libvo/gl_common.c

   1 /*
   2  * common OpenGL routines
   3  *
   4  * copyleft (C) 2005 Reimar Döffinger <Reimar.Doeffinger@stud.uni-karlsruhe.de>
   5  * Special thanks go to the xine team and Matthias Hopf, whose video_out_opengl.c
   6  * gave me lots of good ideas.
   7  *
   8  * This file is part of MPlayer.
   9  *
  10  * MPlayer is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published by
  12  * the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * MPlayer is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  * GNU General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public License along
  21  * with MPlayer; if not, write to the Free Software Foundation, Inc.,
  22  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  23  */
  24
  25 /**
  26  * \file gl_common.c
  27  * \brief OpenGL helper functions used by vo_gl.c and vo_gl2.c
  28  */
  29
  30 #include <stdlib.h>
  31 #include <stdio.h>
  32 #include <string.h>
  33 #include <ctype.h>
  34 #include <math.h>
  35 #include "old_vo_defines.h"
  36 #include "gl_common.h"
  37 #include "libavutil/common.h"
  38
  39 /**
  40  * \defgroup glextfunctions OpenGL extension functions
  41  *
  42  * the pointers to these functions are acquired when the OpenGL
  43  * context is created
  44  * \{
  45  */
  46 void (APIENTRY *GenBuffers)(GLsizei, GLuint *);
  47 void (APIENTRY *DeleteBuffers)(GLsizei, const GLuint *);
  48 void (APIENTRY *BindBuffer)(GLenum, GLuint);
  49 GLvoid* (APIENTRY *MapBuffer)(GLenum, GLenum);
  50 GLboolean (APIENTRY *UnmapBuffer)(GLenum);
  51 void (APIENTRY *BufferData)(GLenum, intptr_t, const GLvoid *, GLenum);
  52 void (APIENTRY *CombinerParameterfv)(GLenum, const GLfloat *);
  53 void (APIENTRY *CombinerParameteri)(GLenum, GLint);
  54 void (APIENTRY *CombinerInput)(GLenum, GLenum, GLenum, GLenum, GLenum,
  55                                GLenum);
  56 void (APIENTRY *CombinerOutput)(GLenum, GLenum, GLenum, GLenum, GLenum,
  57                                 GLenum, GLenum, GLboolean, GLboolean,
  58                                 GLboolean);
  59 void (APIENTRY *BeginFragmentShader)(void);
  60 void (APIENTRY *EndFragmentShader)(void);
  61 void (APIENTRY *SampleMap)(GLuint, GLuint, GLenum);
  62 void (APIENTRY *ColorFragmentOp2)(GLenum, GLuint, GLuint, GLuint, GLuint,
  63                                   GLuint, GLuint, GLuint, GLuint, GLuint);
  64 void (APIENTRY *ColorFragmentOp3)(GLenum, GLuint, GLuint, GLuint, GLuint,
  65                                   GLuint, GLuint, GLuint, GLuint, GLuint,
  66                                   GLuint, GLuint, GLuint);
  67 void (APIENTRY *SetFragmentShaderConstant)(GLuint, const GLfloat *);
  68 void (APIENTRY *ActiveTexture)(GLenum);
  69 void (APIENTRY *BindTexture)(GLenum, GLuint);
  70 void (APIENTRY *MultiTexCoord2f)(GLenum, GLfloat, GLfloat);
  71 void (APIENTRY *GenPrograms)(GLsizei, GLuint *);
  72 void (APIENTRY *DeletePrograms)(GLsizei, const GLuint *);
  73 void (APIENTRY *BindProgram)(GLenum, GLuint);
  74 void (APIENTRY *ProgramString)(GLenum, GLenum, GLsizei, const GLvoid *);
  75 void (APIENTRY *GetProgramiv)(GLenum, GLenum, GLint *);
  76 void (APIENTRY *ProgramEnvParameter4f)(GLenum, GLuint, GLfloat, GLfloat,
  77                                        GLfloat, GLfloat);
  78 int (APIENTRY *SwapInterval)(int);
  79 void (APIENTRY *TexImage3D)(GLenum, GLint, GLenum, GLsizei, GLsizei, GLsizei,
  80                             GLint, GLenum, GLenum, const GLvoid *);
  81 void* (APIENTRY *AllocateMemoryMESA)(void *, int, size_t, float, float, float);
  82 void (APIENTRY *FreeMemoryMESA)(void *, int, void *);
  83 /** \} */ // end of glextfunctions group
  84
  85 //! \defgroup glgeneral OpenGL general helper functions
  86
  87 //! \defgroup glcontext OpenGL context management helper functions
  88
  89 //! \defgroup gltexture OpenGL texture handling helper functions
  90
  91 //! \defgroup glconversion OpenGL conversion helper functions
  92
  93 static GLint hqtexfmt;
  94
  95 /**
  96  * \brief adjusts the GL_UNPACK_ALIGNMENT to fit the stride.
  97  * \param stride number of bytes per line for which alignment should fit.
  98  * \ingroup glgeneral
  99  */
 100 void glAdjustAlignment(int stride) {
 101   GLint gl_alignment;
 102   if (stride % 8 == 0)
 103     gl_alignment=8;
 104   else if (stride % 4 == 0)
 105     gl_alignment=4;
 106   else if (stride % 2 == 0)
 107     gl_alignment=2;
 108   else
 109     gl_alignment=1;
 110   glPixelStorei (GL_UNPACK_ALIGNMENT, gl_alignment);
 111 }
 112
 113 struct gl_name_map_struct {
 114   GLint value;
 115   const char *name;
 116 };
 117
 118 #undef MAP
 119 #define MAP(a) {a, #a}
 120 //! mapping table for the glValName function
 121 static const struct gl_name_map_struct gl_name_map[] = {
 122   // internal format
 123   MAP(GL_R3_G3_B2), MAP(GL_RGB4), MAP(GL_RGB5), MAP(GL_RGB8),
 124   MAP(GL_RGB10), MAP(GL_RGB12), MAP(GL_RGB16), MAP(GL_RGBA2),
 125   MAP(GL_RGBA4), MAP(GL_RGB5_A1), MAP(GL_RGBA8), MAP(GL_RGB10_A2),
 126   MAP(GL_RGBA12), MAP(GL_RGBA16), MAP(GL_LUMINANCE8),
 127
 128   // format
 129   MAP(GL_RGB), MAP(GL_RGBA), MAP(GL_RED), MAP(GL_GREEN), MAP(GL_BLUE),
 130   MAP(GL_ALPHA), MAP(GL_LUMINANCE), MAP(GL_LUMINANCE_ALPHA),
 131   MAP(GL_COLOR_INDEX),
 132   // rest 1.2 only
 133   MAP(GL_BGR), MAP(GL_BGRA),
 134
 135   //type
 136   MAP(GL_BYTE), MAP(GL_UNSIGNED_BYTE), MAP(GL_SHORT), MAP(GL_UNSIGNED_SHORT),
 137   MAP(GL_INT), MAP(GL_UNSIGNED_INT), MAP(GL_FLOAT), MAP(GL_DOUBLE),
 138   MAP(GL_2_BYTES), MAP(GL_3_BYTES), MAP(GL_4_BYTES),
 139   // rest 1.2 only
 140   MAP(GL_UNSIGNED_BYTE_3_3_2), MAP(GL_UNSIGNED_BYTE_2_3_3_REV),
 141   MAP(GL_UNSIGNED_SHORT_5_6_5), MAP(GL_UNSIGNED_SHORT_5_6_5_REV),
 142   MAP(GL_UNSIGNED_SHORT_4_4_4_4), MAP(GL_UNSIGNED_SHORT_4_4_4_4_REV),
 143   MAP(GL_UNSIGNED_SHORT_5_5_5_1), MAP(GL_UNSIGNED_SHORT_1_5_5_5_REV),
 144   MAP(GL_UNSIGNED_INT_8_8_8_8), MAP(GL_UNSIGNED_INT_8_8_8_8_REV),
 145   MAP(GL_UNSIGNED_INT_10_10_10_2), MAP(GL_UNSIGNED_INT_2_10_10_10_REV),
 146   {0, 0}
 147 };
 148 #undef MAP
 149
 150 /**
 151  * \brief return the name of an OpenGL constant
 152  * \param value the constant
 153  * \return name of the constant or "Unknown format!"
 154  * \ingroup glgeneral
 155  */
 156 const char *glValName(GLint value)
 157 {
 158   int i = 0;
 159
 160   while (gl_name_map[i].name) {
 161     if (gl_name_map[i].value == value)
 162       return gl_name_map[i].name;
 163     i++;
 164   }
 165   return "Unknown format!";
 166 }
 167
 168 //! always return this format as internal texture format in glFindFormat
 169 #define TEXTUREFORMAT_ALWAYS GL_RGB8
 170 #undef TEXTUREFORMAT_ALWAYS
 171
 172 /**
 173  * \brief find the OpenGL settings coresponding to format.
 174  *
 175  * All parameters may be NULL.
 176  * \param fmt MPlayer format to analyze.
 177  * \param bpp [OUT] bits per pixel of that format.
 178  * \param gl_texfmt [OUT] internal texture format that fits the
 179  * image format, not necessarily the best for performance.
 180  * \param gl_format [OUT] OpenGL format for this image format.
 181  * \param gl_type [OUT] OpenGL type for this image format.
 182  * \return 1 if format is supported by OpenGL, 0 if not.
