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_RGB24:
 200       *gl_format = GL_RGB;
 201       *gl_type = GL_UNSIGNED_BYTE;
 202       break;
 203     case IMGFMT_RGBA:
 204       *gl_texfmt = 4;
 205       *gl_format = GL_RGBA;
 206       *gl_type = GL_UNSIGNED_BYTE;
 207       break;
 208     case IMGFMT_YV12:
 209       supported = 0; // no native YV12 support
 210     case IMGFMT_Y800:
 211     case IMGFMT_Y8:
 212       *gl_texfmt = 1;
 213       *bpp = 8;
 214       *gl_format = GL_LUMINANCE;
 215       *gl_type = GL_UNSIGNED_BYTE;
 216       break;
 217     case IMGFMT_UYVY:
 218     case IMGFMT_YUY2:
 219       *gl_texfmt = GL_YCBCR_MESA;
 220       *bpp = 16;
 221       *gl_format = GL_YCBCR_MESA;
 222       *gl_type = fmt == IMGFMT_UYVY ? GL_UNSIGNED_SHORT_8_8 : GL_UNSIGNED_SHORT_8_8_REV;
 223       break;
 224 #if 0
 225     // we do not support palettized formats, although the format the
 226     // swscale produces works
 227     case IMGFMT_RGB8:
 228       gl_format = GL_RGB;
 229       gl_type = GL_UNSIGNED_BYTE_2_3_3_REV;
 230       break;
 231 #endif
 232     case IMGFMT_RGB15:
 233       *gl_format = GL_RGBA;
 234       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 235       break;
 236     case IMGFMT_RGB16:
 237       *gl_format = GL_RGB;
 238       *gl_type = GL_UNSIGNED_SHORT_5_6_5_REV;
 239       break;
 240 #if 0
 241     case IMGFMT_BGR8:
 242       // special case as red and blue have a differen number of bits.
 243       // GL_BGR and GL_UNSIGNED_BYTE_3_3_2 isn't supported at least
 244       // by nVidia drivers, and in addition would give more bits to
 245       // blue than to red, which isn't wanted
 246       gl_format = GL_RGB;
 247       gl_type = GL_UNSIGNED_BYTE_3_3_2;
 248       break;
 249 #endif
 250     case IMGFMT_BGR15:
 251       *gl_format = GL_BGRA;
 252       *gl_type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
 253       break;
 254     case IMGFMT_BGR16:
 255       *gl_format = GL_RGB;
 256       *gl_type = GL_UNSIGNED_SHORT_5_6_5;
 257       break;
 258     case IMGFMT_BGR24:
 259       *gl_format = GL_BGR;
 260       *gl_type = GL_UNSIGNED_BYTE;
 261       break;
 262     case IMGFMT_BGRA:
 263       *gl_texfmt = 4;
 264       *gl_format = GL_BGRA;
 265       *gl_type = GL_UNSIGNED_BYTE;
 266       break;
 267     default:
 268       *gl_texfmt = 4;
 269       *gl_format = GL_RGBA;
 270       *gl_type = GL_UNSIGNED_BYTE;
 271       supported = 0;
 272   }
 273 #ifdef TEXTUREFORMAT_ALWAYS
 274   *gl_texfmt = TEXTUREFORMAT_ALWAYS;
 275 #endif
 276   return supported;
 277 }
 278
 279 static void *setNull(const GLubyte *s) {
 280   return NULL;
 281 }
 282
 283 typedef struct {
 284   void *funcptr;
 285   const char *extstr;
 286   const char *funcnames[7];
 287 } extfunc_desc_t;
 288
 289 static const extfunc_desc_t extfuncs[] = {
 290   {&GenBuffers, NULL, {"glGenBuffers", "glGenBuffersARB", NULL}},
 291   {&DeleteBuffers, NULL, {"glDeleteBuffers", "glDeleteBuffersARB", NULL}},
 292   {&BindBuffer, NULL, {"glBindBuffer", "glBindBufferARB", NULL}},
 293   {&MapBuffer, NULL, {"glMapBuffer", "glMapBufferARB", NULL}},
 294   {&UnmapBuffer, NULL, {"glUnmapBuffer", "glUnmapBufferARB", NULL}},
 295   {&BufferData, NULL, {"glBufferData", "glBufferDataARB", NULL}},
 296   {&CombinerParameterfv, "NV_register_combiners", {"glCombinerParameterfv", "glCombinerParameterfvNV", NULL}},
 297   {&CombinerParameteri, "NV_register_combiners", {"glCombinerParameteri", "glCombinerParameteriNV", NULL}},
 298   {&CombinerInput, "NV_register_combiners", {"glCombinerInput", "glCombinerInputNV", NULL}},
 299   {&CombinerOutput, "NV_register_combiners", {"glCombinerOutput", "glCombinerOutputNV", NULL}},
 300   {&BeginFragmentShader, "ATI_fragment_shader", {"glBeginFragmentShaderATI", NULL}},
 301   {&EndFragmentShader, "ATI_fragment_shader", {"glEndFragmentShaderATI", NULL}},
 302   {&SampleMap, "ATI_fragment_shader", {"glSampleMapATI", NULL}},
 303   {&ColorFragmentOp2, "ATI_fragment_shader", {"glColorFragmentOp2ATI", NULL}},
 304   {&ColorFragmentOp3, "ATI_fragment_shader", {"glColorFragmentOp3ATI", NULL}},
 305   {&SetFragmentShaderConstant, "ATI_fragment_shader", {"glSetFragmentShaderConstantATI", NULL}},
 306   {&ActiveTexture, NULL, {"glActiveTexture", "glActiveTextureARB", NULL}},
 307   {&BindTexture, NULL, {"glBindTexture", "glBindTextureARB", "glBindTextureEXT", NULL}},
 308   {&MultiTexCoord2f, NULL, {"glMultiTexCoord2f", "glMultiTexCoord2fARB", NULL}},
 309   {&GenPrograms, "_program", {"glGenProgramsARB", NULL}},
 310   {&DeletePrograms, "_program", {"glDeleteProgramsARB", NULL}},
 311   {&BindProgram, "_program", {"glBindProgramARB", NULL}},
 312   {&ProgramString, "_program", {"glProgramStringARB", NULL}},
 313   {&GetProgramiv, "_program", {"glGetProgramivARB", NULL}},
 314   {&ProgramEnvParameter4f, "_program", {"glProgramEnvParameter4fARB", NULL}},
 315   {&SwapInterval, "_swap_control", {"glXSwapInterval", "glXSwapIntervalEXT", "glXSwapIntervalSGI", "wglSwapInterval", "wglSwapIntervalEXT", "wglSwapIntervalSGI", NULL}},
 316   {&TexImage3D, NULL, {"glTexImage3D", NULL}},
 317   {&AllocateMemoryMESA, "GLX_MESA_allocate_memory", {"glXAllocateMemoryMESA", NULL}},
 318   {&FreeMemoryMESA, "GLX_MESA_allocate_memory", {"glXFreeMemoryMESA", NULL}},
 319   {NULL}
 320 };
 321
 322 /**
 323  * \brief find the function pointers of some useful OpenGL extensions
 324  * \param getProcAddress function to resolve function names, may be NULL
 325  * \param ext2 an extra extension string
 326  */
 327 static void getFunctions(void *(*getProcAddress)(const GLubyte *),
 328                          const char *ext2) {
 329   const extfunc_desc_t *dsc;
 330   const char *extensions = (const char *)glGetString(GL_EXTENSIONS);
 331   char *allexts;
 332   if (!extensions) extensions = "";
 333   if (!ext2) ext2 = "";
 334   allexts = malloc(strlen(extensions) + strlen(ext2) + 2);
 335   strcpy(allexts, extensions);
 336   strcat(allexts, " ");
 337   strcat(allexts, ext2);
 338   mp_msg(MSGT_VO, MSGL_V, "OpenGL extensions string:\n%s\n", allexts);
 339   if (!getProcAddress)
 340     getProcAddress = setNull;
 341   for (dsc = extfuncs; dsc->funcptr; dsc++) {
 342     void *ptr = NULL;
 343     int i;
 344     if (!dsc->extstr || strstr(allexts, dsc->extstr)) {
 345       for (i = 0; !ptr && dsc->funcnames[i]; i++)
 346         ptr = getProcAddress((const GLubyte *)dsc->funcnames[i]);
 347     }
 348     *(void **)dsc->funcptr = ptr;
 349   }
 350   if (strstr(allexts, "_texture_float"))
 351     hqtexfmt = GL_RGB32F;
 352   else if (strstr(allexts, "NV_float_buffer"))
 353     hqtexfmt = GL_FLOAT_RGB32_NV;
 354   else
 355     hqtexfmt = GL_RGB16;
 356   free(allexts);
 357 }
 358
 359 /**
 360  * \brief create a texture and set some defaults
 361  * \param target texture taget, usually GL_TEXTURE_2D
 362  * \param fmt internal texture format
 363  * \param format texture host data format
 364  * \param type texture host data type
 365  * \param filter filter used for scaling, e.g. GL_LINEAR
 366  * \param w texture width
 367  * \param h texture height
 368  * \param val luminance value to fill texture with
 369  * \ingroup gltexture
 370  */
 371 void glCreateClearTex(GLenum target, GLenum fmt, GLenum format, GLenum type, GLint filter,
 372                       int w, int h, unsigned char val) {
 373   GLfloat fval = (GLfloat)val / 255.0;
 374   GLfloat border[4] = {fval, fval, fval, fval};
 375   int stride = w * glFmt2bpp(format, type);
 376   char *init;
 377   if (!stride) return;
 378   init = malloc(stride * h);
 379   memset(init, val, stride * h);
 380   glAdjustAlignment(stride);
 381   glPixelStorei(GL_UNPACK_ROW_LENGTH, w);
 382   glTexImage2D(target, 0, fmt, w, h, 0, format, type, init);
 383   glTexParameterf(target, GL_TEXTURE_PRIORITY, 1.0);
 384   glTexParameteri(target, GL_TEXTURE_MIN_FILTER, filter);
 385   glTexParameteri(target, GL_TEXTURE_MAG_FILTER, filter);
 386   glTexParameteri(target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 387   glTexParameteri(target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 388   // Border texels should not be used with CLAMP_TO_EDGE
 389   // We set a sane default anyway.
