dlls/wined3d/glsl_shader.c

   1 /*
   2  * GLSL pixel and vertex shader implementation
   3  *
   4  * Copyright 2006 Jason Green
   5  * Copyright 2006-2007 Henri Verbeet
   6  *
   7  * This library is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * This library is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with this library; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  20  */
  21
  22 /*
  23  * D3D shader asm has swizzles on source parameters, and write masks for
  24  * destination parameters. GLSL uses swizzles for both. The result of this is
  25  * that for example "mov dst.xw, src.zyxw" becomes "dst.xw = src.zw" in GLSL.
  26  * Ie, to generate a proper GLSL source swizzle, we need to take the D3D write
  27  * mask for the destination parameter into account.
  28  */
  29
  30 #include "config.h"
  31 #include <stdio.h>
  32 #include "wined3d_private.h"
  33
  34 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  35 WINE_DECLARE_DEBUG_CHANNEL(d3d_constants);
  36
  37 #define GLINFO_LOCATION      (*gl_info)
  38
  39 typedef struct {
  40     char reg_name[50];
  41     char mask_str[6];
  42 } glsl_dst_param_t;
  43
  44 typedef struct {
  45     char reg_name[50];
  46     char param_str[100];
  47 } glsl_src_param_t;
  48
  49 typedef struct {
  50     const char *name;
  51     DWORD coord_mask;
  52 } glsl_sample_function_t;
  53
  54 /** Prints the GLSL info log which will contain error messages if they exist */
  55 void print_glsl_info_log(WineD3D_GL_Info *gl_info, GLhandleARB obj) {
  56
  57     int infologLength = 0;
  58     char *infoLog;
  59
  60     GL_EXTCALL(glGetObjectParameterivARB(obj,
  61                GL_OBJECT_INFO_LOG_LENGTH_ARB,
  62                &infologLength));
  63
  64     /* A size of 1 is just a null-terminated string, so the log should be bigger than
  65      * that if there are errors. */
  66     if (infologLength > 1)
  67     {
  68         infoLog = (char *)HeapAlloc(GetProcessHeap(), 0, infologLength);
  69         GL_EXTCALL(glGetInfoLogARB(obj, infologLength, NULL, infoLog));
  70         FIXME("Error received from GLSL shader #%u: %s\n", obj, debugstr_a(infoLog));
  71         HeapFree(GetProcessHeap(), 0, infoLog);
  72     }
  73 }
  74
  75 /**
  76  * Loads (pixel shader) samplers
  77  */
  78 static void shader_glsl_load_psamplers(
  79     WineD3D_GL_Info *gl_info,
  80     IWineD3DStateBlock* iface) {
  81
  82     IWineD3DStateBlockImpl* stateBlock = (IWineD3DStateBlockImpl*) iface;
  83     GLhandleARB programId = stateBlock->glsl_program->programId;
  84     GLhandleARB name_loc;
  85     int i;
  86     char sampler_name[20];
  87
  88     for (i=0; i< GL_LIMITS(samplers); ++i) {
  89         if (stateBlock->textures[i] != NULL) {
  90            snprintf(sampler_name, sizeof(sampler_name), "Psampler%d", i);
  91            name_loc = GL_EXTCALL(glGetUniformLocationARB(programId, sampler_name));
  92            if (name_loc != -1) {
  93                TRACE("Loading %s for texture %d\n", sampler_name, i);
  94                GL_EXTCALL(glUniform1iARB(name_loc, i));
  95                checkGLcall("glUniform1iARB");
  96            }
  97         }
  98     }
  99 }
 100
 101 /**
 102  * Loads floating point constants (aka uniforms) into the currently set GLSL program.
 103  * When constant_list == NULL, it will load all the constants.
 104  */
 105 static void shader_glsl_load_constantsF(IWineD3DBaseShaderImpl* This, WineD3D_GL_Info *gl_info,
 106         unsigned int max_constants, float* constants, GLhandleARB *constant_locations,
 107         struct list *constant_list) {
 108     constants_entry *constant;
 109     local_constant* lconst;
 110     GLhandleARB tmp_loc;
 111     DWORD i, j;
 112     DWORD *idx;
 113
 114     if (TRACE_ON(d3d_shader)) {
 115         LIST_FOR_EACH_ENTRY(constant, constant_list, constants_entry, entry) {
 116             idx = constant->idx;
 117             j = constant->count;
 118             while (j--) {
 119                 i = *idx++;
 120                 tmp_loc = constant_locations[i];
 121                 if (tmp_loc != -1) {
 122                     TRACE_(d3d_constants)("Loading constants %i: %f, %f, %f, %f\n", i,
 123                             constants[i * 4 + 0], constants[i * 4 + 1],
 124                             constants[i * 4 + 2], constants[i * 4 + 3]);
 125                 }
 126             }
 127         }
 128     }
 129     LIST_FOR_EACH_ENTRY(constant, constant_list, constants_entry, entry) {
 130         idx = constant->idx;
 131         j = constant->count;
 132         while (j--) {
 133             i = *idx++;
 134             tmp_loc = constant_locations[i];
 135             if (tmp_loc != -1) {
 136                 /* We found this uniform name in the program - go ahead and send the data */
 137                 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, constants + (i * 4)));
 138             }
 139         }
 140     }
 141     checkGLcall("glUniform4fvARB()");
 142
 143     /* Load immediate constants */
 144     if (TRACE_ON(d3d_shader)) {
 145         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 146             tmp_loc = constant_locations[lconst->idx];
 147             if (tmp_loc != -1) {
 148                 GLfloat* values = (GLfloat*)lconst->value;
 149                 TRACE_(d3d_constants)("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
 150                         values[0], values[1], values[2], values[3]);
 151             }
 152         }
 153     }
 154     LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 155         tmp_loc = constant_locations[lconst->idx];
 156         if (tmp_loc != -1) {
 157             /* We found this uniform name in the program - go ahead and send the data */
 158             GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, (GLfloat*)lconst->value));
 159         }
 160     }
 161     checkGLcall("glUniform4fvARB()");
 162 }
 163
 164 /**
 165  * Loads integer constants (aka uniforms) into the currently set GLSL program.
 166  * When @constants_set == NULL, it will load all the constants.
 167  */
 168 static void shader_glsl_load_constantsI(
 169     IWineD3DBaseShaderImpl* This,
 170     WineD3D_GL_Info *gl_info,
 171     GLhandleARB programId,
 172     unsigned max_constants,
 173     int* constants,
 174     BOOL* constants_set) {
 175
 176     GLhandleARB tmp_loc;
 177     int i;
 178     char tmp_name[8];
 179     char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
 180     const char* prefix = is_pshader? "PI":"VI";
 181     struct list* ptr;
 182
 183     for (i=0; i<max_constants; ++i) {
 184         if (NULL == constants_set || constants_set[i]) {
 185
 186             TRACE_(d3d_constants)("Loading constants %i: %i, %i, %i, %i\n",
 187                   i, constants[i*4], constants[i*4+1], constants[i*4+2], constants[i*4+3]);
 188
 189             /* TODO: Benchmark and see if it would be beneficial to store the
 190              * locations of the constants to avoid looking up each time */
 191             snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
 192             tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
 193             if (tmp_loc != -1) {
 194                 /* We found this uniform name in the program - go ahead and send the data */
 195                 GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, &constants[i*4]));
 196                 checkGLcall("glUniform4ivARB");
 197             }
 198         }
 199     }
 200
 201     /* Load immediate constants */
 202     ptr = list_head(&This->baseShader.constantsI);
 203     while (ptr) {
 204         local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
 205         unsigned int idx = lconst->idx;
 206         GLint* values = (GLint*) lconst->value;
 207
 208         TRACE_(d3d_constants)("Loading local constants %i: %i, %i, %i, %i\n", idx,
 209             values[0], values[1], values[2], values[3]);
 210
 211         snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
 212         tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
 213         if (tmp_loc != -1) {
 214             /* We found this uniform name in the program - go ahead and send the data */
 215             GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, values));
 216             checkGLcall("glUniform4ivARB");
 217         }
 218         ptr = list_next(&This->baseShader.constantsI, ptr);
 219     }
 220 }
 221
 222 /**
 223  * Loads boolean constants (aka uniforms) into the currently set GLSL program.
 224  * When @constants_set == NULL, it will load all the constants.
 225  */
 226 static void shader_glsl_load_constantsB(
 227     IWineD3DBaseShaderImpl* This,
 228     WineD3D_GL_Info *gl_info,
 229     GLhandleARB programId,
 230     unsigned max_constants,
 231     BOOL* constants,
 232     BOOL* constants_set) {
 233
 234     GLhandleARB tmp_loc;
 235     int i;
 236     char tmp_name[8];
 237     char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
 238     const char* prefix = is_pshader? "PB":"VB";
 239     struct list* ptr;
 240
 241     for (i=0; i<max_constants; ++i) {
 242         if (NULL == constants_set || constants_set[i]) {
 243
 244             TRACE_(d3d_constants)("Loading constants %i: %i;\n", i, constants[i*4]);
 245
 246             /* TODO: Benchmark and see if it would be beneficial to store the
 247              * locations of the constants to avoid looking up each time */
 248             snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
 249             tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
 250             if (tmp_loc != -1) {
 251                 /* We found this uniform name in the program - go ahead and send the data */
 252                 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, &constants[i*4]));
 253                 checkGLcall("glUniform1ivARB");
 254             }
 255         }
 256     }
 257
 258     /* Load immediate constants */
 259     ptr = list_head(&This->baseShader.constantsB);
 260     while (ptr) {
 261         local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
 262         unsigned int idx = lconst->idx;
 263         GLint* values = (GLint*) lconst->value;
 264
 265         TRACE_(d3d_constants)("Loading local constants %i: %i\n", idx, values[0]);
 266
 267         snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
 268         tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
 269         if (tmp_loc != -1) {
 270             /* We found this uniform name in the program - go ahead and send the data */
 271             GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, values));
 272             checkGLcall("glUniform1ivARB");
 273         }
 274         ptr = list_next(&This->baseShader.constantsB, ptr);
 275     }
 276 }
 277
 278
 279
 280 /**
 281  * Loads the app-supplied constants into the currently set GLSL program.
 282  */
 283 void shader_glsl_load_constants(
 284     IWineD3DDevice* device,
 285     char usePixelShader,
 286     char useVertexShader) {
 287
 288     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
 289     IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
 290     WineD3D_GL_Info *gl_info = &((IWineD3DImpl*) deviceImpl->wineD3D)->gl_info;
 291
 292     GLhandleARB *constant_locations;
 293     struct list *constant_list;
 294     GLhandleARB programId;
 295     GLint pos;
 296
 297     if (!stateBlock->glsl_program) {
 298         /* No GLSL program set - nothing to do. */
 299         return;
 300     }
 301     programId = stateBlock->glsl_program->programId;
 302
 303     if (useVertexShader) {
 304         IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
 305         GLint pos;
 306
 307         constant_locations = stateBlock->glsl_program->vuniformF_locations;
 308         constant_list = &stateBlock->set_vconstantsF;
 309
 310         /* Load DirectX 9 float constants/uniforms for vertex shader */
 311         shader_glsl_load_constantsF(vshader, gl_info, GL_LIMITS(vshader_constantsF),
 312                 stateBlock->vertexShaderConstantF, constant_locations, constant_list);
 313
 314         /* Load DirectX 9 integer constants/uniforms for vertex shader */
 315         shader_glsl_load_constantsI(vshader, gl_info, programId, MAX_CONST_I,
 316                                     stateBlock->vertexShaderConstantI,
 317                                     stateBlock->set.vertexShaderConstantsI);
 318
 319         /* Load DirectX 9 boolean constants/uniforms for vertex shader */
 320         shader_glsl_load_constantsB(vshader, gl_info, programId, MAX_CONST_B,
 321                                     stateBlock->vertexShaderConstantB,
 322                                     stateBlock->set.vertexShaderConstantsB);
 323
 324         /* Upload the position fixup params */
 325         pos = GL_EXTCALL(glGetUniformLocationARB(programId, "posFixup"));
 326         checkGLcall("glGetUniformLocationARB");
 327         GL_EXTCALL(glUniform4fvARB(pos, 1, &deviceImpl->posFixup[0]));
 328         checkGLcall("glUniform4fvARB");
 329     }
 330
 331     if (usePixelShader) {
 332
 333         IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
 334
 335         constant_locations = stateBlock->glsl_program->puniformF_locations;
 336         constant_list = &stateBlock->set_pconstantsF;
 337
 338         /* Load pixel shader samplers */
 339         shader_glsl_load_psamplers(gl_info, (IWineD3DStateBlock*) stateBlock);
 340
 341         /* Load DirectX 9 float constants/uniforms for pixel shader */
 342         shader_glsl_load_constantsF(pshader, gl_info, GL_LIMITS(pshader_constantsF),
 343                 stateBlock->pixelShaderConstantF, constant_locations, constant_list);
 344
 345         /* Load DirectX 9 integer constants/uniforms for pixel shader */
 346         shader_glsl_load_constantsI(pshader, gl_info, programId, MAX_CONST_I,
 347                                     stateBlock->pixelShaderConstantI,
 348                                     stateBlock->set.pixelShaderConstantsI);
 349
 350         /* Load DirectX 9 boolean constants/uniforms for pixel shader */
 351         shader_glsl_load_constantsB(pshader, gl_info, programId, MAX_CONST_B,
 352                                     stateBlock->pixelShaderConstantB,
 353                                     stateBlock->set.pixelShaderConstantsB);
 354
 355         /* Upload the environment bump map matrix if needed. The needsbumpmat member specifies the texture stage to load the matrix from.
