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
  37 #define GLINFO_LOCATION      (*gl_info)
  38
  39 /********************************************************
  40  * ARB_[vertex/fragment]_program helper functions follow
  41  ********************************************************/
  42
  43 /**
  44  * Loads floating point constants into the currently set ARB_vertex/fragment_program.
  45  * When constant_list == NULL, it will load all the constants.
  46  *
  47  * @target_type should be either GL_VERTEX_PROGRAM_ARB (for vertex shaders)
  48  *  or GL_FRAGMENT_PROGRAM_ARB (for pixel shaders)
  49  */
  50 static void shader_arb_load_constantsF(IWineD3DBaseShaderImpl* This, WineD3D_GL_Info *gl_info, GLuint target_type,
  51         unsigned int max_constants, float* constants, struct list *constant_list) {
  52     constant_entry *constant;
  53     local_constant* lconst;
  54     int i;
  55
  56     if (!constant_list) {
  57         if (TRACE_ON(d3d_shader)) {
  58             for (i = 0; i < max_constants; ++i) {
  59                 TRACE("Loading constants %i: %f, %f, %f, %f\n", i,
  60                         constants[i * 4 + 0], constants[i * 4 + 1],
  61                         constants[i * 4 + 2], constants[i * 4 + 3]);
  62             }
  63         }
  64         for (i = 0; i < max_constants; ++i) {
  65             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, constants + (i * 4)));
  66         }
  67         checkGLcall("glProgramEnvParameter4fvARB()");
  68     } else {
  69         if (TRACE_ON(d3d_shader)) {
  70             LIST_FOR_EACH_ENTRY(constant, constant_list, constant_entry, entry) {
  71                 i = constant->idx;
  72                 TRACE("Loading constants %i: %f, %f, %f, %f\n", i,
  73                         constants[i * 4 + 0], constants[i * 4 + 1],
  74                         constants[i * 4 + 2], constants[i * 4 + 3]);
  75             }
  76         }
  77         LIST_FOR_EACH_ENTRY(constant, constant_list, constant_entry, entry) {
  78             i = constant->idx;
  79             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, constants + (i * 4)));
  80         }
  81         checkGLcall("glProgramEnvParameter4fvARB()");
  82     }
  83
  84     /* Load immediate constants */
  85     if (TRACE_ON(d3d_shader)) {
  86         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
  87             GLfloat* values = (GLfloat*)lconst->value;
  88             TRACE("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
  89                     values[0], values[1], values[2], values[3]);
  90         }
  91     }
  92     LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
  93         GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, lconst->idx, (GLfloat*)lconst->value));
  94     }
  95     checkGLcall("glProgramEnvParameter4fvARB()");
  96 }
  97
  98 /**
  99  * Loads the app-supplied constants into the currently set ARB_[vertex/fragment]_programs.
 100  *
 101  * We only support float constants in ARB at the moment, so don't
 102  * worry about the Integers or Booleans
 103  */
 104 void shader_arb_load_constants(
 105     IWineD3DDevice* device,
 106     char usePixelShader,
 107     char useVertexShader) {
 108
 109     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
 110     IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
 111     WineD3D_GL_Info *gl_info = &((IWineD3DImpl*)deviceImpl->wineD3D)->gl_info;
 112
 113     if (useVertexShader) {
 114         IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
 115         IWineD3DVertexShaderImpl* vshader_impl = (IWineD3DVertexShaderImpl*) stateBlock->vertexShader;
 116         IWineD3DVertexDeclarationImpl* vertexDeclaration =
 117             (IWineD3DVertexDeclarationImpl*) vshader_impl->vertexDeclaration;
 118
 119         if (NULL != vertexDeclaration && NULL != vertexDeclaration->constants) {
 120             /* Load DirectX 8 float constants for vertex shader */
 121             shader_arb_load_constantsF(vshader, gl_info, GL_VERTEX_PROGRAM_ARB,
 122                                        GL_LIMITS(vshader_constantsF),
 123                                        vertexDeclaration->constants, NULL);
 124         }
 125
 126         /* Load DirectX 9 float constants for vertex shader */
 127         shader_arb_load_constantsF(vshader, gl_info, GL_VERTEX_PROGRAM_ARB,