 390   glTexParameterfv(target, GL_TEXTURE_BORDER_COLOR, border);
 391   free(init);
 392 }
 393
 394 /**
 395  * \brief skips whitespace and comments
 396  * \param f file to read from
 397  */
 398 static void ppm_skip(FILE *f) {
 399   int c, comment = 0;
 400   do {
 401     c = fgetc(f);
 402     if (c == '#')
 403       comment = 1;
 404     if (c == '\n')
 405       comment = 0;
 406   } while (c != EOF && (isspace(c) || comment));
 407   if (c != EOF)
 408     ungetc(c, f);
 409 }
 410
 411 #define MAXDIM (16 * 1024)
 412
 413 /**
 414  * \brief creates a texture from a PPM file
 415  * \param target texture taget, usually GL_TEXTURE_2D
 416  * \param fmt internal texture format, 0 for default
 417  * \param filter filter used for scaling, e.g. GL_LINEAR
 418  * \param f file to read PPM from
 419  * \param width [out] width of texture
 420  * \param height [out] height of texture
 421  * \param maxval [out] maxval value from PPM file
 422  * \return 0 on error, 1 otherwise
 423  * \ingroup gltexture
 424  */
 425 int glCreatePPMTex(GLenum target, GLenum fmt, GLint filter,
 426                    FILE *f, int *width, int *height, int *maxval) {
 427   unsigned w, h, m, val, bpp;
 428   char *data;
 429   GLenum type;
 430   ppm_skip(f);
 431   if (fgetc(f) != 'P' || fgetc(f) != '6')
 432     return 0;
 433   ppm_skip(f);
 434   if (fscanf(f, "%u", &w) != 1)
 435     return 0;
 436   ppm_skip(f);
 437   if (fscanf(f, "%u", &h) != 1)
 438     return 0;
 439   ppm_skip(f);
 440   if (fscanf(f, "%u", &m) != 1)
 441     return 0;
 442   val = fgetc(f);
 443   if (!isspace(val))
 444     return 0;
 445   if (w > MAXDIM || h > MAXDIM)
 446     return 0;
 447   bpp = (m > 255) ? 6 : 3;
 448   data = malloc(w * h * bpp);
 449   if (fread(data, w * bpp, h, f) != h)
 450     return 0;
 451   if (!fmt) {
 452     fmt = (m > 255) ? hqtexfmt : 3;
 453     if (fmt == GL_FLOAT_RGB32_NV && target != GL_TEXTURE_RECTANGLE)
 454       fmt = GL_RGB16;
 455   }
 456   type = m > 255 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 457   glCreateClearTex(target, fmt, GL_RGB, type, filter, w, h, 0);
 458   glUploadTex(target, GL_RGB, type,
 459               data, w * bpp, 0, 0, w, h, 0);
 460   free(data);
 461   if (width) *width = w;
 462   if (height) *height = h;
 463   if (maxval) *maxval = m;
 464   return 1;
 465 }
 466
 467 /**
 468  * \brief return the number of bytes per pixel for the given format
 469  * \param format OpenGL format
 470  * \param type OpenGL type
 471  * \return bytes per pixel
 472  * \ingroup glgeneral
 473  *
 474  * Does not handle all possible variants, just those used by MPlayer
 475  */
 476 int glFmt2bpp(GLenum format, GLenum type) {
 477   int component_size = 0;
 478   switch (type) {
 479     case GL_UNSIGNED_BYTE_3_3_2:
 480     case GL_UNSIGNED_BYTE_2_3_3_REV:
 481       return 1;
 482     case GL_UNSIGNED_SHORT_5_5_5_1:
 483     case GL_UNSIGNED_SHORT_1_5_5_5_REV:
 484     case GL_UNSIGNED_SHORT_5_6_5:
 485     case GL_UNSIGNED_SHORT_5_6_5_REV:
 486       return 2;
 487     case GL_UNSIGNED_BYTE:
 488       component_size = 1;
 489       break;
 490     case GL_UNSIGNED_SHORT:
 491       component_size = 2;
 492       break;
 493   }
 494   switch (format) {
 495     case GL_LUMINANCE:
 496     case GL_ALPHA:
 497       return component_size;
 498     case GL_YCBCR_MESA:
 499       return 2;
 500     case GL_RGB:
 501     case GL_BGR:
 502       return 3 * component_size;
 503     case GL_RGBA:
 504     case GL_BGRA:
 505       return 4 * component_size;
 506   }
 507   return 0; // unknown
 508 }
 509
 510 /**
 511  * \brief upload a texture, handling things like stride and slices
 512  * \param target texture target, usually GL_TEXTURE_2D
 513  * \param format OpenGL format of data
 514  * \param type OpenGL type of data
 515  * \param dataptr data to upload
 516  * \param stride data stride
 517  * \param x x offset in texture
 518  * \param y y offset in texture
 519  * \param w width of the texture part to upload
 520  * \param h height of the texture part to upload
 521  * \param slice height of an upload slice, 0 for all at once
 522  * \ingroup gltexture
 523  */
 524 void glUploadTex(GLenum target, GLenum format, GLenum type,
 525                  const void *dataptr, int stride,
 526                  int x, int y, int w, int h, int slice) {
 527   const uint8_t *data = dataptr;
 528   int y_max = y + h;
 529   if (w <= 0 || h <= 0) return;
 530   if (slice <= 0)
 531     slice = h;
 532   if (stride < 0) {
 533     data += (h - 1) * stride;
 534     stride = -stride;
 535   }
 536   // this is not always correct, but should work for MPlayer
 537   glAdjustAlignment(stride);
 538   glPixelStorei(GL_UNPACK_ROW_LENGTH, stride / glFmt2bpp(format, type));
 539   for (; y + slice <= y_max; y += slice) {
 540     glTexSubImage2D(target, 0, x, y, w, slice, format, type, data);
 541     data += stride * slice;
 542   }
 543   if (y < y_max)
 544     glTexSubImage2D(target, 0, x, y, w, y_max - y, format, type, data);
 545 }
 546
 547 static void fillUVcoeff(GLfloat *ucoef, GLfloat *vcoef,
 548                         float uvcos, float uvsin) {
 549   int i;
 550   ucoef[0] = 0 * uvcos + 1.403 * uvsin;
 551   vcoef[0] = 0 * uvsin + 1.403 * uvcos;
 552   ucoef[1] = -0.344 * uvcos + -0.714 * uvsin;
 553   vcoef[1] = -0.344 * uvsin + -0.714 * uvcos;
 554   ucoef[2] = 1.770 * uvcos + 0 * uvsin;
 555   vcoef[2] = 1.770 * uvsin + 0 * uvcos;
 556   ucoef[3] = 0;
 557   vcoef[3] = 0;
 558   // Coefficients (probably) must be in [0, 1] range, whereas they originally
 559   // are in [-2, 2] range, so here comes the trick:
 560   // First put them in the [-0.5, 0.5] range, then add 0.5.
 561   // This can be undone with the HALF_BIAS and SCALE_BY_FOUR arguments
 562   // for CombinerInput and CombinerOutput (or the respective ATI variants)
 563   for (i = 0; i < 4; i++) {
 564     ucoef[i] = ucoef[i] * 0.25 + 0.5;
 565     vcoef[i] = vcoef[i] * 0.25 + 0.5;
 566   }
 567 }
 568
 569 /**
 570  * \brief Setup register combiners for YUV to RGB conversion.