 356          * It can't be 0 for a valid texbem instruction.
 357          */
 358         if(((IWineD3DPixelShaderImpl *) pshader)->needsbumpmat != -1) {
 359             float *data = (float *) &stateBlock->textureState[(int) ((IWineD3DPixelShaderImpl *) pshader)->needsbumpmat][WINED3DTSS_BUMPENVMAT00];
 360             pos = GL_EXTCALL(glGetUniformLocationARB(programId, "bumpenvmat"));
 361             checkGLcall("glGetUniformLocationARB");
 362             GL_EXTCALL(glUniformMatrix2fvARB(pos, 1, 0, data));
 363             checkGLcall("glUniform4fvARB");
 364         }
 365     }
 366 }
 367
 368 /** Generate the variable & register declarations for the GLSL output target */
 369 void shader_generate_glsl_declarations(
 370     IWineD3DBaseShader *iface,
 371     shader_reg_maps* reg_maps,
 372     SHADER_BUFFER* buffer,
 373     WineD3D_GL_Info* gl_info) {
 374
 375     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 376     int i;
 377
 378     /* There are some minor differences between pixel and vertex shaders */
 379     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 380     char prefix = pshader ? 'P' : 'V';
 381
 382     /* Prototype the subroutines */
 383     for (i = 0; i < This->baseShader.limits.label; i++) {
 384         if (reg_maps->labels[i])
 385             shader_addline(buffer, "void subroutine%lu();\n", i);
 386     }
 387
 388     /* Declare the constants (aka uniforms) */
 389     if (This->baseShader.limits.constant_float > 0) {
 390         unsigned max_constantsF = min(This->baseShader.limits.constant_float,
 391                 (pshader ? GL_LIMITS(pshader_constantsF) : GL_LIMITS(vshader_constantsF)));
 392         shader_addline(buffer, "uniform vec4 %cC[%u];\n", prefix, max_constantsF);
 393     }
 394
 395     if (This->baseShader.limits.constant_int > 0)
 396         shader_addline(buffer, "uniform ivec4 %cI[%u];\n", prefix, This->baseShader.limits.constant_int);
 397
 398     if (This->baseShader.limits.constant_bool > 0)
 399         shader_addline(buffer, "uniform bool %cB[%u];\n", prefix, This->baseShader.limits.constant_bool);
 400
 401     if(!pshader)
 402         shader_addline(buffer, "uniform vec4 posFixup;\n");
 403     else if(reg_maps->bumpmat != -1)
 404         shader_addline(buffer, "uniform mat2 bumpenvmat;\n");
 405
 406     /* Declare texture samplers */
 407     for (i = 0; i < This->baseShader.limits.sampler; i++) {
 408         if (reg_maps->samplers[i]) {
 409
 410             DWORD stype = reg_maps->samplers[i] & WINED3DSP_TEXTURETYPE_MASK;
 411             switch (stype) {
 412
 413                 case WINED3DSTT_1D:
 414                     shader_addline(buffer, "uniform sampler1D %csampler%lu;\n", prefix, i);
 415                     break;
 416                 case WINED3DSTT_2D:
 417                     shader_addline(buffer, "uniform sampler2D %csampler%lu;\n", prefix, i);
 418                     break;
 419                 case WINED3DSTT_CUBE:
 420                     shader_addline(buffer, "uniform samplerCube %csampler%lu;\n", prefix, i);
 421                     break;
 422                 case WINED3DSTT_VOLUME:
 423                     shader_addline(buffer, "uniform sampler3D %csampler%lu;\n", prefix, i);
 424                     break;
 425                 default:
 426                     shader_addline(buffer, "uniform unsupported_sampler %csampler%lu;\n", prefix, i);
 427                     FIXME("Unrecognized sampler type: %#x\n", stype);
 428                     break;
 429             }
 430         }
 431     }
 432
 433     /* Declare address variables */
 434     for (i = 0; i < This->baseShader.limits.address; i++) {
 435         if (reg_maps->address[i])
 436             shader_addline(buffer, "ivec4 A%d;\n", i);
 437     }
 438
 439     /* Declare texture coordinate temporaries and initialize them */
 440     for (i = 0; i < This->baseShader.limits.texcoord; i++) {
 441         if (reg_maps->texcoord[i])
 442             shader_addline(buffer, "vec4 T%lu = gl_TexCoord[%lu];\n", i, i);
 443     }
 444
 445     /* Declare input register temporaries */
 446     for (i=0; i < This->baseShader.limits.packed_input; i++) {
 447         if (reg_maps->packed_input[i])
 448             shader_addline(buffer, "vec4 IN%lu;\n", i);
 449     }
 450
 451     /* Declare output register temporaries */
 452     for (i = 0; i < This->baseShader.limits.packed_output; i++) {
 453         if (reg_maps->packed_output[i])
 454             shader_addline(buffer, "vec4 OUT%lu;\n", i);
 455     }
 456
 457     /* Declare temporary variables */
 458     for(i = 0; i < This->baseShader.limits.temporary; i++) {
 459         if (reg_maps->temporary[i])
 460             shader_addline(buffer, "vec4 R%lu;\n", i);
 461     }
 462
 463     /* Declare attributes */
 464     for (i = 0; i < This->baseShader.limits.attributes; i++) {
 465         if (reg_maps->attributes[i])
 466             shader_addline(buffer, "attribute vec4 attrib%i;\n", i);
 467     }
 468
 469     /* Declare loop register aL */
 470     if (reg_maps->loop) {
 471         shader_addline(buffer, "int aL;\n");
 472         shader_addline(buffer, "int tmpInt;\n");
 473     }
 474
 475     /* Temporary variables for matrix operations */
 476     shader_addline(buffer, "vec4 tmp0;\n");
 477     shader_addline(buffer, "vec4 tmp1;\n");
 478
 479     /* Start the main program */
 480     shader_addline(buffer, "void main() {\n");
 481 }
 482
 483 /*****************************************************************************
 484  * Functions to generate GLSL strings from DirectX Shader bytecode begin here.
 485  *
 486  * For more information, see http://wiki.winehq.org/DirectX-Shaders
 487  ****************************************************************************/
 488
 489 /* Prototypes */
 490 static void shader_glsl_add_src_param(SHADER_OPCODE_ARG* arg, const DWORD param,
 491         const DWORD addr_token, DWORD mask, glsl_src_param_t *src_param);
 492
 493 /** Used for opcode modifiers - They multiply the result by the specified amount */
 494 static const char * const shift_glsl_tab[] = {
 495     "",           /*  0 (none) */
 496     "2.0 * ",     /*  1 (x2)   */
 497     "4.0 * ",     /*  2 (x4)   */
 498     "8.0 * ",     /*  3 (x8)   */
 499     "16.0 * ",    /*  4 (x16)  */
 500     "32.0 * ",    /*  5 (x32)  */
 501     "",           /*  6 (x64)  */
 502     "",           /*  7 (x128) */
 503     "",           /*  8 (d256) */
 504     "",           /*  9 (d128) */
 505     "",           /* 10 (d64)  */
 506     "",           /* 11 (d32)  */
 507     "0.0625 * ",  /* 12 (d16)  */
 508     "0.125 * ",   /* 13 (d8)   */
 509     "0.25 * ",    /* 14 (d4)   */
 510     "0.5 * "      /* 15 (d2)   */
 511 };
 512
 513 /* Generate a GLSL parameter that does the input modifier computation and return the input register/mask to use */
 514 static void shader_glsl_gen_modifier (
 515     const DWORD instr,
 516     const char *in_reg,
 517     const char *in_regswizzle,
 518     char *out_str) {
 519
 520     out_str[0] = 0;
 521
 522     if (instr == WINED3DSIO_TEXKILL)
 523         return;
 524
 525     switch (instr & WINED3DSP_SRCMOD_MASK) {
 526     case WINED3DSPSM_DZ: /* Need to handle this in the instructions itself (texld & texcrd). */
 527     case WINED3DSPSM_DW:
 528     case WINED3DSPSM_NONE:
 529         sprintf(out_str, "%s%s", in_reg, in_regswizzle);
 530         break;
 531     case WINED3DSPSM_NEG:
 532         sprintf(out_str, "-%s%s", in_reg, in_regswizzle);
 533         break;
 534     case WINED3DSPSM_NOT:
 535         sprintf(out_str, "!%s%s", in_reg, in_regswizzle);
 536         break;
 537     case WINED3DSPSM_BIAS:
 538         sprintf(out_str, "(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
 539         break;
 540     case WINED3DSPSM_BIASNEG:
 541         sprintf(out_str, "-(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
 542         break;
 543     case WINED3DSPSM_SIGN:
 544         sprintf(out_str, "(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
 545         break;
 546     case WINED3DSPSM_SIGNNEG:
 547         sprintf(out_str, "-(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
 548         break;
 549     case WINED3DSPSM_COMP:
 550         sprintf(out_str, "(1.0 - %s%s)", in_reg, in_regswizzle);
 551         break;
 552     case WINED3DSPSM_X2:
 553         sprintf(out_str, "(2.0 * %s%s)", in_reg, in_regswizzle);
 554         break;
 555     case WINED3DSPSM_X2NEG:
 556         sprintf(out_str, "-(2.0 * %s%s)", in_reg, in_regswizzle);
 557         break;
 558     case WINED3DSPSM_ABS:
 559         sprintf(out_str, "abs(%s%s)", in_reg, in_regswizzle);
 560         break;
 561     case WINED3DSPSM_ABSNEG:
 562         sprintf(out_str, "-abs(%s%s)", in_reg, in_regswizzle);
 563         break;
 564     default:
 565         FIXME("Unhandled modifier %u\n", (instr & WINED3DSP_SRCMOD_MASK));
 566         sprintf(out_str, "%s%s", in_reg, in_regswizzle);
 567     }
 568 }
 569
 570 /** Writes the GLSL variable name that corresponds to the register that the
 571  * DX opcode parameter is trying to access */
 572 static void shader_glsl_get_register_name(
 573     const DWORD param,
 574     const DWORD addr_token,
 575     char* regstr,
 576     BOOL* is_color,
 577     SHADER_OPCODE_ARG* arg) {
 578
 579     /* oPos, oFog and oPts in D3D */
 580     static const char * const hwrastout_reg_names[] = { "gl_Position", "gl_FogFragCoord", "gl_PointSize" };
 581
 582     DWORD reg = param & WINED3DSP_REGNUM_MASK;
 583     DWORD regtype = shader_get_regtype(param);
 584     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) arg->shader;
 585     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
 586     WineD3D_GL_Info* gl_info = &((IWineD3DImpl*)deviceImpl->wineD3D)->gl_info;
 587
 588     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 589     char tmpStr[50];
 590
 591     *is_color = FALSE;
 592
 593     switch (regtype) {
 594     case WINED3DSPR_TEMP:
 595         sprintf(tmpStr, "R%u", reg);
 596     break;
 597     case WINED3DSPR_INPUT:
 598         if (pshader) {
 599             /* Pixel shaders >= 3.0 */
 600             if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
 601                 sprintf(tmpStr, "IN%u", reg);
 602              else {
 603                 if (reg==0)
 604                     strcpy(tmpStr, "gl_Color");
 605                 else
 606                     strcpy(tmpStr, "gl_SecondaryColor");
 607             }
 608         } else {
 609             if (vshader_input_is_color((IWineD3DVertexShader*) This, reg))
 610                *is_color = TRUE;
 611             sprintf(tmpStr, "attrib%u", reg);
 612         }
 613         break;
 614     case WINED3DSPR_CONST:
 615     {
 616         const char* prefix = pshader? "PC":"VC";
 617
 618         /* Relative addressing */
 619         if (param & WINED3DSHADER_ADDRMODE_RELATIVE) {
 620
 621            /* Relative addressing on shaders 2.0+ have a relative address token,
 622             * prior to that, it was hard-coded as "A0.x" because there's only 1 register */
 623            if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2)  {
 624                glsl_src_param_t rel_param;
 625                shader_glsl_add_src_param(arg, addr_token, 0, WINED3DSP_WRITEMASK_0, &rel_param);
 626                sprintf(tmpStr, "%s[%s + %u]", prefix, rel_param.param_str, reg);
