libvo/gl_common.c

   1 /*
   2  * common OpenGL routines
   3  *
   4  * copyleft (C) 2005 Reimar Döffinger <Reimar.Doeffinger@stud.uni-karlsruhe.de>
   5  * Special thanks go to the xine team and Matthias Hopf, whose video_out_opengl.c
   6  * gave me lots of good ideas.
   7  *
   8  * This file is part of MPlayer.
   9  *
  10  * MPlayer is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published by
  12  * the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * MPlayer is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  * GNU General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public License along
  21  * with MPlayer; if not, write to the Free Software Foundation, Inc.,
  22  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  23  */
  24
  25 /**
  26  * \file gl_common.c
  27  * \brief OpenGL helper functions used by vo_gl.c and vo_gl2.c
  28  */
  29
  30 #include <stdlib.h>
  31 #include <stdio.h>
  32 #include <string.h>
  33 #include <ctype.h>
  34 #include <math.h>
  35 #include "gl_common.h"
  36 #include "libavutil/common.h"
  37
  38 /**
  39  * \defgroup glextfunctions OpenGL extension functions
  40  *
  41  * the pointers to these functions are acquired when the OpenGL
  42  * context is created
  43  * \{
  44  */
  45 void (APIENTRY *GenBuffers)(GLsizei, GLuint *);
  46 void (APIENTRY *DeleteBuffers)(GLsizei, const GLuint *);
  47 void (APIENTRY *BindBuffer)(GLenum, GLuint);
  48 GLvoid* (APIENTRY *MapBuffer)(GLenum, GLenum);
  49 GLboolean (APIENTRY *UnmapBuffer)(GLenum);
  50 void (APIENTRY *BufferData)(GLenum, intptr_t, const GLvoid *, GLenum);
  51 void (APIENTRY *CombinerParameterfv)(GLenum, const GLfloat *);
  52 void (APIENTRY *CombinerParameteri)(GLenum, GLint);
  53 void (APIENTRY *CombinerInput)(GLenum, GLenum, GLenum, GLenum, GLenum,
  54                                GLenum);
  55 void (APIENTRY *CombinerOutput)(GLenum, GLenum, GLenum, GLenum, GLenum,
  56                                 GLenum, GLenum, GLboolean, GLboolean,
  57                                 GLboolean);
  58 void (APIENTRY *BeginFragmentShader)(void);
  59 void (APIENTRY *EndFragmentShader)(void);
  60 void (APIENTRY *SampleMap)(GLuint, GLuint, GLenum);
  61 void (APIENTRY *ColorFragmentOp2)(GLenum, GLuint, GLuint, GLuint, GLuint,
  62                                   GLuint, GLuint, GLuint, GLuint, GLuint);
  63 void (APIENTRY *ColorFragmentOp3)(GLenum, GLuint, GLuint, GLuint, GLuint,
  64                                   GLuint, GLuint, GLuint, GLuint, GLuint,
  65                                   GLuint, GLuint, GLuint);
  66 void (APIENTRY *SetFragmentShaderConstant)(GLuint, const GLfloat *);
  67 void (APIENTRY *ActiveTexture)(GLenum);
  68 void (APIENTRY *BindTexture)(GLenum, GLuint);
  69 void (APIENTRY *MultiTexCoord2f)(GLenum, GLfloat, GLfloat);
  70 void (APIENTRY *GenPrograms)(GLsizei, GLuint *);
  71 void (APIENTRY *DeletePrograms)(GLsizei, const GLuint *);
  72 void (APIENTRY *BindProgram)(GLenum, GLuint);
  73 void (APIENTRY *ProgramString)(GLenum, GLenum, GLsizei, const GLvoid *);
  74 void (APIENTRY *GetProgramiv)(GLenum, GLenum, GLint *);
  75 void (APIENTRY *ProgramEnvParameter4f)(GLenum, GLuint, GLfloat, GLfloat,
  76                                        GLfloat, GLfloat);
  77 int (APIENTRY *SwapInterval)(int);
  78 void (APIENTRY *TexImage3D)(GLenum, GLint, GLenum, GLsizei, GLsizei, GLsizei,
  79                             GLint, GLenum, GLenum, const GLvoid *);
  80 void* (APIENTRY *AllocateMemoryMESA)(void *, int, size_t, float, float, float);
  81 void (APIENTRY *FreeMemoryMESA)(void *, int, void *);
  82 /** \} */ // end of glextfunctions group
  83
  84 //! \defgroup glgeneral OpenGL general helper functions
  85
  86 //! \defgroup glcontext OpenGL context management helper functions
  87
  88 //! \defgroup gltexture OpenGL texture handling helper functions
  89
  90 //! \defgroup glconversion OpenGL conversion helper functions
  91
  92 static GLint hqtexfmt;
  93
  94 /**
  95  * \brief adjusts the GL_UNPACK_ALIGNMENT to fit the stride.
  96  * \param stride number of bytes per line for which alignment should fit.
  97  * \ingroup glgeneral
  98  */
  99 void glAdjustAlignment(int stride) {
 100   GLint gl_alignment;
 101   if (stride % 8 == 0)
 102     gl_alignment=8;
 103   else if (stride % 4 == 0)
 104     gl_alignment=4;
 105   else if (stride % 2 == 0)
 106     gl_alignment=2;
 107   else
 108     gl_alignment=1;
 109   glPixelStorei (GL_UNPACK_ALIGNMENT, gl_alignment);
 110 }
 111
 112 struct gl_name_map_struct {
 113   GLint value;
 114   const char *name;
 115 };
 116
 117 #undef MAP
 118 #define MAP(a) {a, #a}
 119 //! mapping table for the glValName function
 120 static const struct gl_name_map_struct gl_name_map[] = {
 121   // internal format
 122   MAP(GL_R3_G3_B2), MAP(GL_RGB4), MAP(GL_RGB5), MAP(GL_RGB8),
 123   MAP(GL_RGB10), MAP(GL_RGB12), MAP(GL_RGB16), MAP(GL_RGBA2),
 124   MAP(GL_RGBA4), MAP(GL_RGB5_A1), MAP(GL_RGBA8), MAP(GL_RGB10_A2),
 125   MAP(GL_RGBA12), MAP(GL_RGBA16), MAP(GL_LUMINANCE8),
 126
 127   // format
 128   MAP(GL_RGB), MAP(GL_RGBA), MAP(GL_RED), MAP(GL_GREEN), MAP(GL_BLUE),
 129   MAP(GL_ALPHA), MAP(GL_LUMINANCE), MAP(GL_LUMINANCE_ALPHA),
 130   MAP(GL_COLOR_INDEX),
 131   // rest 1.2 only
 132   MAP(GL_BGR), MAP(GL_BGRA),
 133
 134   //type
 135   MAP(GL_BYTE), MAP(GL_UNSIGNED_BYTE), MAP(GL_SHORT), MAP(GL_UNSIGNED_SHORT),
 136   MAP(GL_INT), MAP(GL_UNSIGNED_INT), MAP(GL_FLOAT), MAP(GL_DOUBLE),
 137   MAP(GL_2_BYTES), MAP(GL_3_BYTES), MAP(GL_4_BYTES),
 138   // rest 1.2 only
 139   MAP(GL_UNSIGNED_BYTE_3_3_2), MAP(GL_UNSIGNED_BYTE_2_3_3_REV),
 140   MAP(GL_UNSIGNED_SHORT_5_6_5), MAP(GL_UNSIGNED_SHORT_5_6_5_REV),
 141   MAP(GL_UNSIGNED_SHORT_4_4_4_4), MAP(GL_UNSIGNED_SHORT_4_4_4_4_REV),
 142   MAP(GL_UNSIGNED_SHORT_5_5_5_1), MAP(GL_UNSIGNED_SHORT_1_5_5_5_REV),
 143   MAP(GL_UNSIGNED_INT_8_8_8_8), MAP(GL_UNSIGNED_INT_8_8_8_8_REV),
 144   MAP(GL_UNSIGNED_INT_10_10_10_2), MAP(GL_UNSIGNED_INT_2_10_10_10_REV),
 145   {0, 0}
 146 };
 147 #undef MAP
 148
 149 /**
 150  * \brief return the name of an OpenGL constant
 151  * \param value the constant
 152  * \return name of the constant or "Unknown format!"
 153  * \ingroup glgeneral
 154  */
 155 const char *glValName(GLint value)
 156 {
 157   int i = 0;
 158
 159   while (gl_name_map[i].name) {
 160     if (gl_name_map[i].value == value)
 161       return gl_name_map[i].name;
 162     i++;
 163   }
 164   return "Unknown format!";
 165 }
 166
 167 //! always return this format as internal texture format in glFindFormat
 168 #define TEXTUREFORMAT_ALWAYS GL_RGB8
 169 #undef TEXTUREFORMAT_ALWAYS
 170
 171 /**
 172  * \brief find the OpenGL settings coresponding to format.
 173  *
 174  * All parameters may be NULL.
 175  * \param fmt MPlayer format to analyze.
 176  * \param bpp [OUT] bits per pixel of that format.
 177  * \param gl_texfmt [OUT] internal texture format that fits the
 178  * image format, not necessarily the best for performance.
 179  * \param gl_format [OUT] OpenGL format for this image format.
 180  * \param gl_type [OUT] OpenGL type for this image format.
 181  * \return 1 if format is supported by OpenGL, 0 if not.