 183  * \ingroup gltexture
 184  */
 185 int glFindFormat(uint32_t fmt, int *bpp, GLint *gl_texfmt,
 186                   GLenum *gl_format, GLenum *gl_type)
 187 {
 188   int supported = 1;
 189   int dummy1;
 190   GLenum dummy2;
 191   GLint dummy3;
 192   if (!bpp) bpp = &dummy1;
 193   if (!gl_texfmt) gl_texfmt = &dummy3;
 194   if (!gl_format) gl_format = &dummy2;
 195   if (!gl_type) gl_type = &dummy2;
 196
 197   *bpp = IMGFMT_IS_BGR(fmt)?IMGFMT_BGR_DEPTH(fmt):IMGFMT_RGB_DEPTH(fmt);
 198   *gl_texfmt = 3;
 199   switch (fmt) {
 200     case IMGFMT_RGB24:
 201       *gl_format = GL_RGB;
 202       *gl_type = GL_UNSIGNED_BYTE;
 203       break;
 204     case IMGFMT_RGBA:
 205       *gl_texfmt = 4;
 206       *gl_format = GL_RGBA;
 207       *gl_type = GL_UNSIGNED_BYTE;
 208       break;
 209     case IMGFMT_YV12:
 210       supported = 0; // no native YV12 support
 211     case IMGFMT_Y800:
 212     case IMGFMT_Y8:
 213       *gl_texfmt = 1;
 214       *bpp = 8;
 215       *gl_format = GL_LUMINANCE;
 216       *gl_type = GL_UNSIGNED_BYTE;
 217       break;
 218     case IMGFMT_UYVY:
 219     case IMGFMT_YUY2:
 220       *gl_texfmt = GL_YCBCR_MESA;
 221       *bpp = 16;
 222       *gl_format = GL_YCBCR_MESA;
 223       *gl_type = fmt == IMGFMT_UYVY ? GL_UNSIGNED_SHORT_8_8 : GL_UNSIGNED_SHORT_8_8_REV;
 224       break;
 225 #if 0
 226     // we do not support palettized formats, although the format the
 227     // swscale produces works
 228     case IMGFMT_RGB8:
 229       gl_format = GL_RGB;
 230       gl_type = GL_UNSIGNED_BYTE_2_3_3_REV;
 231       break;
 232 #endif
 233     case IMGFMT_RGB15:
 234       *gl_format = GL_RGBA;
 235       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 236       break;
 237     case IMGFMT_RGB16:
 238       *gl_format = GL_RGB;
 239       *gl_type = GL_UNSIGNED_SHORT_5_6_5_REV;
 240       break;
 241 #if 0
 242     case IMGFMT_BGR8:
 243       // special case as red and blue have a differen number of bits.
 244       // GL_BGR and GL_UNSIGNED_BYTE_3_3_2 isn't supported at least
 245       // by nVidia drivers, and in addition would give more bits to
 246       // blue than to red, which isn't wanted
 247       gl_format = GL_RGB;
 248       gl_type = GL_UNSIGNED_BYTE_3_3_2;
 249       break;
 250 #endif
 251     case IMGFMT_BGR15:
 252       *gl_format = GL_BGRA;
 253       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 254       break;
 255     case IMGFMT_BGR16:
 256       *gl_format = GL_RGB;
 257       *gl_type = GL_UNSIGNED_SHORT_5_6_5;
 258       break;
 259     case IMGFMT_BGR24:
 260       *gl_format = GL_BGR;
 261       *gl_type = GL_UNSIGNED_BYTE;
 262       break;
 263     case IMGFMT_BGRA:
 264       *gl_texfmt = 4;
 265       *gl_format = GL_BGRA;
 266       *gl_type = GL_UNSIGNED_BYTE;
 267       break;
 268     default:
 269       *gl_texfmt = 4;
 270       *gl_format = GL_RGBA;
 271       *gl_type = GL_UNSIGNED_BYTE;
 272       supported = 0;
 273   }
 274 #ifdef TEXTUREFORMAT_ALWAYS
 275   *gl_texfmt = TEXTUREFORMAT_ALWAYS;
 276 #endif
 277   return supported;
 278 }
 279
 280 static void *setNull(const GLubyte *s) {
 281   return NULL;
 282 }
 283
 284 typedef struct {
 285   void *funcptr;
 286   const char *extstr;
 287   const char *funcnames[7];
 288 } extfunc_desc_t;
 289
 290 static const extfunc_desc_t extfuncs[] = {
 291   {&GenBuffers, NULL, {"glGenBuffers", "glGenBuffersARB", NULL}},
 292   {&DeleteBuffers, NULL, {"glDeleteBuffers", "glDeleteBuffersARB", NULL}},
 293   {&BindBuffer, NULL, {"glBindBuffer", "glBindBufferARB", NULL}},
 294   {&MapBuffer, NULL, {"glMapBuffer", "glMapBufferARB", NULL}},
 295   {&UnmapBuffer, NULL, {"glUnmapBuffer", "glUnmapBufferARB", NULL}},
 296   {&BufferData, NULL, {"glBufferData", "glBufferDataARB", NULL}},
 297   {&CombinerParameterfv, "NV_register_combiners", {"glCombinerParameterfv", "glCombinerParameterfvNV", NULL}},
 298   {&CombinerParameteri, "NV_register_combiners", {"glCombinerParameteri", "glCombinerParameteriNV", NULL}},
 299   {&CombinerInput, "NV_register_combiners", {"glCombinerInput", "glCombinerInputNV", NULL}},
 300   {&CombinerOutput, "NV_register_combiners", {"glCombinerOutput", "glCombinerOutputNV", NULL}},
 301   {&BeginFragmentShader, "ATI_fragment_shader", {"glBeginFragmentShaderATI", NULL}},
 302   {&EndFragmentShader, "ATI_fragment_shader", {"glEndFragmentShaderATI", NULL}},
 303   {&SampleMap, "ATI_fragment_shader", {"glSampleMapATI", NULL}},
 304   {&ColorFragmentOp2, "ATI_fragment_shader", {"glColorFragmentOp2ATI", NULL}},
 305   {&ColorFragmentOp3, "ATI_fragment_shader", {"glColorFragmentOp3ATI", NULL}},
 306   {&SetFragmentShaderConstant, "ATI_fragment_shader", {"glSetFragmentShaderConstantATI", NULL}},
 307   {&ActiveTexture, NULL, {"glActiveTexture", "glActiveTextureARB", NULL}},
 308   {&BindTexture, NULL, {"glBindTexture", "glBindTextureARB", "glBindTextureEXT", NULL}},
 309   {&MultiTexCoord2f, NULL, {"glMultiTexCoord2f", "glMultiTexCoord2fARB", NULL}},
 310   {&GenPrograms, "_program", {"glGenProgramsARB", NULL}},
 311   {&DeletePrograms, "_program", {"glDeleteProgramsARB", NULL}},
 312   {&BindProgram, "_program", {"glBindProgramARB", NULL}},
 313   {&ProgramString, "_program", {"glProgramStringARB", NULL}},
 314   {&GetProgramiv, "_program", {"glGetProgramivARB", NULL}},
 315   {&ProgramEnvParameter4f, "_program", {"glProgramEnvParameter4fARB", NULL}},
 316   {&SwapInterval, "_swap_control", {"glXSwapInterval", "glXSwapIntervalEXT", "glXSwapIntervalSGI", "wglSwapInterval", "wglSwapIntervalEXT", "wglSwapIntervalSGI", NULL}},
 317   {&TexImage3D, NULL, {"glTexImage3D", NULL}},
 318   {&AllocateMemoryMESA, "GLX_MESA_allocate_memory", {"glXAllocateMemoryMESA", NULL}},
 319   {&FreeMemoryMESA, "GLX_MESA_allocate_memory", {"glXFreeMemoryMESA", NULL}},
 320   {NULL}
 321 };
 322
 323 /**
 324  * \brief find the function pointers of some useful OpenGL extensions
 325  * \param getProcAddress function to resolve function names, may be NULL
 326  * \param ext2 an extra extension string
 327  */
 328 static void getFunctions(void *(*getProcAddress)(const GLubyte *),
 329                          const char *ext2) {
 330   const extfunc_desc_t *dsc;
 331   const char *extensions = (const char *)glGetString(GL_EXTENSIONS);
 332   char *allexts;
 333   if (!extensions) extensions = "";
 334   if (!ext2) ext2 = "";
 335   allexts = malloc(strlen(extensions) + strlen(ext2) + 2);
 336   strcpy(allexts, extensions);
 337   strcat(allexts, " ");
 338   strcat(allexts, ext2);
 339   mp_msg(MSGT_VO, MSGL_V, "OpenGL extensions string:\n%s\n", allexts);
 340   if (!getProcAddress)
 341     getProcAddress = setNull;
 342   for (dsc = extfuncs; dsc->funcptr; dsc++) {
 343     void *ptr = NULL;
 344     int i;
 345     if (!dsc->extstr || strstr(allexts, dsc->extstr)) {
 346       for (i = 0; !ptr && dsc->funcnames[i]; i++)
 347         ptr = getProcAddress((const GLubyte *)dsc->funcnames[i]);
 348     }
 349     *(void **)dsc->funcptr = ptr;
 350   }
 351   if (strstr(allexts, "_texture_float"))
 352     hqtexfmt = GL_RGB32F;
 353   else if (strstr(allexts, "NV_float_buffer"))
 354     hqtexfmt = GL_FLOAT_RGB32_NV;
 355   else
 356     hqtexfmt = GL_RGB16;
 357   free(allexts);
 358 }
 359
 360 /**
 361  * \brief create a texture and set some defaults
 362  * \param target texture taget, usually GL_TEXTURE_2D
 363  * \param fmt internal texture format
 364  * \param format texture host data format
 365  * \param type texture host data type
 366  * \param filter filter used for scaling, e.g. GL_LINEAR
 367  * \param w texture width
 368  * \param h texture height
 369  * \param val luminance value to fill texture with
 370  * \ingroup gltexture
 371  */
 372 void glCreateClearTex(GLenum target, GLenum fmt, GLenum format, GLenum type, GLint filter,
 373                       int w, int h, unsigned char val) {
 374   GLfloat fval = (GLfloat)val / 255.0;
 375   GLfloat border[4] = {fval, fval, fval, fval};
 376   int stride = w * glFmt2bpp(format, type);
 377   char *init;
 378   if (!stride) return;
 379   init = malloc(stride * h);
 380   memset(init, val, stride * h);
 381   glAdjustAlignment(stride);
 382   glPixelStorei(GL_UNPACK_ROW_LENGTH, w);
 383   glTexImage2D(target, 0, fmt, w, h, 0, format, type, init);
 384   glTexParameterf(target, GL_TEXTURE_PRIORITY, 1.0);
 385   glTexParameteri(target, GL_TEXTURE_MIN_FILTER, filter);
 386   glTexParameteri(target, GL_TEXTURE_MAG_FILTER, filter);
 387   glTexParameteri(target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 388   glTexParameteri(target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 389   // Border texels should not be used with CLAMP_TO_EDGE
 390   // We set a sane default anyway.