 571  * \param uvcos used for saturation and hue adjustment
 572  * \param uvsin used for saturation and hue adjustment
 573  */
 574 static void glSetupYUVCombiners(float uvcos, float uvsin) {
 575   GLfloat ucoef[4];
 576   GLfloat vcoef[4];
 577   GLint i;
 578   if (!CombinerInput || !CombinerOutput ||
 579       !CombinerParameterfv || !CombinerParameteri) {
 580     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner functions missing!\n");
 581     return;
 582   }
 583   glGetIntegerv(GL_MAX_GENERAL_COMBINERS_NV, &i);
 584   if (i < 2)
 585     mp_msg(MSGT_VO, MSGL_ERR,
 586            "[gl] 2 general combiners needed for YUV combiner support (found %i)\n", i);
 587   glGetIntegerv(GL_MAX_TEXTURE_UNITS, &i);
 588   if (i < 3)
 589     mp_msg(MSGT_VO, MSGL_ERR,
 590            "[gl] 3 texture units needed for YUV combiner support (found %i)\n", i);
 591   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 592   CombinerParameterfv(GL_CONSTANT_COLOR0_NV, ucoef);
 593   CombinerParameterfv(GL_CONSTANT_COLOR1_NV, vcoef);
 594
 595   // UV first, like this green component cannot overflow
 596   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_A_NV,
 597                 GL_TEXTURE1, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 598   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_B_NV,
 599                 GL_CONSTANT_COLOR0_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 600   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_C_NV,
 601                 GL_TEXTURE2, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 602   CombinerInput(GL_COMBINER0_NV, GL_RGB, GL_VARIABLE_D_NV,
 603                 GL_CONSTANT_COLOR1_NV, GL_HALF_BIAS_NORMAL_NV, GL_RGB);
 604   CombinerOutput(GL_COMBINER0_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 605                  GL_SPARE0_NV, GL_SCALE_BY_FOUR_NV, GL_NONE, GL_FALSE,
 606                  GL_FALSE, GL_FALSE);
 607
 608   // stage 2
 609   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_A_NV, GL_SPARE0_NV,
 610                 GL_SIGNED_IDENTITY_NV, GL_RGB);
 611   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_B_NV, GL_ZERO,
 612                  GL_UNSIGNED_INVERT_NV, GL_RGB);
 613   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_C_NV,
 614                 GL_TEXTURE0, GL_SIGNED_IDENTITY_NV, GL_RGB);
 615   CombinerInput(GL_COMBINER1_NV, GL_RGB, GL_VARIABLE_D_NV, GL_ZERO,
 616                 GL_UNSIGNED_INVERT_NV, GL_RGB);
 617   CombinerOutput(GL_COMBINER1_NV, GL_RGB, GL_DISCARD_NV, GL_DISCARD_NV,
 618                  GL_SPARE0_NV, GL_NONE, GL_NONE, GL_FALSE,
 619                  GL_FALSE, GL_FALSE);
 620
 621   // leave final combiner stage in default mode
 622   CombinerParameteri(GL_NUM_GENERAL_COMBINERS_NV, 2);
 623 }
 624
 625 /**
 626  * \brief Setup ATI version of register combiners for YUV to RGB conversion.
 627  * \param uvcos used for saturation and hue adjustment
 628  * \param uvsin used for saturation and hue adjustment
 629  *
 630  * ATI called this fragment shader, but the name is confusing in the
 631  * light of a very different OpenGL 2.0 extension with the same name
 632  */
 633 static void glSetupYUVCombinersATI(float uvcos, float uvsin) {
 634   GLfloat ucoef[4];
 635   GLfloat vcoef[4];
 636   GLint i;
 637   if (!BeginFragmentShader || !EndFragmentShader ||
 638       !SetFragmentShaderConstant || !SampleMap ||
 639       !ColorFragmentOp2 || !ColorFragmentOp3) {
 640     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Combiner (ATI) functions missing!\n");
 641     return;
 642   }
 643   glGetIntegerv(GL_NUM_FRAGMENT_REGISTERS_ATI, &i);
 644   if (i < 3)
 645     mp_msg(MSGT_VO, MSGL_ERR,
 646            "[gl] 3 registers needed for YUV combiner (ATI) support (found %i)\n", i);
 647   glGetIntegerv (GL_MAX_TEXTURE_UNITS, &i);
 648   if (i < 3)
 649     mp_msg(MSGT_VO, MSGL_ERR,
 650            "[gl] 3 texture units needed for YUV combiner (ATI) support (found %i)\n", i);
 651   fillUVcoeff(ucoef, vcoef, uvcos, uvsin);
 652   BeginFragmentShader();
 653   SetFragmentShaderConstant(GL_CON_0_ATI, ucoef);
 654   SetFragmentShaderConstant(GL_CON_1_ATI, vcoef);
 655   SampleMap(GL_REG_0_ATI, GL_TEXTURE0, GL_SWIZZLE_STR_ATI);
 656   SampleMap(GL_REG_1_ATI, GL_TEXTURE1, GL_SWIZZLE_STR_ATI);
 657   SampleMap(GL_REG_2_ATI, GL_TEXTURE2, GL_SWIZZLE_STR_ATI);
 658   // UV first, like this green component cannot overflow
 659   ColorFragmentOp2(GL_MUL_ATI, GL_REG_1_ATI, GL_NONE, GL_NONE,
 660                    GL_REG_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 661                    GL_CON_0_ATI, GL_NONE, GL_BIAS_BIT_ATI);
 662   ColorFragmentOp3(GL_MAD_ATI, GL_REG_2_ATI, GL_NONE, GL_4X_BIT_ATI,
 663                    GL_REG_2_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 664                    GL_CON_1_ATI, GL_NONE, GL_BIAS_BIT_ATI,
 665                    GL_REG_1_ATI, GL_NONE, GL_NONE);
 666   ColorFragmentOp2(GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 667                    GL_REG_0_ATI, GL_NONE, GL_NONE,
 668                    GL_REG_2_ATI, GL_NONE, GL_NONE);
 669   EndFragmentShader();
 670 }
 671
 672 /**
 673  * \brief helper function for gen_spline_lookup_tex
 674  * \param x subpixel-position ((0,1) range) to calculate weights for
 675  * \param dst where to store transformed weights, must provide space for 4 GLfloats
 676  *
 677  * calculates the weights and stores them after appropriate transformation
 678  * for the scaler fragment program.
 679  */
 680 static void store_weights(float x, GLfloat *dst) {
 681   float w0 = (((-1 * x + 3) * x - 3) * x + 1) / 6;
 682   float w1 = ((( 3 * x - 6) * x + 0) * x + 4) / 6;
 683   float w2 = (((-3 * x + 3) * x + 3) * x + 1) / 6;
 684   float w3 = ((( 1 * x + 0) * x + 0) * x + 0) / 6;
 685   *dst++ = 1 + x - w1 / (w0 + w1);
 686   *dst++ = 1 - x + w3 / (w2 + w3);
 687   *dst++ = w0 + w1;
 688   *dst++ = 0;
 689 }
 690
 691 //! to avoid artefacts this should be rather large
 692 #define LOOKUP_BSPLINE_RES (2 * 1024)
 693 /**
 694  * \brief creates the 1D lookup texture needed for fast higher-order filtering
 695  * \param unit texture unit to attach texture to
 696  */
 697 static void gen_spline_lookup_tex(GLenum unit) {
 698   GLfloat tex[4 * LOOKUP_BSPLINE_RES];
 699   GLfloat *tp = tex;
 700   int i;
 701   for (i = 0; i < LOOKUP_BSPLINE_RES; i++) {
 702     float x = (float)(i + 0.5) / LOOKUP_BSPLINE_RES;
 703     store_weights(x, tp);
 704     tp += 4;
 705   }
 706   store_weights(0, tex);
 707   store_weights(1, &tex[4 * (LOOKUP_BSPLINE_RES - 1)]);
 708   ActiveTexture(unit);
 709   glTexImage1D(GL_TEXTURE_1D, 0, GL_RGBA16, LOOKUP_BSPLINE_RES, 0, GL_RGBA, GL_FLOAT, tex);
 710   glTexParameterf(GL_TEXTURE_1D, GL_TEXTURE_PRIORITY, 1.0);