 182  * \ingroup gltexture
 183  */
 184 int glFindFormat(uint32_t fmt, int *bpp, GLint *gl_texfmt,
 185                   GLenum *gl_format, GLenum *gl_type)
 186 {
 187   int supported = 1;
 188   int dummy1;
 189   GLenum dummy2;
 190   GLint dummy3;
 191   if (!bpp) bpp = &dummy1;
 192   if (!gl_texfmt) gl_texfmt = &dummy3;
 193   if (!gl_format) gl_format = &dummy2;
 194   if (!gl_type) gl_type = &dummy2;
 195
 196   *bpp = IMGFMT_IS_BGR(fmt)?IMGFMT_BGR_DEPTH(fmt):IMGFMT_RGB_DEPTH(fmt);
 197   *gl_texfmt = 3;
 198   switch (fmt) {
 199     case IMGFMT_RGB48NE:
 200       *gl_format = GL_RGB;
 201       *gl_type = GL_UNSIGNED_SHORT;
 202       break;
 203     case IMGFMT_RGB24:
 204       *gl_format = GL_RGB;
 205       *gl_type = GL_UNSIGNED_BYTE;
 206       break;
 207     case IMGFMT_RGBA:
 208       *gl_texfmt = 4;
 209       *gl_format = GL_RGBA;
 210       *gl_type = GL_UNSIGNED_BYTE;
 211       break;
 212     case IMGFMT_YV12:
 213       supported = 0; // no native YV12 support
 214     case IMGFMT_Y800:
 215     case IMGFMT_Y8:
 216       *gl_texfmt = 1;
 217       *bpp = 8;
 218       *gl_format = GL_LUMINANCE;
 219       *gl_type = GL_UNSIGNED_BYTE;
 220       break;
 221     case IMGFMT_UYVY:
 222     case IMGFMT_YUY2:
 223       *gl_texfmt = GL_YCBCR_MESA;
 224       *bpp = 16;
 225       *gl_format = GL_YCBCR_MESA;
 226       *gl_type = fmt == IMGFMT_UYVY ? GL_UNSIGNED_SHORT_8_8 : GL_UNSIGNED_SHORT_8_8_REV;
 227       break;
 228 #if 0
 229     // we do not support palettized formats, although the format the
 230     // swscale produces works
 231     case IMGFMT_RGB8:
 232       gl_format = GL_RGB;
 233       gl_type = GL_UNSIGNED_BYTE_2_3_3_REV;
 234       break;
 235 #endif
 236     case IMGFMT_RGB15:
 237       *gl_format = GL_RGBA;
 238       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 239       break;
 240     case IMGFMT_RGB16:
 241       *gl_format = GL_RGB;
 242       *gl_type = GL_UNSIGNED_SHORT_5_6_5_REV;
 243       break;
 244 #if 0
 245     case IMGFMT_BGR8:
 246       // special case as red and blue have a differen number of bits.
 247       // GL_BGR and GL_UNSIGNED_BYTE_3_3_2 isn't supported at least
 248       // by nVidia drivers, and in addition would give more bits to
 249       // blue than to red, which isn't wanted
 250       gl_format = GL_RGB;
 251       gl_type = GL_UNSIGNED_BYTE_3_3_2;
 252       break;
 253 #endif
 254     case IMGFMT_BGR15:
 255       *gl_format = GL_BGRA;
 256       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 257       break;
 258     case IMGFMT_BGR16:
 259       *gl_format = GL_RGB;
 260       *gl_type = GL_UNSIGNED_SHORT_5_6_5;
 261       break;
 262     case IMGFMT_BGR24:
 263       *gl_format = GL_BGR;
 264       *gl_type = GL_UNSIGNED_BYTE;
 265       break;
 266     case IMGFMT_BGRA:
 267       *gl_texfmt = 4;
 268       *gl_format = GL_BGRA;
 269       *gl_type = GL_UNSIGNED_BYTE;
 270       break;
 271     default:
 272       *gl_texfmt = 4;
 273       *gl_format = GL_RGBA;
 274       *gl_type = GL_UNSIGNED_BYTE;
 275       supported = 0;
 276   }
 277 #ifdef TEXTUREFORMAT_ALWAYS
 278   *gl_texfmt = TEXTUREFORMAT_ALWAYS;
 279 #endif
 280   return supported;
 281 }
 282
 283 static void *setNull(const GLubyte *s) {
 284   return NULL;
 285 }
 286
 287 typedef struct {
 288   void *funcptr;
 289   const char *extstr;
 290   const char *funcnames[7];
 291 } extfunc_desc_t;
 292
 293 static const extfunc_desc_t extfuncs[] = {
 294   {&GenBuffers, NULL, {"glGenBuffers", "glGenBuffersARB", NULL}},
 295   {&DeleteBuffers, NULL, {"glDeleteBuffers", "glDeleteBuffersARB", NULL}},
 296   {&BindBuffer, NULL, {"glBindBuffer", "glBindBufferARB", NULL}},
 297   {&MapBuffer, NULL, {"glMapBuffer", "glMapBufferARB", NULL}},
 298   {&UnmapBuffer, NULL, {"glUnmapBuffer", "glUnmapBufferARB", NULL}},
 299   {&BufferData, NULL, {"glBufferData", "glBufferDataARB", NULL}},
 300   {&CombinerParameterfv, "NV_register_combiners", {"glCombinerParameterfv", "glCombinerParameterfvNV", NULL}},
 301   {&CombinerParameteri, "NV_register_combiners", {"glCombinerParameteri", "glCombinerParameteriNV", NULL}},
 302   {&CombinerInput, "NV_register_combiners", {"glCombinerInput", "glCombinerInputNV", NULL}},
 303   {&CombinerOutput, "NV_register_combiners", {"glCombinerOutput", "glCombinerOutputNV", NULL}},
 304   {&BeginFragmentShader, "ATI_fragment_shader", {"glBeginFragmentShaderATI", NULL}},
 305   {&EndFragmentShader, "ATI_fragment_shader", {"glEndFragmentShaderATI", NULL}},
 306   {&SampleMap, "ATI_fragment_shader", {"glSampleMapATI", NULL}},
 307   {&ColorFragmentOp2, "ATI_fragment_shader", {"glColorFragmentOp2ATI", NULL}},
 308   {&ColorFragmentOp3, "ATI_fragment_shader", {"glColorFragmentOp3ATI", NULL}},
 309   {&SetFragmentShaderConstant, "ATI_fragment_shader", {"glSetFragmentShaderConstantATI", NULL}},
 310   {&ActiveTexture, NULL, {"glActiveTexture", "glActiveTextureARB", NULL}},
 311   {&BindTexture, NULL, {"glBindTexture", "glBindTextureARB", "glBindTextureEXT", NULL}},
 312   {&MultiTexCoord2f, NULL, {"glMultiTexCoord2f", "glMultiTexCoord2fARB", NULL}},
 313   {&GenPrograms, "_program", {"glGenProgramsARB", NULL}},
 314   {&DeletePrograms, "_program", {"glDeleteProgramsARB", NULL}},
 315   {&BindProgram, "_program", {"glBindProgramARB", NULL}},
 316   {&ProgramString, "_program", {"glProgramStringARB", NULL}},
 317   {&GetProgramiv, "_program", {"glGetProgramivARB", NULL}},
 318   {&ProgramEnvParameter4f, "_program", {"glProgramEnvParameter4fARB", NULL}},
 319   {&SwapInterval, "_swap_control", {"glXSwapInterval", "glXSwapIntervalEXT", "glXSwapIntervalSGI", "wglSwapInterval", "wglSwapIntervalEXT", "wglSwapIntervalSGI", NULL}},
 320   {&TexImage3D, NULL, {"glTexImage3D", NULL}},
 321   {&AllocateMemoryMESA, "GLX_MESA_allocate_memory", {"glXAllocateMemoryMESA", NULL}},
 322   {&FreeMemoryMESA, "GLX_MESA_allocate_memory", {"glXFreeMemoryMESA", NULL}},
 323   {NULL}
 324 };
 325
 326 /**
 327  * \brief find the function pointers of some useful OpenGL extensions
 328  * \param getProcAddress function to resolve function names, may be NULL
 329  * \param ext2 an extra extension string
 330  */
 331 static void getFunctions(void *(*getProcAddress)(const GLubyte *),
 332                          const char *ext2) {
 333   const extfunc_desc_t *dsc;
 334   const char *extensions = (const char *)glGetString(GL_EXTENSIONS);
 335   char *allexts;
 336   if (!extensions) extensions = "";
 337   if (!ext2) ext2 = "";
 338   allexts = malloc(strlen(extensions) + strlen(ext2) + 2);
 339   strcpy(allexts, extensions);
 340   strcat(allexts, " ");
 341   strcat(allexts, ext2);
 342   mp_msg(MSGT_VO, MSGL_DBG2, "OpenGL extensions string:\n%s\n", allexts);
 343   if (!getProcAddress)
 344     getProcAddress = setNull;
 345   for (dsc = extfuncs; dsc->funcptr; dsc++) {
 346     void *ptr = NULL;
 347     int i;
 348     if (!dsc->extstr || strstr(allexts, dsc->extstr)) {
 349       for (i = 0; !ptr && dsc->funcnames[i]; i++)
 350         ptr = getProcAddress((const GLubyte *)dsc->funcnames[i]);
 351     }
 352     *(void **)dsc->funcptr = ptr;
 353   }
 354   if (strstr(allexts, "_texture_float"))
 355     hqtexfmt = GL_RGB32F;
 356   else if (strstr(allexts, "NV_float_buffer"))
 357     hqtexfmt = GL_FLOAT_RGB32_NV;
 358   else
 359     hqtexfmt = GL_RGB16;
 360   free(allexts);
 361 }
 362
 363 /**
 364  * \brief create a texture and set some defaults
 365  * \param target texture taget, usually GL_TEXTURE_2D
 366  * \param fmt internal texture format
 367  * \param format texture host data format
 368  * \param type texture host data type
 369  * \param filter filter used for scaling, e.g. GL_LINEAR
 370  * \param w texture width
 371  * \param h texture height
 372  * \param val luminance value to fill texture with
 373  * \ingroup gltexture
 374  */
 375 void glCreateClearTex(GLenum target, GLenum fmt, GLenum format, GLenum type, GLint filter,
 376                       int w, int h, unsigned char val) {
 377   GLfloat fval = (GLfloat)val / 255.0;
 378   GLfloat border[4] = {fval, fval, fval, fval};
 379   int stride = w * glFmt2bpp(format, type);
 380   char *init;
 381   if (!stride) return;
 382   init = malloc(stride * h);
 383   memset(init, val, stride * h);
 384   glAdjustAlignment(stride);
 385   glPixelStorei(GL_UNPACK_ROW_LENGTH, w);
 386   glTexImage2D(target, 0, fmt, w, h, 0, format, type, init);
 387   glTexParameterf(target, GL_TEXTURE_PRIORITY, 1.0);
 388   glTexParameteri(target, GL_TEXTURE_MIN_FILTER, filter);
 389   glTexParameteri(target, GL_TEXTURE_MAG_FILTER, filter);
 390   glTexParameteri(target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 391   glTexParameteri(target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 392   // Border texels should not be used with CLAMP_TO_EDGE
 393   // We set a sane default anyway.
