flip) {
1686     y += h;
1687     h = -h;
1688   }
1689   mpglBegin(GL_QUADS);
1690   mpglTexCoord2f(tx, ty);
1691   if (is_yv12) {
1692     mpglMultiTexCoord2f(GL_TEXTURE1, tx2, ty2);
1693     mpglMultiTexCoord2f(GL_TEXTURE2, tx2, ty2);
1694   }
1695   mpglVertex2f(x, y);
1696   mpglTexCoord2f(tx, ty + th);
1697   if (is_yv12) {
1698     mpglMultiTexCoord2f(GL_TEXTURE1, tx2, ty2 + th2);
1699     mpglMultiTexCoord2f(GL_TEXTURE2, tx2, ty2 + th2);
1700   }
1701   mpglVertex2f(x, y + h);
1702   mpglTexCoord2f(tx + tw, ty + th);
1703   if (is_yv12) {
1704     mpglMultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2 + th2);
1705     mpglMultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2 + th2);
1706   }
1707   mpglVertex2f(x + w, y + h);
1708   mpglTexCoord2f(tx + tw, ty);
1709   if (is_yv12) {
1710     mpglMultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2);
1711     mpglMultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2);
1712   }
1713   mpglVertex2f(x + w, y);
1714   mpglEnd();
1715 }
1716
1717 #ifdef CONFIG_GL_WIN32
1718 #include "w32_common.h"
1719 /**
1720  * \brief little helper since wglGetProcAddress definition does not fit our
1721  *        getProcAddress
1722  * \param procName name of function to look up
1723  * \return function pointer returned by wglGetProcAddress
1724  */
1725 static void *w32gpa(const GLubyte *procName) {
1726   HMODULE oglmod;
1727   void *res = wglGetProcAddress(procName);
1728   if (res) return res;
1729   oglmod = GetModuleHandle("opengl32.dll");
1730   return GetProcAddress(oglmod, procName);
1731 }
1732
1733 static int setGlWindow_w32(MPGLContext *ctx)
1734 {
1735   HWND win = vo_w32_window;
1736   int *vinfo = &ctx->vinfo.w32;
1737   HGLRC *context = &ctx->context.w32;
1738   int new_vinfo;
1739   HDC windc = vo_w32_get_dc(win);
1740   HGLRC new_context = 0;
1741   int keep_context = 0;
1742   int res = SET_WINDOW_FAILED;
1743
1744   // should only be needed when keeping context, but not doing glFinish
1745   // can cause flickering even when we do not keep it.
1746   if (*context)
1747   mpglFinish();
1748   new_vinfo = GetPixelFormat(windc);
1749   if (*context && *vinfo && new_vinfo && *vinfo == new_vinfo) {
1750       // we can keep the wglContext
1751       new_context = *context;
1752       keep_context = 1;
1753   } else {
1754     // create a context
1755     new_context = wglCreateContext(windc);
1756     if (!new_context) {
1757       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GL context!\n");
1758       goto out;
1759     }
1760   }
1761
1762   // set context
1763   if (!wglMakeCurrent(windc, new_context)) {
1764     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GL context!\n");
1765     if (!keep_context) {
1766       wglDeleteContext(new_context);
1767     }
1768     goto out;
1769   }
1770
1771   // set new values
1772   vo_w32_window = win;
1773   {
1774     RECT rect;
1775     GetClientRect(win, &rect);
1776     vo_dwidth = rect.right;
1777     vo_dheight = rect.bottom;
1778   }
1779   if (!keep_context) {
1780     if (*context)
1781       wglDeleteContext(*context);
1782     *context = new_context;
1783     *vinfo = new_vinfo;
1784     getFunctions(w32gpa, NULL);
1785
1786     // and inform that reinit is neccessary
1787     res = SET_WINDOW_REINIT;
1788   } else
1789     res = SET_WINDOW_OK;
1790
1791 out:
1792   vo_w32_release_dc(win, windc);
1793   return res;
1794 }
1795
1796 static void releaseGlContext_w32(MPGLContext *ctx) {
1797   int *vinfo = &ctx->vinfo.w32;
1798   HGLRC *context = &ctx->context.w32;
1799   *vinfo = 0;
1800   if (*context) {
1801     wglMakeCurrent(0, 0);
1802     wglDeleteContext(*context);
1803   }
1804   *context = 0;
1805 }
1806
1807 static void swapGlBuffers_w32(MPGLContext *ctx) {
1808   HDC vo_hdc = vo_w32_get_dc(vo_w32_window);
1809   SwapBuffers(vo_hdc);
1810   vo_w32_release_dc(vo_w32_window, vo_hdc);
1811 }
1812 #endif
1813 #ifdef CONFIG_GL_X11
1814 #include "x11_common.h"
1815
1816 /**
1817  * \brief Returns the XVisualInfo associated with Window win.