 391   glTexParameterfv(target, GL_TEXTURE_BORDER_COLOR, border);
 392   free(init);
 393 }
 394
 395 /**
 396  * \brief skips whitespace and comments
 397  * \param f file to read from
 398  */
 399 static void ppm_skip(FILE *f) {
 400   int c, comment = 0;
 401   do {
 402     c = fgetc(f);
 403     if (c == '#')
 404       comment = 1;
 405     if (c == '\n')
 406       comment = 0;
 407   } while (c != EOF && (isspace(c) || comment));
 408   if (c != EOF)
 409     ungetc(c, f);
 410 }
 411
 412 #define MAXDIM (16 * 1024)
 413
 414 /**
 415  * \brief creates a texture from a PPM file
 416  * \param target texture taget, usually GL_TEXTURE_2D
 417  * \param fmt internal texture format, 0 for default
 418  * \param filter filter used for scaling, e.g. GL_LINEAR
 419  * \param f file to read PPM from
 420  * \param width [out] width of texture
 421  * \param height [out] height of texture
 422  * \param maxval [out] maxval value from PPM file
 423  * \return 0 on error, 1 otherwise
 424  * \ingroup gltexture
 425  */
 426 int glCreatePPMTex(GLenum target, GLenum fmt, GLint filter,
 427                    FILE *f, int *width, int *height, int *maxval) {
 428   unsigned w, h, m, val, bpp;
 429   char *data;
 430   GLenum type;
 431   ppm_skip(f);
 432   if (fgetc(f) != 'P' || fgetc(f) != '6')
 433     return 0;
 434   ppm_skip(f);
 435   if (fscanf(f, "%u", &w) != 1)
 436     return 0;
 437   ppm_skip(f);
 438   if (fscanf(f, "%u", &h) != 1)
 439     return 0;
 440   ppm_skip(f);
 441   if (fscanf(f, "%u", &m) != 1)
 442     return 0;
 443   val = fgetc(f);
 444   if (!isspace(val))
 445     return 0;
 446   if (w > MAXDIM || h > MAXDIM)
 447     return 0;
 448   bpp = (m > 255) ? 6 : 3;
 449   data = malloc(w * h * bpp);
 450   if (fread(data, w * bpp, h, f) != h)
 451     return 0;
 452   if (!fmt) {
 453     fmt = (m > 255) ? hqtexfmt : 3;
 454     if (fmt == GL_FLOAT_RGB32_NV && target != GL_TEXTURE_RECTANGLE)
 455       fmt = GL_RGB16;
 456   }
 457   type = m > 255 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 458   glCreateClearTex(target, fmt, GL_RGB, type, filter, w, h, 0);
 459   glUploadTex(target, GL_RGB, type,
 460               data, w * bpp, 0, 0, w, h, 0);
 461   free(data);
 462   if (width) *width = w;
 463   if (height) *height = h;
 464   if (maxval) *maxval = m;
 465   return 1;
 466 }
 467
 468 /**
 469  * \brief return the number of bytes per pixel for the given format
 470  * \param format OpenGL format
 471  * \param type OpenGL type
 472  * \return bytes per pixel
 473  * \ingroup glgeneral
 474  *
 475  * Does not handle all possible variants, just those used by MPlayer
 476  */
 477 int glFmt2bpp(GLenum format, GLenum type) {
 478   int component_size = 0;
 479   switch (type) {
 480     case GL_UNSIGNED_BYTE_3_3_2:
 481     case GL_UNSIGNED_BYTE_2_3_3_REV:
 482       return 1;
 483     case GL_UNSIGNED_SHORT_5_5_5_1:
 484     case GL_UNSIGNED_SHORT_1_5_5_5_REV:
 485     case GL_UNSIGNED_SHORT_5_6_5:
 486     case GL_UNSIGNED_SHORT_5_6_5_REV:
 487       return 2;
 488     case GL_UNSIGNED_BYTE:
 489       component_size = 1;
 490       break;
 491     case GL_UNSIGNED_SHORT:
 492       component_size = 2;
 493       break;
 494   }
 495   switch (format) {
 496     case GL_LUMINANCE:
 497     case GL_ALPHA:
 498       return component_size;
 499     case GL_YCBCR_MESA:
 500       return 2;
 501     case GL_RGB:
 502     case GL_BGR:
 503       return 3 * component_size;
 504     case GL_RGBA:
 505     case GL_BGRA:
 506       return 4 * component_size;
 507   }
 508   return 0; // unknown
 509 }
 510
 511 /**
 512  * \brief upload a texture, handling things like stride and slices
 513  * \param target texture target, usually GL_TEXTURE_2D
 514  * \param format OpenGL format of data
 515  * \param type OpenGL type of data
 516  * \param dataptr data to upload
 517  * \param stride data stride
 518  * \param x x offset in texture
 519  * \param y y offset in texture
 520  * \param w width of the texture part to upload
 521  * \param h height of the texture part to upload
 522  * \param slice height of an upload slice, 0 for all at once
 523  * \ingroup gltexture
 524  */
 525 void glUploadTex(GLenum target, GLenum format, GLenum type,
 526                  const void *dataptr, int stride,
 527                  int x, int y, int w, int h, int slice) {
 528   const uint8_t *data = dataptr;
 529   int y_max = y + h;
 530   if (w <= 0 || h <= 0) return;
 531   if (slice <= 0)
 532     slice = h;
 533   if (stride < 0) {
 534     data += (h - 1) * stride;
 535     stride = -stride;
 536   }
 537   // this is not always correct, but should work for MPlayer
 538   glAdjustAlignment(stride);
 539   glPixelStorei(GL_UNPACK_ROW_LENGTH, stride / glFmt2bpp(format, type));
 540   for (; y + slice <= y_max; y += slice) {
 541     glTexSubImage2D(target, 0, x, y, w, slice, format, type, data);
 542     data += stride * slice;
 543   }
 544   if (y < y_max)
 545     glTexSubImage2D(target, 0, x, y, w, y_max - y, format, type, data);
 546 }
 547
 548 static void fillUVcoeff(GLfloat *ucoef, GLfloat *vcoef,
 549                         float uvcos, float uvsin) {
 550   int i;
 551   ucoef[0] = 0 * uvcos + 1.403 * uvsin;
 552   vcoef[0] = 0 * uvsin + 1.403 * uvcos;
 553   ucoef[1] = -0.344 * uvcos + -0.714 * uvsin;
 554   vcoef[1] = -0.344 * uvsin + -0.714 * uvcos;
 555   ucoef[2] = 1.770 * uvcos + 0 * uvsin;
 556   vcoef[2] = 1.770 * uvsin + 0 * uvcos;
 557   ucoef[3] = 0;
 558   vcoef[3] = 0;
 559   // Coefficients (probably) must be in [0, 1] range, whereas they originally
 560   // are in [-2, 2] range, so here comes the trick:
 561   // First put them in the [-0.5, 0.5] range, then add 0.5.
 562   // This can be undone with the HALF_BIAS and SCALE_BY_FOUR arguments
 563   // for CombinerInput and CombinerOutput (or the respective ATI variants)
 564   for (i = 0; i < 4; i++) {
 565     ucoef[i] = ucoef[i] * 0.25 + 0.5;
 566     vcoef[i] = vcoef[i] * 0.25 + 0.5;
 567   }
 568 }
 569
 570 /**
 571  * \brief Setup register combiners for YUV to RGB conversion.
 572  * \param uvcos used for saturation and hue adjustment
 573  * \param uvsin used for saturation and hue adjustment
 574  */
 575 static void glSetupYUVCombiners(float uvcos, float uvsin) {
 576   GLfloat ucoef[4];
 577   GLfloat vcoef[4];
 578   GLint i;
 579   if (!CombinerInput || !CombinerOutput ||
 580       !CombinerParameterfv || !CombinerParameteri) {
 581     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner functions missing!\n");
 582     return;
 583   }
 584   glGetIntegerv(GL_MAX_GENERAL_COMBINERS_NV, &i);
 585   if (i < 2)
 586     mp_msg(MSGT_VO, MSGL_ERR,
 587            "[gl] 2 general combiners needed for YUV combiner support (found %i)\n", i);
 588   glGetIntegerv(GL_MAX_TEXTURE_UNITS, &i);
 589   if (i < 3)
 590     mp_msg(MSGT_VO, MSGL_ERR,
 591            "[gl] 3 texture units needed for YUV combiner support (found %i)\n", i);
 592   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 593   CombinerParameterfv(GL_CONSTANT_COLOR0_NV, ucoef);
 594   CombinerParameterfv(GL_CONSTANT_COLOR1_NV, vcoef);
 595
 596   // UV first, like this green component cannot overflow
 597   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_A_NV,
 598                 GL_TEXTURE1, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 599   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_B_NV,
 600                 GL_CONSTANT_COLOR0_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 601   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_C_NV,
 602                 GL_TEXTURE2, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 603   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_D_NV,
 604                 GL_CONSTANT_COLOR1_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 605   CombinerOutput(GL_COMBINER0_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 606                  GL_SPARE0_NV, GL_SCALE_BY_FOUR_NV, GL_NONE, GL_FALSE,
 607                  GL_FALSE, GL_FALSE);
 608
 609   // stage 2
 610   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_A_NV, GL_SPARE0_NV,
 611                 GL_SIGNED_IDENTITY_NV, GL_RGB);
 612   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_B_NV, GL_ZERO,
 613                  GL_UNSIGNED_INVERT_NV, GL_RGB);
 614   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_C_NV,
 615                 GL_TEXTURE0, GL_SIGNED_IDENTITY_NV, GL_RGB);
 616   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_D_NV, GL_ZERO,
 617                 GL_UNSIGNED_INVERT_NV, GL_RGB);
 618   CombinerOutput(GL_COMBINER1_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 619                  GL_SPARE0_NV, GL_NONE, GL_NONE, GL_FALSE,
 620                  GL_FALSE, GL_FALSE);
 621
 622   // leave final combiner stage in default mode
 623   CombinerParameteri(GL_NUM_GENERAL_COMBINERS_NV, 2);
 624 }
 625
 626 /**
 627  * \brief Setup ATI version of register combiners for YUV to RGB conversion.