 627            } else
 628                sprintf(tmpStr, "%s[A0.x + %u]", prefix, reg);
 629
 630         } else
 631              sprintf(tmpStr, "%s[%u]", prefix, reg);
 632
 633         break;
 634     }
 635     case WINED3DSPR_CONSTINT:
 636         if (pshader)
 637             sprintf(tmpStr, "PI[%u]", reg);
 638         else
 639             sprintf(tmpStr, "VI[%u]", reg);
 640         break;
 641     case WINED3DSPR_CONSTBOOL:
 642         if (pshader)
 643             sprintf(tmpStr, "PB[%u]", reg);
 644         else
 645             sprintf(tmpStr, "VB[%u]", reg);
 646         break;
 647     case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
 648         if (pshader) {
 649             sprintf(tmpStr, "T%u", reg);
 650         } else {
 651             sprintf(tmpStr, "A%u", reg);
 652         }
 653     break;
 654     case WINED3DSPR_LOOP:
 655         sprintf(tmpStr, "aL");
 656     break;
 657     case WINED3DSPR_SAMPLER:
 658         if (pshader)
 659             sprintf(tmpStr, "Psampler%u", reg);
 660         else
 661             sprintf(tmpStr, "Vsampler%u", reg);
 662     break;
 663     case WINED3DSPR_COLOROUT:
 664         if (reg >= GL_LIMITS(buffers)) {
 665             WARN("Write to render target %u, only %d supported\n", reg, 4);
 666         }
 667         if (GL_SUPPORT(ARB_DRAW_BUFFERS)) {
 668             sprintf(tmpStr, "gl_FragData[%u]", reg);
 669         } else { /* On older cards with GLSL support like the GeforceFX there's only one buffer. */
 670             sprintf(tmpStr, "gl_FragColor");
 671         }
 672     break;
 673     case WINED3DSPR_RASTOUT:
 674         sprintf(tmpStr, "%s", hwrastout_reg_names[reg]);
 675     break;
 676     case WINED3DSPR_DEPTHOUT:
 677         sprintf(tmpStr, "gl_FragDepth");
 678     break;
 679     case WINED3DSPR_ATTROUT:
 680         if (reg == 0) {
 681             sprintf(tmpStr, "gl_FrontColor");
 682         } else {
 683             sprintf(tmpStr, "gl_FrontSecondaryColor");
 684         }
 685     break;
 686     case WINED3DSPR_TEXCRDOUT:
 687         /* Vertex shaders >= 3.0: WINED3DSPR_OUTPUT */
 688         if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
 689             sprintf(tmpStr, "OUT%u", reg);
 690         else
 691             sprintf(tmpStr, "gl_TexCoord[%u]", reg);
 692     break;
 693     default:
 694         FIXME("Unhandled register name Type(%d)\n", regtype);
 695         sprintf(tmpStr, "unrecognized_register");
 696     break;
 697     }
 698
 699     strcat(regstr, tmpStr);
 700 }
 701
 702 /* Get the GLSL write mask for the destination register */
 703 static DWORD shader_glsl_get_write_mask(const DWORD param, char *write_mask) {
 704     char *ptr = write_mask;
 705     DWORD mask = param & WINED3DSP_WRITEMASK_ALL;
 706     DWORD reg_type = shader_get_regtype(param);
 707
 708     /* gl_FogFragCoord, gl_PointSize and gl_FragDepth are floats, fixup the write mask. */
 709     if (((reg_type == WINED3DSPR_RASTOUT) && ((param & WINED3DSP_REGNUM_MASK) != 0))
 710             || reg_type == WINED3DSPR_DEPTHOUT) {
 711         mask = WINED3DSP_WRITEMASK_0;
 712     } else {
 713         *ptr++ = '.';
 714         if (param & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
 715         if (param & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
 716         if (param & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
 717         if (param & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
 718     }
 719
 720     *ptr = '\0';
 721
 722     return mask;
 723 }
 724
 725 static size_t shader_glsl_get_write_mask_size(DWORD write_mask) {
 726     size_t size = 0;
 727
 728     if (write_mask & WINED3DSP_WRITEMASK_0) ++size;
 729     if (write_mask & WINED3DSP_WRITEMASK_1) ++size;
 730     if (write_mask & WINED3DSP_WRITEMASK_2) ++size;
 731     if (write_mask & WINED3DSP_WRITEMASK_3) ++size;
 732
 733     return size;
 734 }
 735
 736 static void shader_glsl_get_swizzle(const DWORD param, BOOL fixup, DWORD mask, char *swizzle_str) {
 737     /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
 738      * but addressed as "rgba". To fix this we need to swap the register's x
 739      * and z components. */
 740     DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
 741     const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
 742     char *ptr = swizzle_str;
 743
 744     *ptr++ = '.';
 745     /* swizzle bits fields: wwzzyyxx */
 746     if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = swizzle_chars[swizzle & 0x03];
 747     if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = swizzle_chars[(swizzle >> 2) & 0x03];
 748     if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = swizzle_chars[(swizzle >> 4) & 0x03];
 749     if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = swizzle_chars[(swizzle >> 6) & 0x03];
 750
 751     *ptr = '\0';
 752 }
 753
 754 /* From a given parameter token, generate the corresponding GLSL string.
 755  * Also, return the actual register name and swizzle in case the
 756  * caller needs this information as well. */
 757 static void shader_glsl_add_src_param(SHADER_OPCODE_ARG* arg, const DWORD param,
 758         const DWORD addr_token, DWORD mask, glsl_src_param_t *src_param) {
 759     BOOL is_color = FALSE;
 760     char swizzle_str[6];
 761
 762     src_param->reg_name[0] = '\0';
 763     src_param->param_str[0] = '\0';
 764     swizzle_str[0] = '\0';
 765
 766     shader_glsl_get_register_name(param, addr_token, src_param->reg_name, &is_color, arg);
 767
 768     shader_glsl_get_swizzle(param, is_color, mask, swizzle_str);
 769     shader_glsl_gen_modifier(param, src_param->reg_name, swizzle_str, src_param->param_str);
 770 }
 771
 772 /* From a given parameter token, generate the corresponding GLSL string.
 773  * Also, return the actual register name and swizzle in case the
 774  * caller needs this information as well. */
 775 static DWORD shader_glsl_add_dst_param(SHADER_OPCODE_ARG* arg, const DWORD param,
 776         const DWORD addr_token, glsl_dst_param_t *dst_param) {
 777     BOOL is_color = FALSE;
 778
 779     dst_param->mask_str[0] = '\0';
 780     dst_param->reg_name[0] = '\0';
 781
 782     shader_glsl_get_register_name(param, addr_token, dst_param->reg_name, &is_color, arg);
 783     return shader_glsl_get_write_mask(param, dst_param->mask_str);
 784 }
 785
 786 /* Append the destination part of the instruction to the buffer, return the effective write mask */
 787 static DWORD shader_glsl_append_dst(SHADER_BUFFER *buffer, SHADER_OPCODE_ARG *arg) {
 788     glsl_dst_param_t dst_param;
 789     DWORD mask;
 790     int shift;
 791
 792     shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
 793     mask = shader_glsl_add_dst_param(arg, arg->dst, arg->dst_addr, &dst_param);
 794     shader_addline(buffer, "%s%s = %s(", dst_param.reg_name, dst_param.mask_str, shift_glsl_tab[shift]);
 795
 796     return mask;
 797 }
 798
 799 /** Process GLSL instruction modifiers */
 800 void shader_glsl_add_instruction_modifiers(SHADER_OPCODE_ARG* arg) {
 801
 802     DWORD mask = arg->dst & WINED3DSP_DSTMOD_MASK;
 803
 804     if (arg->opcode->dst_token && mask != 0) {
 805         glsl_dst_param_t dst_param;
 806
 807         shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
 808
 809         if (mask & WINED3DSPDM_SATURATE) {
 810             /* _SAT means to clamp the value of the register to between 0 and 1 */
 811             shader_addline(arg->buffer, "%s%s = clamp(%s%s, 0.0, 1.0);\n", dst_param.reg_name,
 812                     dst_param.mask_str, dst_param.reg_name, dst_param.mask_str);
 813         }
 814         if (mask & WINED3DSPDM_MSAMPCENTROID) {
 815             FIXME("_centroid modifier not handled\n");
 816         }
 817         if (mask & WINED3DSPDM_PARTIALPRECISION) {
 818             /* MSDN says this modifier can be safely ignored, so that's what we'll do. */
 819         }
 820     }
 821 }
 822
 823 static inline const char* shader_get_comp_op(
 824     const DWORD opcode) {
 825
 826     DWORD op = (opcode & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
 827     switch (op) {
 828         case COMPARISON_GT: return ">";
 829         case COMPARISON_EQ: return "==";
 830         case COMPARISON_GE: return ">=";
 831         case COMPARISON_LT: return "<";
 832         case COMPARISON_NE: return "!=";
 833         case COMPARISON_LE: return "<=";
 834         default:
 835             FIXME("Unrecognized comparison value: %u\n", op);
 836             return "(\?\?)";
 837     }
 838 }
 839
 840 static void shader_glsl_get_sample_function(DWORD sampler_type, BOOL projected, glsl_sample_function_t *sample_function) {
 841     /* Note that there's no such thing as a projected cube texture. */
 842     switch(sampler_type) {
 843         case WINED3DSTT_1D:
 844             sample_function->name = projected ? "texture1DProj" : "texture1D";
 845             sample_function->coord_mask = WINED3DSP_WRITEMASK_0;
 846             break;
 847         case WINED3DSTT_2D:
 848             sample_function->name = projected ? "texture2DProj" : "texture2D";
 849             sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1;
 850             break;
 851         case WINED3DSTT_CUBE:
 852             sample_function->name = "textureCube";
 853             sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
 854             break;
 855         case WINED3DSTT_VOLUME:
 856             sample_function->name = projected ? "texture3DProj" : "texture3D";
 857             sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
 858             break;
 859         default:
 860             sample_function->name = "";
 861             FIXME("Unrecognized sampler type: %#x;\n", sampler_type);
 862             break;
 863     }
 864 }
 865
 866
 867 /*****************************************************************************
 868  *
 869  * Begin processing individual instruction opcodes
 870  *
 871  ****************************************************************************/
 872
 873 /* Generate GLSL arithmetic functions (dst = src1 + src2) */
 874 void shader_glsl_arith(SHADER_OPCODE_ARG* arg) {
 875     CONST SHADER_OPCODE* curOpcode = arg->opcode;
 876     SHADER_BUFFER* buffer = arg->buffer;
 877     glsl_src_param_t src0_param;
 878     glsl_src_param_t src1_param;
 879     DWORD write_mask;
 880     char op;
 881
 882     /* Determine the GLSL operator to use based on the opcode */
 883     switch (curOpcode->opcode) {
 884         case WINED3DSIO_MUL: op = '*'; break;
 885         case WINED3DSIO_ADD: op = '+'; break;
 886         case WINED3DSIO_SUB: op = '-'; break;
 887         default:
 888             op = ' ';
 889             FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
 890             break;
 891     }
 892
 893     write_mask = shader_glsl_append_dst(buffer, arg);
 894     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
 895     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
 896     shader_addline(buffer, "%s %c %s);\n", src0_param.param_str, op, src1_param.param_str);
 897 }
 898
 899 /* Process the WINED3DSIO_MOV opcode using GLSL (dst = src) */
 900 void shader_glsl_mov(SHADER_OPCODE_ARG* arg) {
 901     IWineD3DBaseShaderImpl* shader = (IWineD3DBaseShaderImpl*) arg->shader;
 902     SHADER_BUFFER* buffer = arg->buffer;
 903     glsl_src_param_t src0_param;
 904     DWORD write_mask;
 905