1818  * \param win Window whose XVisualInfo is returne.
1819  * \return XVisualInfo of the window. Caller must use XFree to free it.
1820  */
1821 static XVisualInfo *getWindowVisualInfo(Window win) {
1822   XWindowAttributes xw_attr;
1823   XVisualInfo vinfo_template;
1824   int tmp;
1825   XGetWindowAttributes(mDisplay, win, &xw_attr);
1826   vinfo_template.visualid = XVisualIDFromVisual(xw_attr.visual);
1827   return XGetVisualInfo(mDisplay, VisualIDMask, &vinfo_template, &tmp);
1828 }
1829
1830 static void appendstr(char **dst, const char *str)
1831 {
1832     int newsize;
1833     char *newstr;
1834     if (!str)
1835         return;
1836     newsize = strlen(*dst) + 1 + strlen(str) + 1;
1837     newstr = realloc(*dst, newsize);
1838     if (!newstr)
1839         return;
1840     *dst = newstr;
1841     strcat(*dst, " ");
1842     strcat(*dst, str);
1843 }
1844
1845 /**
1846  * \brief Changes the window in which video is displayed.
1847  * If possible only transfers the context to the new window, otherwise
1848  * creates a new one, which must be initialized by the caller.
1849  * \param vinfo Currently used visual.
1850  * \param context Currently used context.
1851  * \param win window that should be used for drawing.
1852  * \return one of SET_WINDOW_FAILED, SET_WINDOW_OK or SET_WINDOW_REINIT.
1853  * In case of SET_WINDOW_REINIT the context could not be transfered
1854  * and the caller must initialize it correctly.
1855  * \ingroup glcontext
1856  */
1857 static int setGlWindow_x11(MPGLContext *ctx)
1858 {
1859   XVisualInfo **vinfo = &ctx->vinfo.x11;
1860   GLXContext *context = &ctx->context.x11;
1861   Window win = vo_window;
1862   XVisualInfo *new_vinfo;
1863   GLXContext new_context = NULL;
1864   int keep_context = 0;
1865
1866   // should only be needed when keeping context, but not doing glFinish
1867   // can cause flickering even when we do not keep it.
1868   if (*context)
1869   mpglFinish();
1870   new_vinfo = getWindowVisualInfo(win);
1871   if (*context && *vinfo && new_vinfo &&
1872       (*vinfo)->visualid == new_vinfo->visualid) {
1873       // we can keep the GLXContext
1874       new_context = *context;
1875       XFree(new_vinfo);
1876       new_vinfo = *vinfo;
1877       keep_context = 1;
1878   } else {
1879     // create a context
1880     new_context = glXCreateContext(mDisplay, new_vinfo, NULL, True);
1881     if (!new_context) {
1882       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GLX context!\n");
1883       XFree(new_vinfo);
1884       return SET_WINDOW_FAILED;
1885     }
1886   }
1887
1888   // set context
1889   if (!glXMakeCurrent(mDisplay, vo_window, new_context)) {
1890     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GLX context!\n");
1891     if (!keep_context) {
1892       glXDestroyContext(mDisplay, new_context);
1893       XFree(new_vinfo);
1894     }
1895     return SET_WINDOW_FAILED;
1896   }
1897
1898   // set new values
1899   vo_window = win;
1900   vo_x11_update_geometry();
1901   if (!keep_context) {
1902     void *(*getProcAddress)(const GLubyte *);
1903     const char *(*glXExtStr)(Display *, int);
1904     char *glxstr = strdup("");
1905     if (*context)