 394   glTexParameterfv(target, GL_TEXTURE_BORDER_COLOR, border);
 395   free(init);
 396 }
 397
 398 /**
 399  * \brief skips whitespace and comments
 400  * \param f file to read from
 401  */
 402 static void ppm_skip(FILE *f) {
 403   int c, comment = 0;
 404   do {
 405     c = fgetc(f);
 406     if (c == '#')
 407       comment = 1;
 408     if (c == '\n')
 409       comment = 0;
 410   } while (c != EOF && (isspace(c) || comment));
 411   if (c != EOF)
 412     ungetc(c, f);
 413 }
 414
 415 #define MAXDIM (16 * 1024)
 416
 417 /**
 418  * \brief creates a texture from a PPM file
 419  * \param target texture taget, usually GL_TEXTURE_2D
 420  * \param fmt internal texture format, 0 for default
 421  * \param filter filter used for scaling, e.g. GL_LINEAR
 422  * \param f file to read PPM from
 423  * \param width [out] width of texture
 424  * \param height [out] height of texture
 425  * \param maxval [out] maxval value from PPM file
 426  * \return 0 on error, 1 otherwise
 427  * \ingroup gltexture
 428  */
 429 int glCreatePPMTex(GLenum target, GLenum fmt, GLint filter,
 430                    FILE *f, int *width, int *height, int *maxval) {
 431   unsigned w, h, m, val, bpp;
 432   char *data;
 433   GLenum type;
 434   ppm_skip(f);
 435   if (fgetc(f) != 'P' || fgetc(f) != '6')
 436     return 0;
 437   ppm_skip(f);
 438   if (fscanf(f, "%u", &w) != 1)
 439     return 0;
 440   ppm_skip(f);
 441   if (fscanf(f, "%u", &h) != 1)
 442     return 0;
 443   ppm_skip(f);
 444   if (fscanf(f, "%u", &m) != 1)
 445     return 0;
 446   val = fgetc(f);
 447   if (!isspace(val))
 448     return 0;
 449   if (w > MAXDIM || h > MAXDIM)
 450     return 0;
 451   bpp = (m > 255) ? 6 : 3;
 452   data = malloc(w * h * bpp);
 453   if (fread(data, w * bpp, h, f) != h)
 454     return 0;
 455   if (!fmt) {
 456     fmt = (m > 255) ? hqtexfmt : 3;
 457     if (fmt == GL_FLOAT_RGB32_NV && target != GL_TEXTURE_RECTANGLE)
 458       fmt = GL_RGB16;
 459   }
 460   type = m > 255 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 461   glCreateClearTex(target, fmt, GL_RGB, type, filter, w, h, 0);
 462   glUploadTex(target, GL_RGB, type,
 463               data, w * bpp, 0, 0, w, h, 0);
 464   free(data);
 465   if (width) *width = w;
 466   if (height) *height = h;
 467   if (maxval) *maxval = m;
 468   return 1;
 469 }
 470
 471 /**
 472  * \brief return the number of bytes per pixel for the given format
 473  * \param format OpenGL format
 474  * \param type OpenGL type
 475  * \return bytes per pixel
 476  * \ingroup glgeneral
 477  *
 478  * Does not handle all possible variants, just those used by MPlayer
 479  */
 480 int glFmt2bpp(GLenum format, GLenum type) {
 481   int component_size = 0;
 482   switch (type) {
 483     case GL_UNSIGNED_BYTE_3_3_2:
 484     case GL_UNSIGNED_BYTE_2_3_3_REV:
 485       return 1;
 486     case GL_UNSIGNED_SHORT_5_5_5_1:
 487     case GL_UNSIGNED_SHORT_1_5_5_5_REV:
 488     case GL_UNSIGNED_SHORT_5_6_5:
 489     case GL_UNSIGNED_SHORT_5_6_5_REV:
 490       return 2;
 491     case GL_UNSIGNED_BYTE:
 492       component_size = 1;
 493       break;
 494     case GL_UNSIGNED_SHORT:
 495       component_size = 2;
 496       break;
 497   }
 498   switch (format) {
 499     case GL_LUMINANCE:
 500     case GL_ALPHA:
 501       return component_size;
 502     case GL_YCBCR_MESA:
 503       return 2;
 504     case GL_RGB:
 505     case GL_BGR:
 506       return 3 * component_size;
 507     case GL_RGBA:
 508     case GL_BGRA:
 509       return 4 * component_size;
 510   }
 511   return 0; // unknown
 512 }
 513
 514 /**
 515  * \brief upload a texture, handling things like stride and slices
 516  * \param target texture target, usually GL_TEXTURE_2D
 517  * \param format OpenGL format of data
 518  * \param type OpenGL type of data
 519  * \param dataptr data to upload
 520  * \param stride data stride
 521  * \param x x offset in texture
 522  * \param y y offset in texture
 523  * \param w width of the texture part to upload
 524  * \param h height of the texture part to upload
 525  * \param slice height of an upload slice, 0 for all at once
 526  * \ingroup gltexture
 527  */
 528 void glUploadTex(GLenum target, GLenum format, GLenum type,
 529                  const void *dataptr, int stride,
 530                  int x, int y, int w, int h, int slice) {
 531   const uint8_t *data = dataptr;
 532   int y_max = y + h;
 533   if (w <= 0 || h <= 0) return;
 534   if (slice <= 0)
 535     slice = h;
 536   if (stride < 0) {
 537     data += (h - 1) * stride;
 538     stride = -stride;
 539   }
 540   // this is not always correct, but should work for MPlayer
 541   glAdjustAlignment(stride);
 542   glPixelStorei(GL_UNPACK_ROW_LENGTH, stride / glFmt2bpp(format, type));
 543   for (; y + slice <= y_max; y += slice) {
 544     glTexSubImage2D(target, 0, x, y, w, slice, format, type, data);
 545     data += stride * slice;
 546   }
 547   if (y < y_max)
 548     glTexSubImage2D(target, 0, x, y, w, y_max - y, format, type, data);
 549 }
 550
 551 static void fillUVcoeff(GLfloat *ucoef, GLfloat *vcoef,
 552                         float uvcos, float uvsin) {
 553   int i;
 554   ucoef[0] = 0 * uvcos + 1.403 * uvsin;
 555   vcoef[0] = 0 * uvsin + 1.403 * uvcos;
 556   ucoef[1] = -0.344 * uvcos + -0.714 * uvsin;
 557   vcoef[1] = -0.344 * uvsin + -0.714 * uvcos;
 558   ucoef[2] = 1.770 * uvcos + 0 * uvsin;
 559   vcoef[2] = 1.770 * uvsin + 0 * uvcos;
 560   ucoef[3] = 0;
 561   vcoef[3] = 0;
 562   // Coefficients (probably) must be in [0, 1] range, whereas they originally
 563   // are in [-2, 2] range, so here comes the trick:
 564   // First put them in the [-0.5, 0.5] range, then add 0.5.
 565   // This can be undone with the HALF_BIAS and SCALE_BY_FOUR arguments
 566   // for CombinerInput and CombinerOutput (or the respective ATI variants)
 567   for (i = 0; i < 4; i++) {
 568     ucoef[i] = ucoef[i] * 0.25 + 0.5;
 569     vcoef[i] = vcoef[i] * 0.25 + 0.5;
 570   }
 571 }
 572
 573 /**
 574  * \brief Setup register combiners for YUV to RGB conversion.
 575  * \param uvcos used for saturation and hue adjustment
 576  * \param uvsin used for saturation and hue adjustment
 577  */
 578 static void glSetupYUVCombiners(float uvcos, float uvsin) {
 579   GLfloat ucoef[4];
 580   GLfloat vcoef[4];
 581   GLint i;
 582   if (!CombinerInput || !CombinerOutput ||
 583       !CombinerParameterfv || !CombinerParameteri) {
 584     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner functions missing!\n");
 585     return;
 586   }
 587   glGetIntegerv(GL_MAX_GENERAL_COMBINERS_NV, &i);
 588   if (i < 2)
 589     mp_msg(MSGT_VO, MSGL_ERR,
 590            "[gl] 2 general combiners needed for YUV combiner support (found %i)\n", i);
 591   glGetIntegerv(GL_MAX_TEXTURE_UNITS, &i);
 592   if (i < 3)
 593     mp_msg(MSGT_VO, MSGL_ERR,
 594            "[gl] 3 texture units needed for YUV combiner support (found %i)\n", i);
 595   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 596   CombinerParameterfv(GL_CONSTANT_COLOR0_NV, ucoef);
 597   CombinerParameterfv(GL_CONSTANT_COLOR1_NV, vcoef);
 598
 599   // UV first, like this green component cannot overflow
 600   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_A_NV,
 601                 GL_TEXTURE1, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 602   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_B_NV,
 603                 GL_CONSTANT_COLOR0_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 604   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_C_NV,
 605                 GL_TEXTURE2, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 606   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_D_NV,
 607                 GL_CONSTANT_COLOR1_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 608   CombinerOutput(GL_COMBINER0_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 609                  GL_SPARE0_NV, GL_SCALE_BY_FOUR_NV, GL_NONE, GL_FALSE,
 610                  GL_FALSE, GL_FALSE);
 611
 612   // stage 2
 613   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_A_NV, GL_SPARE0_NV,
 614                 GL_SIGNED_IDENTITY_NV, GL_RGB);
 615   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_B_NV, GL_ZERO,
 616                  GL_UNSIGNED_INVERT_NV, GL_RGB);
 617   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_C_NV,
 618                 GL_TEXTURE0, GL_SIGNED_IDENTITY_NV, GL_RGB);
 619   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_D_NV, GL_ZERO,
 620                 GL_UNSIGNED_INVERT_NV, GL_RGB);
 621   CombinerOutput(GL_COMBINER1_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 622                  GL_SPARE0_NV, GL_NONE, GL_NONE, GL_FALSE,
 623                  GL_FALSE, GL_FALSE);
 624
 625   // leave final combiner stage in default mode
 626   CombinerParameteri(GL_NUM_GENERAL_COMBINERS_NV, 2);
 627 }
 628
 629 /**
 630  * \brief Setup ATI version of register combiners for YUV to RGB conversion.
