dlls/wined3d/arb_program_shader.c

   1 /*
   2  * Pixel and vertex shaders implementation using ARB_vertex_program
   3  * and ARB_fragment_program GL extensions.
   4  *
   5  * Copyright 2002-2003 Jason Edmeades
   6  * Copyright 2002-2003 Raphael Junqueira
   7  * Copyright 2004 Christian Costa
   8  * Copyright 2005 Oliver Stieber
   9  * Copyright 2006 Ivan Gyurdiev
  10  * Copyright 2006 Jason Green
  11  * Copyright 2006 Henri Verbeet
  12  *
  13  * This library is free software; you can redistribute it and/or
  14  * modify it under the terms of the GNU Lesser General Public
  15  * License as published by the Free Software Foundation; either
  16  * version 2.1 of the License, or (at your option) any later version.
  17  *
  18  * This library is distributed in the hope that it will be useful,
  19  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  20  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  21  * Lesser General Public License for more details.
  22  *
  23  * You should have received a copy of the GNU Lesser General Public
  24  * License along with this library; if not, write to the Free Software
  25  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  26  */
  27
  28 #include "config.h"
  29
  30 #include <math.h>
  31 #include <stdio.h>
  32
  33 #include "wined3d_private.h"
  34
  35 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  36 WINE_DECLARE_DEBUG_CHANNEL(d3d_constants);
  37
  38 #define GLINFO_LOCATION      (*gl_info)
  39
  40 /********************************************************
  41  * ARB_[vertex/fragment]_program helper functions follow
  42  ********************************************************/
  43
  44 /**
  45  * Loads floating point constants into the currently set ARB_vertex/fragment_program.
  46  * When constant_list == NULL, it will load all the constants.
  47  *
  48  * @target_type should be either GL_VERTEX_PROGRAM_ARB (for vertex shaders)
  49  *  or GL_FRAGMENT_PROGRAM_ARB (for pixel shaders)
  50  */
  51 static void shader_arb_load_constantsF(IWineD3DBaseShaderImpl* This, WineD3D_GL_Info *gl_info, GLuint target_type,
  52         unsigned int max_constants, float* constants, struct list *constant_list) {
  53     constants_entry *constant;
  54     local_constant* lconst;
  55     DWORD i, j;
  56     DWORD *idx;
  57
  58     if (TRACE_ON(d3d_shader)) {
  59         LIST_FOR_EACH_ENTRY(constant, constant_list, constants_entry, entry) {
  60             idx = constant->idx;
  61             j = constant->count;
  62             while (j--) {
  63                 i = *idx++;
  64                 TRACE_(d3d_constants)("Loading constants %i: %f, %f, %f, %f\n", i,
  65                         constants[i * 4 + 0], constants[i * 4 + 1],
  66                         constants[i * 4 + 2], constants[i * 4 + 3]);
  67             }
  68         }
  69     }
  70     LIST_FOR_EACH_ENTRY(constant, constant_list, constants_entry, entry) {
  71         idx = constant->idx;
  72         j = constant->count;
  73         while (j--) {
  74             i = *idx++;
  75             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, constants + (i * 4)));
  76         }
  77     }
  78     checkGLcall("glProgramEnvParameter4fvARB()");
  79
  80     /* Load immediate constants */
  81     if (TRACE_ON(d3d_shader)) {
  82         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
  83             GLfloat* values = (GLfloat*)lconst->value;
  84             TRACE_(d3d_constants)("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
  85                     values[0], values[1], values[2], values[3]);
  86         }
  87     }
  88     LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
  89         GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, lconst->idx, (GLfloat*)lconst->value));
  90     }
  91     checkGLcall("glProgramEnvParameter4fvARB()");
  92 }
  93
  94 /**
  95  * Loads the app-supplied constants into the currently set ARB_[vertex/fragment]_programs.
  96  *
  97  * We only support float constants in ARB at the moment, so don't
  98  * worry about the Integers or Booleans
  99  */
 100 void shader_arb_load_constants(
 101     IWineD3DDevice* device,
 102     char usePixelShader,
 103     char useVertexShader) {
 104
 105     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
 106     IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
 107     WineD3D_GL_Info *gl_info = &deviceImpl->adapter->gl_info;
 108
 109     if (useVertexShader) {
 110         IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
 111
 112         /* Load DirectX 9 float constants for vertex shader */
 113         shader_arb_load_constantsF(vshader, gl_info, GL_VERTEX_PROGRAM_ARB,
 114                                    GL_LIMITS(vshader_constantsF),
 115                                    stateBlock->vertexShaderConstantF,
 116                                    &stateBlock->set_vconstantsF);
 117
 118         /* Upload the position fixup */
 119         GL_EXTCALL(glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, ARB_SHADER_PRIVCONST_POS, deviceImpl->posFixup));
 120     }
 121
 122     if (usePixelShader) {
 123
 124         IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
 125
 126         /* Load DirectX 9 float constants for pixel shader */
 127         shader_arb_load_constantsF(pshader, gl_info, GL_FRAGMENT_PROGRAM_ARB,
 128                                    GL_LIMITS(pshader_constantsF),
 129                                    stateBlock->pixelShaderConstantF,
 130                                    &stateBlock->set_pconstantsF);
 131         if(((IWineD3DPixelShaderImpl *) pshader)->bumpenvmatconst) {
 132             /* needsbumpmat stores the stage number from where to load the matrix. bumpenvmatconst stores the
 133              * number of the constant to load the matrix into.