 128                                    GL_LIMITS(vshader_constantsF),
 129                                    stateBlock->vertexShaderConstantF,
 130                                    &stateBlock->set_vconstantsF);
 131
 132         /* Upload the position fixup */
 133         GL_EXTCALL(glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, ARB_SHADER_PRIVCONST_POS, deviceImpl->posFixup));
 134     }
 135
 136     if (usePixelShader) {
 137
 138         IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
 139
 140         /* Load DirectX 9 float constants for pixel shader */
 141         shader_arb_load_constantsF(pshader, gl_info, GL_FRAGMENT_PROGRAM_ARB,
 142                                    GL_LIMITS(pshader_constantsF),
 143                                    stateBlock->pixelShaderConstantF,
 144                                    &stateBlock->set_pconstantsF);
 145     }
 146 }
 147
 148 /* Generate the variable & register declarations for the ARB_vertex_program output target */
 149 void shader_generate_arb_declarations(
 150     IWineD3DBaseShader *iface,
 151     shader_reg_maps* reg_maps,
 152     SHADER_BUFFER* buffer,
 153     WineD3D_GL_Info* gl_info) {
 154
 155     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 156     DWORD i;
 157     char pshader = shader_is_pshader_version(This->baseShader.hex_version);
 158     unsigned max_constantsF = min(This->baseShader.limits.constant_float,
 159             (pshader ? GL_LIMITS(pshader_constantsF) : GL_LIMITS(vshader_constantsF)));
 160
 161     /* Temporary Output register */
 162     shader_addline(buffer, "TEMP TMP_OUT;\n");
 163
 164     for(i = 0; i < This->baseShader.limits.temporary; i++) {
 165         if (reg_maps->temporary[i])
 166             shader_addline(buffer, "TEMP R%u;\n", i);
 167     }
 168
 169     for (i = 0; i < This->baseShader.limits.address; i++) {
 170         if (reg_maps->address[i])
 171             shader_addline(buffer, "ADDRESS A%d;\n", i);
 172     }
 173
 174     for(i = 0; i < This->baseShader.limits.texcoord; i++) {
 175         if (reg_maps->texcoord[i])
 176             shader_addline(buffer,"TEMP T%u;\n", i);
 177     }
 178
 179     /* Texture coordinate registers must be pre-loaded */
 180     for (i = 0; i < This->baseShader.limits.texcoord; i++) {
 181         if (reg_maps->texcoord[i])
 182             shader_addline(buffer, "MOV T%u, fragment.texcoord[%u];\n", i, i);
 183     }
 184
 185     /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
 186     shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
 187                    max_constantsF, max_constantsF - 1);
 188 }
 189
 190 static const char* shift_tab[] = {
 191     "dummy",     /*  0 (none) */
 192     "coefmul.x", /*  1 (x2)   */
 193     "coefmul.y", /*  2 (x4)   */
 194     "coefmul.z", /*  3 (x8)   */
 195     "coefmul.w", /*  4 (x16)  */
 196     "dummy",     /*  5 (x32)  */
 197     "dummy",     /*  6 (x64)  */
 198     "dummy",     /*  7 (x128) */
 199     "dummy",     /*  8 (d256) */
 200     "dummy",     /*  9 (d128) */
 201     "dummy",     /* 10 (d64)  */
 202     "dummy",     /* 11 (d32)  */
 203     "coefdiv.w", /* 12 (d16)  */
 204     "coefdiv.z", /* 13 (d8)   */
 205     "coefdiv.y", /* 14 (d4)   */
 206     "coefdiv.x"  /* 15 (d2)   */
 207 };
 208
 209 static void pshader_get_write_mask(const DWORD output_reg, char *write_mask) {
 210     *write_mask = 0;
 211     if ((output_reg & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
 212         strcat(write_mask, ".");
 213         if (output_reg & WINED3DSP_WRITEMASK_0) strcat(write_mask, "r");
 214         if (output_reg & WINED3DSP_WRITEMASK_1) strcat(write_mask, "g");
 215         if (output_reg & WINED3DSP_WRITEMASK_2) strcat(write_mask, "b");
 216         if (output_reg & WINED3DSP_WRITEMASK_3) strcat(write_mask, "a");
 217     }
 218 }
 219
 220 /* TODO: merge with pixel shader */
 221 static void vshader_program_add_output_param_swizzle(const DWORD param, int is_color, char *hwLine) {
 222     /** operand output */
 223     if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
 224       strcat(hwLine, ".");
 225       if (param & WINED3DSP_WRITEMASK_0) { strcat(hwLine, "x"); }
 226       if (param & WINED3DSP_WRITEMASK_1) { strcat(hwLine, "y"); }
 227       if (param & WINED3DSP_WRITEMASK_2) { strcat(hwLine, "z"); }