 631  * \param uvcos used for saturation and hue adjustment
 632  * \param uvsin used for saturation and hue adjustment
 633  *
 634  * ATI called this fragment shader, but the name is confusing in the
 635  * light of a very different OpenGL 2.0 extension with the same name
 636  */
 637 static void glSetupYUVCombinersATI(float uvcos, float uvsin) {
 638   GLfloat ucoef[4];
 639   GLfloat vcoef[4];
 640   GLint i;
 641   if (!BeginFragmentShader || !EndFragmentShader ||
 642       !SetFragmentShaderConstant || !SampleMap ||
 643       !ColorFragmentOp2 || !ColorFragmentOp3) {
 644     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner (ATI) functions missing!\n");
 645     return;
 646   }
 647   glGetIntegerv(GL_NUM_FRAGMENT_REGISTERS_ATI, &i);
 648   if (i < 3)
 649     mp_msg(MSGT_VO, MSGL_ERR,
 650            "[gl] 3 registers needed for YUV combiner (ATI) support (found %i)\n", i);
 651   glGetIntegerv (GL_MAX_TEXTURE_UNITS, &i);
 652   if (i < 3)
 653     mp_msg(MSGT_VO, MSGL_ERR,
 654            "[gl] 3 texture units needed for YUV combiner (ATI) support (found %i)\n", i);
 655   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 656   BeginFragmentShader();
 657   SetFragmentShaderConstant(GL_CON_0_ATI, ucoef);
 658   SetFragmentShaderConstant(GL_CON_1_ATI, vcoef);
 659   SampleMap(GL_REG_0_ATI, GL_TEXTURE0, GL_SWIZZLE_STR_ATI);
 660   SampleMap(GL_REG_1_ATI, GL_TEXTURE1, GL_SWIZZLE_STR_ATI);
 661   SampleMap(GL_REG_2_ATI, GL_TEXTURE2, GL_SWIZZLE_STR_ATI);
 662   // UV first, like this green component cannot overflow
 663   ColorFragmentOp2(GL_MUL_ATI, GL_REG_1_ATI, GL_NONE, GL_NONE,
 664                    GL_REG_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 665                    GL_CON_0_ATI, GL_NONE, GL_BIAS_BIT_ATI);
 666   ColorFragmentOp3(GL_MAD_ATI, GL_REG_2_ATI, GL_NONE, GL_4X_BIT_ATI,
 667                    GL_REG_2_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 668                    GL_CON_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 669                    GL_REG_1_ATI, GL_NONE, GL_NONE);
 670   ColorFragmentOp2(GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 671                    GL_REG_0_ATI, GL_NONE, GL_NONE,
 672                    GL_REG_2_ATI, GL_NONE, GL_NONE);
 673   EndFragmentShader();
 674 }
 675
 676 /**
 677  * \brief helper function for gen_spline_lookup_tex
 678  * \param x subpixel-position ((0,1) range) to calculate weights for
 679  * \param dst where to store transformed weights, must provide space for 4 GLfloats
 680  *
 681  * calculates the weights and stores them after appropriate transformation
 682  * for the scaler fragment program.
 683  */
 684 static void store_weights(float x, GLfloat *dst) {
 685   float w0 = (((-1 * x + 3) * x - 3) * x + 1) / 6;
 686   float w1 = ((( 3 * x - 6) * x + 0) * x + 4) / 6;
 687   float w2 = (((-3 * x + 3) * x + 3) * x + 1) / 6;
 688   float w3 = ((( 1 * x + 0) * x + 0) * x + 0) / 6;
 689   *dst++ = 1 + x - w1 / (w0 + w1);
 690   *dst++ = 1 - x + w3 / (w2 + w3);
 691   *dst++ = w0 + w1;
 692   *dst++ = 0;
 693 }
 694
 695 //! to avoid artefacts this should be rather large
 696 #define LOOKUP_BSPLINE_RES (2 * 1024)
 697 /**
 698  * \brief creates the 1D lookup texture needed for fast higher-order filtering
 699  * \param unit texture unit to attach texture to
 700  */
 701 static void gen_spline_lookup_tex(GLenum unit) {
 702   GLfloat *tex = calloc(4 * LOOKUP_BSPLINE_RES, sizeof(*tex));
 703   GLfloat *tp = tex;
 704   int i;
 705   for (i = 0; i < LOOKUP_BSPLINE_RES; i++) {
 706     float x = (float)(i + 0.5) / LOOKUP_BSPLINE_RES;
 707     store_weights(x, tp);
 708     tp += 4;
 709   }
 710   store_weights(0, tex);
 711   store_weights(1, &tex[4 * (LOOKUP_BSPLINE_RES - 1)]);
 712   ActiveTexture(unit);
 713   glTexImage1D(GL_TEXTURE_1D, 0, GL_RGBA16, LOOKUP_BSPLINE_RES, 0, GL_RGBA, GL_FLOAT, tex);
 714   glTexParameterf(GL_TEXTURE_1D, GL_TEXTURE_PRIORITY, 1.0);
 715   glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 716   glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 717   glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_REPEAT);
 718   ActiveTexture(GL_TEXTURE0);
 719   free(tex);
 720 }
 721
 722 static const char *bilin_filt_template =
 723   "TEX yuv.%c, fragment.texcoord[%c], texture[%c], %s;";
 724
 725 #define BICUB_FILT_MAIN(textype) \
 726   /* first y-interpolation */ \
 727   "ADD coord, fragment.texcoord[%c].xyxy, cdelta.xyxw;" \
 728   "ADD coord2, fragment.texcoord[%c].xyxy, cdelta.zyzw;" \
 729   "TEX a.r, coord.xyxy, texture[%c], "textype";" \
 730   "TEX a.g, coord.zwzw, texture[%c], "textype";" \
 731   /* second y-interpolation */ \
 732   "TEX b.r, coord2.xyxy, texture[%c], "textype";" \
 733   "TEX b.g, coord2.zwzw, texture[%c], "textype";" \
 734   "LRP a.b, parmy.b, a.rrrr, a.gggg;" \
 735   "LRP a.a, parmy.b, b.rrrr, b.gggg;" \
 736   /* x-interpolation */ \
 737   "LRP yuv.%c, parmx.b, a.bbbb, a.aaaa;"
 738
 739 static const char *bicub_filt_template_2D =
 740   "MAD coord.xy, fragment.texcoord[%c], {%f, %f}, {0.5, 0.5};"
 741   "TEX parmx, coord.x, texture[%c], 1D;"
 742   "MUL cdelta.xz, parmx.rrgg, {-%f, 0, %f, 0};"
 743   "TEX parmy, coord.y, texture[%c], 1D;"
 744   "MUL cdelta.yw, parmy.rrgg, {0, -%f, 0, %f};"
 745   BICUB_FILT_MAIN("2D");
 746
 747 static const char *bicub_filt_template_RECT =
 748   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};"
 749   "TEX parmx, coord.x, texture[%c], 1D;"
 750   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};"
 751   "TEX parmy, coord.y, texture[%c], 1D;"
 752   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};"
 753   BICUB_FILT_MAIN("RECT");
 754
 755 #define CALCWEIGHTS(t, s) \
 756   "MAD "t", {-0.5, 0.1666, 0.3333, -0.3333}, "s", {1, 0, -0.5, 0.5};" \
 757   "MAD "t", "t", "s", {0, 0, -0.5, 0.5};" \
 758   "MAD "t", "t", "s", {-0.6666, 0, 0.8333, 0.1666};" \
 759   "RCP a.x, "t".z;" \
 760   "RCP a.y, "t".w;" \
 761   "MAD "t".xy, "t".xyxy, a.xyxy, {1, 1, 0, 0};" \
 762   "ADD "t".x, "t".xxxx, "s";" \
 763   "SUB "t".y, "t".yyyy, "s";"
 764
 765 static const char *bicub_notex_filt_template_2D =
 766   "MAD coord.xy, fragment.texcoord[%c], {%f, %f}, {0.5, 0.5};"
 767   "FRC coord.xy, coord.xyxy;"
 768   CALCWEIGHTS("parmx", "coord.xxxx")
 769   "MUL cdelta.xz, parmx.rrgg, {-%f, 0, %f, 0};"
 770   CALCWEIGHTS("parmy", "coord.yyyy")
 771   "MUL cdelta.yw, parmy.rrgg, {0, -%f, 0, %f};"
 772   BICUB_FILT_MAIN("2D");
 773
 774 static const char *bicub_notex_filt_template_RECT =
 775   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};"
 776   "FRC coord.xy, coord.xyxy;"
 777   CALCWEIGHTS("parmx", "coord.xxxx")
 778   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};"
 779   CALCWEIGHTS("parmy", "coord.yyyy")
 780   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};"
 781   BICUB_FILT_MAIN("RECT");
 782
 783 #define BICUB_X_FILT_MAIN(textype) \
 784   "ADD coord.xy, fragment.texcoord[%c].xyxy, cdelta.xyxy;" \
 785   "ADD coord2.xy, fragment.texcoord[%c].xyxy, cdelta.zyzy;" \
 786   "TEX a.r, coord, texture[%c], "textype";" \
 787   "TEX b.r, coord2, texture[%c], "textype";" \
 788   /* x-interpolation */ \
 789   "LRP yuv.%c, parmx.b, a.rrrr, b.rrrr;"
 790
 791 static const char *bicub_x_filt_template_2D =
 792   "MAD coord.x, fragment.texcoord[%c], {%f}, {0.5};"
 793   "TEX parmx, coord, texture[%c], 1D;"
 794   "MUL cdelta.xyz, parmx.rrgg, {-%f, 0, %f};"
 795   BICUB_X_FILT_MAIN("2D");
 796
 797 static const char *bicub_x_filt_template_RECT =
 798   "ADD coord.x, fragment.texcoord[%c], {0.5};"
 799   "TEX parmx, coord, texture[%c], 1D;"
 800   "MUL cdelta.xyz, parmx.rrgg, {-1, 0, 1};"