 134              * The state manager takes care that this function is always called if the bump env matrix changes
 135              */
 136             IWineD3DPixelShaderImpl *psi = (IWineD3DPixelShaderImpl *) pshader;
 137             float *data = (float *) &stateBlock->textureState[(int) psi->needsbumpmat][WINED3DTSS_BUMPENVMAT00];
 138             GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->bumpenvmatconst, data));
 139         }
 140     }
 141 }
 142
 143 /* Generate the variable & register declarations for the ARB_vertex_program output target */
 144 void shader_generate_arb_declarations(
 145     IWineD3DBaseShader *iface,
 146     shader_reg_maps* reg_maps,
 147     SHADER_BUFFER* buffer,
 148     WineD3D_GL_Info* gl_info) {
 149
 150     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 151     DWORD i;
 152     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 153     unsigned max_constantsF = min(This->baseShader.limits.constant_float,
 154             (pshader ? GL_LIMITS(pshader_constantsF) : GL_LIMITS(vshader_constantsF)));
 155
 156     /* Temporary Output register */
 157     shader_addline(buffer, "TEMP TMP_OUT;\n");
 158
 159     for(i = 0; i < This->baseShader.limits.temporary; i++) {
 160         if (reg_maps->temporary[i])
 161             shader_addline(buffer, "TEMP R%u;\n", i);
 162     }
 163
 164     for (i = 0; i < This->baseShader.limits.address; i++) {
 165         if (reg_maps->address[i])
 166             shader_addline(buffer, "ADDRESS A%d;\n", i);
 167     }
 168
 169     for(i = 0; i < This->baseShader.limits.texcoord; i++) {
 170         if (reg_maps->texcoord[i])
 171             shader_addline(buffer,"TEMP T%u;\n", i);
 172     }
 173
 174     /* Texture coordinate registers must be pre-loaded */
 175     for (i = 0; i < This->baseShader.limits.texcoord; i++) {
 176         if (reg_maps->texcoord[i])
 177             shader_addline(buffer, "MOV T%u, fragment.texcoord[%u];\n", i, i);
 178     }
 179
 180     if(reg_maps->bumpmat != -1 /* Only a pshader can use texbem */) {
 181         /* If the shader does not use all available constants, use the next free constant to load the bump mapping environment matrix from
 182          * the stateblock into the shader. If no constant is available don't load, texbem will then just sample the texture without applying
 183          * bump mapping.
 184          */
 185         if(max_constantsF < GL_LIMITS(pshader_constantsF)) {
 186             ((IWineD3DPixelShaderImpl *)This)->bumpenvmatconst = max_constantsF;
 187             shader_addline(buffer, "PARAM bumpenvmat = program.env[%d];\n", ((IWineD3DPixelShaderImpl *)This)->bumpenvmatconst);
 188         } else {
 189             FIXME("No free constant found to load environemnt bump mapping matrix into the shader. texbem instruction will not apply bump mapping\n");
 190         }
 191     }
 192
 193     /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
 194     shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
 195                    max_constantsF, max_constantsF - 1);
 196 }
 197
 198 static const char * const shift_tab[] = {
 199     "dummy",     /*  0 (none) */
 200     "coefmul.x", /*  1 (x2)   */
 201     "coefmul.y", /*  2 (x4)   */
 202     "coefmul.z", /*  3 (x8)   */
 203     "coefmul.w", /*  4 (x16)  */
 204     "dummy",     /*  5 (x32)  */
 205     "dummy",     /*  6 (x64)  */
 206     "dummy",     /*  7 (x128) */
 207     "dummy",     /*  8 (d256) */
 208     "dummy",     /*  9 (d128) */
 209     "dummy",     /* 10 (d64)  */
 210     "dummy",     /* 11 (d32)  */
 211     "coefdiv.w", /* 12 (d16)  */
 212     "coefdiv.z", /* 13 (d8)   */
 213     "coefdiv.y", /* 14 (d4)   */
 214     "coefdiv.x"  /* 15 (d2)   */
 215 };
 216
 217 static void shader_arb_get_write_mask(const DWORD param, char *write_mask) {
 218     char *ptr = write_mask;
 219
 220     if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
 221         *ptr++ = '.';
 222         if (param & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
 223         if (param & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
 224         if (param & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
 225         if (param & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
 226     }
 227
 228     *ptr = '\0';
 229 }
 230
 231 static void shader_arb_get_swizzle(const DWORD param, BOOL fixup, char *swizzle_str) {
 232     /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
 233      * but addressed as "rgba". To fix this we need to swap the register's x
 234      * and z components. */
 235     const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
 236     char *ptr = swizzle_str;
 237
 238     /* swizzle bits fields: wwzzyyxx */
 239     DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
 240     DWORD swizzle_x = swizzle & 0x03;
 241     DWORD swizzle_y = (swizzle >> 2) & 0x03;
 242     DWORD swizzle_z = (swizzle >> 4) & 0x03;
 243     DWORD swizzle_w = (swizzle >> 6) & 0x03;
 244
 245     /* If the swizzle is the default swizzle (ie, "xyzw"), we don't need to
 246      * generate a swizzle string. Unless we need to our own swizzling. */
 247     if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) != swizzle || fixup) {
 248         *ptr++ = '.';
 249         if (swizzle_x == swizzle_y && swizzle_x == swizzle_z && swizzle_x == swizzle_w) {