 906     write_mask = shader_glsl_append_dst(buffer, arg);
 907     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
 908
 909     /* In vs_1_1 WINED3DSIO_MOV can write to the address register. In later
 910      * shader versions WINED3DSIO_MOVA is used for this. */
 911     if ((WINED3DSHADER_VERSION_MAJOR(shader->baseShader.hex_version) == 1 &&
 912             !shader_is_pshader_version(shader->baseShader.hex_version) &&
 913             shader_get_regtype(arg->dst) == WINED3DSPR_ADDR) ||
 914             arg->opcode->opcode == WINED3DSIO_MOVA) {
 915         /* We need to *round* to the nearest int here. */
 916         size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
 917         if (mask_size > 1) {
 918             shader_addline(buffer, "ivec%d(floor(abs(%s) + vec%d(0.5)) * sign(%s)));\n", mask_size, src0_param.param_str, mask_size, src0_param.param_str);
 919         } else {
 920             shader_addline(buffer, "int(floor(abs(%s) + 0.5) * sign(%s)));\n", src0_param.param_str, src0_param.param_str);
 921         }
 922     } else {
 923         shader_addline(buffer, "%s);\n", src0_param.param_str);
 924     }
 925 }
 926
 927 /* Process the dot product operators DP3 and DP4 in GLSL (dst = dot(src0, src1)) */
 928 void shader_glsl_dot(SHADER_OPCODE_ARG* arg) {
 929     CONST SHADER_OPCODE* curOpcode = arg->opcode;
 930     SHADER_BUFFER* buffer = arg->buffer;
 931     glsl_src_param_t src0_param;
 932     glsl_src_param_t src1_param;
 933     DWORD dst_write_mask, src_write_mask;
 934     size_t dst_size = 0;
 935
 936     dst_write_mask = shader_glsl_append_dst(buffer, arg);
 937     dst_size = shader_glsl_get_write_mask_size(dst_write_mask);
 938
 939     /* dp3 works on vec3, dp4 on vec4 */
 940     if (curOpcode->opcode == WINED3DSIO_DP4) {
 941         src_write_mask = WINED3DSP_WRITEMASK_ALL;
 942     } else {
 943         src_write_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
 944     }
 945
 946     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_write_mask, &src0_param);
 947     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_write_mask, &src1_param);
 948
 949     if (dst_size > 1) {
 950         shader_addline(buffer, "vec%d(dot(%s, %s)));\n", dst_size, src0_param.param_str, src1_param.param_str);
 951     } else {
 952         shader_addline(buffer, "dot(%s, %s));\n", src0_param.param_str, src1_param.param_str);
 953     }
 954 }
 955
 956 /* Note that this instruction has some restrictions. The destination write mask
 957  * can't contain the w component, and the source swizzles have to be .xyzw */
 958 void shader_glsl_cross(SHADER_OPCODE_ARG *arg) {
 959     DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
 960     glsl_src_param_t src0_param;
 961     glsl_src_param_t src1_param;
 962     char dst_mask[6];
 963
 964     shader_glsl_get_write_mask(arg->dst, dst_mask);
 965     shader_glsl_append_dst(arg->buffer, arg);
 966     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
 967     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_mask, &src1_param);
 968     shader_addline(arg->buffer, "cross(%s, %s).%s);\n", src0_param.param_str, src1_param.param_str, dst_mask);
 969 }
 970
 971 /* Map the opcode 1-to-1 to the GL code (arg->dst = instruction(src0, src1, ...) */
 972 void shader_glsl_map2gl(SHADER_OPCODE_ARG* arg) {
 973     CONST SHADER_OPCODE* curOpcode = arg->opcode;
 974     SHADER_BUFFER* buffer = arg->buffer;
 975     glsl_src_param_t src_param;
 976     const char *instruction;
 977     char arguments[256];
 978     DWORD write_mask;
 979     unsigned i;
 980
 981     /* Determine the GLSL function to use based on the opcode */
 982     /* TODO: Possibly make this a table for faster lookups */
 983     switch (curOpcode->opcode) {
 984         case WINED3DSIO_MIN: instruction = "min"; break;
 985         case WINED3DSIO_MAX: instruction = "max"; break;
 986         case WINED3DSIO_RSQ: instruction = "inversesqrt"; break;
 987         case WINED3DSIO_ABS: instruction = "abs"; break;
 988         case WINED3DSIO_FRC: instruction = "fract"; break;
 989         case WINED3DSIO_POW: instruction = "pow"; break;
 990         case WINED3DSIO_NRM: instruction = "normalize"; break;
 991         case WINED3DSIO_LOGP:
 992         case WINED3DSIO_LOG: instruction = "log2"; break;
 993         case WINED3DSIO_EXP: instruction = "exp2"; break;
 994         case WINED3DSIO_SGN: instruction = "sign"; break;
 995         default: instruction = "";
 996             FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
 997             break;
 998     }
 999
1000     write_mask = shader_glsl_append_dst(buffer, arg);
1001
1002     arguments[0] = '\0';
1003     if (curOpcode->num_params > 0) {
1004         shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src_param);
1005         strcat(arguments, src_param.param_str);
1006         for (i = 2; i < curOpcode->num_params; ++i) {
1007             strcat(arguments, ", ");
1008             shader_glsl_add_src_param(arg, arg->src[i-1], arg->src_addr[i-1], write_mask, &src_param);
1009             strcat(arguments, src_param.param_str);
1010         }
1011     }
1012
1013     shader_addline(buffer, "%s(%s));\n", instruction, arguments);
1014 }
1015
1016 /** Process the WINED3DSIO_EXPP instruction in GLSL:
1017  * For shader model 1.x, do the following (and honor the writemask, so use a temporary variable):
1018  *   dst.x = 2^(floor(src))
1019  *   dst.y = src - floor(src)
1020  *   dst.z = 2^src   (partial precision is allowed, but optional)
1021  *   dst.w = 1.0;
1022  * For 2.0 shaders, just do this (honoring writemask and swizzle):
1023  *   dst = 2^src;    (partial precision is allowed, but optional)
1024  */
1025 void shader_glsl_expp(SHADER_OPCODE_ARG* arg) {
1026     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)arg->shader;
1027     glsl_src_param_t src_param;
1028
1029     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src_param);
1030
1031     if (shader->baseShader.hex_version < WINED3DPS_VERSION(2,0)) {
1032         char dst_mask[6];
1033
1034         shader_addline(arg->buffer, "tmp0.x = exp2(floor(%s));\n", src_param.param_str);
1035         shader_addline(arg->buffer, "tmp0.y = %s - floor(%s);\n", src_param.param_str, src_param.param_str);
1036         shader_addline(arg->buffer, "tmp0.z = exp2(%s);\n", src_param.param_str);
1037         shader_addline(arg->buffer, "tmp0.w = 1.0;\n");
1038
1039         shader_glsl_append_dst(arg->buffer, arg);
1040         shader_glsl_get_write_mask(arg->dst, dst_mask);
1041         shader_addline(arg->buffer, "tmp0%s);\n", dst_mask);
1042     } else {
1043         DWORD write_mask;
1044         size_t mask_size;
1045
1046         write_mask = shader_glsl_append_dst(arg->buffer, arg);
1047         mask_size = shader_glsl_get_write_mask_size(write_mask);
1048
1049         if (mask_size > 1) {
1050             shader_addline(arg->buffer, "vec%d(exp2(%s)));\n", mask_size, src_param.param_str);
1051         } else {
1052             shader_addline(arg->buffer, "exp2(%s));\n", src_param.param_str);
1053         }
1054     }
1055 }
1056
1057 /** Process the RCP (reciprocal or inverse) opcode in GLSL (dst = 1 / src) */
1058 void shader_glsl_rcp(SHADER_OPCODE_ARG* arg) {
1059     glsl_src_param_t src_param;
1060     DWORD write_mask;
1061     size_t mask_size;
1062
1063     write_mask = shader_glsl_append_dst(arg->buffer, arg);
1064     mask_size = shader_glsl_get_write_mask_size(write_mask);
1065     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src_param);
1066
1067     if (mask_size > 1) {
1068         shader_addline(arg->buffer, "vec%d(1.0 / %s));\n", mask_size, src_param.param_str);
1069     } else {
1070         shader_addline(arg->buffer, "1.0 / %s);\n", src_param.param_str);
1071     }
1072 }
1073
1074 /** Process signed comparison opcodes in GLSL. */
1075 void shader_glsl_compare(SHADER_OPCODE_ARG* arg) {
1076     glsl_src_param_t src0_param;
1077     glsl_src_param_t src1_param;
1078     DWORD write_mask;
1079     size_t mask_size;
1080
1081     write_mask = shader_glsl_append_dst(arg->buffer, arg);
1082     mask_size = shader_glsl_get_write_mask_size(write_mask);
1083     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
1084     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1085
1086     if (mask_size > 1) {
1087         const char *compare;
1088
1089         switch(arg->opcode->opcode) {
1090             case WINED3DSIO_SLT: compare = "lessThan"; break;
1091             case WINED3DSIO_SGE: compare = "greaterThanEqual"; break;
1092             default: compare = "";
1093                 FIXME("Can't handle opcode %s\n", arg->opcode->name);
1094         }
1095
1096         shader_addline(arg->buffer, "vec%d(%s(%s, %s)));\n", mask_size, compare,
1097                 src0_param.param_str, src1_param.param_str);
1098     } else {
1099         const char *compare;
1100
1101         switch(arg->opcode->opcode) {
1102             case WINED3DSIO_SLT: compare = "<"; break;
1103             case WINED3DSIO_SGE: compare = ">="; break;
1104             default: compare = "";
1105                 FIXME("Can't handle opcode %s\n", arg->opcode->name);
1106         }
1107
1108         shader_addline(arg->buffer, "(%s %s %s) ? 1.0 : 0.0);\n",
1109                 src0_param.param_str, compare, src1_param.param_str);
1110     }
1111 }
1112
1113 /** Process CMP instruction in GLSL (dst = src0 >= 0.0 ? src1 : src2), per channel */
1114 void shader_glsl_cmp(SHADER_OPCODE_ARG* arg) {
1115     glsl_src_param_t src0_param;
1116     glsl_src_param_t src1_param;
1117     glsl_src_param_t src2_param;
1118     DWORD write_mask;
1119     size_t mask_size;
1120
1121     write_mask = shader_glsl_append_dst(arg->buffer, arg);
1122     mask_size = shader_glsl_get_write_mask_size(write_mask);
1123
1124     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
1125     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1126     shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1127
1128     if (mask_size > 1) {
1129         shader_addline(arg->buffer, "mix(%s, %s, vec%d(lessThan(%s, vec%d(0.0)))));\n",
1130                 src1_param.param_str, src2_param.param_str, mask_size, src0_param.param_str, mask_size);
1131     } else {
1132         shader_addline(arg->buffer, "%s >= 0.0 ? %s : %s);\n",
1133                 src0_param.param_str, src1_param.param_str, src2_param.param_str);
1134     }
1135 }
1136
1137 /** Process the CND opcode in GLSL (dst = (src0 < 0.5) ? src1 : src2) */
1138 /* For ps 1.1-1.3, only a single component of src0 is used. For ps 1.4
1139  * the compare is done per component of src0. */
1140 void shader_glsl_cnd(SHADER_OPCODE_ARG* arg) {
1141     IWineD3DBaseShaderImpl* shader = (IWineD3DBaseShaderImpl*) arg->shader;
1142     glsl_src_param_t src0_param;
1143     glsl_src_param_t src1_param;
1144     glsl_src_param_t src2_param;
1145     DWORD write_mask;
1146     size_t mask_size;
1147
1148     write_mask = shader_glsl_append_dst(arg->buffer, arg);
1149
1150     if (shader->baseShader.hex_version < WINED3DPS_VERSION(1, 4)) {
1151         mask_size = 1;
1152         shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1153     } else {
1154         mask_size = shader_glsl_get_write_mask_size(write_mask);
1155         shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
1156     }
1157
1158     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1159     shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1160