 801   BICUB_X_FILT_MAIN("RECT");
 802
 803 static const char *unsharp_filt_template =
 804   "PARAM dcoord%c = {%f, %f, %f, %f};"
 805   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;"
 806   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;"
 807   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;"
 808   "TEX b.r, coord.xyxy, texture[%c], %s;"
 809   "TEX b.g, coord.zwzw, texture[%c], %s;"
 810   "ADD b.r, b.r, b.g;"
 811   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 812   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 813   "DP3 b, b, {0.25, 0.25, 0.25};"
 814   "SUB b.r, a.r, b.r;"
 815   "MAD yuv.%c, b.r, {%f}, a.r;";
 816
 817 static const char *unsharp_filt_template2 =
 818   "PARAM dcoord%c = {%f, %f, %f, %f};"
 819   "PARAM dcoord2%c = {%f, 0, 0, %f};"
 820   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;"
 821   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;"
 822   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;"
 823   "TEX b.r, coord.xyxy, texture[%c], %s;"
 824   "TEX b.g, coord.zwzw, texture[%c], %s;"
 825   "ADD b.r, b.r, b.g;"
 826   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 827   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 828   "ADD b.r, b.r, b.b;"
 829   "ADD b.a, b.r, b.g;"
 830   "ADD coord, fragment.texcoord[%c].xyxy, dcoord2%c;"
 831   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord2%c;"
 832   "TEX b.r, coord.xyxy, texture[%c], %s;"
 833   "TEX b.g, coord.zwzw, texture[%c], %s;"
 834   "ADD b.r, b.r, b.g;"
 835   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 836   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 837   "DP4 b.r, b, {-0.1171875, -0.1171875, -0.1171875, -0.09765625};"
 838   "MAD b.r, a.r, {0.859375}, b.r;"
 839   "MAD yuv.%c, b.r, {%f}, a.r;";
 840
 841 static const char *yuv_prog_template =
 842   "PARAM ycoef = {%.4f, %.4f, %.4f};"
 843   "PARAM ucoef = {%.4f, %.4f, %.4f};"
 844   "PARAM vcoef = {%.4f, %.4f, %.4f};"
 845   "PARAM offsets = {%.4f, %.4f, %.4f};"
 846   "TEMP res;"
 847   "MAD res.rgb, yuv.rrrr, ycoef, offsets;"
 848   "MAD res.rgb, yuv.gggg, ucoef, res;"
 849   "MAD result.color.rgb, yuv.bbbb, vcoef, res;"
 850   "END";
 851
 852 static const char *yuv_pow_prog_template =
 853   "PARAM ycoef = {%.4f, %.4f, %.4f};"
 854   "PARAM ucoef = {%.4f, %.4f, %.4f};"
 855   "PARAM vcoef = {%.4f, %.4f, %.4f};"
 856   "PARAM offsets = {%.4f, %.4f, %.4f};"
 857   "PARAM gamma = {%.4f, %.4f, %.4f};"
 858   "TEMP res;"
 859   "MAD res.rgb, yuv.rrrr, ycoef, offsets;"
 860   "MAD res.rgb, yuv.gggg, ucoef, res;"
 861   "MAD_SAT res.rgb, yuv.bbbb, vcoef, res;"
 862   "POW result.color.r, res.r, gamma.r;"
 863   "POW result.color.g, res.g, gamma.g;"
 864   "POW result.color.b, res.b, gamma.b;"
 865   "END";
 866
 867 static const char *yuv_lookup_prog_template =
 868   "PARAM ycoef = {%.4f, %.4f, %.4f, 0};"
 869   "PARAM ucoef = {%.4f, %.4f, %.4f, 0};"
 870   "PARAM vcoef = {%.4f, %.4f, %.4f, 0};"
 871   "PARAM offsets = {%.4f, %.4f, %.4f, 0.125};"
 872   "TEMP res;"
 873   "MAD res, yuv.rrrr, ycoef, offsets;"
 874   "MAD res.rgb, yuv.gggg, ucoef, res;"
 875   "MAD res.rgb, yuv.bbbb, vcoef, res;"
 876   "TEX result.color.r, res.raaa, texture[%c], 2D;"
 877   "ADD res.a, res.a, 0.25;"
 878   "TEX result.color.g, res.gaaa, texture[%c], 2D;"
 879   "ADD res.a, res.a, 0.25;"
 880   "TEX result.color.b, res.baaa, texture[%c], 2D;"
 881   "END";
 882
 883 static const char *yuv_lookup3d_prog_template =
 884   "TEX result.color, yuv, texture[%c], 3D;"
 885   "END";
 886
 887 /**
 888  * \brief creates and initializes helper textures needed for scaling texture read
 889  * \param scaler scaler type to create texture for
 890  * \param texu contains next free texture unit number
 891  * \param texs texture unit ids for the scaler are stored in this array
 892  */
 893 static void create_scaler_textures(int scaler, int *texu, char *texs) {
 894   switch (scaler) {
 895     case YUV_SCALER_BILIN:
 896     case YUV_SCALER_BICUB_NOTEX:
 897     case YUV_SCALER_UNSHARP:
 898     case YUV_SCALER_UNSHARP2:
 899       break;
 900     case YUV_SCALER_BICUB:
 901     case YUV_SCALER_BICUB_X:
 902       texs[0] = (*texu)++;
 903       gen_spline_lookup_tex(GL_TEXTURE0 + texs[0]);
 904       texs[0] += '0';
 905       break;
 906     default:
 907       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown scaler type %i\n", scaler);
 908   }
 909 }
 910
 911 static void gen_gamma_map(unsigned char *map, int size, float gamma);
 912
 913 #define ROW_R 0
 914 #define ROW_G 1
 915 #define ROW_B 2
 916 #define COL_Y 0
 917 #define COL_U 1
 918 #define COL_V 2
 919 #define COL_C 3
 920
 921 static void get_yuv2rgb_coeffs(gl_conversion_params_t *params, float yuv2rgb[3][4]) {
 922   float uvcos = params->saturation * cos(params->hue);
 923   float uvsin = params->saturation * sin(params->hue);
 924   int i;
 925   float uv_coeffs[3][2] = {
 926     { 0.000,  1.596},
 927     {-0.391, -0.813},
 928     { 2.018,  0.000}
 929   };
 930   for (i = 0; i < 3; i++) {
 931     yuv2rgb[i][COL_C]  = params->brightness;
 932     yuv2rgb[i][COL_Y]  = 1.164 * params->contrast;
 933     yuv2rgb[i][COL_C] += (-16 / 255.0) * yuv2rgb[i][COL_Y];
 934     yuv2rgb[i][COL_U]  = uv_coeffs[i][0] * uvcos + uv_coeffs[i][1] * uvsin;
 935     yuv2rgb[i][COL_C] += (-128 / 255.0) * yuv2rgb[i][COL_U];
 936     yuv2rgb[i][COL_V]  = uv_coeffs[i][0] * uvsin + uv_coeffs[i][1] * uvcos;
 937     yuv2rgb[i][COL_C] += (-128 / 255.0) * yuv2rgb[i][COL_V];
 938     // this "centers" contrast control so that e.g. a contrast of 0
 939     // leads to a grey image, not a black one
 940     yuv2rgb[i][COL_C] += 0.5 - params->contrast / 2.0;
 941   }
 942 }
 943
 944 //! size of gamma map use to avoid slow exp function in gen_yuv2rgb_map
 945 #define GMAP_SIZE (1024)
 946 /**
 947  * \brief generate a 3D YUV -> RGB map
 948  * \param params struct containing parameters like brightness, gamma, ...
 949  * \param map where to store map. Must provide space for (size + 2)^3 elements
 950  * \param size size of the map, excluding border
 951  */
 952 static void gen_yuv2rgb_map(gl_conversion_params_t *params, unsigned char *map, int size) {
 953   int i, j, k, l;
 954   float step = 1.0 / size;
 955   float y, u, v;
 956   float yuv2rgb[3][4];
 957   unsigned char gmaps[3][GMAP_SIZE];
 958   gen_gamma_map(gmaps[0], GMAP_SIZE, params->rgamma);
 959   gen_gamma_map(gmaps[1], GMAP_SIZE, params->ggamma);
 960   gen_gamma_map(gmaps[2], GMAP_SIZE, params->bgamma);
 961   get_yuv2rgb_coeffs(params, yuv2rgb);
 962   for (i = 0; i < 3; i++)
 963     for (j = 0; j < 4; j++)
 964       yuv2rgb[i][j] *= GMAP_SIZE - 1;
 965   v = 0;
 966   for (i = -1; i <= size; i++) {
 967     u = 0;
 968     for (j = -1; j <= size; j++) {
 969       y = 0;
 970       for (k = -1; k <= size; k++) {
 971         for (l = 0; l < 3; l++) {
 972           float rgb = yuv2rgb[l][COL_Y] * y + yuv2rgb[l][COL_U] * u + yuv2rgb[l][COL_V] * v + yuv2rgb[l][COL_C];
 973           *map++ = gmaps[l][av_clip(rgb, 0, GMAP_SIZE - 1)];
 974         }
 975         y += (k == -1 || k == size - 1) ? step / 2 : step;
 976       }
 977       u += (j == -1 || j == size - 1) ? step / 2 : step;
 978     }
 979     v += (i == -1 || i == size - 1) ? step / 2 : step;
 980   }
 981 }
 982
 983 //! resolution of texture for gamma lookup table
 984 #define LOOKUP_RES 512
 985 //! resolution for 3D yuv->rgb conversion lookup table
 986 #define LOOKUP_3DRES 32
 987 /**
 988  * \brief creates and initializes helper textures needed for yuv conversion
 989  * \param params struct containing parameters like brightness, gamma, ...