 250             *ptr++ = swizzle_chars[swizzle_x];
 251         } else {
 252             *ptr++ = swizzle_chars[swizzle_x];
 253             *ptr++ = swizzle_chars[swizzle_y];
 254             *ptr++ = swizzle_chars[swizzle_z];
 255             *ptr++ = swizzle_chars[swizzle_w];
 256         }
 257     }
 258
 259     *ptr = '\0';
 260 }
 261
 262 static void pshader_get_register_name(
 263     const DWORD param, char* regstr) {
 264
 265     DWORD reg = param & WINED3DSP_REGNUM_MASK;
 266     DWORD regtype = shader_get_regtype(param);
 267
 268     switch (regtype) {
 269     case WINED3DSPR_TEMP:
 270         sprintf(regstr, "R%u", reg);
 271     break;
 272     case WINED3DSPR_INPUT:
 273         if (reg==0) {
 274             strcpy(regstr, "fragment.color.primary");
 275         } else {
 276             strcpy(regstr, "fragment.color.secondary");
 277         }
 278     break;
 279     case WINED3DSPR_CONST:
 280         sprintf(regstr, "C[%u]", reg);
 281     break;
 282     case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
 283         sprintf(regstr,"T%u", reg);
 284     break;
 285     case WINED3DSPR_COLOROUT:
 286         if (reg == 0)
 287             sprintf(regstr, "TMP_COLOR");
 288         else {
 289             /* TODO: See GL_ARB_draw_buffers */
 290             FIXME("Unsupported write to render target %u\n", reg);
 291             sprintf(regstr, "unsupported_register");
 292         }
 293     break;
 294     case WINED3DSPR_DEPTHOUT:
 295         sprintf(regstr, "result.depth");
 296     break;
 297     case WINED3DSPR_ATTROUT:
 298         sprintf(regstr, "oD[%u]", reg);
 299     break;
 300     case WINED3DSPR_TEXCRDOUT:
 301         sprintf(regstr, "oT[%u]", reg);
 302     break;
 303     default:
 304         FIXME("Unhandled register name Type(%d)\n", regtype);
 305         sprintf(regstr, "unrecognized_register");
 306     break;
 307     }
 308 }
 309
 310 /* TODO: merge with pixel shader */
 311 static void vshader_program_add_param(SHADER_OPCODE_ARG *arg, const DWORD param, BOOL is_input, char *hwLine) {
 312
 313   IWineD3DVertexShaderImpl* This = (IWineD3DVertexShaderImpl*) arg->shader;
 314
 315   /* oPos, oFog and oPts in D3D */
 316   static const char * const hwrastout_reg_names[] = { "TMP_OUT", "result.fogcoord", "result.pointsize" };
 317
 318   DWORD reg = param & WINED3DSP_REGNUM_MASK;
 319   DWORD regtype = shader_get_regtype(param);
 320   char  tmpReg[255];
 321   BOOL is_color = FALSE;
 322
 323   if ((param & WINED3DSP_SRCMOD_MASK) == WINED3DSPSM_NEG) {
 324       strcat(hwLine, " -");
 325   } else {
 326       strcat(hwLine, " ");
 327   }
 328
 329   switch (regtype) {
 330   case WINED3DSPR_TEMP:
 331     sprintf(tmpReg, "R%u", reg);
 332     strcat(hwLine, tmpReg);
 333     break;
 334   case WINED3DSPR_INPUT:
 335
 336     if (vshader_input_is_color((IWineD3DVertexShader*) This, reg))
 337         is_color = TRUE;
 338
 339     sprintf(tmpReg, "vertex.attrib[%u]", reg);
 340     strcat(hwLine, tmpReg);
 341     break;
 342   case WINED3DSPR_CONST:
 343     sprintf(tmpReg, "C[%s%u]", (param & WINED3DSHADER_ADDRMODE_RELATIVE) ? "A0.x + " : "", reg);
 344     strcat(hwLine, tmpReg);
 345     break;
 346   case WINED3DSPR_ADDR: /*case D3DSPR_TEXTURE:*/
 347     sprintf(tmpReg, "A%u", reg);
 348     strcat(hwLine, tmpReg);
 349     break;
 350   case WINED3DSPR_RASTOUT:
 351     sprintf(tmpReg, "%s", hwrastout_reg_names[reg]);
 352     strcat(hwLine, tmpReg);
 353     break;
 354   case WINED3DSPR_ATTROUT:
 355     if (reg==0) {
 356        strcat(hwLine, "result.color.primary");
 357     } else {
 358        strcat(hwLine, "result.color.secondary");
 359     }
 360     break;
 361   case WINED3DSPR_TEXCRDOUT:
 362     sprintf(tmpReg, "result.texcoord[%u]", reg);
 363     strcat(hwLine, tmpReg);
 364     break;
 365   default:
 366     FIXME("Unknown reg type %d %d\n", regtype, reg);
 367     strcat(hwLine, "unrecognized_register");
 368     break;
 369   }
 370
 371   if (!is_input) {
 372     char write_mask[6];
 373     shader_arb_get_write_mask(param, write_mask);
 374     strcat(hwLine, write_mask);
 375   } else {
 376     char swizzle[6];
 377     shader_arb_get_swizzle(param, is_color, swizzle);
 378     strcat(hwLine, swizzle);
 379   }
 380 }
 381
 382 static void shader_hw_sample(SHADER_OPCODE_ARG* arg, DWORD sampler_idx, const char *dst_str, const char *coord_reg, BOOL projective) {
 383     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 384     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
 385     IWineD3DBaseTextureImpl *texture = (IWineD3DBaseTextureImpl *) deviceImpl->stateBlock->textures[sampler_idx];
 386     WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)This->baseShader.device)->adapter->gl_info;
 387
 388     SHADER_BUFFER* buffer = arg->buffer;
 389     DWORD sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
 390     const char *tex_type;
 391
 392     switch(sampler_type) {
 393         case WINED3DSTT_1D:
 394             tex_type = "1D";
 395             break;
 396
 397         case WINED3DSTT_2D:
 398             tex_type = "2D";
 399             break;
 400
 401         case WINED3DSTT_VOLUME:
 402             tex_type = "3D";
 403             break;
 404
 405         case WINED3DSTT_CUBE:
 406             tex_type = "CUBE";
 407             break;
 408
 409         default:
 410             ERR("Unexpected texture type %d\n", sampler_type);
 411             tex_type = "";
 412     }
 413
 414     if (projective && deviceImpl->stateBlock->textureState[sampler_idx][WINED3DTSS_TEXTURETRANSFORMFLAGS] & WINED3DTTFF_PROJECTED) {
 415         shader_addline(buffer, "TXP %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 416     } else {
 417         shader_addline(buffer, "TEX %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 418     }
 419
 420     /* Signedness correction */
 421     if(!GL_SUPPORT(NV_TEXTURE_SHADER3) /* Provides signed formats */ && texture) {
 422         WINED3DFORMAT format = texture->resource.format;
 423
 424         if((format == WINED3DFMT_V8U8 && !GL_SUPPORT(ATI_ENVMAP_BUMPMAP)) ||
 425             format == WINED3DFMT_Q8W8V8U8 ||
 426             format == WINED3DFMT_V16U16) {
 427             shader_addline(buffer, "MAD %s, %s, coefmul.x, -one;\n", dst_str, dst_str);
 428         } else if(format == WINED3DFMT_X8L8V8U8) {
 429             shader_addline(buffer, "MAD %s.rg, %s, coefmul.x, -one;\n", dst_str, dst_str);
 430         }
 431     }
 432 }
 433
 434
 435 static void pshader_gen_input_modifier_line (
 436     SHADER_BUFFER* buffer,
 437     const DWORD instr,
 438     int tmpreg,
 439     char *outregstr) {
 440
 441     /* Generate a line that does the input modifier computation and return the input register to use */
 442     char regstr[256];
 443     char swzstr[20];
 444     int insert_line;
 445
 446     /* Assume a new line will be added */
 447     insert_line = 1;
 448
 449     /* Get register name */
 450     pshader_get_register_name(instr, regstr);
 451     shader_arb_get_swizzle(instr, FALSE, swzstr);
 452
 453     switch (instr & WINED3DSP_SRCMOD_MASK) {
 454     case WINED3DSPSM_NONE:
 455         sprintf(outregstr, "%s%s", regstr, swzstr);
 456         insert_line = 0;
 457         break;
 458     case WINED3DSPSM_NEG:
 459         sprintf(outregstr, "-%s%s", regstr, swzstr);
 460         insert_line = 0;
 461         break;
 462     case WINED3DSPSM_BIAS:
 463         shader_addline(buffer, "ADD T%c, %s, -coefdiv.x;\n", 'A' + tmpreg, regstr);
 464         break;
 465     case WINED3DSPSM_BIASNEG:
 466         shader_addline(buffer, "ADD T%c, -%s, coefdiv.x;\n", 'A' + tmpreg, regstr);
 467         break;
 468     case WINED3DSPSM_SIGN:
 469         shader_addline(buffer, "MAD T%c, %s, coefmul.x, -one.x;\n", 'A' + tmpreg, regstr);
 470         break;
 471     case WINED3DSPSM_SIGNNEG:
 472         shader_addline(buffer, "MAD T%c, %s, -coefmul.x, one.x;\n", 'A' + tmpreg, regstr);
 473         break;
 474     case WINED3DSPSM_COMP:
 475         shader_addline(buffer, "SUB T%c, one.x, %s;\n", 'A' + tmpreg, regstr);
 476         break;
 477     case WINED3DSPSM_X2:
 478         shader_addline(buffer, "ADD T%c, %s, %s;\n", 'A' + tmpreg, regstr, regstr);
 479         break;
 480     case WINED3DSPSM_X2NEG:
 481         shader_addline(buffer, "ADD T%c, -%s, -%s;\n", 'A' + tmpreg, regstr, regstr);
 482         break;
 483     case WINED3DSPSM_DZ:
 484         shader_addline(buffer, "RCP T%c, %s.z;\n", 'A' + tmpreg, regstr);
 485         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 486         break;
 487     case WINED3DSPSM_DW:
 488         shader_addline(buffer, "RCP T%c, %s.w;\n", 'A' + tmpreg, regstr);
 489         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 490         break;
 491     default:
 492         sprintf(outregstr, "%s%s", regstr, swzstr);
 493         insert_line = 0;
 494     }
 495
 496     /* Return modified or original register, with swizzle */
 497     if (insert_line)
 498         sprintf(outregstr, "T%c%s", 'A' + tmpreg, swzstr);
 499 }
 500
 501 static inline void pshader_gen_output_modifier_line(
 502     SHADER_BUFFER* buffer,
 503     int saturate,
 504     char *write_mask,
 505     int shift,
 506     char *regstr) {
 507
 508     /* Generate a line that does the output modifier computation */
 509     shader_addline(buffer, "MUL%s %s%s, %s, %s;\n", saturate ? "_SAT" : "",
 510         regstr, write_mask, regstr, shift_tab[shift]);
 511 }
 512
 513 void pshader_hw_bem(SHADER_OPCODE_ARG* arg) {
 514     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 515
 516     SHADER_BUFFER* buffer = arg->buffer;
 517     char dst_name[50];
 518     char src_name[2][50];
 519     char dst_wmask[20];
 520
 521     pshader_get_register_name(arg->dst, dst_name);
 522     shader_arb_get_write_mask(arg->dst, dst_wmask);
 523     strcat(dst_name, dst_wmask);
 524
 525     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src_name[0]);
 526     pshader_gen_input_modifier_line(buffer, arg->src[1], 1, src_name[1]);
 527
 528     if(This->bumpenvmatconst != -1) {
 529         /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
 530         shader_addline(buffer, "SWZ TMP2, bumpenvmat, x, z, 0, 0;\n");
 531         shader_addline(buffer, "DP3 TMP.r, TMP2, %s;\n", src_name[1]);
 532         shader_addline(buffer, "SWZ TMP2, bumpenvmat, y, w, 0, 0;\n");
 533         shader_addline(buffer, "DP3 TMP.g, TMP2, %s;\n", src_name[1]);
 534
 535         shader_addline(buffer, "ADD %s, %s, TMP;\n", dst_name, src_name[0]);
 536     } else {
 537         shader_addline(buffer, "MOV %s, %s;\n", dst_name, src_name[0]);
 538     }
 539 }
 540
 541 void pshader_hw_cnd(SHADER_OPCODE_ARG* arg) {
 542