 711   glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 712   glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 713   glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_REPEAT);
 714   ActiveTexture(GL_TEXTURE0);
 715 }
 716
 717 static const char *bilin_filt_template =
 718   "TEX yuv.%c, fragment.texcoord[%c], texture[%c], %s;";
 719
 720 #define BICUB_FILT_MAIN(textype) \
 721   /* first y-interpolation */ \
 722   "ADD coord, fragment.texcoord[%c].xyxy, cdelta.xyxw;" \
 723   "ADD coord2, fragment.texcoord[%c].xyxy, cdelta.zyzw;" \
 724   "TEX a.r, coord.xyxy, texture[%c], "textype";" \
 725   "TEX a.g, coord.zwzw, texture[%c], "textype";" \
 726   /* second y-interpolation */ \
 727   "TEX b.r, coord2.xyxy, texture[%c], "textype";" \
 728   "TEX b.g, coord2.zwzw, texture[%c], "textype";" \
 729   "LRP a.b, parmy.b, a.rrrr, a.gggg;" \
 730   "LRP a.a, parmy.b, b.rrrr, b.gggg;" \
 731   /* x-interpolation */ \
 732   "LRP yuv.%c, parmx.b, a.bbbb, a.aaaa;"
 733
 734 static const char *bicub_filt_template_2D =
 735   "MAD coord.xy, fragment.texcoord[%c], {%f, %f}, {0.5, 0.5};"
 736   "TEX parmx, coord.x, texture[%c], 1D;"
 737   "MUL cdelta.xz, parmx.rrgg, {-%f, 0, %f, 0};"
 738   "TEX parmy, coord.y, texture[%c], 1D;"
 739   "MUL cdelta.yw, parmy.rrgg, {0, -%f, 0, %f};"
 740   BICUB_FILT_MAIN("2D");
 741
 742 static const char *bicub_filt_template_RECT =
 743   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};"
 744   "TEX parmx, coord.x, texture[%c], 1D;"
 745   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};"
 746   "TEX parmy, coord.y, texture[%c], 1D;"
 747   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};"
 748   BICUB_FILT_MAIN("RECT");
 749
 750 #define CALCWEIGHTS(t, s) \
 751   "MAD "t", {-0.5, 0.1666, 0.3333, -0.3333}, "s", {1, 0, -0.5, 0.5};" \
 752   "MAD "t", "t", "s", {0, 0, -0.5, 0.5};" \
 753   "MAD "t", "t", "s", {-0.6666, 0, 0.8333, 0.1666};" \
 754   "RCP a.x, "t".z;" \
 755   "RCP a.y, "t".w;" \
 756   "MAD "t".xy, "t".xyxy, a.xyxy, {1, 1, 0, 0};" \
 757   "ADD "t".x, "t".xxxx, "s";" \
 758   "SUB "t".y, "t".yyyy, "s";"
 759
 760 static const char *bicub_notex_filt_template_2D =
 761   "MAD coord.xy, fragment.texcoord[%c], {%f, %f}, {0.5, 0.5};"
 762   "FRC coord.xy, coord.xyxy;"
 763   CALCWEIGHTS("parmx", "coord.xxxx")
 764   "MUL cdelta.xz, parmx.rrgg, {-%f, 0, %f, 0};"
 765   CALCWEIGHTS("parmy", "coord.yyyy")
 766   "MUL cdelta.yw, parmy.rrgg, {0, -%f, 0, %f};"
 767   BICUB_FILT_MAIN("2D");
 768
 769 static const char *bicub_notex_filt_template_RECT =
 770   "ADD coord, fragment.texcoord[%c], {0.5, 0.5};"
 771   "FRC coord.xy, coord.xyxy;"
 772   CALCWEIGHTS("parmx", "coord.xxxx")
 773   "MUL cdelta.xz, parmx.rrgg, {-1, 0, 1, 0};"
 774   CALCWEIGHTS("parmy", "coord.yyyy")
 775   "MUL cdelta.yw, parmy.rrgg, {0, -1, 0, 1};"
 776   BICUB_FILT_MAIN("RECT");
 777
 778 #define BICUB_X_FILT_MAIN(textype) \
 779   "ADD coord.xy, fragment.texcoord[%c].xyxy, cdelta.xyxy;" \
 780   "ADD coord2.xy, fragment.texcoord[%c].xyxy, cdelta.zyzy;" \
 781   "TEX a.r, coord, texture[%c], "textype";" \
 782   "TEX b.r, coord2, texture[%c], "textype";" \
 783   /* x-interpolation */ \
 784   "LRP yuv.%c, parmx.b, a.rrrr, b.rrrr;"
 785
 786 static const char *bicub_x_filt_template_2D =
 787   "MAD coord.x, fragment.texcoord[%c], {%f}, {0.5};"
 788   "TEX parmx, coord, texture[%c], 1D;"
 789   "MUL cdelta.xyz, parmx.rrgg, {-%f, 0, %f};"
 790   BICUB_X_FILT_MAIN("2D");
 791
 792 static const char *bicub_x_filt_template_RECT =
 793   "ADD coord.x, fragment.texcoord[%c], {0.5};"
 794   "TEX parmx, coord, texture[%c], 1D;"
 795   "MUL cdelta.xyz, parmx.rrgg, {-1, 0, 1};"
 796   BICUB_X_FILT_MAIN("RECT");
 797
 798 static const char *unsharp_filt_template =
 799   "PARAM dcoord%c = {%f, %f, %f, %f};"
 800   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;"
 801   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;"
 802   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;"
 803   "TEX b.r, coord.xyxy, texture[%c], %s;"
 804   "TEX b.g, coord.zwzw, texture[%c], %s;"
 805   "ADD b.r, b.r, b.g;"
 806   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 807   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 808   "DP3 b, b, {0.25, 0.25, 0.25};"
 809   "SUB b.r, a.r, b.r;"
 810   "MAD yuv.%c, b.r, {%f}, a.r;";
 811
 812 static const char *unsharp_filt_template2 =
 813   "PARAM dcoord%c = {%f, %f, %f, %f};"
 814   "PARAM dcoord2%c = {%f, 0, 0, %f};"
 815   "ADD coord, fragment.texcoord[%c].xyxy, dcoord%c;"
 816   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord%c;"
 817   "TEX a.r, fragment.texcoord[%c], texture[%c], %s;"
 818   "TEX b.r, coord.xyxy, texture[%c], %s;"
 819   "TEX b.g, coord.zwzw, texture[%c], %s;"
 820   "ADD b.r, b.r, b.g;"
 821   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 822   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 823   "ADD b.r, b.r, b.b;"
 824   "ADD b.a, b.r, b.g;"
 825   "ADD coord, fragment.texcoord[%c].xyxy, dcoord2%c;"
 826   "SUB coord2, fragment.texcoord[%c].xyxy, dcoord2%c;"
 827   "TEX b.r, coord.xyxy, texture[%c], %s;"
 828   "TEX b.g, coord.zwzw, texture[%c], %s;"
 829   "ADD b.r, b.r, b.g;"
 830   "TEX b.b, coord2.xyxy, texture[%c], %s;"
 831   "TEX b.g, coord2.zwzw, texture[%c], %s;"
 832   "DP4 b.r, b, {-0.1171875, -0.1171875, -0.1171875, -0.09765625};"
 833   "MAD b.r, a.r, {0.859375}, b.r;"
 834   "MAD yuv.%c, b.r, {%f}, a.r;";
 835
 836 static const char *yuv_prog_template =
 837   "PARAM ycoef = {%.4f, %.4f, %.4f};"
 838   "PARAM ucoef = {%.4f, %.4f, %.4f};"
 839   "PARAM vcoef = {%.4f, %.4f, %.4f};"
 840   "PARAM offsets = {%.4f, %.4f, %.4f};"
 841   "TEMP res;"
 842   "MAD res.rgb, yuv.rrrr, ycoef, offsets;"
 843   "MAD res.rgb, yuv.gggg, ucoef, res;"
 844   "MAD result.color.rgb, yuv.bbbb, vcoef, res;"
 845   "END";
 846
 847 static const char *yuv_pow_prog_template =
 848   "PARAM ycoef = {%.4f, %.4f, %.4f};"
 849   "PARAM ucoef = {%.4f, %.4f, %.4f};"
 850   "PARAM vcoef = {%.4f, %.4f, %.4f};"
 851   "PARAM offsets = {%.4f, %.4f, %.4f};"
 852   "PARAM gamma = {%.4f, %.4f, %.4f};"
 853   "TEMP res;"
 854   "MAD res.rgb, yuv.rrrr, ycoef, offsets;"
 855   "MAD res.rgb, yuv.gggg, ucoef, res;"
 856   "MAD_SAT res.rgb, yuv.bbbb, vcoef, res;"
 857   "POW result.color.r, res.r, gamma.r;"
 858   "POW result.color.g, res.g, gamma.g;"
 859   "POW result.color.b, res.b, gamma.b;"
 860   "END";
 861
 862 static const char *yuv_lookup_prog_template =
 863   "PARAM ycoef = {%.4f, %.4f, %.4f, 0};"
 864   "PARAM ucoef = {%.4f, %.4f, %.4f, 0};"
 865   "PARAM vcoef = {%.4f, %.4f, %.4f, 0};"
 866   "PARAM offsets = {%.4f, %.4f, %.4f, 0.125};"
 867   "TEMP res;"
 868   "MAD res, yuv.rrrr, ycoef, offsets;"