 228       if (param & WINED3DSP_WRITEMASK_3) { strcat(hwLine, "w"); }
 229     }
 230 }
 231
 232 static void pshader_get_input_register_swizzle(const DWORD instr, char *swzstring) {
 233     static const char swizzle_reg_chars[] = "rgba";
 234     DWORD swizzle = (instr & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
 235     DWORD swizzle_x = swizzle & 0x03;
 236     DWORD swizzle_y = (swizzle >> 2) & 0x03;
 237     DWORD swizzle_z = (swizzle >> 4) & 0x03;
 238     DWORD swizzle_w = (swizzle >> 6) & 0x03;
 239     /**
 240      * swizzle bits fields:
 241      *  WWZZYYXX
 242      */
 243     *swzstring = 0;
 244     if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) != swizzle) {
 245         if (swizzle_x == swizzle_y &&
 246         swizzle_x == swizzle_z &&
 247         swizzle_x == swizzle_w) {
 248             sprintf(swzstring, ".%c", swizzle_reg_chars[swizzle_x]);
 249         } else {
 250             sprintf(swzstring, ".%c%c%c%c",
 251                 swizzle_reg_chars[swizzle_x],
 252                 swizzle_reg_chars[swizzle_y],
 253                 swizzle_reg_chars[swizzle_z],
 254                 swizzle_reg_chars[swizzle_w]);
 255         }
 256     }
 257 }
 258
 259 /* TODO: merge with pixel shader */
 260 static void vshader_program_add_input_param_swizzle(const DWORD param, int is_color, char *hwLine) {
 261     static const char swizzle_reg_chars_color_fix[] = "zyxw";
 262     static const char swizzle_reg_chars[] = "xyzw";
 263     const char* swizzle_regs = NULL;
 264     char  tmpReg[255];
 265
 266     /** operand input */
 267     DWORD swizzle = (param & WINED3DVS_SWIZZLE_MASK) >> WINED3DVS_SWIZZLE_SHIFT;
 268     DWORD swizzle_x = swizzle & 0x03;
 269     DWORD swizzle_y = (swizzle >> 2) & 0x03;
 270     DWORD swizzle_z = (swizzle >> 4) & 0x03;
 271     DWORD swizzle_w = (swizzle >> 6) & 0x03;
 272
 273     if (is_color) {
 274       swizzle_regs = swizzle_reg_chars_color_fix;
 275     } else {
 276       swizzle_regs = swizzle_reg_chars;
 277     }
 278
 279     /**
 280      * swizzle bits fields:
 281      *  WWZZYYXX
 282      */
 283     if ((WINED3DVS_NOSWIZZLE >> WINED3DVS_SWIZZLE_SHIFT) == swizzle) {
 284       if (is_color) {
 285         sprintf(tmpReg, ".%c%c%c%c",
 286                 swizzle_regs[swizzle_x],
 287                 swizzle_regs[swizzle_y],
 288                 swizzle_regs[swizzle_z],
 289                 swizzle_regs[swizzle_w]);
 290         strcat(hwLine, tmpReg);
 291       }
 292       return ;
 293     }
 294     if (swizzle_x == swizzle_y &&
 295         swizzle_x == swizzle_z &&
 296         swizzle_x == swizzle_w)
 297     {
 298       sprintf(tmpReg, ".%c", swizzle_regs[swizzle_x]);
 299       strcat(hwLine, tmpReg);
 300     } else {
 301       sprintf(tmpReg, ".%c%c%c%c",
 302               swizzle_regs[swizzle_x],
 303               swizzle_regs[swizzle_y],
 304               swizzle_regs[swizzle_z],
 305               swizzle_regs[swizzle_w]);
 306       strcat(hwLine, tmpReg);
 307     }
 308 }
 309
 310 static void pshader_get_register_name(
 311     const DWORD param, char* regstr) {
 312
 313     DWORD reg = param & WINED3DSP_REGNUM_MASK;
 314     DWORD regtype = shader_get_regtype(param);
 315
 316     switch (regtype) {
 317     case WINED3DSPR_TEMP:
 318         sprintf(regstr, "R%u", reg);
 319     break;
 320     case WINED3DSPR_INPUT:
 321         if (reg==0) {
 322             strcpy(regstr, "fragment.color.primary");
 323         } else {
 324             strcpy(regstr, "fragment.color.secondary");
 325         }
 326     break;
 327     case WINED3DSPR_CONST:
 328         sprintf(regstr, "C[%u]", reg);
 329     break;
 330     case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
 331         sprintf(regstr,"T%u", reg);
 332     break;
 333     case WINED3DSPR_COLOROUT:
 334         if (reg == 0)
 335             sprintf(regstr, "result.color");
 336         else {
 337             /* TODO: See GL_ARB_draw_buffers */
 338             FIXME("Unsupported write to render target %u\n", reg);
 339             sprintf(regstr, "unsupported_register");
 340         }
 341     break;
 342     case WINED3DSPR_DEPTHOUT:
 343         sprintf(regstr, "result.depth");
 344     break;
 345     case WINED3DSPR_ATTROUT:
 346         sprintf(regstr, "oD[%u]", reg);
 347     break;
 348     case WINED3DSPR_TEXCRDOUT:
 349         sprintf(regstr, "oT[%u]", reg);
 350     break;
 351     default:
 352         FIXME("Unhandled register name Type(%d)\n", regtype);
 353         sprintf(regstr, "unrecognized_register");
 354     break;
 355     }
 356 }
 357
 358 /* TODO: merge with pixel shader */
 359 static void vshader_program_add_param(SHADER_OPCODE_ARG *arg, const DWORD param, BOOL is_input, char *hwLine) {
 360
 361   IWineD3DVertexShaderImpl* This = (IWineD3DVertexShaderImpl*) arg->shader;
 362
 363   /* oPos, oFog and oPts in D3D */
 364   static const char* hwrastout_reg_names[] = { "TMP_OUT", "TMP_FOG", "result.pointsize" };
 365
 366   DWORD reg = param & WINED3DSP_REGNUM_MASK;
 367   DWORD regtype = shader_get_regtype(param);
 368   char  tmpReg[255];
 369   BOOL is_color = FALSE;
 370
 371   if ((param & WINED3DSP_SRCMOD_MASK) == WINED3DSPSM_NEG) {
 372       strcat(hwLine, " -");
 373   } else {
 374       strcat(hwLine, " ");
 375   }
 376
 377   switch (regtype) {
 378   case WINED3DSPR_TEMP:
 379     sprintf(tmpReg, "R%u", reg);
 380     strcat(hwLine, tmpReg);
 381     break;
 382   case WINED3DSPR_INPUT:
 383
 384     if (vshader_input_is_color((IWineD3DVertexShader*) This, reg))
 385         is_color = TRUE;
 386
 387     sprintf(tmpReg, "vertex.attrib[%u]", reg);
 388     strcat(hwLine, tmpReg);
 389     break;
 390   case WINED3DSPR_CONST:
 391     sprintf(tmpReg, "C[%s%u]", (param & WINED3DSHADER_ADDRMODE_RELATIVE) ? "A0.x + " : "", reg);
 392     strcat(hwLine, tmpReg);
 393     break;
 394   case WINED3DSPR_ADDR: /*case D3DSPR_TEXTURE:*/
 395     sprintf(tmpReg, "A%u", reg);
 396     strcat(hwLine, tmpReg);
 397     break;
 398   case WINED3DSPR_RASTOUT:
 399     sprintf(tmpReg, "%s", hwrastout_reg_names[reg]);
 400     strcat(hwLine, tmpReg);
 401     break;
 402   case WINED3DSPR_ATTROUT:
 403     if (reg==0) {
 404        strcat(hwLine, "result.color.primary");
 405     } else {
 406        strcat(hwLine, "result.color.secondary");
 407     }
 408     break;
 409   case WINED3DSPR_TEXCRDOUT:
 410     sprintf(tmpReg, "result.texcoord[%u]", reg);
 411     strcat(hwLine, tmpReg);
 412     break;
 413   default:
 414     FIXME("Unknown reg type %d %d\n", regtype, reg);
 415     strcat(hwLine, "unrecognized_register");
 416     break;
 417   }
 418
 419   if (!is_input) {
 420     vshader_program_add_output_param_swizzle(param, is_color, hwLine);
 421   } else {
 422     vshader_program_add_input_param_swizzle(param, is_color, hwLine);
 423   }
 424 }
 425
 426 static void pshader_gen_input_modifier_line (
 427     SHADER_BUFFER* buffer,
 428     const DWORD instr,
 429     int tmpreg,
 430     char *outregstr) {
 431
 432     /* Generate a line that does the input modifier computation and return the input register to use */
 433     char regstr[256];
 434     char swzstr[20];
 435     int insert_line;
 436
 437     /* Assume a new line will be added */
 438     insert_line = 1;
 439
 440     /* Get register name */
 441     pshader_get_register_name(instr, regstr);
 442     pshader_get_input_register_swizzle(instr, swzstr);
 443
 444     switch (instr & WINED3DSP_SRCMOD_MASK) {
 445     case WINED3DSPSM_NONE:
 446         sprintf(outregstr, "%s%s", regstr, swzstr);
 447         insert_line = 0;
 448         break;
 449     case WINED3DSPSM_NEG:
 450         sprintf(outregstr, "-%s%s", regstr, swzstr);
 451         insert_line = 0;
 452         break;
 453     case WINED3DSPSM_BIAS:
 454         shader_addline(buffer, "ADD T%c, %s, -coefdiv.x;\n", 'A' + tmpreg, regstr);
 455         break;
 456     case WINED3DSPSM_BIASNEG:
 457         shader_addline(buffer, "ADD T%c, -%s, coefdiv.x;\n", 'A' + tmpreg, regstr);
 458         break;
 459     case WINED3DSPSM_SIGN:
 460         shader_addline(buffer, "MAD T%c, %s, coefmul.x, -one.x;\n", 'A' + tmpreg, regstr);
 461         break;
 462     case WINED3DSPSM_SIGNNEG:
 463         shader_addline(buffer, "MAD T%c, %s, -coefmul.x, one.x;\n", 'A' + tmpreg, regstr);
 464         break;
 465     case WINED3DSPSM_COMP:
 466         shader_addline(buffer, "SUB T%c, one.x, %s;\n", 'A' + tmpreg, regstr);
 467         break;
 468     case WINED3DSPSM_X2:
 469         shader_addline(buffer, "ADD T%c, %s, %s;\n", 'A' + tmpreg, regstr, regstr);
 470         break;
 471     case WINED3DSPSM_X2NEG:
 472         shader_addline(buffer, "ADD T%c, -%s, -%s;\n", 'A' + tmpreg, regstr, regstr);