 543     SHADER_BUFFER* buffer = arg->buffer;
 544     char dst_wmask[20];
 545     char dst_name[50];
 546     char src_name[3][50];
 547
 548     /* FIXME: support output modifiers */
 549
 550     /* Handle output register */
 551     pshader_get_register_name(arg->dst, dst_name);
 552     shader_arb_get_write_mask(arg->dst, dst_wmask);
 553     strcat(dst_name, dst_wmask);
 554
 555     /* Generate input register names (with modifiers) */
 556     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src_name[0]);
 557     pshader_gen_input_modifier_line(buffer, arg->src[1], 1, src_name[1]);
 558     pshader_gen_input_modifier_line(buffer, arg->src[2], 2, src_name[2]);
 559
 560     shader_addline(buffer, "ADD TMP, -%s, coefdiv.x;\n", src_name[0]);
 561     shader_addline(buffer, "CMP %s, TMP, %s, %s;\n", dst_name, src_name[1], src_name[2]);
 562 }
 563
 564 void pshader_hw_cmp(SHADER_OPCODE_ARG* arg) {
 565
 566     SHADER_BUFFER* buffer = arg->buffer;
 567     char dst_wmask[20];
 568     char dst_name[50];
 569     char src_name[3][50];
 570
 571     /* FIXME: support output modifiers */
 572
 573     /* Handle output register */
 574     pshader_get_register_name(arg->dst, dst_name);
 575     shader_arb_get_write_mask(arg->dst, dst_wmask);
 576     strcat(dst_name, dst_wmask);
 577
 578     /* Generate input register names (with modifiers) */
 579     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src_name[0]);
 580     pshader_gen_input_modifier_line(buffer, arg->src[1], 1, src_name[1]);
 581     pshader_gen_input_modifier_line(buffer, arg->src[2], 2, src_name[2]);
 582
 583     shader_addline(buffer, "CMP %s, %s, %s, %s;\n", dst_name,
 584         src_name[0], src_name[2], src_name[1]);
 585 }
 586
 587 /* Map the opcode 1-to-1 to the GL code */
 588 void pshader_hw_map2gl(SHADER_OPCODE_ARG* arg) {
 589
 590      CONST SHADER_OPCODE* curOpcode = arg->opcode;
 591      SHADER_BUFFER* buffer = arg->buffer;
 592      DWORD dst = arg->dst;
 593      DWORD* src = arg->src;
 594
 595      unsigned int i;
 596      char tmpLine[256];
 597
 598      /* Output token related */
 599      char output_rname[256];
 600      char output_wmask[20];
 601      BOOL saturate = FALSE;
 602      BOOL centroid = FALSE;
 603      BOOL partialprecision = FALSE;
 604      DWORD shift;
 605
 606      strcpy(tmpLine, curOpcode->glname);
 607
 608      /* Process modifiers */
 609      if (0 != (dst & WINED3DSP_DSTMOD_MASK)) {
 610          DWORD mask = dst & WINED3DSP_DSTMOD_MASK;
 611
 612          saturate = mask & WINED3DSPDM_SATURATE;
 613          centroid = mask & WINED3DSPDM_MSAMPCENTROID;
 614          partialprecision = mask & WINED3DSPDM_PARTIALPRECISION;
 615          mask &= ~(WINED3DSPDM_MSAMPCENTROID | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_SATURATE);
 616          if (mask)
 617             FIXME("Unrecognized modifier(0x%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
 618
 619          if (centroid)
 620              FIXME("Unhandled modifier(0x%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
 621      }
 622      shift = (dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
 623
 624       /* Generate input and output registers */
 625       if (curOpcode->num_params > 0) {
 626           char operands[4][100];
 627
 628           /* Generate input register names (with modifiers) */
 629           for (i = 1; i < curOpcode->num_params; ++i)
 630               pshader_gen_input_modifier_line(buffer, src[i-1], i-1, operands[i]);
 631
 632           /* Handle output register */
 633           pshader_get_register_name(dst, output_rname);
 634           strcpy(operands[0], output_rname);
 635           shader_arb_get_write_mask(dst, output_wmask);
 636           strcat(operands[0], output_wmask);
 637
 638           if (saturate && (shift == 0))
 639              strcat(tmpLine, "_SAT");
 640           strcat(tmpLine, " ");
 641           strcat(tmpLine, operands[0]);
 642           for (i = 1; i < curOpcode->num_params; i++) {
 643               strcat(tmpLine, ", ");
 644               strcat(tmpLine, operands[i]);
 645           }
 646           strcat(tmpLine,";\n");
 647           shader_addline(buffer, tmpLine);
 648
 649           /* A shift requires another line. */
 650           if (shift != 0)
 651               pshader_gen_output_modifier_line(buffer, saturate, output_wmask, shift, output_rname);
 652       }
 653 }
 654
 655 void pshader_hw_tex(SHADER_OPCODE_ARG* arg) {
 656
 657     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 658
 659     DWORD dst = arg->dst;
 660     DWORD* src = arg->src;
 661     SHADER_BUFFER* buffer = arg->buffer;
 662     DWORD hex_version = This->baseShader.hex_version;
 663
 664     char reg_dest[40];
 665     char reg_coord[40];
 666     DWORD reg_dest_code;
 667     DWORD reg_sampler_code;
 668
 669     /* All versions have a destination register */
 670     reg_dest_code = dst & WINED3DSP_REGNUM_MASK;
 671     pshader_get_register_name(dst, reg_dest);
 672
 673     /* 1.0-1.3: Use destination register as coordinate source.
 674        1.4+: Use provided coordinate source register. */
 675    if (hex_version < WINED3DPS_VERSION(1,4))
 676       strcpy(reg_coord, reg_dest);
 677    else
 678       pshader_gen_input_modifier_line(buffer, src[0], 0, reg_coord);
 679
 680   /* 1.0-1.4: Use destination register number as texture code.