1161     if (mask_size > 1) {
1162         shader_addline(arg->buffer, "mix(%s, %s, vec%d(lessThan(%s, vec%d(0.5)))));\n",
1163                 src2_param.param_str, src1_param.param_str, mask_size, src0_param.param_str, mask_size);
1164     } else {
1165         shader_addline(arg->buffer, "%s < 0.5 ? %s : %s);\n",
1166                 src0_param.param_str, src1_param.param_str, src2_param.param_str);
1167     }
1168 }
1169
1170 /** GLSL code generation for WINED3DSIO_MAD: Multiply the first 2 opcodes, then add the last */
1171 void shader_glsl_mad(SHADER_OPCODE_ARG* arg) {
1172     glsl_src_param_t src0_param;
1173     glsl_src_param_t src1_param;
1174     glsl_src_param_t src2_param;
1175     DWORD write_mask;
1176
1177     write_mask = shader_glsl_append_dst(arg->buffer, arg);
1178     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
1179     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1180     shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1181     shader_addline(arg->buffer, "(%s * %s) + %s);\n",
1182             src0_param.param_str, src1_param.param_str, src2_param.param_str);
1183 }
1184
1185 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1186     Vertex shaders to GLSL codes */
1187 void shader_glsl_mnxn(SHADER_OPCODE_ARG* arg) {
1188     int i;
1189     int nComponents = 0;
1190     SHADER_OPCODE_ARG tmpArg;
1191
1192     memset(&tmpArg, 0, sizeof(SHADER_OPCODE_ARG));
1193
1194     /* Set constants for the temporary argument */
1195     tmpArg.shader      = arg->shader;
1196     tmpArg.buffer      = arg->buffer;
1197     tmpArg.src[0]      = arg->src[0];
1198     tmpArg.src_addr[0] = arg->src_addr[0];
1199     tmpArg.src_addr[1] = arg->src_addr[1];
1200     tmpArg.reg_maps = arg->reg_maps;
1201
1202     switch(arg->opcode->opcode) {
1203         case WINED3DSIO_M4x4:
1204             nComponents = 4;
1205             tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
1206             break;
1207         case WINED3DSIO_M4x3:
1208             nComponents = 3;
1209             tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
1210             break;
1211         case WINED3DSIO_M3x4:
1212             nComponents = 4;
1213             tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1214             break;
1215         case WINED3DSIO_M3x3:
1216             nComponents = 3;
1217             tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1218             break;
1219         case WINED3DSIO_M3x2:
1220             nComponents = 2;
1221             tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1222             break;
1223         default:
1224             break;
1225     }
1226
1227     for (i = 0; i < nComponents; i++) {
1228         tmpArg.dst = ((arg->dst) & ~WINED3DSP_WRITEMASK_ALL)|(WINED3DSP_WRITEMASK_0<<i);
1229         tmpArg.src[1]      = arg->src[1]+i;
1230         shader_glsl_dot(&tmpArg);
1231     }
1232 }
1233
1234 /**
1235     The LRP instruction performs a component-wise linear interpolation
1236     between the second and third operands using the first operand as the
1237     blend factor.  Equation:  (dst = src2 + src0 * (src1 - src2))
1238     This is equivalent to mix(src2, src1, src0);
1239 */
1240 void shader_glsl_lrp(SHADER_OPCODE_ARG* arg) {
1241     glsl_src_param_t src0_param;
1242     glsl_src_param_t src1_param;
1243     glsl_src_param_t src2_param;
1244     DWORD write_mask;
1245
1246     write_mask = shader_glsl_append_dst(arg->buffer, arg);
1247
1248     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
1249     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1250     shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1251
1252     shader_addline(arg->buffer, "mix(%s, %s, %s));\n",
1253             src2_param.param_str, src1_param.param_str, src0_param.param_str);
1254 }
1255
1256 /** Process the WINED3DSIO_LIT instruction in GLSL:
1257  * dst.x = dst.w = 1.0
1258  * dst.y = (src0.x > 0) ? src0.x
1259  * dst.z = (src0.x > 0) ? ((src0.y > 0) ? pow(src0.y, src.w) : 0) : 0
1260  *                                        where src.w is clamped at +- 128
1261  */
1262 void shader_glsl_lit(SHADER_OPCODE_ARG* arg) {
1263     glsl_src_param_t src0_param;
1264     glsl_src_param_t src1_param;
1265     glsl_src_param_t src3_param;
1266     char dst_mask[6];
1267
1268     shader_glsl_append_dst(arg->buffer, arg);
1269     shader_glsl_get_write_mask(arg->dst, dst_mask);
1270
1271     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1272     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_1, &src1_param);
1273     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, &src3_param);
1274
1275     shader_addline(arg->buffer, "vec4(1.0, (%s > 0.0 ? %s : 0.0), (%s > 0.0 ? ((%s > 0.0) ? pow(%s, clamp(%s, -128.0, 128.0)) : 0.0) : 0.0), 1.0)%s);\n",
1276         src0_param.param_str, src0_param.param_str, src0_param.param_str, src1_param.param_str, src1_param.param_str, src3_param.param_str, dst_mask);
1277 }
1278
1279 /** Process the WINED3DSIO_DST instruction in GLSL:
1280  * dst.x = 1.0
1281  * dst.y = src0.x * src0.y
1282  * dst.z = src0.z
1283  * dst.w = src1.w
1284  */
1285 void shader_glsl_dst(SHADER_OPCODE_ARG* arg) {
1286     glsl_src_param_t src0y_param;
1287     glsl_src_param_t src0z_param;
1288     glsl_src_param_t src1y_param;
1289     glsl_src_param_t src1w_param;
1290     char dst_mask[6];
1291
1292     shader_glsl_append_dst(arg->buffer, arg);
1293     shader_glsl_get_write_mask(arg->dst, dst_mask);
1294
1295     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_1, &src0y_param);
1296     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_2, &src0z_param);
1297     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_1, &src1y_param);
1298     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_3, &src1w_param);
1299
1300     shader_addline(arg->buffer, "vec4(1.0, %s * %s, %s, %s))%s;\n",
1301             src0y_param.param_str, src1y_param.param_str, src0z_param.param_str, src1w_param.param_str, dst_mask);
1302 }
1303
1304 /** Process the WINED3DSIO_SINCOS instruction in GLSL:
1305  * VS 2.0 requires that specific cosine and sine constants be passed to this instruction so the hardware
1306  * can handle it.  But, these functions are built-in for GLSL, so we can just ignore the last 2 params.
1307  *
1308  * dst.x = cos(src0.?)
1309  * dst.y = sin(src0.?)
1310  * dst.z = dst.z
1311  * dst.w = dst.w
1312  */
1313 void shader_glsl_sincos(SHADER_OPCODE_ARG* arg) {
1314     glsl_src_param_t src0_param;
1315     DWORD write_mask;
1316
1317     write_mask = shader_glsl_append_dst(arg->buffer, arg);
1318     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1319
1320     switch (write_mask) {
1321         case WINED3DSP_WRITEMASK_0:
1322             shader_addline(arg->buffer, "cos(%s));\n", src0_param.param_str);
1323             break;
1324
1325         case WINED3DSP_WRITEMASK_1:
1326             shader_addline(arg->buffer, "sin(%s));\n", src0_param.param_str);
1327             break;
1328
1329         case (WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1):
1330             shader_addline(arg->buffer, "vec2(cos(%s), sin(%s)));\n", src0_param.param_str, src0_param.param_str);
1331             break;
1332
1333         default:
1334             ERR("Write mask should be .x, .y or .xy\n");
1335             break;
1336     }
1337 }
1338
1339 /** Process the WINED3DSIO_LOOP instruction in GLSL:
1340  * Start a for() loop where src1.y is the initial value of aL,
1341  *  increment aL by src1.z for a total of src1.x iterations.
1342  *  Need to use a temporary variable for this operation.
1343  */
1344 /* FIXME: I don't think nested loops will work correctly this way. */
1345 void shader_glsl_loop(SHADER_OPCODE_ARG* arg) {
1346     glsl_src_param_t src1_param;
1347
1348     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_ALL, &src1_param);
1349
1350     shader_addline(arg->buffer, "for (tmpInt = 0, aL = %s.y; tmpInt < %s.x; tmpInt++, aL += %s.z) {\n",
1351             src1_param.reg_name, src1_param.reg_name, src1_param.reg_name);
1352 }
1353
1354 void shader_glsl_end(SHADER_OPCODE_ARG* arg) {
1355     shader_addline(arg->buffer, "}\n");
1356 }
1357
1358 void shader_glsl_rep(SHADER_OPCODE_ARG* arg) {
1359     glsl_src_param_t src0_param;
1360
1361     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1362     shader_addline(arg->buffer, "for (tmpInt = 0; tmpInt < %s; tmpInt++) {\n", src0_param.param_str);
1363 }
1364
1365 void shader_glsl_if(SHADER_OPCODE_ARG* arg) {
1366     glsl_src_param_t src0_param;
1367
1368     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1369     shader_addline(arg->buffer, "if (%s) {\n", src0_param.param_str);
1370 }
1371
1372 void shader_glsl_ifc(SHADER_OPCODE_ARG* arg) {
1373     glsl_src_param_t src0_param;
1374     glsl_src_param_t src1_param;
1375
1376     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1377     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, &src1_param);
1378
1379     shader_addline(arg->buffer, "if (%s %s %s) {\n",
1380             src0_param.param_str, shader_get_comp_op(arg->opcode_token), src1_param.param_str);
1381 }
1382
1383 void shader_glsl_else(SHADER_OPCODE_ARG* arg) {
1384     shader_addline(arg->buffer, "} else {\n");
1385 }
1386
1387 void shader_glsl_break(SHADER_OPCODE_ARG* arg) {
1388     shader_addline(arg->buffer, "break;\n");
1389 }
1390
1391 /* FIXME: According to MSDN the compare is done per component. */
1392 void shader_glsl_breakc(SHADER_OPCODE_ARG* arg) {
1393     glsl_src_param_t src0_param;
1394     glsl_src_param_t src1_param;
1395
1396     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1397     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, &src1_param);
1398
1399     shader_addline(arg->buffer, "if (%s %s %s) break;\n",
1400             src0_param.param_str, shader_get_comp_op(arg->opcode_token), src1_param.param_str);
1401 }
1402
1403 void shader_glsl_label(SHADER_OPCODE_ARG* arg) {
1404
1405     DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1406     shader_addline(arg->buffer, "}\n");
1407     shader_addline(arg->buffer, "void subroutine%lu () {\n",  snum);
1408 }
1409
1410 void shader_glsl_call(SHADER_OPCODE_ARG* arg) {
1411     DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1412     shader_addline(arg->buffer, "subroutine%lu();\n", snum);
1413 }
1414
1415 void shader_glsl_callnz(SHADER_OPCODE_ARG* arg) {
1416     glsl_src_param_t src1_param;
1417
1418     DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1419     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, &src1_param);
1420     shader_addline(arg->buffer, "if (%s) subroutine%lu();\n", src1_param.param_str, snum);
1421 }
1422
1423 /*********************************************
1424  * Pixel Shader Specific Code begins here
1425  ********************************************/
1426 void pshader_glsl_tex(SHADER_OPCODE_ARG* arg) {
1427     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1428     DWORD hex_version = This->baseShader.hex_version;
1429     char dst_swizzle[6];
1430     glsl_sample_function_t sample_function;
1431     DWORD sampler_type;
1432     DWORD sampler_idx;
1433     BOOL projected;
1434     DWORD mask = 0;
1435
1436     /* All versions have a destination register */
1437     shader_glsl_append_dst(arg->buffer, arg);
1438
1439     /* 1.0-1.4: Use destination register as sampler source.