 628  * \param uvcos used for saturation and hue adjustment
 629  * \param uvsin used for saturation and hue adjustment
 630  *
 631  * ATI called this fragment shader, but the name is confusing in the
 632  * light of a very different OpenGL 2.0 extension with the same name
 633  */
 634 static void glSetupYUVCombinersATI(float uvcos, float uvsin) {
 635   GLfloat ucoef[4];
 636   GLfloat vcoef[4];
 637   GLint i;
 638   if (!BeginFragmentShader || !EndFragmentShader ||
 639       !SetFragmentShaderConstant || !SampleMap ||
 640       !ColorFragmentOp2 || !ColorFragmentOp3) {
 641     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner (ATI) functions missing!\n");
 642     return;
 643   }
 644   glGetIntegerv(GL_NUM_FRAGMENT_REGISTERS_ATI, &i);
 645   if (i < 3)
 646     mp_msg(MSGT_VO, MSGL_ERR,
 647            "[gl] 3 registers needed for YUV combiner (ATI) support (found %i)\n", i);
 648   glGetIntegerv (GL_MAX_TEXTURE_UNITS, &i);
 649   if (i < 3)
 650     mp_msg(MSGT_VO, MSGL_ERR,
 651            "[gl] 3 texture units needed for YUV combiner (ATI) support (found %i)\n", i);
 652   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 653   BeginFragmentShader();
 654   SetFragmentShaderConstant(GL_CON_0_ATI, ucoef);
 655   SetFragmentShaderConstant(GL_CON_1_ATI, vcoef);
 656   SampleMap(GL_REG_0_ATI, GL_TEXTURE0, GL_SWIZZLE_STR_ATI);
 657   SampleMap(GL_REG_1_ATI, GL_TEXTURE1, GL_SWIZZLE_STR_ATI);
 658   SampleMap(GL_REG_2_ATI, GL_TEXTURE2, GL_SWIZZLE_STR_ATI);
 659   // UV first, like this green component cannot overflow
 660   ColorFragmentOp2(GL_MUL_ATI, GL_REG_1_ATI, GL_NONE, GL_NONE,
 661                    GL_REG_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 662                    GL_CON_0_ATI, GL_NONE, GL_BIAS_BIT_ATI);
 663   ColorFragmentOp3(GL_MAD_ATI, GL_REG_2_ATI, GL_NONE, GL_4X_BIT_ATI,
 664                    GL_REG_2_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 665                    GL_CON_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 666                    GL_REG_1_ATI, GL_NONE, GL_NONE);
 667   ColorFragmentOp2(GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 668                    GL_REG_0_ATI, GL_NONE, GL_NONE,
 669                    GL_REG_2_ATI, GL_NONE, GL_NONE);
 670   EndFragmentShader();
 671 }
 672
 673 /**
 674  * \brief helper function for gen_spline_lookup_tex
 675  * \param x subpixel-position ((0,1) range) to calculate weights for
 676  * \param dst where to store transformed weights, must provide space for 4 GLfloats
 677  *
 678  * calculates the weights and stores them after appropriate transformation
 679  * for the scaler fragment program.
 680  */
 681 static void store_weights(float x, GLfloat *dst) {
 682   float w0 = (((-1 * x + 3) * x - 3) * x + 1) / 6;
 683   float w1 = ((( 3 * x - 6) * x + 0) * x + 4) / 6;
 684   float w2 = (((-3 * x + 3) * x + 3) * x + 1) / 6;
 685   float w3 = ((( 1 * x + 0) * x + 0) * x + 0) / 6;
 686   *dst++ = 1 + x - w1 / (w0 + w1);
 687   *dst++ = 1 - x + w3 / (w2 + w3);
 688   *dst++ = w0 + w1;
 689   *dst++ = 0;
 690 }
 691
 692 //! to avoid artefacts this should be rather large
 693 #define LOOKUP_BSPLINE_RES (2 * 1024)
 694 /**
 695  * \brief creates the 1D lookup texture needed for fast higher-order filtering
 696  * \param unit texture unit to attach texture to
 697  */
 698 static void gen_spline_lookup_tex(GLenum unit) {
 699   GLfloat tex[4 * LOOKUP_BSPLINE_RES];
 700   GLfloat *tp = tex;
 701   int i;
 702   for (i = 0; i < LOOKUP_BSPLINE_RES; i++) {
 703     float x = (float)(i + 0.5) / LOOKUP_BSPLINE_RES;
 704     store_weights(x, tp);
 705     tp += 4;
 706   }
 707   store_weights(0, tex);
 708   store_weights(1, &tex[4 * (LOOKUP_BSPLINE_RES - 1)]);
 709   ActiveTexture(unit);
 710   glTexImage1D(GL_TEXTURE_1D, 0, GL_RGBA16, LOOKUP_BSPLINE_RES, 0, GL_RGBA, GL_FLOAT, tex);
 711   glTexParameterf(GL_TEXTURE_1D, GL_TEXTURE_PRIORITY, 1.0);
 712   glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 713   glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 714   glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_REPEAT);
 715   ActiveTexture(GL_TEXTURE0);
 716 }
 717
 718 static const char *bilin_filt_template =
 719   "TEX yuv.%c, fragment.texcoord[%c], texture[%c], %s;";
 720
 721 #define BICUB_FILT_MAIN(textype) \
 722   /* first y-interpolation */ \
 723   "ADD coord, fragment.texcoord[%c].xyxy, cdelta.xyxw;" \
 724   "ADD coord2, fragment.texcoord[%c].xyxy, cdelta.zyzw;" \
 725   "TEX a.r, coord.xyxy, texture[%c], "textype";" \
 726   "TEX a.g, coord.zwzw, texture[%c], "textype";" \
 727   /* second y-interpolation */ \
 728   "TEX b.r, coord2.xyxy, texture[%c], "textype";" \
 729   "TEX b.g, coord2.zwzw, texture[%c], "textype";" \
 730   "LRP a.b, parmy.b, a.rrrr, a.gggg;" \
 731   "LRP a.a, parmy.b, b.rrrr, b.gggg;" \
 732   /* x-interpolation */ \
 733   "LRP yuv.%c, parmx.b, a.bbbb, a.aaaa;"
 734
 735 static const char *bicub_filt_template_2D =
 736   "MAD coord.xy, fragment.texcoord[%c], {%f, %f}, {0.5, 0.5};"
 737   "TEX parmx, coord.x, texture[%c], 1D;"
 738   "MUL cdelta.xz, parmx.rrgg, {-%f, 0, %f, 0};"
 739   "TEX parmy, coord.y, texture[%c], 1D;"
 740   "MUL cdelta.yw, parmy.rrgg, {0, -%f, 0, %f};"
 741   BICUB_FILT_MAIN("2D");
 742
 743 static const char *bicub_filt_template_RECT =
 744   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};"
 745   "TEX parmx, coord.x, texture[%c], 1D;"
 746   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};"
 747   "TEX parmy, coord.y, texture[%c], 1D;"
 748   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};"
 749   BICUB_FILT_MAIN("RECT");
 750
 751 #define CALCWEIGHTS(t, s) \
 752   "MAD "t", {-0.5, 0.1666, 0.3333, -0.3333}, "s", {1, 0, -0.5, 0.5};" \
 753   "MAD "t", "t", "s", {0, 0, -0.5, 0.5};" \
 754   "MAD "t", "t", "s", {-0.6666, 0, 0.8333, 0.1666};" \
 755   "RCP a.x, "t".z;" \
 756   "RCP a.y, "t".w;" \
 757   "MAD "t".xy, "t".xyxy, a.xyxy, {1, 1, 0, 0};" \
 758   "ADD "t".x, "t".xxxx, "s";" \
 759   "SUB "t".y, "t".yyyy, "s";"
 760
 761 static const char *bicub_notex_filt_template_2D =
 762   "MAD coord.xy, fragment.texcoord[%c], {%f, %f}, {0.5, 0.5};"
 763   "FRC coord.xy, coord.xyxy;"
 764   CALCWEIGHTS("parmx", "coord.xxxx")
 765   "MUL cdelta.xz, parmx.rrgg, {-%f, 0, %f, 0};"
 766   CALCWEIGHTS("parmy", "coord.yyyy")
 767   "MUL cdelta.yw, parmy.rrgg, {0, -%f, 0, %f};"
 768   BICUB_FILT_MAIN("2D");
 769
 770 static const char *bicub_notex_filt_template_RECT =
 771   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};"
 772   "FRC coord.xy, coord.xyxy;"
 773   CALCWEIGHTS("parmx", "coord.xxxx")
 774   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};"
 775   CALCWEIGHTS("parmy", "coord.yyyy")
 776   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};"
 777   BICUB_FILT_MAIN("RECT");
 778
 779 #define BICUB_X_FILT_MAIN(textype) \
 780   "ADD coord.xy, fragment.texcoord[%c].xyxy, cdelta.xyxy;" \
 781   "ADD coord2.xy, fragment.texcoord[%c].xyxy, cdelta.zyzy;" \
 782   "TEX a.r, coord, texture[%c], "textype";" \
 783   "TEX b.r, coord2, texture[%c], "textype";" \
 784   /* x-interpolation */ \
 785   "LRP yuv.%c, parmx.b, a.rrrr, b.rrrr;"
 786
 787 static const char *bicub_x_filt_template_2D =
 788   "MAD coord.x, fragment.texcoord[%c], {%f}, {0.5};"
 789   "TEX parmx, coord, texture[%c], 1D;"
 790   "MUL cdelta.xyz, parmx.rrgg, {-%f, 0, %f};"
 791   BICUB_X_FILT_MAIN("2D");
 792
 793 static const char *bicub_x_filt_template_RECT =
 794   "ADD coord.x, fragment.texcoord[%c], {0.5};"
 795   "TEX parmx, coord, texture[%c], 1D;"
 796   "MUL cdelta.xyz, parmx.rrgg, {-1, 0, 1};"