 681      2.0+: Use provided sampler number as texure code. */
 682   if (hex_version < WINED3DPS_VERSION(2,0))
 683      reg_sampler_code = reg_dest_code;
 684   else
 685      reg_sampler_code = src[1] & WINED3DSP_REGNUM_MASK;
 686
 687   shader_hw_sample(arg, reg_sampler_code, reg_dest, reg_coord, TRUE);
 688 }
 689
 690 void pshader_hw_texcoord(SHADER_OPCODE_ARG* arg) {
 691
 692     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 693     DWORD dst = arg->dst;
 694     SHADER_BUFFER* buffer = arg->buffer;
 695     DWORD hex_version = This->baseShader.hex_version;
 696
 697     char tmp[20];
 698     shader_arb_get_write_mask(dst, tmp);
 699     if (hex_version != WINED3DPS_VERSION(1,4)) {
 700         DWORD reg = dst & WINED3DSP_REGNUM_MASK;
 701         shader_addline(buffer, "MOV_SAT T%u%s, fragment.texcoord[%u];\n", reg, tmp, reg);
 702     } else {
 703         DWORD reg1 = dst & WINED3DSP_REGNUM_MASK;
 704         char reg_src[40];
 705
 706         pshader_gen_input_modifier_line(buffer, arg->src[0], 0, reg_src);
 707         shader_addline(buffer, "MOV R%u%s, %s;\n", reg1, tmp, reg_src);
 708    }
 709 }
 710
 711 void pshader_hw_texreg2ar(SHADER_OPCODE_ARG* arg) {
 712
 713      SHADER_BUFFER* buffer = arg->buffer;
 714
 715      DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
 716      DWORD reg2 = arg->src[0] & WINED3DSP_REGNUM_MASK;
 717      char dst_str[8];
 718
 719      sprintf(dst_str, "T%u", reg1);
 720      shader_addline(buffer, "MOV TMP.r, T%u.a;\n", reg2);
 721      shader_addline(buffer, "MOV TMP.g, T%u.r;\n", reg2);
 722      shader_hw_sample(arg, reg1, dst_str, "TMP", TRUE);
 723 }
 724
 725 void pshader_hw_texreg2gb(SHADER_OPCODE_ARG* arg) {
 726
 727      SHADER_BUFFER* buffer = arg->buffer;
 728
 729      DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
 730      DWORD reg2 = arg->src[0] & WINED3DSP_REGNUM_MASK;
 731      char dst_str[8];
 732
 733      sprintf(dst_str, "T%u", reg1);
 734      shader_addline(buffer, "MOV TMP.r, T%u.g;\n", reg2);
 735      shader_addline(buffer, "MOV TMP.g, T%u.b;\n", reg2);
 736      shader_hw_sample(arg, reg1, dst_str, "TMP", TRUE);
 737 }
 738
 739 void pshader_hw_texbem(SHADER_OPCODE_ARG* arg) {
 740     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 741
 742     DWORD dst = arg->dst;
 743     DWORD src = arg->src[0] & WINED3DSP_REGNUM_MASK;
 744     SHADER_BUFFER* buffer = arg->buffer;
 745
 746     char reg_coord[40];
 747     DWORD reg_dest_code;
 748
 749     /* All versions have a destination register */
 750     reg_dest_code = dst & WINED3DSP_REGNUM_MASK;
 751     /* Can directly use the name because texbem is only valid for <= 1.3 shaders */
 752     pshader_get_register_name(dst, reg_coord);
 753
 754     if(This->bumpenvmatconst != -1) {
 755         /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
 756
 757         shader_addline(buffer, "SWZ TMP2, bumpenvmat, x, z, 0, 0;\n");
 758         shader_addline(buffer, "DP3 TMP.r, TMP2, T%u;\n", src);
 759         shader_addline(buffer, "SWZ TMP2, bumpenvmat, y, w, 0, 0;\n");
 760         shader_addline(buffer, "DP3 TMP.g, TMP2, T%u;\n", src);
 761
 762         /* with projective textures, texbem only divides the static texture coord, not the displacement,
 763          * so we can't let the GL handle this.
 764          */
 765         if (((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS]
 766               & WINED3DTTFF_PROJECTED) {
 767             shader_addline(buffer, "RCP TMP2.a, %s.a;\n", reg_coord);
 768             shader_addline(buffer, "MUL TMP2.rg, %s, TMP2.a;\n", reg_coord);
 769             shader_addline(buffer, "ADD TMP.rg, TMP, TMP2;\n");
 770         } else {
 771             shader_addline(buffer, "ADD TMP.rg, TMP, %s;\n", reg_coord);
 772         }
 773
 774         shader_hw_sample(arg, reg_dest_code, reg_coord, "TMP", FALSE);
 775     } else {
 776         /* Without a bump matrix loaded, just sample with the unmodified coordinates */
 777         shader_hw_sample(arg, reg_dest_code, reg_coord, reg_coord, TRUE);
 778     }
 779 }
 780
 781 void pshader_hw_texm3x2pad(SHADER_OPCODE_ARG* arg) {
 782
 783     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
 784     SHADER_BUFFER* buffer = arg->buffer;
 785     char src0_name[50];
 786
 787     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src0_name);
 788     shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", reg, src0_name);
 789 }
 790
 791 void pshader_hw_texm3x2tex(SHADER_OPCODE_ARG* arg) {
 792
 793     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
 794     SHADER_BUFFER* buffer = arg->buffer;
 795     char dst_str[8];
 796     char src0_name[50];
 797
 798     sprintf(dst_str, "T%u", reg);
 799     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src0_name);
 800     shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", reg, src0_name);
 801     shader_hw_sample(arg, reg, dst_str, "TMP", TRUE);
 802 }
 803
 804 void pshader_hw_texm3x3pad(SHADER_OPCODE_ARG* arg) {
 805
 806     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 807     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
 808     SHADER_BUFFER* buffer = arg->buffer;
 809     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
 810     char src0_name[50];
 811
 812     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src0_name);
 813     shader_addline(buffer, "DP3 TMP.%c, T%u, %s;\n", 'x' + current_state->current_row, reg, src0_name);
 814     current_state->texcoord_w[current_state->current_row++] = reg;
 815 }
 816
 817 void pshader_hw_texm3x3tex(SHADER_OPCODE_ARG* arg) {
 818
 819     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 820     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
 821     SHADER_BUFFER* buffer = arg->buffer;
 822     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
 823     char dst_str[8];