1440      * 2.0+: Use provided sampler source. */
1441     if (hex_version < WINED3DPS_VERSION(1,4)) {
1442         IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1443         DWORD flags;
1444
1445         sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1446         flags = deviceImpl->stateBlock->textureState[sampler_idx][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1447
1448         if (flags & WINED3DTTFF_PROJECTED) {
1449             projected = TRUE;
1450             switch (flags & ~WINED3DTTFF_PROJECTED) {
1451                 case WINED3DTTFF_COUNT1: FIXME("WINED3DTTFF_PROJECTED with WINED3DTTFF_COUNT1?\n"); break;
1452                 case WINED3DTTFF_COUNT2: mask = WINED3DSP_WRITEMASK_1; break;
1453                 case WINED3DTTFF_COUNT3: mask = WINED3DSP_WRITEMASK_2; break;
1454                 case WINED3DTTFF_COUNT4:
1455                 case WINED3DTTFF_DISABLE: mask = WINED3DSP_WRITEMASK_3; break;
1456             }
1457         } else {
1458             projected = FALSE;
1459         }
1460     } else if (hex_version < WINED3DPS_VERSION(2,0)) {
1461         DWORD src_mod = arg->src[0] & WINED3DSP_SRCMOD_MASK;
1462         sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1463
1464         if (src_mod == WINED3DSPSM_DZ) {
1465             projected = TRUE;
1466             mask = WINED3DSP_WRITEMASK_2;
1467         } else if (src_mod == WINED3DSPSM_DW) {
1468             projected = TRUE;
1469             mask = WINED3DSP_WRITEMASK_3;
1470         } else {
1471             projected = FALSE;
1472         }
1473     } else {
1474         sampler_idx = arg->src[1] & WINED3DSP_REGNUM_MASK;
1475         if(arg->opcode_token & WINED3DSI_TEXLD_PROJECT) {
1476                 /* ps 2.0 texldp instruction always divides by the fourth component. */
1477                 projected = TRUE;
1478                 mask = WINED3DSP_WRITEMASK_3;
1479         } else {
1480             projected = FALSE;
1481         }
1482     }
1483
1484     sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1485     shader_glsl_get_sample_function(sampler_type, projected, &sample_function);
1486     mask |= sample_function.coord_mask;
1487
1488     if (hex_version < WINED3DPS_VERSION(2,0)) {
1489         shader_glsl_get_write_mask(arg->dst, dst_swizzle);
1490     } else {
1491         shader_glsl_get_swizzle(arg->src[1], FALSE, arg->dst, dst_swizzle);
1492     }
1493
1494     /* 1.0-1.3: Use destination register as coordinate source.
1495        1.4+: Use provided coordinate source register. */
1496     if (hex_version < WINED3DPS_VERSION(1,4)) {
1497         char coord_mask[6];
1498         shader_glsl_get_write_mask(mask, coord_mask);
1499         shader_addline(arg->buffer, "%s(Psampler%u, T%u%s)%s);\n",
1500                 sample_function.name, sampler_idx, sampler_idx, coord_mask, dst_swizzle);
1501     } else {
1502         glsl_src_param_t coord_param;
1503         shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], mask, &coord_param);
1504         shader_addline(arg->buffer, "%s(Psampler%u, %s)%s);\n",
1505                 sample_function.name, sampler_idx, coord_param.param_str, dst_swizzle);
1506     }
1507 }
1508
1509 void pshader_glsl_texcoord(SHADER_OPCODE_ARG* arg) {
1510
1511     /* FIXME: Make this work for more than just 2D textures */
1512
1513     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1514     SHADER_BUFFER* buffer = arg->buffer;
1515     DWORD hex_version = This->baseShader.hex_version;
1516     DWORD write_mask;
1517     char dst_mask[6];
1518
1519     write_mask = shader_glsl_append_dst(arg->buffer, arg);
1520     shader_glsl_get_write_mask(write_mask, dst_mask);
1521
1522     if (hex_version != WINED3DPS_VERSION(1,4)) {
1523         DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1524         shader_addline(buffer, "clamp(gl_TexCoord[%u], 0.0, 1.0)%s);\n", reg, dst_mask);
1525     } else {
1526         DWORD reg = arg->src[0] & WINED3DSP_REGNUM_MASK;
1527         DWORD src_mod = arg->src[0] & WINED3DSP_SRCMOD_MASK;
1528         char dst_swizzle[6];
1529
1530         shader_glsl_get_swizzle(arg->src[0], FALSE, write_mask, dst_swizzle);
1531
1532         if (src_mod == WINED3DSPSM_DZ) {
1533             glsl_src_param_t div_param;
1534             size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
1535             shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_2, &div_param);
1536
1537             if (mask_size > 1) {
1538                 shader_addline(buffer, "gl_TexCoord[%u]%s / vec%d(%s));\n", reg, dst_swizzle, mask_size, div_param.param_str);
1539             } else {
1540                 shader_addline(buffer, "gl_TexCoord[%u]%s / %s);\n", reg, dst_swizzle, div_param.param_str);
1541             }
1542         } else if (src_mod == WINED3DSPSM_DW) {
1543             glsl_src_param_t div_param;
1544             size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
1545             shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, &div_param);
1546
1547             if (mask_size > 1) {
1548                 shader_addline(buffer, "gl_TexCoord[%u]%s / vec%d(%s));\n", reg, dst_swizzle, mask_size, div_param.param_str);
1549             } else {
1550                 shader_addline(buffer, "gl_TexCoord[%u]%s / %s);\n", reg, dst_swizzle, div_param.param_str);
1551             }
1552         } else {
1553             shader_addline(buffer, "gl_TexCoord[%u]%s);\n", reg, dst_swizzle);
1554         }
1555     }
1556 }
1557
1558 /** Process the WINED3DSIO_TEXDP3TEX instruction in GLSL:
1559  * Take a 3-component dot product of the TexCoord[dstreg] and src,
1560  * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1561 void pshader_glsl_texdp3tex(SHADER_OPCODE_ARG* arg) {
1562     glsl_src_param_t src0_param;
1563     char dst_mask[6];
1564     DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1565     DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1566
1567     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1568
1569     shader_glsl_append_dst(arg->buffer, arg);
1570     shader_glsl_get_write_mask(arg->dst, dst_mask);
1571     shader_addline(arg->buffer, "texture1D(Psampler%u, dot(gl_TexCoord[%u].xyz, %s))%s);\n",
1572             sampler_idx, sampler_idx, src0_param.param_str, dst_mask);
1573 }
1574
1575 /** Process the WINED3DSIO_TEXDP3 instruction in GLSL:
1576  * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1577 void pshader_glsl_texdp3(SHADER_OPCODE_ARG* arg) {
1578     glsl_src_param_t src0_param;
1579     DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1580     DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1581     DWORD dst_mask;
1582     size_t mask_size;
1583
1584     dst_mask = shader_glsl_append_dst(arg->buffer, arg);
1585     mask_size = shader_glsl_get_write_mask_size(dst_mask);
1586     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1587
1588     if (mask_size > 1) {
1589         shader_addline(arg->buffer, "vec%d(dot(T%u.xyz, %s)));\n", mask_size, dstreg, src0_param.param_str);
1590     } else {
1591         shader_addline(arg->buffer, "dot(T%u.xyz, %s));\n", dstreg, src0_param.param_str);
1592     }
1593 }
1594
1595 /** Process the WINED3DSIO_TEXDEPTH instruction in GLSL:
1596  * Calculate the depth as dst.x / dst.y   */
1597 void pshader_glsl_texdepth(SHADER_OPCODE_ARG* arg) {
1598     glsl_dst_param_t dst_param;
1599
1600     shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
1601
1602     shader_addline(arg->buffer, "gl_FragDepth = %s.x / %s.y;\n", dst_param.reg_name, dst_param.reg_name);
1603 }
1604
1605 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in GLSL:
1606  * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1607  * Calculate tmp0.y = TexCoord[dstreg] . src.xyz;  (tmp0.x has already been calculated)
1608  * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1609  */
1610 void pshader_glsl_texm3x2depth(SHADER_OPCODE_ARG* arg) {
1611     DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1612     DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1613     glsl_src_param_t src0_param;
1614
1615     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1616
1617     shader_addline(arg->buffer, "tmp0.y = dot(T%u.xyz, %s);\n", dstreg, src0_param.param_str);
1618     shader_addline(arg->buffer, "gl_FragDepth = (tmp0.y == 0.0) ? 1.0 : clamp(tmp0.x / tmp0.y, 0.0, 1.0);\n");
1619 }
1620
1621 /** Process the WINED3DSIO_TEXM3X2PAD instruction in GLSL
1622  * Calculate the 1st of a 2-row matrix multiplication. */
1623 void pshader_glsl_texm3x2pad(SHADER_OPCODE_ARG* arg) {
1624     DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1625     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1626     SHADER_BUFFER* buffer = arg->buffer;
1627     glsl_src_param_t src0_param;
1628
1629     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1630     shader_addline(buffer, "tmp0.x = dot(T%u.xyz, %s);\n", reg, src0_param.param_str);
1631 }
1632
1633 /** Process the WINED3DSIO_TEXM3X3PAD instruction in GLSL
1634  * Calculate the 1st or 2nd row of a 3-row matrix multiplication. */
1635 void pshader_glsl_texm3x3pad(SHADER_OPCODE_ARG* arg) {
1636
1637     IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1638     DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1639     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1640     SHADER_BUFFER* buffer = arg->buffer;
1641     SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1642     glsl_src_param_t src0_param;
1643
1644     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1645     shader_addline(buffer, "tmp0.%c = dot(T%u.xyz, %s);\n", 'x' + current_state->current_row, reg, src0_param.param_str);
1646     current_state->texcoord_w[current_state->current_row++] = reg;
1647 }
1648
1649 void pshader_glsl_texm3x2tex(SHADER_OPCODE_ARG* arg) {
1650     DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1651     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1652     SHADER_BUFFER* buffer = arg->buffer;
1653     glsl_src_param_t src0_param;
1654     char dst_mask[6];
1655
1656     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1657     shader_addline(buffer, "tmp0.y = dot(T%u.xyz, %s);\n", reg, src0_param.param_str);
1658
1659     shader_glsl_append_dst(buffer, arg);
1660     shader_glsl_get_write_mask(arg->dst, dst_mask);
1661
1662     /* Sample the texture using the calculated coordinates */
1663     shader_addline(buffer, "texture2D(Psampler%u, tmp0.xy)%s);\n", reg, dst_mask);
1664 }
1665
1666 /** Process the WINED3DSIO_TEXM3X3TEX instruction in GLSL
1667  * Perform the 3rd row of a 3x3 matrix multiply, then sample the texture using the calculated coordinates */
1668 void pshader_glsl_texm3x3tex(SHADER_OPCODE_ARG* arg) {
1669     DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1670     glsl_src_param_t src0_param;
1671     char dst_mask[6];
1672     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1673     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1674     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1675     DWORD sampler_type = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1676     glsl_sample_function_t sample_function;
1677
1678     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1679     shader_addline(arg->buffer, "tmp0.z = dot(T%u.xyz, %s);\n", reg, src0_param.param_str);
1680
1681     shader_glsl_append_dst(arg->buffer, arg);
1682     shader_glsl_get_write_mask(arg->dst, dst_mask);
1683     shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1684
1685     /* Sample the texture using the calculated coordinates */
1686     shader_addline(arg->buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1687
1688     current_state->current_row = 0;
1689 }
1690
1691 /** Process the WINED3DSIO_TEXM3X3 instruction in GLSL
1692  * Perform the 3rd row of a 3x3 matrix multiply */
1693 void pshader_glsl_texm3x3(SHADER_OPCODE_ARG* arg) {
1694     DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1695     glsl_src_param_t src0_param;
1696     char dst_mask[6];
1697     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1698     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1699     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1700
1701     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1702
1703     shader_glsl_append_dst(arg->buffer, arg);
1704     shader_glsl_get_write_mask(arg->dst, dst_mask);
1705     shader_addline(arg->buffer, "vec4(tmp.xy, dot(T%u.xyz, %s), 1.0)%s);\n", reg, src0_param.param_str, dst_mask);
1706
1707     current_state->current_row = 0;
1708 }
1709
1710 /** Process the WINED3DSIO_TEXM3X3SPEC instruction in GLSL
1711  * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1712 void pshader_glsl_texm3x3spec(SHADER_OPCODE_ARG* arg) {
1713
1714     IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1715     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1716     glsl_src_param_t src0_param;
1717     glsl_src_param_t src1_param;
1718     char dst_mask[6];
1719     SHADER_BUFFER* buffer = arg->buffer;
1720     SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1721     DWORD stype = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1722     DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1723     glsl_sample_function_t sample_function;
1724
1725     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1726     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_mask, &src1_param);
1727
1728     /* Perform the last matrix multiply operation */
1729     shader_addline(buffer, "tmp0.z = dot(T%u.xyz, %s);\n", reg, src0_param.param_str);
1730
1731     /* Calculate reflection vector, 2*(tmp0.src1)*tmp0-src1
1732      * This is equivalent to reflect(-src1, tmp0); */
1733     shader_addline(buffer, "tmp0.xyz = reflect(-(%s), tmp0.xyz);\n", src1_param.param_str);
1734
1735     shader_glsl_append_dst(buffer, arg);
1736     shader_glsl_get_write_mask(arg->dst, dst_mask);
1737     shader_glsl_get_sample_function(stype, FALSE, &sample_function);
1738
1739     /* Sample the texture */
1740     shader_addline(buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1741
1742     current_state->current_row = 0;
1743 }
1744
1745 /** Process the WINED3DSIO_TEXM3X3VSPEC instruction in GLSL
1746  * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1747 void pshader_glsl_texm3x3vspec(SHADER_OPCODE_ARG* arg) {
1748
1749     IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1750     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1751     SHADER_BUFFER* buffer = arg->buffer;
1752     SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1753     glsl_src_param_t src0_param;
1754     char dst_mask[6];
1755     DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1756     DWORD sampler_type = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1757     glsl_sample_function_t sample_function;
1758
1759     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1760
1761     /* Perform the last matrix multiply operation */
1762     shader_addline(buffer, "tmp0.z = dot(vec3(T%u), vec3(%s));\n", reg, src0_param.param_str);
1763
1764     /* Construct the eye-ray vector from w coordinates */
1765     shader_addline(buffer, "tmp1.xyz = normalize(vec3(gl_TexCoord[%u].w, gl_TexCoord[%u].w, gl_TexCoord[%u].w));\n",
1766             current_state->texcoord_w[0], current_state->texcoord_w[1], reg);
1767
1768     /* Calculate reflection vector (Assume normal is normalized): RF = 2*(tmp0.tmp1)*tmp0-tmp1
1769      * This is equivalent to reflect(-tmp1, tmp0); */
1770     shader_addline(buffer, "tmp0.xyz = reflect(-tmp1.xyz, tmp0.xyz);\n");
1771
1772     shader_glsl_append_dst(buffer, arg);
1773     shader_glsl_get_write_mask(arg->dst, dst_mask);
1774     shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1775
1776     /* Sample the texture using the calculated coordinates */
1777     shader_addline(buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1778
1779     current_state->current_row = 0;
1780 }
1781
1782 /** Process the WINED3DSIO_TEXBEM instruction in GLSL.