 473         break;
 474     case WINED3DSPSM_DZ:
 475         shader_addline(buffer, "RCP T%c, %s.z;\n", 'A' + tmpreg, regstr);
 476         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 477         break;
 478     case WINED3DSPSM_DW:
 479         shader_addline(buffer, "RCP T%c, %s.w;\n", 'A' + tmpreg, regstr);
 480         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 481         break;
 482     default:
 483         sprintf(outregstr, "%s%s", regstr, swzstr);
 484         insert_line = 0;
 485     }
 486
 487     /* Return modified or original register, with swizzle */
 488     if (insert_line)
 489         sprintf(outregstr, "T%c%s", 'A' + tmpreg, swzstr);
 490 }
 491
 492 inline static void pshader_gen_output_modifier_line(
 493     SHADER_BUFFER* buffer,
 494     int saturate,
 495     char *write_mask,
 496     int shift,
 497     char *regstr) {
 498
 499     /* Generate a line that does the output modifier computation */
 500     shader_addline(buffer, "MUL%s %s%s, %s, %s;\n", saturate ? "_SAT" : "",
 501         regstr, write_mask, regstr, shift_tab[shift]);
 502 }
 503
 504 void pshader_hw_cnd(SHADER_OPCODE_ARG* arg) {
 505
 506     SHADER_BUFFER* buffer = arg->buffer;
 507     char dst_wmask[20];
 508     char dst_name[50];
 509     char src_name[3][50];
 510
 511     /* FIXME: support output modifiers */
 512
 513     /* Handle output register */
 514     pshader_get_register_name(arg->dst, dst_name);
 515     pshader_get_write_mask(arg->dst, dst_wmask);
 516     strcat(dst_name, dst_wmask);
 517
 518     /* Generate input register names (with modifiers) */
 519     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src_name[0]);
 520     pshader_gen_input_modifier_line(buffer, arg->src[1], 1, src_name[1]);
 521     pshader_gen_input_modifier_line(buffer, arg->src[2], 2, src_name[2]);
 522
 523     shader_addline(buffer, "ADD TMP, -%s, coefdiv.x;\n", src_name[0]);
 524     shader_addline(buffer, "CMP %s, TMP, %s, %s;\n", dst_name, src_name[1], src_name[2]);
 525 }
 526
 527 void pshader_hw_cmp(SHADER_OPCODE_ARG* arg) {
 528
 529     SHADER_BUFFER* buffer = arg->buffer;
 530     char dst_wmask[20];
 531     char dst_name[50];
 532     char src_name[3][50];
 533
 534     /* FIXME: support output modifiers */
 535
 536     /* Handle output register */
 537     pshader_get_register_name(arg->dst, dst_name);
 538     pshader_get_write_mask(arg->dst, dst_wmask);
 539     strcat(dst_name, dst_wmask);
 540
 541     /* Generate input register names (with modifiers) */
 542     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src_name[0]);
 543     pshader_gen_input_modifier_line(buffer, arg->src[1], 1, src_name[1]);
 544     pshader_gen_input_modifier_line(buffer, arg->src[2], 2, src_name[2]);
 545
 546     shader_addline(buffer, "CMP %s, %s, %s, %s;\n", dst_name,
 547         src_name[0], src_name[2], src_name[1]);
 548 }
 549
 550 /* Map the opcode 1-to-1 to the GL code */
 551 void pshader_hw_map2gl(SHADER_OPCODE_ARG* arg) {
 552
 553      CONST SHADER_OPCODE* curOpcode = arg->opcode;
 554      SHADER_BUFFER* buffer = arg->buffer;
 555      DWORD dst = arg->dst;
 556      DWORD* src = arg->src;
 557
 558      unsigned int i;
 559      char tmpLine[256];
 560
 561      /* Output token related */
 562      char output_rname[256];
 563      char output_wmask[20];
 564      BOOL saturate = FALSE;
 565      BOOL centroid = FALSE;
 566      BOOL partialprecision = FALSE;
 567      DWORD shift;
 568
 569      strcpy(tmpLine, curOpcode->glname);
 570
 571      /* Process modifiers */
 572      if (0 != (dst & WINED3DSP_DSTMOD_MASK)) {
 573          DWORD mask = dst & WINED3DSP_DSTMOD_MASK;
 574
 575          saturate = mask & WINED3DSPDM_SATURATE;
 576          centroid = mask & WINED3DSPDM_MSAMPCENTROID;
 577          partialprecision = mask & WINED3DSPDM_PARTIALPRECISION;
 578          mask &= ~(WINED3DSPDM_MSAMPCENTROID | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_SATURATE);
 579          if (mask)
 580             FIXME("Unrecognized modifier(0x%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
 581
 582          if (centroid)
 583              FIXME("Unhandled modifier(0x%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
 584      }
 585      shift = (dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
 586