1906       glXDestroyContext(mDisplay, *context);
1907     *context = new_context;
1908     if (*vinfo)
1909       XFree(*vinfo);
1910     *vinfo = new_vinfo;
1911     getProcAddress = getdladdr("glXGetProcAddress");
1912     if (!getProcAddress)
1913       getProcAddress = getdladdr("glXGetProcAddressARB");
1914     glXExtStr = getdladdr("glXQueryExtensionsString");
1915     if (glXExtStr)
1916         appendstr(&glxstr, glXExtStr(mDisplay, DefaultScreen(mDisplay)));
1917     glXExtStr = getdladdr("glXGetClientString");
1918     if (glXExtStr)
1919         appendstr(&glxstr, glXExtStr(mDisplay, GLX_EXTENSIONS));
1920     glXExtStr = getdladdr("glXGetServerString");
1921     if (glXExtStr)
1922         appendstr(&glxstr, glXExtStr(mDisplay, GLX_EXTENSIONS));
1923
1924     getFunctions(getProcAddress, glxstr);
1925     if (!mpglGenPrograms && mpglGetString &&
1926         getProcAddress &&
1927         strstr(mpglGetString(GL_EXTENSIONS), "GL_ARB_vertex_program")) {
1928       mp_msg(MSGT_VO, MSGL_WARN, "Broken glXGetProcAddress detected, trying workaround\n");
1929       getFunctions(NULL, glxstr);
1930     }
1931     free(glxstr);
1932
1933     // and inform that reinit is neccessary
1934     return SET_WINDOW_REINIT;
1935   }
1936   return SET_WINDOW_OK;
1937 }
1938
1939 /**
1940  * \brief free the VisualInfo and GLXContext of an OpenGL context.
1941  * \ingroup glcontext
1942  */
1943 static void releaseGlContext_x11(MPGLContext *ctx) {
1944   XVisualInfo **vinfo = &ctx->vinfo.x11;
1945   GLXContext *context = &ctx->context.x11;
1946   if (*vinfo)
1947     XFree(*vinfo);
1948   *vinfo = NULL;
1949   if (*context)
1950   {
1951     mpglFinish();
1952     glXMakeCurrent(mDisplay, None, NULL);
1953     glXDestroyContext(mDisplay, *context);
1954   }
1955   *context = 0;
1956 }
1957
1958 static void swapGlBuffers_x11(MPGLContext *ctx) {
1959   glXSwapBuffers(mDisplay, vo_window);
1960 }
1961
1962 static int x11_check_events(void) {
1963   return vo_x11_check_events(mDisplay);
1964 }
1965
1966 static void gl_update_xinerama_info(void)
1967 {
1968     update_xinerama_info(global_vo);
1969 }
1970
1971 static void gl_border(void)
1972 {
1973     vo_x11_border();
1974 }
1975
1976 static void gl_fullscreen(void)
1977 {
1978     vo_x11_fullscreen();
1979 }
1980
1981 static void gl_ontop(void)
1982 {
1983     vo_x11_ontop();
1984 }
1985 #endif
1986
1987 #ifdef CONFIG_GL_SDL
1988 #include "sdl_common.h"
1989
1990 static void swapGlBuffers_sdl(MPGLContext *ctx) {
1991   SDL_GL_SwapBuffers();
1992 }
1993
1994 static void *sdlgpa(const GLubyte *name) {
1995   return SDL_GL_GetProcAddress(name);
1996 }
1997
1998 static int setGlWindow_sdl(MPGLContext *ctx) {
1999   if (sdl_set_mode(0, SDL_OPENGL | SDL_RESIZABLE) < 0)
2000     return SET_WINDOW_FAILED;
2001   SDL_GL_LoadLibrary(NULL);
2002   getFunctions(sdlgpa, NULL);
2003   return SET_WINDOW_OK;
2004 }
2005
2006 static int sdl_check_events(void) {
2007   int res = 0;
2008   SDL_Event event;
2009   while (SDL_PollEvent(&event)) {
2010     res |= sdl_default_handle_event(&event);
2011   }
2012   // poll "events" from within MPlayer code
2013   res |= sdl_default_handle_event(NULL);
2014   if (res & VO_EVENT_RESIZE)