 797   BICUB_X_FILT_MAIN("RECT");
 798
 799 static const char *unsharp_filt_template =
 800   "PARAM dcoord%c = {%f, %f, %f, %f};"
 801   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;"
 802   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;"
 803   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;"
 804   "TEX b.r, coord.xyxy, texture[%c], %s;"
 805   "TEX b.g, coord.zwzw, texture[%c], %s;"
 806   "ADD b.r, b.r, b.g;"
 807   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 808   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 809   "DP3 b, b, {0.25, 0.25, 0.25};"
 810   "SUB b.r, a.r, b.r;"
 811   "MAD yuv.%c, b.r, {%f}, a.r;";
 812
 813 static const char *unsharp_filt_template2 =
 814   "PARAM dcoord%c = {%f, %f, %f, %f};"
 815   "PARAM dcoord2%c = {%f, 0, 0, %f};"
 816   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;"
 817   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;"
 818   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;"
 819   "TEX b.r, coord.xyxy, texture[%c], %s;"
 820   "TEX b.g, coord.zwzw, texture[%c], %s;"
 821   "ADD b.r, b.r, b.g;"
 822   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 823   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 824   "ADD b.r, b.r, b.b;"
 825   "ADD b.a, b.r, b.g;"
 826   "ADD coord, fragment.texcoord[%c].xyxy, dcoord2%c;"
 827   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord2%c;"
 828   "TEX b.r, coord.xyxy, texture[%c], %s;"
 829   "TEX b.g, coord.zwzw, texture[%c], %s;"
 830   "ADD b.r, b.r, b.g;"
 831   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 832   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 833   "DP4 b.r, b, {-0.1171875, -0.1171875, -0.1171875, -0.09765625};"
 834   "MAD b.r, a.r, {0.859375}, b.r;"
 835   "MAD yuv.%c, b.r, {%f}, a.r;";
 836
 837 static const char *yuv_prog_template =
 838   "PARAM ycoef = {%.4f, %.4f, %.4f};"
 839   "PARAM ucoef = {%.4f, %.4f, %.4f};"
 840   "PARAM vcoef = {%.4f, %.4f, %.4f};"
 841   "PARAM offsets = {%.4f, %.4f, %.4f};"
 842   "TEMP res;"
 843   "MAD res.rgb, yuv.rrrr, ycoef, offsets;"
 844   "MAD res.rgb, yuv.gggg, ucoef, res;"
 845   "MAD result.color.rgb, yuv.bbbb, vcoef, res;"
 846   "END";
 847
 848 static const char *yuv_pow_prog_template =
 849   "PARAM ycoef = {%.4f, %.4f, %.4f};"
 850   "PARAM ucoef = {%.4f, %.4f, %.4f};"
 851   "PARAM vcoef = {%.4f, %.4f, %.4f};"
 852   "PARAM offsets = {%.4f, %.4f, %.4f};"
 853   "PARAM gamma = {%.4f, %.4f, %.4f};"
 854   "TEMP res;"
 855   "MAD res.rgb, yuv.rrrr, ycoef, offsets;"
 856   "MAD res.rgb, yuv.gggg, ucoef, res;"
 857   "MAD_SAT res.rgb, yuv.bbbb, vcoef, res;"
 858   "POW result.color.r, res.r, gamma.r;"
 859   "POW result.color.g, res.g, gamma.g;"
 860   "POW result.color.b, res.b, gamma.b;"
 861   "END";
 862
 863 static const char *yuv_lookup_prog_template =
 864   "PARAM ycoef = {%.4f, %.4f, %.4f, 0};"
 865   "PARAM ucoef = {%.4f, %.4f, %.4f, 0};"
 866   "PARAM vcoef = {%.4f, %.4f, %.4f, 0};"
 867   "PARAM offsets = {%.4f, %.4f, %.4f, 0.125};"
 868   "TEMP res;"
 869   "MAD res, yuv.rrrr, ycoef, offsets;"
 870   "MAD res.rgb, yuv.gggg, ucoef, res;"
 871   "MAD res.rgb, yuv.bbbb, vcoef, res;"
 872   "TEX result.color.r, res.raaa, texture[%c], 2D;"
 873   "ADD res.a, res.a, 0.25;"
 874   "TEX result.color.g, res.gaaa, texture[%c], 2D;"
 875   "ADD res.a, res.a, 0.25;"
 876   "TEX result.color.b, res.baaa, texture[%c], 2D;"
 877   "END";
 878
 879 static const char *yuv_lookup3d_prog_template =
 880   "TEX result.color, yuv, texture[%c], 3D;"
 881   "END";
 882
 883 /**
 884  * \brief creates and initializes helper textures needed for scaling texture read
 885  * \param scaler scaler type to create texture for
 886  * \param texu contains next free texture unit number
 887  * \param texs texture unit ids for the scaler are stored in this array
 888  */
 889 static void create_scaler_textures(int scaler, int *texu, char *texs) {
 890   switch (scaler) {
 891     case YUV_SCALER_BILIN:
 892     case YUV_SCALER_BICUB_NOTEX:
 893     case YUV_SCALER_UNSHARP:
 894     case YUV_SCALER_UNSHARP2:
 895       break;
 896     case YUV_SCALER_BICUB:
 897     case YUV_SCALER_BICUB_X:
 898       texs[0] = (*texu)++;
 899       gen_spline_lookup_tex(GL_TEXTURE0 + texs[0]);
 900       texs[0] += '0';
 901       break;
 902     default:
 903       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown scaler type %i\n", scaler);
 904   }
 905 }
 906
 907 static void gen_gamma_map(unsigned char *map, int size, float gamma);
 908
 909 #define ROW_R 0
 910 #define ROW_G 1
 911 #define ROW_B 2
 912 #define COL_Y 0
 913 #define COL_U 1
 914 #define COL_V 2
 915 #define COL_C 3
 916
 917 static void get_yuv2rgb_coeffs(gl_conversion_params_t *params, float yuv2rgb[3][4]) {
 918   float uvcos = params->saturation * cos(params->hue);
 919   float uvsin = params->saturation * sin(params->hue);
 920   int i;
 921   float uv_coeffs[3][2] = {
 922     { 0.000,  1.596},
 923     {-0.391, -0.813},
 924     { 2.018,  0.000}
 925   };
 926   for (i = 0; i < 3; i++) {
 927     yuv2rgb[i][COL_C]  = params->brightness;
 928     yuv2rgb[i][COL_Y]  = 1.164 * params->contrast;
 929     yuv2rgb[i][COL_C] += (-16 / 255.0) * yuv2rgb[i][COL_Y];
 930     yuv2rgb[i][COL_U]  = uv_coeffs[i][0] * uvcos + uv_coeffs[i][1] * uvsin;
 931     yuv2rgb[i][COL_C] += (-128 / 255.0) * yuv2rgb[i][COL_U];
 932     yuv2rgb[i][COL_V]  = uv_coeffs[i][0] * uvsin + uv_coeffs[i][1] * uvcos;
 933     yuv2rgb[i][COL_C] += (-128 / 255.0) * yuv2rgb[i][COL_V];
 934     // this "centers" contrast control so that e.g. a contrast of 0
 935     // leads to a grey image, not a black one
 936     yuv2rgb[i][COL_C] += 0.5 - params->contrast / 2.0;
 937   }
 938 }
 939
 940 //! size of gamma map use to avoid slow exp function in gen_yuv2rgb_map
 941 #define GMAP_SIZE (1024)
 942 /**
 943  * \brief generate a 3D YUV -> RGB map
 944  * \param params struct containing parameters like brightness, gamma, ...
 945  * \param map where to store map. Must provide space for (size + 2)^3 elements
 946  * \param size size of the map, excluding border
 947  */
 948 static void gen_yuv2rgb_map(gl_conversion_params_t *params, unsigned char *map, int size) {
 949   int i, j, k, l;
 950   float step = 1.0 / size;
 951   float y, u, v;
 952   float yuv2rgb[3][4];
 953   unsigned char gmaps[3][GMAP_SIZE];
 954   gen_gamma_map(gmaps[0], GMAP_SIZE, params->rgamma);
 955   gen_gamma_map(gmaps[1], GMAP_SIZE, params->ggamma);
 956   gen_gamma_map(gmaps[2], GMAP_SIZE, params->bgamma);
 957   get_yuv2rgb_coeffs(params, yuv2rgb);
 958   for (i = 0; i < 3; i++)
 959     for (j = 0; j < 4; j++)
 960       yuv2rgb[i][j] *= GMAP_SIZE - 1;
 961   v = 0;
 962   for (i = -1; i <= size; i++) {
 963     u = 0;
 964     for (j = -1; j <= size; j++) {
 965       y = 0;
 966       for (k = -1; k <= size; k++) {
 967         for (l = 0; l < 3; l++) {
 968           float rgb = yuv2rgb[l][COL_Y] * y + yuv2rgb[l][COL_U] * u + yuv2rgb[l][COL_V] * v + yuv2rgb[l][COL_C];
 969           *map++ = gmaps[l][av_clip(rgb, 0, GMAP_SIZE - 1)];
 970         }
 971         y += (k == -1 || k == size - 1) ? step / 2 : step;
 972       }
 973       u += (j == -1 || j == size - 1) ? step / 2 : step;
 974     }
 975     v += (i == -1 || i == size - 1) ? step / 2 : step;
 976   }
 977 }
 978
 979 //! resolution of texture for gamma lookup table
 980 #define LOOKUP_RES 512
 981 //! resolution for 3D yuv->rgb conversion lookup table
 982 #define LOOKUP_3DRES 32
 983 /**
 984  * \brief creates and initializes helper textures needed for yuv conversion
 985  * \param params struct containing parameters like brightness, gamma, ...