 990  * \param texu contains next free texture unit number
 991  * \param texs texture unit ids for the conversion are stored in this array
 992  */
 993 static void create_conv_textures(gl_conversion_params_t *params, int *texu, char *texs) {
 994   unsigned char *lookup_data = NULL;
 995   int conv = YUV_CONVERSION(params->type);
 996   switch (conv) {
 997     case YUV_CONVERSION_FRAGMENT:
 998     case YUV_CONVERSION_FRAGMENT_POW:
 999      break;
1000     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1001       texs[0] = (*texu)++;
1002       ActiveTexture(GL_TEXTURE0 + texs[0]);
1003       lookup_data = malloc(4 * LOOKUP_RES);
1004       gen_gamma_map(lookup_data, LOOKUP_RES, params->rgamma);
1005       gen_gamma_map(&lookup_data[LOOKUP_RES], LOOKUP_RES, params->ggamma);
1006       gen_gamma_map(&lookup_data[2 * LOOKUP_RES], LOOKUP_RES, params->bgamma);
1007       glCreateClearTex(GL_TEXTURE_2D, GL_LUMINANCE8, GL_LUMINANCE, GL_UNSIGNED_BYTE, GL_LINEAR,
1008                        LOOKUP_RES, 4, 0);
1009       glUploadTex(GL_TEXTURE_2D, GL_LUMINANCE, GL_UNSIGNED_BYTE, lookup_data,
1010                   LOOKUP_RES, 0, 0, LOOKUP_RES, 4, 0);
1011       ActiveTexture(GL_TEXTURE0);
1012       texs[0] += '0';
1013       break;
1014     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1015       {
1016         int sz = LOOKUP_3DRES + 2; // texture size including borders
1017         if (!TexImage3D) {
1018           mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing 3D texture function!\n");
1019           break;
1020         }
1021         texs[0] = (*texu)++;
1022         ActiveTexture(GL_TEXTURE0 + texs[0]);
1023         lookup_data = malloc(3 * sz * sz * sz);
1024         gen_yuv2rgb_map(params, lookup_data, LOOKUP_3DRES);
1025         glAdjustAlignment(sz);
1026         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
1027         TexImage3D(GL_TEXTURE_3D, 0, 3, sz, sz, sz, 1,
1028                    GL_RGB, GL_UNSIGNED_BYTE, lookup_data);
1029         glTexParameterf(GL_TEXTURE_3D, GL_TEXTURE_PRIORITY, 1.0);
1030         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1031         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1032         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_S, GL_CLAMP);
1033         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_T, GL_CLAMP);
1034         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_R, GL_CLAMP);
1035         ActiveTexture(GL_TEXTURE0);
1036         texs[0] += '0';
1037       }
1038       break;
1039     default:
1040       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", conv);
1041   }
1042   if (lookup_data)
1043     free(lookup_data);
1044 }
1045
1046 /**
1047  * \brief adds a scaling texture read at the current fragment program position
1048  * \param scaler type of scaler to insert
1049  * \param prog_pos current position in fragment program
1050  * \param remain how many bytes remain in the buffer given by prog_pos
1051  * \param texs array containing the texture unit identifiers for this scaler
1052  * \param in_tex texture unit the scaler should read from
1053  * \param out_comp component of the yuv variable the scaler stores the result in
1054  * \param rect if rectangular (pixel) adressing should be used for in_tex
1055  * \param texw width of the in_tex texture
1056  * \param texh height of the in_tex texture
1057  * \param strength strength of filter effect if the scaler does some kind of filtering
1058  */
1059 static void add_scaler(int scaler, char **prog_pos, int *remain, char *texs,
1060                     char in_tex, char out_comp, int rect, int texw, int texh,
1061                     double strength) {
1062   const char *ttype = rect ? "RECT" : "2D";
1063   const float ptw = rect ? 1.0 : 1.0 / texw;
1064   const float pth = rect ? 1.0 : 1.0 / texh;
1065   switch (scaler) {
1066     case YUV_SCALER_BILIN:
1067       snprintf(*prog_pos, *remain, bilin_filt_template, out_comp, in_tex,
1068                  in_tex, ttype);
1069       break;
1070     case YUV_SCALER_BICUB:
1071       if (rect)
1072         snprintf(*prog_pos, *remain, bicub_filt_template_RECT,
1073                  in_tex, texs[0], texs[0],
1074                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1075       else
1076         snprintf(*prog_pos, *remain, bicub_filt_template_2D,
1077                  in_tex, (float)texw, (float)texh,
1078                  texs[0], ptw, ptw, texs[0], pth, pth,
1079                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1080       break;
1081     case YUV_SCALER_BICUB_X:
1082       if (rect)
1083         snprintf(*prog_pos, *remain, bicub_x_filt_template_RECT,
1084                  in_tex, texs[0],
1085                  in_tex, in_tex, in_tex, in_tex, out_comp);
1086       else
1087         snprintf(*prog_pos, *remain, bicub_x_filt_template_2D,
1088                  in_tex, (float)texw,
1089                  texs[0], ptw, ptw,
1090                  in_tex, in_tex, in_tex, in_tex, out_comp);
1091       break;
1092     case YUV_SCALER_BICUB_NOTEX:
1093       if (rect)
1094         snprintf(*prog_pos, *remain, bicub_notex_filt_template_RECT,
1095                  in_tex,
1096                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1097       else
1098         snprintf(*prog_pos, *remain, bicub_notex_filt_template_2D,
1099                  in_tex, (float)texw, (float)texh, ptw, ptw, pth, pth,
1100                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1101       break;
1102     case YUV_SCALER_UNSHARP:
1103       snprintf(*prog_pos, *remain, unsharp_filt_template,
1104                out_comp, 0.5 * ptw, 0.5 * pth, 0.5 * ptw, -0.5 * pth,
1105                in_tex, out_comp, in_tex, out_comp, in_tex,
1106                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1107                in_tex, ttype, out_comp, strength);
1108       break;
1109     case YUV_SCALER_UNSHARP2:
1110       snprintf(*prog_pos, *remain, unsharp_filt_template2,
1111                out_comp, 1.2 * ptw, 1.2 * pth, 1.2 * ptw, -1.2 * pth,
1112                out_comp, 1.5 * ptw, 1.5 * pth,
1113                in_tex, out_comp, in_tex, out_comp, in_tex,
1114                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1115                in_tex, ttype, in_tex, out_comp, in_tex, out_comp,
1116                in_tex, ttype, in_tex, ttype, in_tex, ttype,
1117                in_tex, ttype, out_comp, strength);
1118       break;
1119   }
1120   *remain -= strlen(*prog_pos);
1121   *prog_pos += strlen(*prog_pos);
1122 }
1123
1124 static const struct {
1125   const char *name;
1126   GLenum cur;
1127   GLenum max;
1128 } progstats[] = {
1129   {"instructions", 0x88A0, 0x88A1},
1130   {"native instructions", 0x88A2, 0x88A3},
1131   {"temporaries", 0x88A4, 0x88A5},
1132   {"native temporaries", 0x88A6, 0x88A7},
1133   {"parameters", 0x88A8, 0x88A9},
1134   {"native parameters", 0x88AA, 0x88AB},
1135   {"attribs", 0x88AC, 0x88AD},
1136   {"native attribs", 0x88AE, 0x88AF},
1137   {"ALU instructions", 0x8805, 0x880B},
1138   {"TEX instructions", 0x8806, 0x880C},
1139   {"TEX indirections", 0x8807, 0x880D},
1140   {"native ALU instructions", 0x8808, 0x880E},
1141   {"native TEX instructions", 0x8809, 0x880F},
1142   {"native TEX indirections", 0x880A, 0x8810},
1143   {NULL, 0, 0}
1144 };
1145
1146 /**
1147  * \brief load the specified GPU Program
1148  * \param target program target to load into, only GL_FRAGMENT_PROGRAM is tested
1149  * \param prog program string
1150  * \return 1 on success, 0 otherwise
1151  */
1152 int loadGPUProgram(GLenum target, char *prog) {
1153   int i;
1154   GLint cur = 0, max = 0, err = 0;
1155   if (!ProgramString) {
1156     mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing GPU program function\n");
1157     return 0;
1158   }
1159   ProgramString(target, GL_PROGRAM_FORMAT_ASCII, strlen(prog), prog);
1160   glGetIntegerv(GL_PROGRAM_ERROR_POSITION, &err);
1161   if (err != -1) {
1162     mp_msg(MSGT_VO, MSGL_ERR,
1163       "[gl] Error compiling fragment program, make sure your card supports\n"
1164       "[gl]   GL_ARB_fragment_program (use glxinfo to check).\n"
1165       "[gl]   Error message:\n  %s at %.10s\n",
1166       glGetString(GL_PROGRAM_ERROR_STRING), &prog[err]);
1167     return 0;
1168   }
1169   if (!GetProgramiv || !mp_msg_test(MSGT_VO, MSGL_DBG2))
1170     return 1;
1171   mp_msg(MSGT_VO, MSGL_V, "[gl] Program statistics:\n");
1172   for (i = 0; progstats[i].name; i++) {
1173     GetProgramiv(target, progstats[i].cur, &cur);
1174     GetProgramiv(target, progstats[i].max, &max);
1175     mp_msg(MSGT_VO, MSGL_V, "[gl]   %s: %i/%i\n", progstats[i].name, cur, max);
1176   }
1177   return 1;
1178 }
1179
1180 #define MAX_PROGSZ (1024*1024)
1181
1182 /**
1183  * \brief setup a fragment program that will do YUV->RGB conversion
1184  * \param parms struct containing parameters like conversion and scaler type,
1185  *              brightness, ...