 869   "MAD res.rgb, yuv.gggg, ucoef, res;"
 870   "MAD res.rgb, yuv.bbbb, vcoef, res;"
 871   "TEX result.color.r, res.raaa, texture[%c], 2D;"
 872   "ADD res.a, res.a, 0.25;"
 873   "TEX result.color.g, res.gaaa, texture[%c], 2D;"
 874   "ADD res.a, res.a, 0.25;"
 875   "TEX result.color.b, res.baaa, texture[%c], 2D;"
 876   "END";
 877
 878 static const char *yuv_lookup3d_prog_template =
 879   "TEX result.color, yuv, texture[%c], 3D;"
 880   "END";
 881
 882 /**
 883  * \brief creates and initializes helper textures needed for scaling texture read
 884  * \param scaler scaler type to create texture for
 885  * \param texu contains next free texture unit number
 886  * \param texs texture unit ids for the scaler are stored in this array
 887  */
 888 static void create_scaler_textures(int scaler, int *texu, char *texs) {
 889   switch (scaler) {
 890     case YUV_SCALER_BILIN:
 891     case YUV_SCALER_BICUB_NOTEX:
 892     case YUV_SCALER_UNSHARP:
 893     case YUV_SCALER_UNSHARP2:
 894       break;
 895     case YUV_SCALER_BICUB:
 896     case YUV_SCALER_BICUB_X:
 897       texs[0] = (*texu)++;
 898       gen_spline_lookup_tex(GL_TEXTURE0 + texs[0]);
 899       texs[0] += '0';
 900       break;
 901     default:
 902       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown scaler type %i\n", scaler);
 903   }
 904 }
 905
 906 static void gen_gamma_map(unsigned char *map, int size, float gamma);
 907
 908 #define ROW_R 0
 909 #define ROW_G 1
 910 #define ROW_B 2
 911 #define COL_Y 0
 912 #define COL_U 1
 913 #define COL_V 2
 914 #define COL_C 3
 915
 916 static void get_yuv2rgb_coeffs(gl_conversion_params_t *params, float yuv2rgb[3][4]) {
 917   float uvcos = params->saturation * cos(params->hue);
 918   float uvsin = params->saturation * sin(params->hue);
 919   int i;
 920   float uv_coeffs[3][2] = {
 921     { 0.000,  1.596},
 922     {-0.391, -0.813},
 923     { 2.018,  0.000}
 924   };
 925   for (i = 0; i < 3; i++) {
 926     yuv2rgb[i][COL_C]  = params->brightness;
 927     yuv2rgb[i][COL_Y]  = 1.164 * params->contrast;
 928     yuv2rgb[i][COL_C] += (-16 / 255.0) * yuv2rgb[i][COL_Y];
 929     yuv2rgb[i][COL_U]  = uv_coeffs[i][0] * uvcos + uv_coeffs[i][1] * uvsin;
 930     yuv2rgb[i][COL_C] += (-128 / 255.0) * yuv2rgb[i][COL_U];
 931     yuv2rgb[i][COL_V]  = uv_coeffs[i][0] * uvsin + uv_coeffs[i][1] * uvcos;
 932     yuv2rgb[i][COL_C] += (-128 / 255.0) * yuv2rgb[i][COL_V];
 933     // this "centers" contrast control so that e.g. a contrast of 0
 934     // leads to a grey image, not a black one
 935     yuv2rgb[i][COL_C] += 0.5 - params->contrast / 2.0;
 936   }
 937 }
 938
 939 //! size of gamma map use to avoid slow exp function in gen_yuv2rgb_map
 940 #define GMAP_SIZE (1024)
 941 /**
 942  * \brief generate a 3D YUV -> RGB map
 943  * \param params struct containing parameters like brightness, gamma, ...
 944  * \param map where to store map. Must provide space for (size + 2)^3 elements
 945  * \param size size of the map, excluding border
 946  */
 947 static void gen_yuv2rgb_map(gl_conversion_params_t *params, unsigned char *map, int size) {
 948   int i, j, k, l;
 949   float step = 1.0 / size;
 950   float y, u, v;
 951   float yuv2rgb[3][4];
 952   unsigned char gmaps[3][GMAP_SIZE];
 953   gen_gamma_map(gmaps[0], GMAP_SIZE, params->rgamma);
 954   gen_gamma_map(gmaps[1], GMAP_SIZE, params->ggamma);
 955   gen_gamma_map(gmaps[2], GMAP_SIZE, params->bgamma);
 956   get_yuv2rgb_coeffs(params, yuv2rgb);
 957   for (i = 0; i < 3; i++)
 958     for (j = 0; j < 4; j++)
 959       yuv2rgb[i][j] *= GMAP_SIZE - 1;
 960   v = 0;
 961   for (i = -1; i <= size; i++) {
 962     u = 0;
 963     for (j = -1; j <= size; j++) {
 964       y = 0;
 965       for (k = -1; k <= size; k++) {
 966         for (l = 0; l < 3; l++) {
 967           float rgb = yuv2rgb[l][COL_Y] * y + yuv2rgb[l][COL_U] * u + yuv2rgb[l][COL_V] * v + yuv2rgb[l][COL_C];
 968           *map++ = gmaps[l][av_clip(rgb, 0, GMAP_SIZE - 1)];
 969         }
 970         y += (k == -1 || k == size - 1) ? step / 2 : step;
 971       }
 972       u += (j == -1 || j == size - 1) ? step / 2 : step;
 973     }
 974     v += (i == -1 || i == size - 1) ? step / 2 : step;
 975   }
 976 }
 977
 978 //! resolution of texture for gamma lookup table
 979 #define LOOKUP_RES 512
 980 //! resolution for 3D yuv->rgb conversion lookup table
 981 #define LOOKUP_3DRES 32
 982 /**
 983  * \brief creates and initializes helper textures needed for yuv conversion
 984  * \param params struct containing parameters like brightness, gamma, ...
 985  * \param texu contains next free texture unit number
 986  * \param texs texture unit ids for the conversion are stored in this array
 987  */
 988 static void create_conv_textures(gl_conversion_params_t *params, int *texu, char *texs) {
 989   unsigned char *lookup_data = NULL;
 990   int conv = YUV_CONVERSION(params->type);
 991   switch (conv) {
 992     case YUV_CONVERSION_FRAGMENT:
 993     case YUV_CONVERSION_FRAGMENT_POW:
 994      break;
 995     case YUV_CONVERSION_FRAGMENT_LOOKUP:
 996       texs[0] = (*texu)++;
 997       ActiveTexture(GL_TEXTURE0 + texs[0]);
 998       lookup_data = malloc(4 * LOOKUP_RES);
 999       gen_gamma_map(lookup_data, LOOKUP_RES, params->rgamma);
1000       gen_gamma_map(&lookup_data[LOOKUP_RES], LOOKUP_RES, params->ggamma);
1001       gen_gamma_map(&lookup_data[2 * LOOKUP_RES], LOOKUP_RES, params->bgamma);
1002       glCreateClearTex(GL_TEXTURE_2D, GL_LUMINANCE8, GL_LUMINANCE, GL_UNSIGNED_BYTE, GL_LINEAR,
1003                        LOOKUP_RES, 4, 0);
1004       glUploadTex(GL_TEXTURE_2D, GL_LUMINANCE, GL_UNSIGNED_BYTE, lookup_data,
1005                   LOOKUP_RES, 0, 0, LOOKUP_RES, 4, 0);
1006       ActiveTexture(GL_TEXTURE0);
1007       texs[0] += '0';
1008       break;
1009     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1010       {
1011         int sz = LOOKUP_3DRES + 2; // texture size including borders
1012         if (!TexImage3D) {
1013           mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing 3D texture function!\n");
1014           break;
1015         }
1016         texs[0] = (*texu)++;
1017         ActiveTexture(GL_TEXTURE0 + texs[0]);
1018         lookup_data = malloc(3 * sz * sz * sz);
1019         gen_yuv2rgb_map(params, lookup_data, LOOKUP_3DRES);
1020         glAdjustAlignment(sz);
1021         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
1022         TexImage3D(GL_TEXTURE_3D, 0, 3, sz, sz, sz, 1,
1023                    GL_RGB, GL_UNSIGNED_BYTE, lookup_data);
1024         glTexParameterf(GL_TEXTURE_3D, GL_TEXTURE_PRIORITY, 1.0);
1025         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1026         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