 587       /* Generate input and output registers */
 588       if (curOpcode->num_params > 0) {
 589           char operands[4][100];
 590
 591           /* Generate input register names (with modifiers) */
 592           for (i = 1; i < curOpcode->num_params; ++i)
 593               pshader_gen_input_modifier_line(buffer, src[i-1], i-1, operands[i]);
 594
 595           /* Handle output register */
 596           pshader_get_register_name(dst, output_rname);
 597           strcpy(operands[0], output_rname);
 598           pshader_get_write_mask(dst, output_wmask);
 599           strcat(operands[0], output_wmask);
 600
 601           if (saturate && (shift == 0))
 602              strcat(tmpLine, "_SAT");
 603           strcat(tmpLine, " ");
 604           strcat(tmpLine, operands[0]);
 605           for (i = 1; i < curOpcode->num_params; i++) {
 606               strcat(tmpLine, ", ");
 607               strcat(tmpLine, operands[i]);
 608           }
 609           strcat(tmpLine,";\n");
 610           shader_addline(buffer, tmpLine);
 611
 612           /* A shift requires another line. */
 613           if (shift != 0)
 614               pshader_gen_output_modifier_line(buffer, saturate, output_wmask, shift, output_rname);
 615       }
 616 }
 617
 618 void pshader_hw_tex(SHADER_OPCODE_ARG* arg) {
 619
 620     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 621     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
 622
 623     DWORD dst = arg->dst;
 624     DWORD* src = arg->src;
 625     SHADER_BUFFER* buffer = arg->buffer;
 626     DWORD hex_version = This->baseShader.hex_version;
 627
 628     char reg_dest[40];
 629     char reg_coord[40];
 630     const char *tex_type;
 631     DWORD reg_dest_code;
 632     DWORD reg_sampler_code;
 633     DWORD sampler_type;
 634
 635     /* All versions have a destination register */
 636     reg_dest_code = dst & WINED3DSP_REGNUM_MASK;
 637     pshader_get_register_name(dst, reg_dest);
 638
 639     /* 1.0-1.3: Use destination register as coordinate source.
 640        1.4+: Use provided coordinate source register. */
 641    if (hex_version < WINED3DPS_VERSION(1,4))
 642       strcpy(reg_coord, reg_dest);
 643    else
 644       pshader_gen_input_modifier_line(buffer, src[0], 0, reg_coord);
 645
 646   /* 1.0-1.4: Use destination register number as texture code.
 647      2.0+: Use provided sampler number as texure code. */
 648   if (hex_version < WINED3DPS_VERSION(2,0))
 649      reg_sampler_code = reg_dest_code;
 650   else
 651      reg_sampler_code = src[1] & WINED3DSP_REGNUM_MASK;
 652
 653   sampler_type = arg->reg_maps->samplers[reg_sampler_code] & WINED3DSP_TEXTURETYPE_MASK;
 654   switch(sampler_type) {
 655      case WINED3DSTT_1D:
 656          tex_type = "1D";
 657          break;
 658
 659      case WINED3DSTT_2D:
 660          tex_type = "2D";
 661          break;
 662
 663      case WINED3DSTT_VOLUME:
 664          tex_type = "3D";
 665          break;
 666
 667      case WINED3DSTT_CUBE:
 668          tex_type = "CUBE";
 669          break;
 670
 671      default:
 672          ERR("Unexpected texture type %d\n", sampler_type);
 673          tex_type = "2D";
 674   }
 675
 676   if (deviceImpl->stateBlock->textureState[reg_sampler_code][WINED3DTSS_TEXTURETRANSFORMFLAGS] & WINED3DTTFF_PROJECTED) {
 677       shader_addline(buffer, "TXP %s, %s, texture[%u], %s;\n",
 678           reg_dest, reg_coord, reg_sampler_code, tex_type);
 679   } else {
 680       shader_addline(buffer, "TEX %s, %s, texture[%u], %s;\n",
 681                      reg_dest, reg_coord, reg_sampler_code, tex_type);
 682   }
 683 }
 684
 685 void pshader_hw_texcoord(SHADER_OPCODE_ARG* arg) {
 686
 687     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 688     DWORD dst = arg->dst;
 689     DWORD* src = arg->src;
 690     SHADER_BUFFER* buffer = arg->buffer;
 691     DWORD hex_version = This->baseShader.hex_version;
 692
 693     char tmp[20];
 694     pshader_get_write_mask(dst, tmp);
 695     if (hex_version != WINED3DPS_VERSION(1,4)) {
 696         DWORD reg = dst & WINED3DSP_REGNUM_MASK;
 697         shader_addline(buffer, "MOV_SAT T%u%s, fragment.texcoord[%u];\n", reg, tmp, reg);
 698     } else {
 699         DWORD reg1 = dst & WINED3DSP_REGNUM_MASK;
 700         DWORD reg2 = src[0] & WINED3DSP_REGNUM_MASK;