1186  */
1187 static void glSetupYUVFragprog(gl_conversion_params_t *params) {
1188   int type = params->type;
1189   int texw = params->texw;
1190   int texh = params->texh;
1191   int rect = params->target == GL_TEXTURE_RECTANGLE;
1192   static const char prog_hdr[] =
1193     "!!ARBfp1.0\n"
1194     "OPTION ARB_precision_hint_fastest;"
1195     // all scaler variables must go here so they aren't defined
1196     // multiple times when the same scaler is used more than once
1197     "TEMP coord, coord2, cdelta, parmx, parmy, a, b, yuv;";
1198   int prog_remain;
1199   char *yuv_prog, *prog_pos;
1200   int cur_texu = 3;
1201   char lum_scale_texs[1];
1202   char chrom_scale_texs[1];
1203   char conv_texs[1];
1204   GLint i;
1205   // this is the conversion matrix, with y, u, v factors
1206   // for red, green, blue and the constant offsets
1207   float yuv2rgb[3][4];
1208   create_conv_textures(params, &cur_texu, conv_texs);
1209   create_scaler_textures(YUV_LUM_SCALER(type), &cur_texu, lum_scale_texs);
1210   if (YUV_CHROM_SCALER(type) == YUV_LUM_SCALER(type))
1211     memcpy(chrom_scale_texs, lum_scale_texs, sizeof(chrom_scale_texs));
1212   else
1213     create_scaler_textures(YUV_CHROM_SCALER(type), &cur_texu, chrom_scale_texs);
1214   glGetIntegerv(GL_MAX_TEXTURE_IMAGE_UNITS, &i);
1215   if (i < cur_texu)
1216     mp_msg(MSGT_VO, MSGL_ERR,
1217            "[gl] %i texture units needed for this type of YUV fragment support (found %i)\n",
1218            cur_texu, i);
1219   if (!ProgramString) {
1220     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] ProgramString function missing!\n");
1221     return;
1222   }
1223   yuv_prog = malloc(MAX_PROGSZ);
1224   strcpy(yuv_prog, prog_hdr);
1225   prog_pos    = yuv_prog + sizeof(prog_hdr) - 1;
1226   prog_remain = MAX_PROGSZ - sizeof(prog_hdr);
1227   add_scaler(YUV_LUM_SCALER(type), &prog_pos, &prog_remain, lum_scale_texs,
1228              '0', 'r', rect, texw, texh, params->filter_strength);
1229   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1230              '1', 'g', rect, texw / 2, texh / 2, params->filter_strength);
1231   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1232              '2', 'b', rect, texw / 2, texh / 2, params->filter_strength);
1233   get_yuv2rgb_coeffs(params, yuv2rgb);
1234   switch (YUV_CONVERSION(type)) {
1235     case YUV_CONVERSION_FRAGMENT:
1236       snprintf(prog_pos, prog_remain, yuv_prog_template,
1237                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1238                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1239                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1240                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C]);
1241       break;
1242     case YUV_CONVERSION_FRAGMENT_POW:
1243       snprintf(prog_pos, prog_remain, yuv_pow_prog_template,
1244                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1245                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1246                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1247                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1248                (float)1.0 / params->rgamma, (float)1.0 / params->bgamma, (float)1.0 / params->bgamma);
1249       break;
1250     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1251       snprintf(prog_pos, prog_remain, yuv_lookup_prog_template,
1252                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1253                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1254                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1255                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1256                conv_texs[0], conv_texs[0], conv_texs[0]);
1257       break;
1258     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1259       snprintf(prog_pos, prog_remain, yuv_lookup3d_prog_template, conv_texs[0]);
1260       break;
1261     default:
1262       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(type));
1263       break;
1264   }
1265   mp_msg(MSGT_VO, MSGL_DBG2, "[gl] generated fragment program:\n%s\n", yuv_prog);
1266   loadGPUProgram(GL_FRAGMENT_PROGRAM, yuv_prog);
1267   free(yuv_prog);
1268 }
1269
1270 /**
1271  * \brief little helper function to create a lookup table for gamma
1272  * \param map buffer to create map into
1273  * \param size size of buffer
1274  * \param gamma gamma value
1275  */
1276 static void gen_gamma_map(unsigned char *map, int size, float gamma) {
1277   int i;
1278   if (gamma == 1.0) {
1279     for (i = 0; i < size; i++)
1280       map[i] = 255 * i / (size - 1);
1281     return;
1282   }
1283   gamma = 1.0 / gamma;
1284   for (i = 0; i < size; i++) {
1285     float tmp = (float)i / (size - 1.0);
1286     tmp = pow(tmp, gamma);
1287     if (tmp > 1.0) tmp = 1.0;
1288     if (tmp < 0.0) tmp = 0.0;
1289     map[i] = 255 * tmp;
1290   }
1291 }
1292
1293 /**
1294  * \brief setup YUV->RGB conversion
1295  * \param parms struct containing parameters like conversion and scaler type,
1296  *              brightness, ...
1297  * \ingroup glconversion
1298  */
1299 void glSetupYUVConversion(gl_conversion_params_t *params) {
1300   float uvcos = params->saturation * cos(params->hue);
1301   float uvsin = params->saturation * sin(params->hue);
1302   switch (YUV_CONVERSION(params->type)) {
1303     case YUV_CONVERSION_COMBINERS:
1304       glSetupYUVCombiners(uvcos, uvsin);
1305       break;
1306     case YUV_CONVERSION_COMBINERS_ATI:
1307       glSetupYUVCombinersATI(uvcos, uvsin);
1308       break;
1309     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1310     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1311     case YUV_CONVERSION_FRAGMENT:
1312     case YUV_CONVERSION_FRAGMENT_POW:
1313       glSetupYUVFragprog(params);
1314       break;
1315     default:
1316       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(params->type));
1317   }
1318 }
1319
1320 /**
1321  * \brief enable the specified YUV conversion
1322  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1323  * \param type type of YUV conversion
1324  * \ingroup glconversion
1325  */
1326 void glEnableYUVConversion(GLenum target, int type) {
1327   if (type <= 0) return;
1328   switch (YUV_CONVERSION(type)) {
1329     case YUV_CONVERSION_COMBINERS:
1330       ActiveTexture(GL_TEXTURE1);
1331       glEnable(target);
1332       ActiveTexture(GL_TEXTURE2);
1333       glEnable(target);
1334       ActiveTexture(GL_TEXTURE0);
1335       glEnable(GL_REGISTER_COMBINERS_NV);
1336       break;
1337     case YUV_CONVERSION_COMBINERS_ATI:
1338       ActiveTexture(GL_TEXTURE1);
1339       glEnable(target);
1340       ActiveTexture(GL_TEXTURE2);
1341       glEnable(target);
1342       ActiveTexture(GL_TEXTURE0);
1343       glEnable(GL_FRAGMENT_SHADER_ATI);
1344       break;
1345     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1346     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1347     case YUV_CONVERSION_FRAGMENT_POW:
1348     case YUV_CONVERSION_FRAGMENT:
1349       glEnable(GL_FRAGMENT_PROGRAM);
1350       break;
1351   }
1352 }
1353
1354 /**
1355  * \brief disable the specified YUV conversion
1356  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1357  * \param type type of YUV conversion
1358  * \ingroup glconversion
1359  */
1360 void glDisableYUVConversion(GLenum target, int type) {
1361   if (type <= 0) return;
1362   switch (YUV_CONVERSION(type)) {
1363     case YUV_CONVERSION_COMBINERS:
1364       ActiveTexture(GL_TEXTURE1);
1365       glDisable(target);
1366       ActiveTexture(GL_TEXTURE2);
1367       glDisable(target);
1368       ActiveTexture(GL_TEXTURE0);
1369       glDisable(GL_REGISTER_COMBINERS_NV);
1370       break;
1371     case YUV_CONVERSION_COMBINERS_ATI:
1372       ActiveTexture(GL_TEXTURE1);
1373       glDisable(target);
1374       ActiveTexture(GL_TEXTURE2);
1375       glDisable(target);
1376       ActiveTexture(GL_TEXTURE0);
1377       glDisable(GL_FRAGMENT_SHADER_ATI);
1378       break;
1379     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1380     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1381     case YUV_CONVERSION_FRAGMENT_POW:
1382     case YUV_CONVERSION_FRAGMENT:
1383       glDisable(GL_FRAGMENT_PROGRAM);
1384       break;
1385   }
1386 }
1387
1388 /**
1389  * \brief draw a texture part at given 2D coordinates
1390  * \param x screen top coordinate
1391  * \param y screen left coordinate
1392  * \param w screen width coordinate
1393  * \param h screen height coordinate
1394  * \param tx texture top coordinate in pixels
1395  * \param ty texture left coordinate in pixels
1396  * \param tw texture part width in pixels
1397  * \param th texture part height in pixels
1398  * \param sx width of texture in pixels
1399  * \param sy height of texture in pixels
1400  * \param rect_tex whether this texture uses texture_rectangle extension
1401  * \param is_yv12 if set, also draw the textures from units 1 and 2
1402  * \param flip flip the texture upside down
1403  * \ingroup gltexture
1404  */
1405 void glDrawTex(GLfloat x, GLfloat y, GLfloat w, GLfloat h,
1406                GLfloat tx, GLfloat ty, GLfloat tw, GLfloat th,
1407                int sx, int sy, int rect_tex, int is_yv12, int flip) {
1408   GLfloat tx2 = tx / 2, ty2 = ty / 2, tw2 = tw / 2, th2 = th / 2;
1409   if (!rect_tex) {
1410     tx /= sx; ty /= sy; tw /= sx; th /= sy;
1411     tx2 = tx, ty2 = ty, tw2 = tw, th2 = th;
1412   }
1413   if (flip) {
1414     y += h;
1415     h = -h;
1416   }
1417   glBegin(GL_QUADS);
1418   glTexCoord2f(tx, ty);
1419   if (is_yv12) {
1420     MultiTexCoord2f(GL_TEXTURE1, tx2, ty2);
1421     MultiTexCoord2f(GL_TEXTURE2, tx2, ty2);
1422   }
1423   glVertex2f(x, y);
1424   glTexCoord2f(tx, ty + th);
1425   if (is_yv12) {
1426     MultiTexCoord2f(GL_TEXTURE1, tx2, ty2 + th2);
1427     MultiTexCoord2f(GL_TEXTURE2, tx2, ty2 + th2);
1428   }
1429   glVertex2f(x, y + h);
1430   glTexCoord2f(tx + tw, ty + th);
1431   if (is_yv12) {
1432     MultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2 + th2);
1433     MultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2 + th2);
1434   }
1435   glVertex2f(x + w, y + h);
1436   glTexCoord2f(tx + tw, ty);
1437   if (is_yv12) {
1438     MultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2);
1439     MultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2);
1440   }
1441   glVertex2f(x + w, y);
1442   glEnd();
1443 }
1444
1445 #ifdef GL_WIN32
1446 #include "w32_common.h"
1447 /**
1448  * \brief little helper since wglGetProcAddress definition does not fit our
1449  *        getProcAddress
1450  * \param procName name of function to look up
1451  * \return function pointer returned by wglGetProcAddress
1452  */
1453 static void *w32gpa(const GLubyte *procName) {
1454   HMODULE oglmod;
1455   void *res = wglGetProcAddress(procName);
1456   if (res) return res;
1457   oglmod = GetModuleHandle("opengl32.dll");
1458   return GetProcAddress(oglmod, procName);
1459 }
1460
1461 int setGlWindow(int *vinfo, HGLRC *context, HWND win)
1462 {
1463   int new_vinfo;
1464   HDC windc = vo_w32_get_dc(win);
1465   HGLRC new_context = 0;
1466   int keep_context = 0;
1467   int res = SET_WINDOW_FAILED;
1468
1469   // should only be needed when keeping context, but not doing glFinish
1470   // can cause flickering even when we do not keep it.