 986  * \param texu contains next free texture unit number
 987  * \param texs texture unit ids for the conversion are stored in this array
 988  */
 989 static void create_conv_textures(gl_conversion_params_t *params, int *texu, char *texs) {
 990   unsigned char *lookup_data = NULL;
 991   int conv = YUV_CONVERSION(params->type);
 992   switch (conv) {
 993     case YUV_CONVERSION_FRAGMENT:
 994     case YUV_CONVERSION_FRAGMENT_POW:
 995      break;
 996     case YUV_CONVERSION_FRAGMENT_LOOKUP:
 997       texs[0] = (*texu)++;
 998       ActiveTexture(GL_TEXTURE0 + texs[0]);
 999       lookup_data = malloc(4 * LOOKUP_RES);
1000       gen_gamma_map(lookup_data, LOOKUP_RES, params->rgamma);
1001       gen_gamma_map(&lookup_data[LOOKUP_RES], LOOKUP_RES, params->ggamma);
1002       gen_gamma_map(&lookup_data[2 * LOOKUP_RES], LOOKUP_RES, params->bgamma);
1003       glCreateClearTex(GL_TEXTURE_2D, GL_LUMINANCE8, GL_LUMINANCE, GL_UNSIGNED_BYTE, GL_LINEAR,
1004                        LOOKUP_RES, 4, 0);
1005       glUploadTex(GL_TEXTURE_2D, GL_LUMINANCE, GL_UNSIGNED_BYTE, lookup_data,
1006                   LOOKUP_RES, 0, 0, LOOKUP_RES, 4, 0);
1007       ActiveTexture(GL_TEXTURE0);
1008       texs[0] += '0';
1009       break;
1010     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1011       {
1012         int sz = LOOKUP_3DRES + 2; // texture size including borders
1013         if (!TexImage3D) {
1014           mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing 3D texture function!\n");
1015           break;
1016         }
1017         texs[0] = (*texu)++;
1018         ActiveTexture(GL_TEXTURE0 + texs[0]);
1019         lookup_data = malloc(3 * sz * sz * sz);
1020         gen_yuv2rgb_map(params, lookup_data, LOOKUP_3DRES);
1021         glAdjustAlignment(sz);
1022         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
1023         TexImage3D(GL_TEXTURE_3D, 0, 3, sz, sz, sz, 1,
1024                    GL_RGB, GL_UNSIGNED_BYTE, lookup_data);
1025         glTexParameterf(GL_TEXTURE_3D, GL_TEXTURE_PRIORITY, 1.0);
1026         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1027         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1028         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_S, GL_CLAMP);
1029         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_T, GL_CLAMP);
1030         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_R, GL_CLAMP);
1031         ActiveTexture(GL_TEXTURE0);
1032         texs[0] += '0';
1033       }
1034       break;
1035     default:
1036       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", conv);
1037   }
1038   if (lookup_data)
1039     free(lookup_data);
1040 }
1041
1042 /**
1043  * \brief adds a scaling texture read at the current fragment program position
1044  * \param scaler type of scaler to insert
1045  * \param prog_pos current position in fragment program
1046  * \param remain how many bytes remain in the buffer given by prog_pos
1047  * \param texs array containing the texture unit identifiers for this scaler
1048  * \param in_tex texture unit the scaler should read from
1049  * \param out_comp component of the yuv variable the scaler stores the result in
1050  * \param rect if rectangular (pixel) adressing should be used for in_tex
1051  * \param texw width of the in_tex texture
1052  * \param texh height of the in_tex texture
1053  * \param strength strength of filter effect if the scaler does some kind of filtering
1054  */
1055 static void add_scaler(int scaler, char **prog_pos, int *remain, char *texs,
1056                     char in_tex, char out_comp, int rect, int texw, int texh,
1057                     double strength) {
1058   const char *ttype = rect ? "RECT" : "2D";
1059   const float ptw = rect ? 1.0 : 1.0 / texw;
1060   const float pth = rect ? 1.0 : 1.0 / texh;
1061   switch (scaler) {
1062     case YUV_SCALER_BILIN:
1063       snprintf(*prog_pos, *remain, bilin_filt_template, out_comp, in_tex,
1064                  in_tex, ttype);
1065       break;
1066     case YUV_SCALER_BICUB:
1067       if (rect)
1068         snprintf(*prog_pos, *remain, bicub_filt_template_RECT,
1069                  in_tex, texs[0], texs[0],
1070                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1071       else
1072         snprintf(*prog_pos, *remain, bicub_filt_template_2D,
1073                  in_tex, (float)texw, (float)texh,
1074                  texs[0], ptw, ptw, texs[0], pth, pth,
1075                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1076       break;
1077     case YUV_SCALER_BICUB_X:
1078       if (rect)
1079         snprintf(*prog_pos, *remain, bicub_x_filt_template_RECT,
1080                  in_tex, texs[0],
1081                  in_tex, in_tex, in_tex, in_tex, out_comp);
1082       else
1083         snprintf(*prog_pos, *remain, bicub_x_filt_template_2D,
1084                  in_tex, (float)texw,
1085                  texs[0], ptw, ptw,
1086                  in_tex, in_tex, in_tex, in_tex, out_comp);
1087       break;
1088     case YUV_SCALER_BICUB_NOTEX:
1089       if (rect)
1090         snprintf(*prog_pos, *remain, bicub_notex_filt_template_RECT,
1091                  in_tex,
1092                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1093       else
1094         snprintf(*prog_pos, *remain, bicub_notex_filt_template_2D,
1095                  in_tex, (float)texw, (float)texh, ptw, ptw, pth, pth,
1096                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1097       break;
1098     case YUV_SCALER_UNSHARP:
1099       snprintf(*prog_pos, *remain, unsharp_filt_template,
1100                out_comp, 0.5 * ptw, 0.5 * pth, 0.5 * ptw, -0.5 * pth,
1101                in_tex, out_comp, in_tex, out_comp, in_tex,
1102                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1103                in_tex, ttype, out_comp, strength);
1104       break;
1105     case YUV_SCALER_UNSHARP2:
1106       snprintf(*prog_pos, *remain, unsharp_filt_template2,
1107                out_comp, 1.2 * ptw, 1.2 * pth, 1.2 * ptw, -1.2 * pth,
1108                out_comp, 1.5 * ptw, 1.5 * pth,
1109                in_tex, out_comp, in_tex, out_comp, in_tex,
1110                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1111                in_tex, ttype, in_tex, out_comp, in_tex, out_comp,
1112                in_tex, ttype, in_tex, ttype, in_tex, ttype,
1113                in_tex, ttype, out_comp, strength);
1114       break;
1115   }
1116   *remain -= strlen(*prog_pos);
1117   *prog_pos += strlen(*prog_pos);
1118 }
1119
1120 static const struct {
1121   const char *name;
1122   GLenum cur;
1123   GLenum max;
1124 } progstats[] = {
1125   {"instructions", 0x88A0, 0x88A1},
1126   {"native instructions", 0x88A2, 0x88A3},
1127   {"temporaries", 0x88A4, 0x88A5},
1128   {"native temporaries", 0x88A6, 0x88A7},
1129   {"parameters", 0x88A8, 0x88A9},
1130   {"native parameters", 0x88AA, 0x88AB},
1131   {"attribs", 0x88AC, 0x88AD},
1132   {"native attribs", 0x88AE, 0x88AF},
1133   {"ALU instructions", 0x8805, 0x880B},
1134   {"TEX instructions", 0x8806, 0x880C},
1135   {"TEX indirections", 0x8807, 0x880D},
1136   {"native ALU instructions", 0x8808, 0x880E},
1137   {"native TEX instructions", 0x8809, 0x880F},
1138   {"native TEX indirections", 0x880A, 0x8810},
1139   {NULL, 0, 0}
1140 };
1141
1142 /**
1143  * \brief load the specified GPU Program
1144  * \param target program target to load into, only GL_FRAGMENT_PROGRAM is tested
1145  * \param prog program string
1146  * \return 1 on success, 0 otherwise
1147  */
1148 int loadGPUProgram(GLenum target, char *prog) {
1149   int i;
1150   GLint cur = 0, max = 0, err = 0;
1151   if (!ProgramString) {
1152     mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing GPU program function\n");
1153     return 0;
1154   }
1155   ProgramString(target, GL_PROGRAM_FORMAT_ASCII, strlen(prog), prog);
1156   glGetIntegerv(GL_PROGRAM_ERROR_POSITION, &err);
1157   if (err != -1) {
1158     mp_msg(MSGT_VO, MSGL_ERR,
1159       "[gl] Error compiling fragment program, make sure your card supports\n"
1160       "[gl]   GL_ARB_fragment_program (use glxinfo to check).\n"
1161       "[gl]   Error message:\n  %s at %.10s\n",
1162       glGetString(GL_PROGRAM_ERROR_STRING), &prog[err]);
1163     return 0;
1164   }
1165   if (!GetProgramiv || !mp_msg_test(MSGT_VO, MSGL_V))
1166     return 1;
1167   mp_msg(MSGT_VO, MSGL_V, "[gl] Program statistics:\n");
1168   for (i = 0; progstats[i].name; i++) {
1169     GetProgramiv(target, progstats[i].cur, &cur);
1170     GetProgramiv(target, progstats[i].max, &max);
1171     mp_msg(MSGT_VO, MSGL_V, "[gl]   %s: %i/%i\n", progstats[i].name, cur, max);
1172   }
1173   return 1;
1174 }
1175
1176 #define MAX_PROGSZ (1024*1024)
1177
1178 /**
1179  * \brief setup a fragment program that will do YUV->RGB conversion
1180  * \param parms struct containing parameters like conversion and scaler type,
1181  *              brightness, ...