 701         shader_addline(buffer, "MOV R%u%s, fragment.texcoord[%u];\n", reg1, tmp, reg2);
 702    }
 703 }
 704
 705 void pshader_hw_texreg2ar(SHADER_OPCODE_ARG* arg) {
 706
 707      SHADER_BUFFER* buffer = arg->buffer;
 708
 709      DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
 710      DWORD reg2 = arg->src[0] & WINED3DSP_REGNUM_MASK;
 711      shader_addline(buffer, "MOV TMP.r, T%u.a;\n", reg2);
 712      shader_addline(buffer, "MOV TMP.g, T%u.r;\n", reg2);
 713      shader_addline(buffer, "TEX T%u, TMP, texture[%u], 2D;\n", reg1, reg1);
 714 }
 715
 716 void pshader_hw_texreg2gb(SHADER_OPCODE_ARG* arg) {
 717
 718      SHADER_BUFFER* buffer = arg->buffer;
 719
 720      DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
 721      DWORD reg2 = arg->src[0] & WINED3DSP_REGNUM_MASK;
 722      shader_addline(buffer, "MOV TMP.r, T%u.g;\n", reg2);
 723      shader_addline(buffer, "MOV TMP.g, T%u.b;\n", reg2);
 724      shader_addline(buffer, "TEX T%u, TMP, texture[%u], 2D;\n", reg1, reg1);
 725 }
 726
 727 void pshader_hw_texbem(SHADER_OPCODE_ARG* arg) {
 728
 729      SHADER_BUFFER* buffer = arg->buffer;
 730      DWORD reg1 = arg->dst  & WINED3DSP_REGNUM_MASK;
 731      DWORD reg2 = arg->src[0] & WINED3DSP_REGNUM_MASK;
 732
 733      /* FIXME: Should apply the BUMPMAPENV matrix */
 734      shader_addline(buffer, "ADD TMP.rg, fragment.texcoord[%u], T%u;\n", reg1, reg2);
 735      shader_addline(buffer, "TEX T%u, TMP, texture[%u], 2D;\n", reg1, reg1);
 736 }
 737
 738 void pshader_hw_texm3x2pad(SHADER_OPCODE_ARG* arg) {
 739
 740     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
 741     SHADER_BUFFER* buffer = arg->buffer;
 742     char src0_name[50];
 743
 744     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src0_name);
 745     shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", reg, src0_name);
 746 }
 747
 748 void pshader_hw_texm3x2tex(SHADER_OPCODE_ARG* arg) {
 749
 750     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
 751     SHADER_BUFFER* buffer = arg->buffer;
 752     char src0_name[50];
 753
 754     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src0_name);
 755     shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", reg, src0_name);
 756     shader_addline(buffer, "TEX T%u, TMP, texture[%u], 2D;\n", reg, reg);
 757 }
 758
 759 void pshader_hw_texm3x3pad(SHADER_OPCODE_ARG* arg) {
 760
 761     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 762     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
 763     SHADER_BUFFER* buffer = arg->buffer;
 764     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
 765     char src0_name[50];
 766
 767     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src0_name);
 768     shader_addline(buffer, "DP3 TMP.%c, T%u, %s;\n", 'x' + current_state->current_row, reg, src0_name);
 769     current_state->texcoord_w[current_state->current_row++] = reg;
 770 }
 771
 772 void pshader_hw_texm3x3tex(SHADER_OPCODE_ARG* arg) {
 773
 774     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 775     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
 776     SHADER_BUFFER* buffer = arg->buffer;
 777     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
 778     char src0_name[50];
 779
 780     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src0_name);
 781     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
 782
 783     /* Cubemap textures will be more used than 3D ones. */
 784     shader_addline(buffer, "TEX T%u, TMP, texture[%u], CUBE;\n", reg, reg);
 785     current_state->current_row = 0;
 786 }
 787
 788 void pshader_hw_texm3x3vspec(SHADER_OPCODE_ARG* arg) {
 789
 790     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 791     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
 792     SHADER_BUFFER* buffer = arg->buffer;
 793     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
 794     char src0_name[50];
 795
 796     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src0_name);
 797     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
 798
 799     /* Construct the eye-ray vector from w coordinates */
 800     shader_addline(buffer, "MOV TMP2.x, fragment.texcoord[%u].w;\n", current_state->texcoord_w[0]);
 801     shader_addline(buffer, "MOV TMP2.y, fragment.texcoord[%u].w;\n", current_state->texcoord_w[1]);
 802     shader_addline(buffer, "MOV TMP2.z, fragment.texcoord[%u].w;\n", reg);