1471   if (*context)
1472   glFinish();
1473   new_vinfo = GetPixelFormat(windc);
1474   if (*context && *vinfo && new_vinfo && *vinfo == new_vinfo) {
1475       // we can keep the wglContext
1476       new_context = *context;
1477       keep_context = 1;
1478   } else {
1479     // create a context
1480     new_context = wglCreateContext(windc);
1481     if (!new_context) {
1482       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GL context!\n");
1483       goto out;
1484     }
1485   }
1486
1487   // set context
1488   if (!wglMakeCurrent(windc, new_context)) {
1489     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GL context!\n");
1490     if (!keep_context) {
1491       wglDeleteContext(new_context);
1492     }
1493     goto out;
1494   }
1495
1496   // set new values
1497   vo_w32_window = win;
1498   {
1499     RECT rect;
1500     GetClientRect(win, &rect);
1501     vo_dwidth = rect.right;
1502     vo_dheight = rect.bottom;
1503   }
1504   if (!keep_context) {
1505     if (*context)
1506       wglDeleteContext(*context);
1507     *context = new_context;
1508     *vinfo = new_vinfo;
1509     getFunctions(w32gpa, NULL);
1510
1511     // and inform that reinit is neccessary
1512     res = SET_WINDOW_REINIT;
1513   } else
1514     res = SET_WINDOW_OK;
1515
1516 out:
1517   vo_w32_release_dc(win, windc);
1518   return res;
1519 }
1520
1521 void releaseGlContext(int *vinfo, HGLRC *context) {
1522   *vinfo = 0;
1523   if (*context) {
1524     wglMakeCurrent(0, 0);
1525     wglDeleteContext(*context);
1526   }
1527   *context = 0;
1528 }
1529
1530 void swapGlBuffers(void) {
1531   HDC vo_hdc = vo_w32_get_dc(vo_w32_window);
1532   SwapBuffers(vo_hdc);
1533   vo_w32_release_dc(vo_w32_window, vo_hdc);
1534 }
1535 #else
1536 #ifdef HAVE_LIBDL
1537 #include <dlfcn.h>
1538 #endif
1539 #include "x11_common.h"
1540 /**
1541  * \brief find address of a linked function
1542  * \param s name of function to find
1543  * \return address of function or NULL if not found
1544  */
1545 static void *getdladdr(const char *s) {
1546   void *ret = NULL;
1547 #ifdef HAVE_LIBDL
1548   void *handle = dlopen(NULL, RTLD_LAZY);
1549   if (!handle)
1550     return NULL;
1551   ret = dlsym(handle, s);
1552   dlclose(handle);
1553 #endif
1554   return ret;
1555 }
1556
1557 /**
1558  * \brief Returns the XVisualInfo associated with Window win.
1559  * \param win Window whose XVisualInfo is returne.
1560  * \return XVisualInfo of the window. Caller must use XFree to free it.
1561  */
1562 static XVisualInfo *getWindowVisualInfo(Window win) {
1563   XWindowAttributes xw_attr;
1564   XVisualInfo vinfo_template;
1565   int tmp;
1566   XGetWindowAttributes(mDisplay, win, &xw_attr);
1567   vinfo_template.visualid = XVisualIDFromVisual(xw_attr.visual);
1568   return XGetVisualInfo(mDisplay, VisualIDMask, &vinfo_template, &tmp);
1569 }
1570
1571 static void appendstr(char **dst, const char *str)
1572 {
1573     int newsize;
1574     char *newstr;
1575     if (!str)
1576         return;
1577     newsize = strlen(*dst) + 1 + strlen(str) + 1;
1578     newstr = realloc(*dst, newsize);
1579     if (!newstr)
1580         return;
1581     *dst = newstr;
1582     strcat(*dst, " ");
1583     strcat(*dst, str);
1584 }
1585
1586 /**
1587  * \brief Changes the window in which video is displayed.
1588  * If possible only transfers the context to the new window, otherwise
1589  * creates a new one, which must be initialized by the caller.
1590  * \param vinfo Currently used visual.
1591  * \param context Currently used context.
1592  * \param win window that should be used for drawing.
1593  * \return one of SET_WINDOW_FAILED, SET_WINDOW_OK or SET_WINDOW_REINIT.
1594  * In case of SET_WINDOW_REINIT the context could not be transfered
1595  * and the caller must initialize it correctly.
1596  * \ingroup glcontext
1597  */
1598 int setGlWindow(XVisualInfo **vinfo, GLXContext *context, Window win)
1599 {
1600   XVisualInfo *new_vinfo;
1601   GLXContext new_context = NULL;
1602   int keep_context = 0;
1603
1604   // should only be needed when keeping context, but not doing glFinish
1605   // can cause flickering even when we do not keep it.
1606   if (*context)
1607   glFinish();
1608   new_vinfo = getWindowVisualInfo(win);
1609   if (*context && *vinfo && new_vinfo &&
1610       (*vinfo)->visualid == new_vinfo->visualid) {
1611       // we can keep the GLXContext
1612       new_context = *context;
1613       XFree(new_vinfo);
1614       new_vinfo = *vinfo;
1615       keep_context = 1;
1616   } else {
1617     // create a context
1618     new_context = glXCreateContext(mDisplay, new_vinfo, NULL, True);
1619     if (!new_context) {
1620       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GLX context!\n");
1621       XFree(new_vinfo);
1622       return SET_WINDOW_FAILED;
1623     }
1624   }
1625
1626   // set context
1627   if (!glXMakeCurrent(mDisplay, vo_window, new_context)) {
1628     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GLX context!\n");
1629     if (!keep_context) {
1630       glXDestroyContext(mDisplay, new_context);
1631       XFree(new_vinfo);
1632     }
1633     return SET_WINDOW_FAILED;
1634   }
1635
1636   // set new values
1637   vo_window = win;
1638   vo_x11_update_geometry();
1639   if (!keep_context) {
1640     void *(*getProcAddress)(const GLubyte *);
1641     const char *(*glXExtStr)(Display *, int);
1642     char *glxstr = strdup("");
1643     if (*context)
1644       glXDestroyContext(mDisplay, *context);
1645     *context = new_context;
1646     if (*vinfo)
1647       XFree(*vinfo);
1648     *vinfo = new_vinfo;
1649       getProcAddress = getdladdr("glXGetProcAddress");
1650     if (!getProcAddress)
1651       getProcAddress = getdladdr("glXGetProcAddressARB");
1652     if (!getProcAddress)
1653       getProcAddress = (void *)getdladdr;
1654     glXExtStr = getdladdr("glXQueryExtensionsString");
1655     if (glXExtStr)
1656         appendstr(&glxstr, glXExtStr(mDisplay, DefaultScreen(mDisplay)));
1657     glXExtStr = getdladdr("glXGetClientString");
1658     if (glXExtStr)
1659         appendstr(&glxstr, glXExtStr(mDisplay, GLX_EXTENSIONS));
1660     glXExtStr = getdladdr("glXGetServerString");
1661     if (glXExtStr)
1662         appendstr(&glxstr, glXExtStr(mDisplay, GLX_EXTENSIONS));
1663
1664     getFunctions(getProcAddress, glxstr);
1665     free(glxstr);
1666
1667     // and inform that reinit is neccessary
1668     return SET_WINDOW_REINIT;
1669   }
1670   return SET_WINDOW_OK;
1671 }
1672
1673 /**
1674  * \brief free the VisualInfo and GLXContext of an OpenGL context.
1675  * \ingroup glcontext
1676  */
1677 void releaseGlContext(XVisualInfo **vinfo, GLXContext *context) {
1678   if (*vinfo)
1679     XFree(*vinfo);
1680   *vinfo = NULL;
1681   if (*context)
1682   {
1683     glFinish();
1684     glXMakeCurrent(mDisplay, None, NULL);
1685     glXDestroyContext(mDisplay, *context);
1686   }
1687   *context = 0;
1688 }
1689
1690 void swapGlBuffers(void) {
1691   glXSwapBuffers(mDisplay, vo_window);
1692 }
1693 #endif
1694