1027         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_S, GL_CLAMP);
1028         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_T, GL_CLAMP);
1029         glTexParameteri(GL_TEXTURE_3D, GL_TEXTURE_WRAP_R, GL_CLAMP);
1030         ActiveTexture(GL_TEXTURE0);
1031         texs[0] += '0';
1032       }
1033       break;
1034     default:
1035       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", conv);
1036   }
1037   if (lookup_data)
1038     free(lookup_data);
1039 }
1040
1041 /**
1042  * \brief adds a scaling texture read at the current fragment program position
1043  * \param scaler type of scaler to insert
1044  * \param prog_pos current position in fragment program
1045  * \param remain how many bytes remain in the buffer given by prog_pos
1046  * \param texs array containing the texture unit identifiers for this scaler
1047  * \param in_tex texture unit the scaler should read from
1048  * \param out_comp component of the yuv variable the scaler stores the result in
1049  * \param rect if rectangular (pixel) adressing should be used for in_tex
1050  * \param texw width of the in_tex texture
1051  * \param texh height of the in_tex texture
1052  * \param strength strength of filter effect if the scaler does some kind of filtering
1053  */
1054 static void add_scaler(int scaler, char **prog_pos, int *remain, char *texs,
1055                     char in_tex, char out_comp, int rect, int texw, int texh,
1056                     double strength) {
1057   const char *ttype = rect ? "RECT" : "2D";
1058   const float ptw = rect ? 1.0 : 1.0 / texw;
1059   const float pth = rect ? 1.0 : 1.0 / texh;
1060   switch (scaler) {
1061     case YUV_SCALER_BILIN:
1062       snprintf(*prog_pos, *remain, bilin_filt_template, out_comp, in_tex,
1063                  in_tex, ttype);
1064       break;
1065     case YUV_SCALER_BICUB:
1066       if (rect)
1067         snprintf(*prog_pos, *remain, bicub_filt_template_RECT,
1068                  in_tex, texs[0], texs[0],
1069                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1070       else
1071         snprintf(*prog_pos, *remain, bicub_filt_template_2D,
1072                  in_tex, (float)texw, (float)texh,
1073                  texs[0], ptw, ptw, texs[0], pth, pth,
1074                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1075       break;
1076     case YUV_SCALER_BICUB_X:
1077       if (rect)
1078         snprintf(*prog_pos, *remain, bicub_x_filt_template_RECT,
1079                  in_tex, texs[0],
1080                  in_tex, in_tex, in_tex, in_tex, out_comp);
1081       else
1082         snprintf(*prog_pos, *remain, bicub_x_filt_template_2D,
1083                  in_tex, (float)texw,
1084                  texs[0], ptw, ptw,
1085                  in_tex, in_tex, in_tex, in_tex, out_comp);
1086       break;
1087     case YUV_SCALER_BICUB_NOTEX:
1088       if (rect)
1089         snprintf(*prog_pos, *remain, bicub_notex_filt_template_RECT,
1090                  in_tex,
1091                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1092       else
1093         snprintf(*prog_pos, *remain, bicub_notex_filt_template_2D,
1094                  in_tex, (float)texw, (float)texh, ptw, ptw, pth, pth,
1095                  in_tex, in_tex, in_tex, in_tex, in_tex, in_tex, out_comp);
1096       break;
1097     case YUV_SCALER_UNSHARP:
1098       snprintf(*prog_pos, *remain, unsharp_filt_template,
1099                out_comp, 0.5 * ptw, 0.5 * pth, 0.5 * ptw, -0.5 * pth,
1100                in_tex, out_comp, in_tex, out_comp, in_tex,
1101                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1102                in_tex, ttype, out_comp, strength);
1103       break;
1104     case YUV_SCALER_UNSHARP2:
1105       snprintf(*prog_pos, *remain, unsharp_filt_template2,
1106                out_comp, 1.2 * ptw, 1.2 * pth, 1.2 * ptw, -1.2 * pth,
1107                out_comp, 1.5 * ptw, 1.5 * pth,
1108                in_tex, out_comp, in_tex, out_comp, in_tex,
1109                in_tex, ttype, in_tex, ttype, in_tex, ttype, in_tex, ttype,
1110                in_tex, ttype, in_tex, out_comp, in_tex, out_comp,
1111                in_tex, ttype, in_tex, ttype, in_tex, ttype,
1112                in_tex, ttype, out_comp, strength);
1113       break;
1114   }
1115   *remain -= strlen(*prog_pos);
1116   *prog_pos += strlen(*prog_pos);
1117 }
1118
1119 static const struct {
1120   const char *name;
1121   GLenum cur;
1122   GLenum max;
1123 } progstats[] = {
1124   {"instructions", 0x88A0, 0x88A1},
1125   {"native instructions", 0x88A2, 0x88A3},
1126   {"temporaries", 0x88A4, 0x88A5},
1127   {"native temporaries", 0x88A6, 0x88A7},
1128   {"parameters", 0x88A8, 0x88A9},
1129   {"native parameters", 0x88AA, 0x88AB},
1130   {"attribs", 0x88AC, 0x88AD},
1131   {"native attribs", 0x88AE, 0x88AF},
1132   {"ALU instructions", 0x8805, 0x880B},
1133   {"TEX instructions", 0x8806, 0x880C},
1134   {"TEX indirections", 0x8807, 0x880D},
1135   {"native ALU instructions", 0x8808, 0x880E},
1136   {"native TEX instructions", 0x8809, 0x880F},
1137   {"native TEX indirections", 0x880A, 0x8810},
1138   {NULL, 0, 0}
1139 };
1140
1141 /**
1142  * \brief load the specified GPU Program
1143  * \param target program target to load into, only GL_FRAGMENT_PROGRAM is tested
1144  * \param prog program string
1145  * \return 1 on success, 0 otherwise
1146  */
1147 int loadGPUProgram(GLenum target, char *prog) {
1148   int i;
1149   GLint cur = 0, max = 0, err = 0;
1150   if (!ProgramString) {
1151     mp_msg(MSGT_VO, MSGL_ERR, "[gl] Missing GPU program function\n");
1152     return 0;
1153   }
1154   ProgramString(target, GL_PROGRAM_FORMAT_ASCII, strlen(prog), prog);
1155   glGetIntegerv(GL_PROGRAM_ERROR_POSITION, &err);
1156   if (err != -1) {
1157     mp_msg(MSGT_VO, MSGL_ERR,
1158       "[gl] Error compiling fragment program, make sure your card supports\n"
1159       "[gl]   GL_ARB_fragment_program (use glxinfo to check).\n"
1160       "[gl]   Error message:\n  %s at %.10s\n",
1161       glGetString(GL_PROGRAM_ERROR_STRING), &prog[err]);
1162     return 0;
1163   }
1164   if (!GetProgramiv || !mp_msg_test(MSGT_VO, MSGL_V))
1165     return 1;
1166   mp_msg(MSGT_VO, MSGL_V, "[gl] Program statistics:\n");
1167   for (i = 0; progstats[i].name; i++) {
1168     GetProgramiv(target, progstats[i].cur, &cur);
1169     GetProgramiv(target, progstats[i].max, &max);
1170     mp_msg(MSGT_VO, MSGL_V, "[gl]   %s: %i/%i\n", progstats[i].name, cur, max);
1171   }
1172   return 1;
1173 }
1174
1175 #define MAX_PROGSZ (1024*1024)
1176
1177 /**
1178  * \brief setup a fragment program that will do YUV->RGB conversion
1179  * \param parms struct containing parameters like conversion and scaler type,
1180  *              brightness, ...