1182  */
1183 static void glSetupYUVFragprog(gl_conversion_params_t *params) {
1184   int type = params->type;
1185   int texw = params->texw;
1186   int texh = params->texh;
1187   int rect = params->target == GL_TEXTURE_RECTANGLE;
1188   static const char prog_hdr[] =
1189     "!!ARBfp1.0\n"
1190     "OPTION ARB_precision_hint_fastest;"
1191     // all scaler variables must go here so they aren't defined
1192     // multiple times when the same scaler is used more than once
1193     "TEMP coord, coord2, cdelta, parmx, parmy, a, b, yuv;";
1194   int prog_remain;
1195   char *yuv_prog, *prog_pos;
1196   int cur_texu = 3;
1197   char lum_scale_texs[1];
1198   char chrom_scale_texs[1];
1199   char conv_texs[1];
1200   GLint i;
1201   // this is the conversion matrix, with y, u, v factors
1202   // for red, green, blue and the constant offsets
1203   float yuv2rgb[3][4];
1204   create_conv_textures(params, &cur_texu, conv_texs);
1205   create_scaler_textures(YUV_LUM_SCALER(type), &cur_texu, lum_scale_texs);
1206   if (YUV_CHROM_SCALER(type) == YUV_LUM_SCALER(type))
1207     memcpy(chrom_scale_texs, lum_scale_texs, sizeof(chrom_scale_texs));
1208   else
1209     create_scaler_textures(YUV_CHROM_SCALER(type), &cur_texu, chrom_scale_texs);
1210   glGetIntegerv(GL_MAX_TEXTURE_IMAGE_UNITS, &i);
1211   if (i < cur_texu)
1212     mp_msg(MSGT_VO, MSGL_ERR,
1213            "[gl] %i texture units needed for this type of YUV fragment support (found %i)\n",
1214            cur_texu, i);
1215   if (!ProgramString) {
1216     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] ProgramString function missing!\n");
1217     return;
1218   }
1219   yuv_prog = malloc(MAX_PROGSZ);
1220   strcpy(yuv_prog, prog_hdr);
1221   prog_pos    = yuv_prog + sizeof(prog_hdr) - 1;
1222   prog_remain = MAX_PROGSZ - sizeof(prog_hdr);
1223   add_scaler(YUV_LUM_SCALER(type), &prog_pos, &prog_remain, lum_scale_texs,
1224              '0', 'r', rect, texw, texh, params->filter_strength);
1225   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1226              '1', 'g', rect, texw / 2, texh / 2, params->filter_strength);
1227   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1228              '2', 'b', rect, texw / 2, texh / 2, params->filter_strength);
1229   get_yuv2rgb_coeffs(params, yuv2rgb);
1230   switch (YUV_CONVERSION(type)) {
1231     case YUV_CONVERSION_FRAGMENT:
1232       snprintf(prog_pos, prog_remain, yuv_prog_template,
1233                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1234                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1235                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1236                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C]);
1237       break;
1238     case YUV_CONVERSION_FRAGMENT_POW:
1239       snprintf(prog_pos, prog_remain, yuv_pow_prog_template,
1240                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1241                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1242                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1243                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1244                (float)1.0 / params->rgamma, (float)1.0 / params->bgamma, (float)1.0 / params->bgamma);
1245       break;
1246     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1247       snprintf(prog_pos, prog_remain, yuv_lookup_prog_template,
1248                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1249                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1250                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1251                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1252                conv_texs[0], conv_texs[0], conv_texs[0]);
1253       break;
1254     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1255       snprintf(prog_pos, prog_remain, yuv_lookup3d_prog_template, conv_texs[0]);
1256       break;
1257     default:
1258       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(type));
1259       break;
1260   }
1261   mp_msg(MSGT_VO, MSGL_V, "[gl] generated fragment program:\n%s\n", yuv_prog);
1262   loadGPUProgram(GL_FRAGMENT_PROGRAM, yuv_prog);
1263   free(yuv_prog);
1264 }
1265
1266 /**
1267  * \brief little helper function to create a lookup table for gamma
1268  * \param map buffer to create map into
1269  * \param size size of buffer
1270  * \param gamma gamma value
1271  */
1272 static void gen_gamma_map(unsigned char *map, int size, float gamma) {
1273   int i;
1274   if (gamma == 1.0) {
1275     for (i = 0; i < size; i++)
1276       map[i] = 255 * i / (size - 1);
1277     return;
1278   }
1279   gamma = 1.0 / gamma;
1280   for (i = 0; i < size; i++) {
1281     float tmp = (float)i / (size - 1.0);
1282     tmp = pow(tmp, gamma);
1283     if (tmp > 1.0) tmp = 1.0;
1284     if (tmp < 0.0) tmp = 0.0;
1285     map[i] = 255 * tmp;
1286   }
1287 }
1288
1289 /**
1290  * \brief setup YUV->RGB conversion
1291  * \param parms struct containing parameters like conversion and scaler type,
1292  *              brightness, ...
1293  * \ingroup glconversion
1294  */
1295 void glSetupYUVConversion(gl_conversion_params_t *params) {
1296   float uvcos = params->saturation * cos(params->hue);
1297   float uvsin = params->saturation * sin(params->hue);
1298   switch (YUV_CONVERSION(params->type)) {
1299     case YUV_CONVERSION_COMBINERS:
1300       glSetupYUVCombiners(uvcos, uvsin);
1301       break;
1302     case YUV_CONVERSION_COMBINERS_ATI:
1303       glSetupYUVCombinersATI(uvcos, uvsin);
1304       break;
1305     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1306     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1307     case YUV_CONVERSION_FRAGMENT:
1308     case YUV_CONVERSION_FRAGMENT_POW:
1309       glSetupYUVFragprog(params);
1310       break;
1311     default:
1312       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(params->type));
1313   }
1314 }
1315
1316 /**
1317  * \brief enable the specified YUV conversion
1318  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1319  * \param type type of YUV conversion
1320  * \ingroup glconversion
1321  */
1322 void glEnableYUVConversion(GLenum target, int type) {
1323   if (type <= 0) return;
1324   switch (YUV_CONVERSION(type)) {
1325     case YUV_CONVERSION_COMBINERS:
1326       ActiveTexture(GL_TEXTURE1);
1327       glEnable(target);
1328       ActiveTexture(GL_TEXTURE2);
1329       glEnable(target);
1330       ActiveTexture(GL_TEXTURE0);
1331       glEnable(GL_REGISTER_COMBINERS_NV);
1332       break;
1333     case YUV_CONVERSION_COMBINERS_ATI:
1334       ActiveTexture(GL_TEXTURE1);
1335       glEnable(target);
1336       ActiveTexture(GL_TEXTURE2);
1337       glEnable(target);
1338       ActiveTexture(GL_TEXTURE0);
1339       glEnable(GL_FRAGMENT_SHADER_ATI);
1340       break;
1341     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1342     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1343     case YUV_CONVERSION_FRAGMENT_POW:
1344     case YUV_CONVERSION_FRAGMENT:
1345       glEnable(GL_FRAGMENT_PROGRAM);
1346       break;
1347   }
1348 }
1349
1350 /**
1351  * \brief disable the specified YUV conversion
1352  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1353  * \param type type of YUV conversion
1354  * \ingroup glconversion
1355  */
1356 void glDisableYUVConversion(GLenum target, int type) {
1357   if (type <= 0) return;
1358   switch (YUV_CONVERSION(type)) {
1359     case YUV_CONVERSION_COMBINERS:
1360       ActiveTexture(GL_TEXTURE1);
1361       glDisable(target);
1362       ActiveTexture(GL_TEXTURE2);
1363       glDisable(target);
1364       ActiveTexture(GL_TEXTURE0);
1365       glDisable(GL_REGISTER_COMBINERS_NV);
1366       break;
1367     case YUV_CONVERSION_COMBINERS_ATI:
1368       ActiveTexture(GL_TEXTURE1);
1369       glDisable(target);
1370       ActiveTexture(GL_TEXTURE2);
1371       glDisable(target);
1372       ActiveTexture(GL_TEXTURE0);
1373       glDisable(GL_FRAGMENT_SHADER_ATI);
1374       break;
1375     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1376     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1377     case YUV_CONVERSION_FRAGMENT_POW:
1378     case YUV_CONVERSION_FRAGMENT:
1379       glDisable(GL_FRAGMENT_PROGRAM);
1380       break;
1381   }
1382 }
1383
1384 /**
1385  * \brief draw a texture part at given 2D coordinates
1386  * \param x screen top coordinate
1387  * \param y screen left coordinate
1388  * \param w screen width coordinate
1389  * \param h screen height coordinate
1390  * \param tx texture top coordinate in pixels
1391  * \param ty texture left coordinate in pixels
1392  * \param tw texture part width in pixels
1393  * \param th texture part height in pixels
1394  * \param sx width of texture in pixels
1395  * \param sy height of texture in pixels
1396  * \param rect_tex whether this texture uses texture_rectangle extension
1397  * \param is_yv12 if set, also draw the textures from units 1 and 2
1398  * \param flip flip the texture upside down
1399  * \ingroup gltexture
1400  */
1401 void glDrawTex(GLfloat x, GLfloat y, GLfloat w, GLfloat h,
1402                GLfloat tx, GLfloat ty, GLfloat tw, GLfloat th,
1403                int sx, int sy, int rect_tex, int is_yv12, int flip) {
1404   GLfloat tx2 = tx / 2, ty2 = ty / 2, tw2 = tw / 2, th2 = th / 2;
1405   if (!rect_tex) {
1406     tx /= sx; ty /= sy; tw /= sx; th /= sy;
1407     tx2 = tx, ty2 = ty, tw2 = tw, th2 = th;
1408   }
1409   if (flip) {
1410     y += h;
1411     h = -h;
1412   }
1413   glBegin(GL_QUADS);
1414   glTexCoord2f(tx, ty);
1415   if (is_yv12) {
1416     MultiTexCoord2f(GL_TEXTURE1, tx2, ty2);
1417     MultiTexCoord2f(GL_TEXTURE2, tx2, ty2);
1418   }
1419   glVertex2f(x, y);
1420   glTexCoord2f(tx, ty + th);
1421   if (is_yv12) {
1422     MultiTexCoord2f(GL_TEXTURE1, tx2, ty2 + th2);
1423     MultiTexCoord2f(GL_TEXTURE2, tx2, ty2 + th2);
1424   }
1425   glVertex2f(x, y + h);
1426   glTexCoord2f(tx + tw, ty + th);
1427   if (is_yv12) {
1428     MultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2 + th2);
1429     MultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2 + th2);
1430   }
1431   glVertex2f(x + w, y + h);
1432   glTexCoord2f(tx + tw, ty);
1433   if (is_yv12) {
1434     MultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2);
1435     MultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2);
1436   }
1437   glVertex2f(x + w, y);
1438   glEnd();
1439 }
1440
1441 #ifdef GL_WIN32
1442 #include "w32_common.h"
1443 /**
1444  * \brief little helper since wglGetProcAddress definition does not fit our
1445  *        getProcAddress
1446  * \param procName name of function to look up
1447  * \return function pointer returned by wglGetProcAddress
1448  */
1449 static void *w32gpa(const GLubyte *procName) {
1450   HMODULE oglmod;
1451   void *res = wglGetProcAddress(procName);
1452   if (res) return res;
1453   oglmod = GetModuleHandle("opengl32.dll");
1454   return GetProcAddress(oglmod, procName);
1455 }
1456
1457 int setGlWindow(int *vinfo, HGLRC *context, HWND win)
1458 {
1459   int new_vinfo;
1460   HDC windc = GetDC(win);
1461   HGLRC new_context = 0;
1462   int keep_context = 0;
1463   int res = SET_WINDOW_FAILED;
1464
1465   // should only be needed when keeping context, but not doing glFinish
1466   // can cause flickering even when we do not keep it.