 824     char src0_name[50];
 825
 826     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src0_name);
 827     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
 828
 829     /* Sample the texture using the calculated coordinates */
 830     sprintf(dst_str, "T%u", reg);
 831     shader_hw_sample(arg, reg, dst_str, "TMP", TRUE);
 832     current_state->current_row = 0;
 833 }
 834
 835 void pshader_hw_texm3x3vspec(SHADER_OPCODE_ARG* arg) {
 836
 837     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 838     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
 839     SHADER_BUFFER* buffer = arg->buffer;
 840     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
 841     char dst_str[8];
 842     char src0_name[50];
 843
 844     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src0_name);
 845     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
 846
 847     /* Construct the eye-ray vector from w coordinates */
 848     shader_addline(buffer, "MOV TMP2.x, fragment.texcoord[%u].w;\n", current_state->texcoord_w[0]);
 849     shader_addline(buffer, "MOV TMP2.y, fragment.texcoord[%u].w;\n", current_state->texcoord_w[1]);
 850     shader_addline(buffer, "MOV TMP2.z, fragment.texcoord[%u].w;\n", reg);
 851
 852     /* Calculate reflection vector (Assume normal is normalized): RF = 2*(N.E)*N -E */
 853     shader_addline(buffer, "DP3 TMP.w, TMP, TMP2;\n");
 854     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
 855     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -TMP2;\n");
 856
 857     /* Sample the texture using the calculated coordinates */
 858     sprintf(dst_str, "T%u", reg);
 859     shader_hw_sample(arg, reg, dst_str, "TMP", TRUE);
 860     current_state->current_row = 0;
 861 }
 862
 863 void pshader_hw_texm3x3spec(SHADER_OPCODE_ARG* arg) {
 864
 865     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 866     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
 867     DWORD reg3 = arg->src[1] & WINED3DSP_REGNUM_MASK;
 868     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
 869     SHADER_BUFFER* buffer = arg->buffer;
 870     char dst_str[8];
 871     char src0_name[50];
 872
 873     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src0_name);
 874     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
 875
 876     /* Calculate reflection vector (Assume normal is normalized): RF = 2*(N.E)*N -E */
 877     shader_addline(buffer, "DP3 TMP.w, TMP, C[%u];\n", reg3);
 878     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
 879     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -C[%u];\n", reg3);
 880
 881     /* Sample the texture using the calculated coordinates */
 882     sprintf(dst_str, "T%u", reg);
 883     shader_hw_sample(arg, reg, dst_str, "TMP", TRUE);
 884     current_state->current_row = 0;
 885 }
 886
 887 /** Handles transforming all WINED3DSIO_M?x? opcodes for
 888     Vertex shaders to ARB_vertex_program codes */
 889 void vshader_hw_mnxn(SHADER_OPCODE_ARG* arg) {
 890
 891     int i;
 892     int nComponents = 0;
 893     SHADER_OPCODE_ARG tmpArg;
 894
 895     memset(&tmpArg, 0, sizeof(SHADER_OPCODE_ARG));
 896
 897     /* Set constants for the temporary argument */
 898     tmpArg.shader      = arg->shader;
 899     tmpArg.buffer      = arg->buffer;
 900     tmpArg.src[0]      = arg->src[0];
 901     tmpArg.src_addr[0] = arg->src_addr[0];
 902     tmpArg.src_addr[1] = arg->src_addr[1];
 903     tmpArg.reg_maps = arg->reg_maps;
 904
 905     switch(arg->opcode->opcode) {
 906     case WINED3DSIO_M4x4:
 907         nComponents = 4;
 908         tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
 909         break;
 910     case WINED3DSIO_M4x3:
 911         nComponents = 3;
 912         tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
 913         break;
 914     case WINED3DSIO_M3x4:
 915         nComponents = 4;
 916         tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
 917         break;
 918     case WINED3DSIO_M3x3:
 919         nComponents = 3;
 920         tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
 921         break;
 922     case WINED3DSIO_M3x2:
 923         nComponents = 2;
 924         tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
 925         break;
 926     default:
 927         break;
 928     }
 929
 930     for (i = 0; i < nComponents; i++) {
 931         tmpArg.dst = ((arg->dst) & ~WINED3DSP_WRITEMASK_ALL)|(WINED3DSP_WRITEMASK_0<<i);
 932         tmpArg.src[1] = arg->src[1]+i;
 933         vshader_hw_map2gl(&tmpArg);
 934     }
 935 }
 936
 937 void vshader_hw_rsq_rcp(SHADER_OPCODE_ARG* arg) {
 938     CONST SHADER_OPCODE* curOpcode = arg->opcode;
 939     SHADER_BUFFER* buffer = arg->buffer;
 940     DWORD dst = arg->dst;
 941     DWORD src = arg->src[0];
 942     DWORD swizzle = (src & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
 943
 944     char tmpLine[256];
 945
 946     strcpy(tmpLine, curOpcode->glname); /* Opcode */
 947     vshader_program_add_param(arg, dst, FALSE, tmpLine); /* Destination */
 948     strcat(tmpLine, ",");
 949     vshader_program_add_param(arg, src, TRUE, tmpLine);
 950     if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) == swizzle) {
 951         /* Dx sdk says .x is used if no swizzle is given */
 952         strcat(tmpLine, ".x");
 953     }
 954
 955     shader_addline(buffer, "%s;\n", tmpLine);
 956 }
 957
 958 /* TODO: merge with pixel shader */
 959 /* Map the opcode 1-to-1 to the GL code */
 960 void vshader_hw_map2gl(SHADER_OPCODE_ARG* arg) {
 961
 962     CONST SHADER_OPCODE* curOpcode = arg->opcode;
 963     SHADER_BUFFER* buffer = arg->buffer;
 964     DWORD dst = arg->dst;
 965     DWORD* src = arg->src;
 966
 967     DWORD dst_regtype = shader_get_regtype(dst);
 968     char tmpLine[256];
 969     unsigned int i;
 970