1181  */
1182 static void glSetupYUVFragprog(gl_conversion_params_t *params) {
1183   int type = params->type;
1184   int texw = params->texw;
1185   int texh = params->texh;
1186   int rect = params->target == GL_TEXTURE_RECTANGLE;
1187   static const char prog_hdr[] =
1188     "!!ARBfp1.0\n"
1189     "OPTION ARB_precision_hint_fastest;"
1190     // all scaler variables must go here so they aren't defined
1191     // multiple times when the same scaler is used more than once
1192     "TEMP coord, coord2, cdelta, parmx, parmy, a, b, yuv;";
1193   int prog_remain;
1194   char *yuv_prog, *prog_pos;
1195   int cur_texu = 3;
1196   char lum_scale_texs[1];
1197   char chrom_scale_texs[1];
1198   char conv_texs[1];
1199   GLint i;
1200   // this is the conversion matrix, with y, u, v factors
1201   // for red, green, blue and the constant offsets
1202   float yuv2rgb[3][4];
1203   create_conv_textures(params, &cur_texu, conv_texs);
1204   create_scaler_textures(YUV_LUM_SCALER(type), &cur_texu, lum_scale_texs);
1205   if (YUV_CHROM_SCALER(type) == YUV_LUM_SCALER(type))
1206     memcpy(chrom_scale_texs, lum_scale_texs, sizeof(chrom_scale_texs));
1207   else
1208     create_scaler_textures(YUV_CHROM_SCALER(type), &cur_texu, chrom_scale_texs);
1209   glGetIntegerv(GL_MAX_TEXTURE_IMAGE_UNITS, &i);
1210   if (i < cur_texu)
1211     mp_msg(MSGT_VO, MSGL_ERR,
1212            "[gl] %i texture units needed for this type of YUV fragment support (found %i)\n",
1213            cur_texu, i);
1214   if (!ProgramString) {
1215     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] ProgramString function missing!\n");
1216     return;
1217   }
1218   yuv_prog = malloc(MAX_PROGSZ);
1219   strcpy(yuv_prog, prog_hdr);
1220   prog_pos    = yuv_prog + sizeof(prog_hdr) - 1;
1221   prog_remain = MAX_PROGSZ - sizeof(prog_hdr);
1222   add_scaler(YUV_LUM_SCALER(type), &prog_pos, &prog_remain, lum_scale_texs,
1223              '0', 'r', rect, texw, texh, params->filter_strength);
1224   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1225              '1', 'g', rect, texw / 2, texh / 2, params->filter_strength);
1226   add_scaler(YUV_CHROM_SCALER(type), &prog_pos, &prog_remain, chrom_scale_texs,
1227              '2', 'b', rect, texw / 2, texh / 2, params->filter_strength);
1228   get_yuv2rgb_coeffs(params, yuv2rgb);
1229   switch (YUV_CONVERSION(type)) {
1230     case YUV_CONVERSION_FRAGMENT:
1231       snprintf(prog_pos, prog_remain, yuv_prog_template,
1232                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1233                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1234                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1235                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C]);
1236       break;
1237     case YUV_CONVERSION_FRAGMENT_POW:
1238       snprintf(prog_pos, prog_remain, yuv_pow_prog_template,
1239                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1240                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1241                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1242                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1243                (float)1.0 / params->rgamma, (float)1.0 / params->bgamma, (float)1.0 / params->bgamma);
1244       break;
1245     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1246       snprintf(prog_pos, prog_remain, yuv_lookup_prog_template,
1247                yuv2rgb[ROW_R][COL_Y], yuv2rgb[ROW_G][COL_Y], yuv2rgb[ROW_B][COL_Y],
1248                yuv2rgb[ROW_R][COL_U], yuv2rgb[ROW_G][COL_U], yuv2rgb[ROW_B][COL_U],
1249                yuv2rgb[ROW_R][COL_V], yuv2rgb[ROW_G][COL_V], yuv2rgb[ROW_B][COL_V],
1250                yuv2rgb[ROW_R][COL_C], yuv2rgb[ROW_G][COL_C], yuv2rgb[ROW_B][COL_C],
1251                conv_texs[0], conv_texs[0], conv_texs[0]);
1252       break;
1253     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1254       snprintf(prog_pos, prog_remain, yuv_lookup3d_prog_template, conv_texs[0]);
1255       break;
1256     default:
1257       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(type));
1258       break;
1259   }
1260   mp_msg(MSGT_VO, MSGL_V, "[gl] generated fragment program:\n%s\n", yuv_prog);
1261   loadGPUProgram(GL_FRAGMENT_PROGRAM, yuv_prog);
1262   free(yuv_prog);
1263 }
1264
1265 /**
1266  * \brief little helper function to create a lookup table for gamma
1267  * \param map buffer to create map into
1268  * \param size size of buffer
1269  * \param gamma gamma value
1270  */
1271 static void gen_gamma_map(unsigned char *map, int size, float gamma) {
1272   int i;
1273   if (gamma == 1.0) {
1274     for (i = 0; i < size; i++)
1275       map[i] = 255 * i / (size - 1);
1276     return;
1277   }
1278   gamma = 1.0 / gamma;
1279   for (i = 0; i < size; i++) {
1280     float tmp = (float)i / (size - 1.0);
1281     tmp = pow(tmp, gamma);
1282     if (tmp > 1.0) tmp = 1.0;
1283     if (tmp < 0.0) tmp = 0.0;
1284     map[i] = 255 * tmp;
1285   }
1286 }
1287
1288 /**
1289  * \brief setup YUV->RGB conversion
1290  * \param parms struct containing parameters like conversion and scaler type,
1291  *              brightness, ...
1292  * \ingroup glconversion
1293  */
1294 void glSetupYUVConversion(gl_conversion_params_t *params) {
1295   float uvcos = params->saturation * cos(params->hue);
1296   float uvsin = params->saturation * sin(params->hue);
1297   switch (YUV_CONVERSION(params->type)) {
1298     case YUV_CONVERSION_COMBINERS:
1299       glSetupYUVCombiners(uvcos, uvsin);
1300       break;
1301     case YUV_CONVERSION_COMBINERS_ATI:
1302       glSetupYUVCombinersATI(uvcos, uvsin);
1303       break;
1304     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1305     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1306     case YUV_CONVERSION_FRAGMENT:
1307     case YUV_CONVERSION_FRAGMENT_POW:
1308       glSetupYUVFragprog(params);
1309       break;
1310     default:
1311       mp_msg(MSGT_VO, MSGL_ERR, "[gl] unknown conversion type %i\n", YUV_CONVERSION(params->type));
1312   }
1313 }
1314
1315 /**
1316  * \brief enable the specified YUV conversion
1317  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1318  * \param type type of YUV conversion
1319  * \ingroup glconversion
1320  */
1321 void glEnableYUVConversion(GLenum target, int type) {
1322   if (type <= 0) return;
1323   switch (YUV_CONVERSION(type)) {
1324     case YUV_CONVERSION_COMBINERS:
1325       ActiveTexture(GL_TEXTURE1);
1326       glEnable(target);
1327       ActiveTexture(GL_TEXTURE2);
1328       glEnable(target);
1329       ActiveTexture(GL_TEXTURE0);
1330       glEnable(GL_REGISTER_COMBINERS_NV);
1331       break;
1332     case YUV_CONVERSION_COMBINERS_ATI:
1333       ActiveTexture(GL_TEXTURE1);
1334       glEnable(target);
1335       ActiveTexture(GL_TEXTURE2);
1336       glEnable(target);
1337       ActiveTexture(GL_TEXTURE0);
1338       glEnable(GL_FRAGMENT_SHADER_ATI);
1339       break;
1340     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1341     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1342     case YUV_CONVERSION_FRAGMENT_POW:
1343     case YUV_CONVERSION_FRAGMENT:
1344       glEnable(GL_FRAGMENT_PROGRAM);
1345       break;
1346   }
1347 }
1348
1349 /**
1350  * \brief disable the specified YUV conversion
1351  * \param target texture target for Y, U and V textures (e.g. GL_TEXTURE_2D)
1352  * \param type type of YUV conversion
1353  * \ingroup glconversion
1354  */
1355 void glDisableYUVConversion(GLenum target, int type) {
1356   if (type <= 0) return;
1357   switch (YUV_CONVERSION(type)) {
1358     case YUV_CONVERSION_COMBINERS:
1359       ActiveTexture(GL_TEXTURE1);
1360       glDisable(target);
1361       ActiveTexture(GL_TEXTURE2);
1362       glDisable(target);
1363       ActiveTexture(GL_TEXTURE0);
1364       glDisable(GL_REGISTER_COMBINERS_NV);
1365       break;
1366     case YUV_CONVERSION_COMBINERS_ATI:
1367       ActiveTexture(GL_TEXTURE1);
1368       glDisable(target);
1369       ActiveTexture(GL_TEXTURE2);
1370       glDisable(target);
1371       ActiveTexture(GL_TEXTURE0);
1372       glDisable(GL_FRAGMENT_SHADER_ATI);
1373       break;
1374     case YUV_CONVERSION_FRAGMENT_LOOKUP3D:
1375     case YUV_CONVERSION_FRAGMENT_LOOKUP:
1376     case YUV_CONVERSION_FRAGMENT_POW:
1377     case YUV_CONVERSION_FRAGMENT:
1378       glDisable(GL_FRAGMENT_PROGRAM);
1379       break;
1380   }
1381 }
1382
1383 /**
1384  * \brief draw a texture part at given 2D coordinates
1385  * \param x screen top coordinate
1386  * \param y screen left coordinate
1387  * \param w screen width coordinate
1388  * \param h screen height coordinate
1389  * \param tx texture top coordinate in pixels
1390  * \param ty texture left coordinate in pixels
1391  * \param tw texture part width in pixels
1392  * \param th texture part height in pixels
1393  * \param sx width of texture in pixels
1394  * \param sy height of texture in pixels
1395  * \param rect_tex whether this texture uses texture_rectangle extension
1396  * \param is_yv12 if set, also draw the textures from units 1 and 2
1397  * \param flip flip the texture upside down
1398  * \ingroup gltexture
1399  */
1400 void glDrawTex(GLfloat x, GLfloat y, GLfloat w, GLfloat h,
1401                GLfloat tx, GLfloat ty, GLfloat tw, GLfloat th,
1402                int sx, int sy, int rect_tex, int is_yv12, int flip) {
1403   GLfloat tx2 = tx / 2, ty2 = ty / 2, tw2 = tw / 2, th2 = th / 2;
1404   if (!rect_tex) {
1405     tx /= sx; ty /= sy; tw /= sx; th /= sy;
1406     tx2 = tx, ty2 = ty, tw2 = tw, th2 = th;
1407   }
1408   if (flip) {
1409     y += h;
1410     h = -h;
1411   }
1412   glBegin(GL_QUADS);
1413   glTexCoord2f(tx, ty);
1414   if (is_yv12) {
1415     MultiTexCoord2f(GL_TEXTURE1, tx2, ty2);
1416     MultiTexCoord2f(GL_TEXTURE2, tx2, ty2);
1417   }
1418   glVertex2f(x, y);
1419   glTexCoord2f(tx, ty + th);
1420   if (is_yv12) {
1421     MultiTexCoord2f(GL_TEXTURE1, tx2, ty2 + th2);
1422     MultiTexCoord2f(GL_TEXTURE2, tx2, ty2 + th2);
1423   }
1424   glVertex2f(x, y + h);
1425   glTexCoord2f(tx + tw, ty + th);
1426   if (is_yv12) {
1427     MultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2 + th2);
1428     MultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2 + th2);
1429   }
1430   glVertex2f(x + w, y + h);
1431   glTexCoord2f(tx + tw, ty);
1432   if (is_yv12) {
1433     MultiTexCoord2f(GL_TEXTURE1, tx2 + tw2, ty2);
1434     MultiTexCoord2f(GL_TEXTURE2, tx2 + tw2, ty2);
1435   }
1436   glVertex2f(x + w, y);
1437   glEnd();
1438 }
1439
1440 #ifdef GL_WIN32
1441 #include "w32_common.h"
1442 /**
1443  * \brief little helper since wglGetProcAddress definition does not fit our
1444  *        getProcAddress
1445  * \param procName name of function to look up
1446  * \return function pointer returned by wglGetProcAddress
1447  */
1448 static void *w32gpa(const GLubyte *procName) {
1449   HMODULE oglmod;
1450   void *res = wglGetProcAddress(procName);
1451   if (res) return res;
1452   oglmod = GetModuleHandle("opengl32.dll");
1453   return GetProcAddress(oglmod, procName);
1454 }
1455
1456 int setGlWindow(int *vinfo, HGLRC *context, HWND win)
1457 {
1458   int new_vinfo;
1459   HDC windc = GetDC(win);
1460   HGLRC new_context = 0;
1461   int keep_context = 0;
1462   int res = SET_WINDOW_FAILED;
1463
1464   // should only be needed when keeping context, but not doing glFinish
1465   // can cause flickering even when we do not keep it.