2015     sdl_set_mode(0, SDL_OPENGL | SDL_RESIZABLE);
2016   return res;
2017 }
2018
2019 #endif
2020
2021 static int setGlWindow_dummy(MPGLContext *ctx) {
2022   getFunctions(NULL, NULL);
2023   return SET_WINDOW_OK;
2024 }
2025
2026 static void releaseGlContext_dummy(MPGLContext *ctx) {
2027 }
2028
2029 static int dummy_check_events(void) {
2030   return 0;
2031 }
2032
2033 static void dummy_update_xinerama_info(void) {
2034   if (vo_screenwidth <= 0 || vo_screenheight <= 0) {
2035     mp_msg(MSGT_VO, MSGL_ERR, "You must specify the screen dimensions "
2036                               "with -screenw and -screenh\n");
2037     vo_screenwidth  = 1280;
2038     vo_screenheight = 768;
2039   }
2040   aspect_save_screenres(vo_screenwidth, vo_screenheight);
2041 }
2042
2043 int init_mpglcontext(MPGLContext *ctx, enum MPGLType type) {
2044   if (type == GLTYPE_AUTO) {
2045     int res = init_mpglcontext(ctx, GLTYPE_W32);
2046     if (res) return res;
2047     res = init_mpglcontext(ctx, GLTYPE_X11);
2048     if (res) return res;
2049     res = init_mpglcontext(ctx, GLTYPE_SDL);
2050     return res;
2051   }
2052   memset(ctx, 0, sizeof(*ctx));
2053   ctx->setGlWindow = setGlWindow_dummy;
2054   ctx->releaseGlContext = releaseGlContext_dummy;
2055   ctx->update_xinerama_info = dummy_update_xinerama_info;
2056   ctx->check_events = dummy_check_events;
2057   ctx->type = type;
2058   switch (ctx->type) {
2059 #ifdef CONFIG_GL_WIN32
2060   case GLTYPE_W32:
2061     ctx->setGlWindow = setGlWindow_w32;
2062     ctx->releaseGlContext = releaseGlContext_w32;
2063     ctx->swapGlBuffers = swapGlBuffers_w32;
2064     ctx->update_xinerama_info = w32_update_xinerama_info;
2065     ctx->border = vo_w32_border;
2066     ctx->check_events = vo_w32_check_events;
2067     ctx->fullscreen = vo_w32_fullscreen;
2068     ctx->ontop = vo_w32_ontop;
2069     return vo_w32_init();
2070 #endif
2071 #ifdef CONFIG_GL_X11
2072   case GLTYPE_X11:
2073     ctx->setGlWindow = setGlWindow_x11;
2074     ctx->releaseGlContext = releaseGlContext_x11;
2075     ctx->swapGlBuffers = swapGlBuffers_x11;
2076     ctx->update_xinerama_info = gl_update_xinerama_info;
2077     ctx->border = gl_border;
2078     ctx->check_events = x11_check_events;
2079     ctx->fullscreen = gl_fullscreen;
2080     ctx->ontop = gl_ontop;
2081     return vo_init();
2082 #endif
2083 #ifdef CONFIG_GL_SDL
2084   case GLTYPE_SDL:
2085     SDL_Init(SDL_INIT_VIDEO);
2086     ctx->setGlWindow = setGlWindow_sdl;
2087     ctx->swapGlBuffers = swapGlBuffers_sdl;
2088     ctx->check_events = sdl_check_events;
2089     ctx->fullscreen = vo_sdl_fullscreen;
2090     return vo_sdl_init();
2091 #endif
2092   default:
2093     return 0;
2094   }
2095 }
2096
2097 void uninit_mpglcontext(MPGLContext *ctx) {
2098   ctx->releaseGlContext(ctx);
2099   switch (ctx->type) {
2100 #ifdef CONFIG_GL_WIN32
2101   case GLTYPE_W32:
2102     vo_w32_uninit();
2103     break;
2104 #endif
2105 #ifdef CONFIG_GL_X11
2106   case GLTYPE_X11:
2107     vo_x11_uninit();
2108     break;
2109 #endif
2110 #ifdef CONFIG_GL_SDL
2111   case GLTYPE_SDL:
2112     vo_sdl_uninit();
2113     break;
2114 #endif
2115   }
2116   memset(ctx, 0, sizeof(*ctx));
2117 }