 803
 804     /* Calculate reflection vector (Assume normal is normalized): RF = 2*(N.E)*N -E */
 805     shader_addline(buffer, "DP3 TMP.w, TMP, TMP2;\n");
 806     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
 807     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -TMP2;\n");
 808
 809     /* Cubemap textures will be more used than 3D ones. */
 810     shader_addline(buffer, "TEX T%u, TMP, texture[%u], CUBE;\n", reg, reg);
 811     current_state->current_row = 0;
 812 }
 813
 814 void pshader_hw_texm3x3spec(SHADER_OPCODE_ARG* arg) {
 815
 816     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 817     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
 818     DWORD reg3 = arg->src[1] & WINED3DSP_REGNUM_MASK;
 819     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
 820     SHADER_BUFFER* buffer = arg->buffer;
 821     char src0_name[50];
 822
 823     pshader_gen_input_modifier_line(buffer, arg->src[0], 0, src0_name);
 824     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
 825
 826     /* Calculate reflection vector (Assume normal is normalized): RF = 2*(N.E)*N -E */
 827     shader_addline(buffer, "DP3 TMP.w, TMP, C[%u];\n", reg3);
 828     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
 829     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -C[%u];\n", reg3);
 830
 831     /* Cubemap textures will be more used than 3D ones. */
 832     shader_addline(buffer, "TEX T%u, TMP, texture[%u], CUBE;\n", reg, reg);
 833     current_state->current_row = 0;
 834 }
 835
 836 /** Handles transforming all WINED3DSIO_M?x? opcodes for
 837     Vertex shaders to ARB_vertex_program codes */
 838 void vshader_hw_mnxn(SHADER_OPCODE_ARG* arg) {
 839
 840     int i;
 841     int nComponents = 0;
 842     SHADER_OPCODE_ARG tmpArg;
 843
 844     memset(&tmpArg, 0, sizeof(SHADER_OPCODE_ARG));
 845
 846     /* Set constants for the temporary argument */
 847     tmpArg.shader      = arg->shader;
 848     tmpArg.buffer      = arg->buffer;
 849     tmpArg.src[0]      = arg->src[0];
 850     tmpArg.src_addr[0] = arg->src_addr[0];
 851     tmpArg.src_addr[1] = arg->src_addr[1];
 852     tmpArg.reg_maps = arg->reg_maps;
 853
 854     switch(arg->opcode->opcode) {
 855     case WINED3DSIO_M4x4:
 856         nComponents = 4;
 857         tmpArg.opcode = &IWineD3DVertexShaderImpl_shader_ins[WINED3DSIO_DP4];
 858         break;
 859     case WINED3DSIO_M4x3:
 860         nComponents = 3;
 861         tmpArg.opcode = &IWineD3DVertexShaderImpl_shader_ins[WINED3DSIO_DP4];
 862         break;
 863     case WINED3DSIO_M3x4:
 864         nComponents = 4;
 865         tmpArg.opcode = &IWineD3DVertexShaderImpl_shader_ins[WINED3DSIO_DP3];
 866         break;
 867     case WINED3DSIO_M3x3:
 868         nComponents = 3;
 869         tmpArg.opcode = &IWineD3DVertexShaderImpl_shader_ins[WINED3DSIO_DP3];
 870         break;
 871     case WINED3DSIO_M3x2:
 872         nComponents = 2;
 873         tmpArg.opcode = &IWineD3DVertexShaderImpl_shader_ins[WINED3DSIO_DP3];
 874         break;
 875     default:
 876         break;
 877     }
 878
 879     for (i = 0; i < nComponents; i++) {
 880         tmpArg.dst = ((arg->dst) & ~WINED3DSP_WRITEMASK_ALL)|(WINED3DSP_WRITEMASK_0<<i);
 881         tmpArg.src[1] = arg->src[1]+i;
 882         vshader_hw_map2gl(&tmpArg);
 883     }
 884 }
 885
 886 /* TODO: merge with pixel shader */
 887 /* Map the opcode 1-to-1 to the GL code */
 888 void vshader_hw_map2gl(SHADER_OPCODE_ARG* arg) {
 889
 890     CONST SHADER_OPCODE* curOpcode = arg->opcode;
 891     SHADER_BUFFER* buffer = arg->buffer;
 892     DWORD dst = arg->dst;
 893     DWORD* src = arg->src;
 894
 895     DWORD dst_regtype = shader_get_regtype(dst);
 896     char tmpLine[256];
 897     unsigned int i;
 898
 899     if ((curOpcode->opcode == WINED3DSIO_MOV && dst_regtype == WINED3DSPR_ADDR) || curOpcode->opcode == WINED3DSIO_MOVA)
 900         strcpy(tmpLine, "ARL");
 901     else
 902         strcpy(tmpLine, curOpcode->glname);
 903
 904     if (curOpcode->num_params > 0) {
 905         vshader_program_add_param(arg, dst, FALSE, tmpLine);
 906         for (i = 1; i < curOpcode->num_params; ++i) {
 907            strcat(tmpLine, ",");
 908            vshader_program_add_param(arg, src[i-1], TRUE, tmpLine);
 909         }
 910     }
 911    shader_addline(buffer, "%s;\n", tmpLine);
 912 }