1783  * Apply a fake bump map transform.
1784  * texbem is pshader <= 1.3 only, this saves a few version checks
1785  */
1786 void pshader_glsl_texbem(SHADER_OPCODE_ARG* arg) {
1787     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1788     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1789     char dst_swizzle[6];
1790     glsl_sample_function_t sample_function;
1791     glsl_src_param_t coord_param;
1792     DWORD sampler_type;
1793     DWORD sampler_idx;
1794     DWORD mask;
1795     DWORD flags;
1796     char coord_mask[6];
1797
1798     sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1799     flags = deviceImpl->stateBlock->textureState[sampler_idx][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1800
1801     sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1802     shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1803     mask = sample_function.coord_mask;
1804
1805     shader_glsl_get_write_mask(arg->dst, dst_swizzle);
1806
1807     shader_glsl_get_write_mask(mask, coord_mask);
1808
1809     /* with projective textures, texbem only divides the static texture coord, not the displacement,
1810          * so we can't let the GL handle this.
1811          */
1812     if (flags & WINED3DTTFF_PROJECTED) {
1813         DWORD div_mask=0;
1814         char coord_div_mask[3];
1815         switch (flags & ~WINED3DTTFF_PROJECTED) {
1816             case WINED3DTTFF_COUNT1: FIXME("WINED3DTTFF_PROJECTED with WINED3DTTFF_COUNT1?\n"); break;
1817             case WINED3DTTFF_COUNT2: div_mask = WINED3DSP_WRITEMASK_1; break;
1818             case WINED3DTTFF_COUNT3: div_mask = WINED3DSP_WRITEMASK_2; break;
1819             case WINED3DTTFF_COUNT4:
1820             case WINED3DTTFF_DISABLE: div_mask = WINED3DSP_WRITEMASK_3; break;
1821         }
1822         shader_glsl_get_write_mask(div_mask, coord_div_mask);
1823         shader_addline(arg->buffer, "T%u%s /= T%u%s;\n", sampler_idx, coord_mask, sampler_idx, coord_div_mask);
1824     }
1825
1826     shader_glsl_append_dst(arg->buffer, arg);
1827     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0|WINED3DSP_WRITEMASK_1, &coord_param);
1828     shader_addline(arg->buffer, "%s(Psampler%u, T%u%s + vec4(bumpenvmat * %s, 0.0, 0.0)%s )%s);\n",
1829                    sample_function.name, sampler_idx, sampler_idx, coord_mask, coord_param.param_str, coord_mask, dst_swizzle);
1830 }
1831
1832 void pshader_glsl_bem(SHADER_OPCODE_ARG* arg) {
1833     glsl_src_param_t src0_param, src1_param;
1834
1835     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0|WINED3DSP_WRITEMASK_1, &src0_param);
1836     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0|WINED3DSP_WRITEMASK_1, &src1_param);
1837
1838     shader_glsl_append_dst(arg->buffer, arg);
1839     shader_addline(arg->buffer, "%s + bumpenvmat * %s);\n",
1840                    src0_param.param_str, src1_param.param_str);
1841 }
1842
1843 /** Process the WINED3DSIO_TEXREG2AR instruction in GLSL
1844  * Sample 2D texture at dst using the alpha & red (wx) components of src as texture coordinates */
1845 void pshader_glsl_texreg2ar(SHADER_OPCODE_ARG* arg) {
1846     glsl_src_param_t src0_param;
1847     DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1848     char dst_mask[6];
1849
1850     shader_glsl_append_dst(arg->buffer, arg);
1851     shader_glsl_get_write_mask(arg->dst, dst_mask);
1852     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_ALL, &src0_param);
1853
1854     shader_addline(arg->buffer, "texture2D(Psampler%u, %s.wx)%s);\n", sampler_idx, src0_param.reg_name, dst_mask);
1855 }
1856
1857 /** Process the WINED3DSIO_TEXREG2GB instruction in GLSL
1858  * Sample 2D texture at dst using the green & blue (yz) components of src as texture coordinates */
1859 void pshader_glsl_texreg2gb(SHADER_OPCODE_ARG* arg) {
1860     glsl_src_param_t src0_param;
1861     DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1862     char dst_mask[6];
1863
1864     shader_glsl_append_dst(arg->buffer, arg);
1865     shader_glsl_get_write_mask(arg->dst, dst_mask);
1866     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_ALL, &src0_param);
1867
1868     shader_addline(arg->buffer, "texture2D(Psampler%u, %s.yz)%s);\n", sampler_idx, src0_param.reg_name, dst_mask);
1869 }
1870
1871 /** Process the WINED3DSIO_TEXREG2RGB instruction in GLSL
1872  * Sample texture at dst using the rgb (xyz) components of src as texture coordinates */
1873 void pshader_glsl_texreg2rgb(SHADER_OPCODE_ARG* arg) {
1874     glsl_src_param_t src0_param;
1875     char dst_mask[6];
1876     DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1877     DWORD sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1878     glsl_sample_function_t sample_function;
1879
1880     shader_glsl_append_dst(arg->buffer, arg);
1881     shader_glsl_get_write_mask(arg->dst, dst_mask);
1882     shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1883     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], sample_function.coord_mask, &src0_param);
1884
1885     shader_addline(arg->buffer, "%s(Psampler%u, %s)%s);\n", sample_function.name, sampler_idx, src0_param.param_str, dst_mask);
1886 }
1887
1888 /** Process the WINED3DSIO_TEXKILL instruction in GLSL.
1889  * If any of the first 3 components are < 0, discard this pixel */
1890 void pshader_glsl_texkill(SHADER_OPCODE_ARG* arg) {
1891     glsl_dst_param_t dst_param;
1892
1893     shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
1894     shader_addline(arg->buffer, "if (any(lessThan(%s.xyz, vec3(0.0)))) discard;\n", dst_param.reg_name);
1895 }
1896
1897 /** Process the WINED3DSIO_DP2ADD instruction in GLSL.