1467   if (*context)
1468   glFinish();
1469   new_vinfo = GetPixelFormat(windc);
1470   if (*context && *vinfo && new_vinfo && *vinfo == new_vinfo) {
1471       // we can keep the wglContext
1472       new_context = *context;
1473       keep_context = 1;
1474   } else {
1475     // create a context
1476     new_context = wglCreateContext(windc);
1477     if (!new_context) {
1478       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GL context!\n");
1479       goto out;
1480     }
1481   }
1482
1483   // set context
1484   if (!wglMakeCurrent(windc, new_context)) {
1485     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GL context!\n");
1486     if (!keep_context) {
1487       wglDeleteContext(new_context);
1488     }
1489     goto out;
1490   }
1491
1492   // set new values
1493   vo_w32_window = win;
1494   {
1495     RECT rect;
1496     GetClientRect(win, &rect);
1497     vo_dwidth = rect.right;
1498     vo_dheight = rect.bottom;
1499   }
1500   if (!keep_context) {
1501     if (*context)
1502       wglDeleteContext(*context);
1503     *context = new_context;
1504     *vinfo = new_vinfo;
1505     getFunctions(w32gpa, NULL);
1506
1507     // and inform that reinit is neccessary
1508     res = SET_WINDOW_REINIT;
1509   } else
1510     res = SET_WINDOW_OK;
1511
1512 out:
1513   ReleaseDC(win, windc);
1514   return res;
1515 }
1516
1517 void releaseGlContext(int *vinfo, HGLRC *context) {
1518   *vinfo = 0;
1519   if (*context) {
1520     wglMakeCurrent(0, 0);
1521     wglDeleteContext(*context);
1522   }
1523   *context = 0;
1524 }
1525
1526 void swapGlBuffers(void) {
1527   HDC vo_hdc = GetDC(vo_w32_window);
1528   SwapBuffers(vo_hdc);
1529   ReleaseDC(vo_w32_window, vo_hdc);
1530 }
1531 #else
1532 #ifdef HAVE_LIBDL
1533 #include <dlfcn.h>
1534 #endif
1535 #include "x11_common.h"
1536 /**
1537  * \brief find address of a linked function
1538  * \param s name of function to find
1539  * \return address of function or NULL if not found
1540  */
1541 static void *getdladdr(const char *s) {
1542   void *ret = NULL;
1543 #ifdef HAVE_LIBDL
1544   void *handle = dlopen(NULL, RTLD_LAZY);
1545   if (!handle)
1546     return NULL;
1547   ret = dlsym(handle, s);
1548   dlclose(handle);
1549 #endif
1550   return ret;
1551 }
1552
1553 /**
1554  * \brief Returns the XVisualInfo associated with Window win.
1555  * \param win Window whose XVisualInfo is returne.
1556  * \return XVisualInfo of the window. Caller must use XFree to free it.
1557  */
1558 static XVisualInfo *getWindowVisualInfo(Window win) {
1559   XWindowAttributes xw_attr;
1560   XVisualInfo vinfo_template;
1561   int tmp;
1562   XGetWindowAttributes(mDisplay, win, &xw_attr);
1563   vinfo_template.visualid = XVisualIDFromVisual(xw_attr.visual);
1564   return XGetVisualInfo(mDisplay, VisualIDMask, &vinfo_template, &tmp);
1565 }
1566
1567 /**
1568  * \brief Changes the window in which video is displayed.
1569  * If possible only transfers the context to the new window, otherwise
1570  * creates a new one, which must be initialized by the caller.
1571  * \param vinfo Currently used visual.
1572  * \param context Currently used context.
1573  * \param win window that should be used for drawing.
1574  * \return one of SET_WINDOW_FAILED, SET_WINDOW_OK or SET_WINDOW_REINIT.
1575  * In case of SET_WINDOW_REINIT the context could not be transfered
1576  * and the caller must initialize it correctly.
1577  * \ingroup glcontext
1578  */
1579 int setGlWindow(XVisualInfo **vinfo, GLXContext *context, Window win)
1580 {
1581   XVisualInfo *new_vinfo;
1582   GLXContext new_context = NULL;
1583   int keep_context = 0;
1584
1585   // should only be needed when keeping context, but not doing glFinish
1586   // can cause flickering even when we do not keep it.
1587   if (*context)
1588   glFinish();
1589   new_vinfo = getWindowVisualInfo(win);
1590   if (*context && *vinfo && new_vinfo &&
1591       (*vinfo)->visualid == new_vinfo->visualid) {
1592       // we can keep the GLXContext
1593       new_context = *context;
1594       XFree(new_vinfo);
1595       new_vinfo = *vinfo;
1596       keep_context = 1;
1597   } else {
1598     // create a context
1599     new_context = glXCreateContext(mDisplay, new_vinfo, NULL, True);
1600     if (!new_context) {
1601       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GLX context!\n");
1602       XFree(new_vinfo);
1603       return SET_WINDOW_FAILED;
1604     }
1605   }
1606
1607   // set context
1608   if (!glXMakeCurrent(mDisplay, vo_window, new_context)) {
1609     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GLX context!\n");
1610     if (!keep_context) {
1611       glXDestroyContext(mDisplay, new_context);
1612       XFree(new_vinfo);
1613     }
1614     return SET_WINDOW_FAILED;
1615   }
1616
1617   // set new values
1618   vo_window = win;
1619   vo_x11_update_geometry();
1620   if (!keep_context) {
1621     void *(*getProcAddress)(const GLubyte *);
1622     const char *(*glXExtStr)(Display *, int);
1623     if (*context)
1624       glXDestroyContext(mDisplay, *context);
1625     *context = new_context;
1626     if (*vinfo)
1627       XFree(*vinfo);
1628     *vinfo = new_vinfo;
1629       getProcAddress = getdladdr("glXGetProcAddress");
1630     if (!getProcAddress)
1631       getProcAddress = getdladdr("glXGetProcAddressARB");
1632     if (!getProcAddress)
1633       getProcAddress = (void *)getdladdr;
1634     glXExtStr = getdladdr("glXQueryExtensionsString");
1635     getFunctions(getProcAddress, !glXExtStr ? NULL :
1636                  glXExtStr(mDisplay, DefaultScreen(mDisplay)));
1637
1638     // and inform that reinit is neccessary
1639     return SET_WINDOW_REINIT;
1640   }
1641   return SET_WINDOW_OK;
1642 }
1643
1644 /**
1645  * \brief free the VisualInfo and GLXContext of an OpenGL context.
1646  * \ingroup glcontext
1647  */
1648 void releaseGlContext(XVisualInfo **vinfo, GLXContext *context) {
1649   if (*vinfo)
1650     XFree(*vinfo);
1651   *vinfo = NULL;
1652   if (*context)
1653   {
1654     glFinish();
1655     glXMakeCurrent(mDisplay, None, NULL);
1656     glXDestroyContext(mDisplay, *context);
1657   }
1658   *context = 0;
1659 }
1660
1661 void swapGlBuffers(void) {
1662   glXSwapBuffers(mDisplay, vo_window);
1663 }
1664 #endif