 971     if ((curOpcode->opcode == WINED3DSIO_MOV && dst_regtype == WINED3DSPR_ADDR) || curOpcode->opcode == WINED3DSIO_MOVA)
 972         strcpy(tmpLine, "ARL");
 973     else
 974         strcpy(tmpLine, curOpcode->glname);
 975
 976     if (curOpcode->num_params > 0) {
 977         vshader_program_add_param(arg, dst, FALSE, tmpLine);
 978         for (i = 1; i < curOpcode->num_params; ++i) {
 979            strcat(tmpLine, ",");
 980            vshader_program_add_param(arg, src[i-1], TRUE, tmpLine);
 981         }
 982     }
 983    shader_addline(buffer, "%s;\n", tmpLine);
 984 }
 985
 986 static GLuint create_arb_blt_vertex_program(WineD3D_GL_Info *gl_info) {
 987     GLuint program_id = 0;
 988     const char *blt_vprogram =
 989         "!!ARBvp1.0\n"
 990         "PARAM c[1] = { { 1, 0.5 } };\n"
 991         "MOV result.position, vertex.position;\n"
 992         "MOV result.color, c[0].x;\n"
 993         "MAD result.texcoord[0].y, -vertex.position, c[0], c[0];\n"
 994         "MAD result.texcoord[0].x, vertex.position, c[0].y, c[0].y;\n"
 995         "END\n";
 996
 997     GL_EXTCALL(glGenProgramsARB(1, &program_id));
 998     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, program_id));
 999     GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_vprogram), blt_vprogram));
1000
1001     if (glGetError() == GL_INVALID_OPERATION) {
1002         GLint pos;
1003         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1004         FIXME("Vertex program error at position %d: %s\n", pos,
1005             debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1006     }
1007
1008     return program_id;
1009 }
1010
1011 static GLuint create_arb_blt_fragment_program(WineD3D_GL_Info *gl_info) {
1012     GLuint program_id = 0;
1013     const char *blt_fprogram =
1014         "!!ARBfp1.0\n"
1015         "TEMP R0;\n"
1016         "TEX R0.x, fragment.texcoord[0], texture[0], 2D;\n"
1017         "MOV result.depth.z, R0.x;\n"
1018         "END\n";
1019
1020     GL_EXTCALL(glGenProgramsARB(1, &program_id));
1021     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, program_id));
1022     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_fprogram), blt_fprogram));
1023
1024     if (glGetError() == GL_INVALID_OPERATION) {
1025         GLint pos;
1026         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1027         FIXME("Fragment program error at position %d: %s\n", pos,
1028             debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1029     }
1030
1031     return program_id;
1032 }
1033
1034 static void shader_arb_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
1035     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1036     WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1037
1038     if (useVS) {
1039         TRACE("Using vertex shader\n");
1040
1041         /* Bind the vertex program */
1042         GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB,
1043             ((IWineD3DVertexShaderImpl *)This->stateBlock->vertexShader)->baseShader.prgId));
1044         checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vertexShader->prgId);");
1045
1046         /* Enable OpenGL vertex programs */
1047         glEnable(GL_VERTEX_PROGRAM_ARB);
1048         checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1049         TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n",
1050             This, ((IWineD3DVertexShaderImpl *)This->stateBlock->vertexShader)->baseShader.prgId);
1051     } else if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
1052         glDisable(GL_VERTEX_PROGRAM_ARB);
1053         checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1054     }
1055
1056     if (usePS) {
1057         TRACE("Using pixel shader\n");
1058
1059         /* Bind the fragment program */
1060         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB,
1061             ((IWineD3DPixelShaderImpl *)This->stateBlock->pixelShader)->baseShader.prgId));
1062         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, pixelShader->prgId);");
1063
1064         /* Enable OpenGL fragment programs */
1065         glEnable(GL_FRAGMENT_PROGRAM_ARB);
1066         checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1067         TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n",
1068             This, ((IWineD3DPixelShaderImpl *)This->stateBlock->pixelShader)->baseShader.prgId);
1069     } else if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM)) {
1070         glDisable(GL_FRAGMENT_PROGRAM_ARB);
1071         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1072     }
1073 }
1074
1075 static void shader_arb_select_depth_blt(IWineD3DDevice *iface) {
1076     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1077     WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1078     static GLuint vprogram_id = 0;
1079     static GLuint fprogram_id = 0;
1080
1081     if (!vprogram_id) vprogram_id = create_arb_blt_vertex_program(gl_info);
1082     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vprogram_id));
1083     glEnable(GL_VERTEX_PROGRAM_ARB);
1084
1085     if (!fprogram_id) fprogram_id = create_arb_blt_fragment_program(gl_info);
1086     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, fprogram_id));
1087     glEnable(GL_FRAGMENT_PROGRAM_ARB);
1088 }
1089
1090 static void shader_arb_cleanup(IWineD3DDevice *iface) {
1091     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1092     WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1093     if (GL_SUPPORT(ARB_VERTEX_PROGRAM)) glDisable(GL_VERTEX_PROGRAM_ARB);
1094     if (GL_SUPPORT(ARB_FRAGMENT_PROGRAM)) glDisable(GL_FRAGMENT_PROGRAM_ARB);
1095 }
1096
1097 const shader_backend_t arb_program_shader_backend = {
1098     &shader_arb_select,
1099     &shader_arb_select_depth_blt,
1100     &shader_arb_load_constants,
1101     &shader_arb_cleanup
1102 };