1898  * dst = dot2(src0, src1) + src2 */
1899 void pshader_glsl_dp2add(SHADER_OPCODE_ARG* arg) {
1900     glsl_src_param_t src0_param;
1901     glsl_src_param_t src1_param;
1902     glsl_src_param_t src2_param;
1903     DWORD write_mask;
1904     size_t mask_size;
1905
1906     write_mask = shader_glsl_append_dst(arg->buffer, arg);
1907     mask_size = shader_glsl_get_write_mask_size(write_mask);
1908
1909     shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1, &src0_param);
1910     shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1, &src1_param);
1911     shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], WINED3DSP_WRITEMASK_0, &src2_param);
1912
1913     shader_addline(arg->buffer, "dot(%s, %s) + %s);\n", src0_param.param_str, src1_param.param_str, src2_param.param_str);
1914 }
1915
1916 void pshader_glsl_input_pack(
1917    SHADER_BUFFER* buffer,
1918    semantic* semantics_in) {
1919
1920    unsigned int i;
1921
1922    for (i = 0; i < MAX_REG_INPUT; i++) {
1923
1924        DWORD usage_token = semantics_in[i].usage;
1925        DWORD register_token = semantics_in[i].reg;
1926        DWORD usage, usage_idx;
1927        char reg_mask[6];
1928
1929        /* Uninitialized */
1930        if (!usage_token) continue;
1931        usage = (usage_token & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
1932        usage_idx = (usage_token & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
1933        shader_glsl_get_write_mask(register_token, reg_mask);
1934
1935        switch(usage) {
1936
1937            case WINED3DDECLUSAGE_COLOR:
1938                if (usage_idx == 0)
1939                    shader_addline(buffer, "IN%u%s = vec4(gl_Color)%s;\n",
1940                        i, reg_mask, reg_mask);
1941                else if (usage_idx == 1)
1942                    shader_addline(buffer, "IN%u%s = vec4(gl_SecondaryColor)%s;\n",
1943                        i, reg_mask, reg_mask);
1944                else
1945                    shader_addline(buffer, "IN%u%s = vec4(unsupported_color_input)%s;\n",
1946                        i, reg_mask, reg_mask);
1947                break;
1948
1949            case WINED3DDECLUSAGE_TEXCOORD:
1950                shader_addline(buffer, "IN%u%s = vec4(gl_TexCoord[%u])%s;\n",
1951                    i, reg_mask, usage_idx, reg_mask );
1952                break;
1953
1954            case WINED3DDECLUSAGE_FOG:
1955                shader_addline(buffer, "IN%u%s = vec4(gl_FogFragCoord)%s;\n",
1956                    i, reg_mask, reg_mask);
1957                break;
1958
1959            default:
1960                shader_addline(buffer, "IN%u%s = vec4(unsupported_input)%s;\n",
1961                    i, reg_mask, reg_mask);
1962         }
1963     }
1964 }
1965
1966 /*********************************************
1967  * Vertex Shader Specific Code begins here
1968  ********************************************/
1969
1970 void vshader_glsl_output_unpack(
1971    SHADER_BUFFER* buffer,
1972    semantic* semantics_out) {
1973
1974    unsigned int i;
1975
1976    for (i = 0; i < MAX_REG_OUTPUT; i++) {
1977
1978        DWORD usage_token = semantics_out[i].usage;
1979        DWORD register_token = semantics_out[i].reg;
1980        DWORD usage, usage_idx;
1981        char reg_mask[6];
1982
1983        /* Uninitialized */
1984        if (!usage_token) continue;
1985
1986        usage = (usage_token & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
1987        usage_idx = (usage_token & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
1988        shader_glsl_get_write_mask(register_token, reg_mask);
1989
1990        switch(usage) {
1991
1992            case WINED3DDECLUSAGE_COLOR:
1993                if (usage_idx == 0)
1994                    shader_addline(buffer, "gl_FrontColor%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1995                else if (usage_idx == 1)
1996                    shader_addline(buffer, "gl_FrontSecondaryColor%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1997                else
1998                    shader_addline(buffer, "unsupported_color_output%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1999                break;
2000
2001            case WINED3DDECLUSAGE_POSITION:
2002                shader_addline(buffer, "gl_Position%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2003                break;
2004
2005            case WINED3DDECLUSAGE_TEXCOORD:
2006                shader_addline(buffer, "gl_TexCoord[%u]%s = OUT%u%s;\n",
2007                    usage_idx, reg_mask, i, reg_mask);
2008                break;
2009
2010            case WINED3DDECLUSAGE_PSIZE:
2011                shader_addline(buffer, "gl_PointSize = OUT%u.x;\n", i);
2012                break;
2013
2014            case WINED3DDECLUSAGE_FOG:
2015                shader_addline(buffer, "gl_FogFragCoord = OUT%u%s;\n", i, reg_mask);
2016                break;
2017
2018            default:
2019                shader_addline(buffer, "unsupported_output%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2020        }
2021     }
2022 }
2023
2024 static void add_glsl_program_entry(IWineD3DDeviceImpl *device, struct glsl_shader_prog_link *entry) {
2025     glsl_program_key_t *key;
2026
2027     key = HeapAlloc(GetProcessHeap(), 0, sizeof(glsl_program_key_t));
2028     key->vshader = entry->vshader;
2029     key->pshader = entry->pshader;
2030
2031     hash_table_put(device->glsl_program_lookup, key, entry);
2032 }
2033
2034 static struct glsl_shader_prog_link *get_glsl_program_entry(IWineD3DDeviceImpl *device,
2035         GLhandleARB vshader, GLhandleARB pshader) {
2036     glsl_program_key_t key;
2037
2038     key.vshader = vshader;
2039     key.pshader = pshader;
2040
2041     return (struct glsl_shader_prog_link *)hash_table_get(device->glsl_program_lookup, &key);
2042 }
2043
2044 void delete_glsl_program_entry(IWineD3DDevice *iface, struct glsl_shader_prog_link *entry) {
2045     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2046     WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2047     glsl_program_key_t *key;
2048
2049     key = HeapAlloc(GetProcessHeap(), 0, sizeof(glsl_program_key_t));
2050     key->vshader = entry->vshader;
2051     key->pshader = entry->pshader;
2052     hash_table_remove(This->glsl_program_lookup, key);
2053
2054     GL_EXTCALL(glDeleteObjectARB(entry->programId));
2055     if (entry->vshader) list_remove(&entry->vshader_entry);
2056     if (entry->pshader) list_remove(&entry->pshader_entry);
2057     HeapFree(GetProcessHeap(), 0, entry->vuniformF_locations);
2058     HeapFree(GetProcessHeap(), 0, entry->puniformF_locations);
2059     HeapFree(GetProcessHeap(), 0, entry);
2060 }
2061
2062 /** Sets the GLSL program ID for the given pixel and vertex shader combination.
2063  * It sets the programId on the current StateBlock (because it should be called
2064  * inside of the DrawPrimitive() part of the render loop).
2065  *
2066  * If a program for the given combination does not exist, create one, and store
2067  * the program in the hash table.  If it creates a program, it will link the
2068  * given objects, too.
2069  */
2070 static void set_glsl_shader_program(IWineD3DDevice *iface, BOOL use_ps, BOOL use_vs) {
2071     IWineD3DDeviceImpl *This               = (IWineD3DDeviceImpl *)iface;
2072     WineD3D_GL_Info *gl_info               = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2073     IWineD3DPixelShader  *pshader          = This->stateBlock->pixelShader;
2074     IWineD3DVertexShader *vshader          = This->stateBlock->vertexShader;
2075     struct glsl_shader_prog_link *entry    = NULL;
2076     GLhandleARB programId                  = 0;
2077     int i;
2078     char glsl_name[8];
2079
2080     GLhandleARB vshader_id = use_vs ? ((IWineD3DBaseShaderImpl*)vshader)->baseShader.prgId : 0;
2081     GLhandleARB pshader_id = use_ps ? ((IWineD3DBaseShaderImpl*)pshader)->baseShader.prgId : 0;
2082     entry = get_glsl_program_entry(This, vshader_id, pshader_id);
2083     if (entry) {
2084         This->stateBlock->glsl_program = entry;
2085         return;
2086     }
2087
2088     /* If we get to this point, then no matching program exists, so we create one */
2089     programId = GL_EXTCALL(glCreateProgramObjectARB());
2090     TRACE("Created new GLSL shader program %u\n", programId);
2091
2092     /* Create the entry */
2093     entry = HeapAlloc(GetProcessHeap(), 0, sizeof(struct glsl_shader_prog_link));
2094     entry->programId = programId;
2095     entry->vshader = vshader_id;
2096     entry->pshader = pshader_id;
2097     /* Add the hash table entry */
2098     add_glsl_program_entry(This, entry);
2099
2100     /* Set the current program */
2101     This->stateBlock->glsl_program = entry;
2102
2103     /* Attach GLSL vshader */
2104     if (vshader_id) {
2105         int max_attribs = 16;   /* TODO: Will this always be the case? It is at the moment... */
2106         char tmp_name[10];
2107
2108         TRACE("Attaching GLSL shader object %u to program %u\n", vshader_id, programId);
2109         GL_EXTCALL(glAttachObjectARB(programId, vshader_id));
2110         checkGLcall("glAttachObjectARB");
2111
2112         /* Bind vertex attributes to a corresponding index number to match
2113          * the same index numbers as ARB_vertex_programs (makes loading
2114          * vertex attributes simpler).  With this method, we can use the
2115          * exact same code to load the attributes later for both ARB and
2116          * GLSL shaders.
2117          *
2118          * We have to do this here because we need to know the Program ID
2119          * in order to make the bindings work, and it has to be done prior
2120          * to linking the GLSL program. */
2121         for (i = 0; i < max_attribs; ++i) {
2122              snprintf(tmp_name, sizeof(tmp_name), "attrib%i", i);
2123              GL_EXTCALL(glBindAttribLocationARB(programId, i, tmp_name));
2124         }
2125         checkGLcall("glBindAttribLocationARB");
2126
2127         list_add_head(&((IWineD3DBaseShaderImpl *)vshader)->baseShader.linked_programs, &entry->vshader_entry);
2128     }
2129
2130     /* Attach GLSL pshader */
2131     if (pshader_id) {
2132         TRACE("Attaching GLSL shader object %u to program %u\n", pshader_id, programId);
2133         GL_EXTCALL(glAttachObjectARB(programId, pshader_id));
2134         checkGLcall("glAttachObjectARB");
2135
2136         list_add_head(&((IWineD3DBaseShaderImpl *)pshader)->baseShader.linked_programs, &entry->pshader_entry);
2137     }
2138
2139     /* Link the program */
2140     TRACE("Linking GLSL shader program %u\n", programId);
2141     GL_EXTCALL(glLinkProgramARB(programId));
2142     print_glsl_info_log(&GLINFO_LOCATION, programId);
2143
2144     entry->vuniformF_locations = HeapAlloc(GetProcessHeap(), 0, sizeof(GLhandleARB) * GL_LIMITS(vshader_constantsF));
2145     for (i = 0; i < GL_LIMITS(vshader_constantsF); ++i) {
2146         snprintf(glsl_name, sizeof(glsl_name), "VC[%i]", i);
2147         entry->vuniformF_locations[i] = GL_EXTCALL(glGetUniformLocationARB(programId, glsl_name));
2148     }
2149     entry->puniformF_locations = HeapAlloc(GetProcessHeap(), 0, sizeof(GLhandleARB) * GL_LIMITS(pshader_constantsF));
2150     for (i = 0; i < GL_LIMITS(pshader_constantsF); ++i) {
2151         snprintf(glsl_name, sizeof(glsl_name), "PC[%i]", i);
2152         entry->puniformF_locations[i] = GL_EXTCALL(glGetUniformLocationARB(programId, glsl_name));
2153     }
2154 }
2155
2156 static GLhandleARB create_glsl_blt_shader(WineD3D_GL_Info *gl_info) {
2157     GLhandleARB program_id;
2158     GLhandleARB vshader_id, pshader_id;
2159     const char *blt_vshader[] = {
2160         "void main(void)\n"
2161         "{\n"
2162         "    gl_Position = gl_Vertex;\n"
2163         "    gl_FrontColor = vec4(1.0);\n"
2164         "    gl_TexCoord[0].x = (gl_Vertex.x * 0.5) + 0.5;\n"
2165         "    gl_TexCoord[0].y = (-gl_Vertex.y * 0.5) + 0.5;\n"
2166         "}\n"
2167     };
2168
2169     const char *blt_pshader[] = {
2170         "uniform sampler2D sampler;\n"
2171         "void main(void)\n"
2172         "{\n"
2173         "    gl_FragDepth = texture2D(sampler, gl_TexCoord[0].xy).x;\n"
2174         "}\n"
2175     };
2176
2177     vshader_id = GL_EXTCALL(glCreateShaderObjectARB(GL_VERTEX_SHADER_ARB));
2178     GL_EXTCALL(glShaderSourceARB(vshader_id, 1, blt_vshader, NULL));
2179     GL_EXTCALL(glCompileShaderARB(vshader_id));
2180
2181     pshader_id = GL_EXTCALL(glCreateShaderObjectARB(GL_FRAGMENT_SHADER_ARB));
2182     GL_EXTCALL(glShaderSourceARB(pshader_id, 1, blt_pshader, NULL));
2183     GL_EXTCALL(glCompileShaderARB(pshader_id));
2184
2185     program_id = GL_EXTCALL(glCreateProgramObjectARB());
2186     GL_EXTCALL(glAttachObjectARB(program_id, vshader_id));
2187     GL_EXTCALL(glAttachObjectARB(program_id, pshader_id));
2188     GL_EXTCALL(glLinkProgramARB(program_id));
2189
2190     print_glsl_info_log(&GLINFO_LOCATION, program_id);
2191
2192     return program_id;
2193 }
2194
2195 static void shader_glsl_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
2196     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2197     WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2198     GLhandleARB program_id = 0;
2199
2200     if (useVS || usePS) set_glsl_shader_program(iface, usePS, useVS);
2201     else This->stateBlock->glsl_program = NULL;
2202
2203     program_id = This->stateBlock->glsl_program ? This->stateBlock->glsl_program->programId : 0;
2204     if (program_id) TRACE("Using GLSL program %u\n", program_id);
2205     GL_EXTCALL(glUseProgramObjectARB(program_id));
2206     checkGLcall("glUseProgramObjectARB");
2207 }
2208
2209 static void shader_glsl_select_depth_blt(IWineD3DDevice *iface) {
2210     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2211     WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2212     static GLhandleARB program_id = 0;
2213     static GLhandleARB loc = -1;
2214
2215     if (!program_id) {
2216         program_id = create_glsl_blt_shader(gl_info);
2217         loc = GL_EXTCALL(glGetUniformLocationARB(program_id, "sampler"));
2218     }
2219
2220     GL_EXTCALL(glUseProgramObjectARB(program_id));
2221     GL_EXTCALL(glUniform1iARB(loc, 0));
2222 }
2223
2224 static void shader_glsl_cleanup(IWineD3DDevice *iface) {
2225     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2226     WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2227     GL_EXTCALL(glUseProgramObjectARB(0));
2228 }
2229
2230 const shader_backend_t glsl_shader_backend = {
2231     &shader_glsl_select,
2232     &shader_glsl_select_depth_blt,
2233     &shader_glsl_load_constants,
2234     &shader_glsl_cleanup
2235 };