1466   if (*context)
1467   glFinish();
1468   new_vinfo = GetPixelFormat(windc);
1469   if (*context && *vinfo && new_vinfo && *vinfo == new_vinfo) {
1470       // we can keep the wglContext
1471       new_context = *context;
1472       keep_context = 1;
1473   } else {
1474     // create a context
1475     new_context = wglCreateContext(windc);
1476     if (!new_context) {
1477       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GL context!\n");
1478       goto out;
1479     }
1480   }
1481
1482   // set context
1483   if (!wglMakeCurrent(windc, new_context)) {
1484     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GL context!\n");
1485     if (!keep_context) {
1486       wglDeleteContext(new_context);
1487     }
1488     goto out;
1489   }
1490
1491   // set new values
1492   vo_w32_window = win;
1493   {
1494     RECT rect;
1495     GetClientRect(win, &rect);
1496     vo_dwidth = rect.right;
1497     vo_dheight = rect.bottom;
1498   }
1499   if (!keep_context) {
1500     if (*context)
1501       wglDeleteContext(*context);
1502     *context = new_context;
1503     *vinfo = new_vinfo;
1504     getFunctions(w32gpa, NULL);
1505
1506     // and inform that reinit is neccessary
1507     res = SET_WINDOW_REINIT;
1508   } else
1509     res = SET_WINDOW_OK;
1510
1511 out:
1512   ReleaseDC(win, windc);
1513   return res;
1514 }
1515
1516 void releaseGlContext(int *vinfo, HGLRC *context) {
1517   *vinfo = 0;
1518   if (*context) {
1519     wglMakeCurrent(0, 0);
1520     wglDeleteContext(*context);
1521   }
1522   *context = 0;
1523 }
1524
1525 void swapGlBuffers(void) {
1526   HDC vo_hdc = GetDC(vo_w32_window);
1527   SwapBuffers(vo_hdc);
1528   ReleaseDC(vo_w32_window, vo_hdc);
1529 }
1530 #else
1531 #ifdef HAVE_LIBDL
1532 #include <dlfcn.h>
1533 #endif
1534 #include "x11_common.h"
1535 /**
1536  * \brief find address of a linked function
1537  * \param s name of function to find
1538  * \return address of function or NULL if not found
1539  */
1540 static void *getdladdr(const char *s) {
1541   void *ret = NULL;
1542 #ifdef HAVE_LIBDL
1543   void *handle = dlopen(NULL, RTLD_LAZY);
1544   if (!handle)
1545     return NULL;
1546   ret = dlsym(handle, s);
1547   dlclose(handle);
1548 #endif
1549   return ret;
1550 }
1551
1552 /**
1553  * \brief Returns the XVisualInfo associated with Window win.
1554  * \param win Window whose XVisualInfo is returne.
1555  * \return XVisualInfo of the window. Caller must use XFree to free it.
1556  */
1557 static XVisualInfo *getWindowVisualInfo(Window win) {
1558   XWindowAttributes xw_attr;
1559   XVisualInfo vinfo_template;
1560   int tmp;
1561   XGetWindowAttributes(mDisplay, win, &xw_attr);
1562   vinfo_template.visualid = XVisualIDFromVisual(xw_attr.visual);
1563   return XGetVisualInfo(mDisplay, VisualIDMask, &vinfo_template, &tmp);
1564 }
1565
1566 /**
1567  * \brief Changes the window in which video is displayed.
1568  * If possible only transfers the context to the new window, otherwise
1569  * creates a new one, which must be initialized by the caller.
1570  * \param vinfo Currently used visual.
1571  * \param context Currently used context.
1572  * \param win window that should be used for drawing.
1573  * \return one of SET_WINDOW_FAILED, SET_WINDOW_OK or SET_WINDOW_REINIT.
1574  * In case of SET_WINDOW_REINIT the context could not be transfered
1575  * and the caller must initialize it correctly.
1576  * \ingroup glcontext
1577  */
1578 int setGlWindow(XVisualInfo **vinfo, GLXContext *context, Window win)
1579 {
1580   XVisualInfo *new_vinfo;
1581   GLXContext new_context = NULL;
1582   int keep_context = 0;
1583
1584   // should only be needed when keeping context, but not doing glFinish
1585   // can cause flickering even when we do not keep it.
1586   if (*context)
1587   glFinish();
1588   new_vinfo = getWindowVisualInfo(win);
1589   if (*context && *vinfo && new_vinfo &&
1590       (*vinfo)->visualid == new_vinfo->visualid) {
1591       // we can keep the GLXContext
1592       new_context = *context;
1593       XFree(new_vinfo);
1594       new_vinfo = *vinfo;
1595       keep_context = 1;
1596   } else {
1597     // create a context
1598     new_context = glXCreateContext(mDisplay, new_vinfo, NULL, True);
1599     if (!new_context) {
1600       mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not create GLX context!\n");
1601       XFree(new_vinfo);
1602       return SET_WINDOW_FAILED;
1603     }
1604   }
1605
1606   // set context
1607   if (!glXMakeCurrent(mDisplay, vo_window, new_context)) {
1608     mp_msg(MSGT_VO, MSGL_FATAL, "[gl] Could not set GLX context!\n");
1609     if (!keep_context) {
1610       glXDestroyContext(mDisplay, new_context);
1611       XFree(new_vinfo);
1612     }
1613     return SET_WINDOW_FAILED;
1614   }
1615
1616   // set new values
1617   vo_window = win;
1618   vo_x11_update_geometry();
1619   if (!keep_context) {
1620     void *(*getProcAddress)(const GLubyte *);
1621     const char *(*glXExtStr)(Display *, int);
1622     if (*context)
1623       glXDestroyContext(mDisplay, *context);
1624     *context = new_context;
1625     if (*vinfo)
1626       XFree(*vinfo);
1627     *vinfo = new_vinfo;
1628       getProcAddress = getdladdr("glXGetProcAddress");
1629     if (!getProcAddress)
1630       getProcAddress = getdladdr("glXGetProcAddressARB");
1631     if (!getProcAddress)
1632       getProcAddress = (void *)getdladdr;
1633     glXExtStr = getdladdr("glXQueryExtensionsString");
1634     getFunctions(getProcAddress, !glXExtStr ? NULL :
1635                  glXExtStr(mDisplay, DefaultScreen(mDisplay)));
1636
1637     // and inform that reinit is neccessary
1638     return SET_WINDOW_REINIT;
1639   }
1640   return SET_WINDOW_OK;
1641 }
1642
1643 /**
1644  * \brief free the VisualInfo and GLXContext of an OpenGL context.
1645  * \ingroup glcontext
1646  */
1647 void releaseGlContext(XVisualInfo **vinfo, GLXContext *context) {
1648   if (*vinfo)
1649     XFree(*vinfo);
1650   *vinfo = NULL;
1651   if (*context)
1652   {
1653     glFinish();
1654     glXMakeCurrent(mDisplay, None, NULL);
1655     glXDestroyContext(mDisplay, *context);
1656   }
1657   *context = 0;
1658 }
1659
1660 void swapGlBuffers(void) {
1661   glXSwapBuffers(mDisplay, vo_window);
1662 }
1663 #endif
1664