git.oblomov.eu Git - wine/blob - dlls/wined3d/ati_fragment_shader.c

   1 /*
   2  * Fixed function pipeline replacement using GL_ATI_fragment_shader
   3  *
   4  * Copyright 2008 Stefan Dösinger(for CodeWeavers)
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  19  */
  20
  21 #include "config.h"
  22
  23 #include <math.h>
  24 #include <stdio.h>
  25
  26 #include "wined3d_private.h"
  27
  28 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  29
  30 /* Some private defines, Constant associations, etc
  31  * Env bump matrix and per stage constant should be independent,
  32  * a stage that bumpmaps can't read the per state constant
  33  */
  34 #define ATI_FFP_CONST_BUMPMAT(i) (GL_CON_0_ATI + i)
  35 #define ATI_FFP_CONST_CONSTANT0 GL_CON_0_ATI
  36 #define ATI_FFP_CONST_CONSTANT1 GL_CON_1_ATI
  37 #define ATI_FFP_CONST_CONSTANT2 GL_CON_2_ATI
  38 #define ATI_FFP_CONST_CONSTANT3 GL_CON_3_ATI
  39 #define ATI_FFP_CONST_CONSTANT4 GL_CON_4_ATI
  40 #define ATI_FFP_CONST_CONSTANT5 GL_CON_5_ATI
  41 #define ATI_FFP_CONST_TFACTOR   GL_CON_6_ATI
  42
  43 /* GL_ATI_fragment_shader specific fixed function pipeline description. "Inherits" from the common one */
  44 struct atifs_ffp_desc
  45 {
  46     struct ffp_desc parent;
  47     GLuint shader;
  48 };
  49
  50 struct atifs_private_data
  51 {
  52     struct shader_arb_priv parent;
  53     struct list fragment_shaders; /* A linked list to track fragment pipeline replacement shaders */
  54
  55 };
  56
  57 static const char *debug_dstmod(GLuint mod) {
  58     switch(mod) {
  59         case GL_NONE:               return "GL_NONE";
  60         case GL_2X_BIT_ATI:         return "GL_2X_BIT_ATI";
  61         case GL_4X_BIT_ATI:         return "GL_4X_BIT_ATI";
  62         case GL_8X_BIT_ATI:         return "GL_8X_BIT_ATI";
  63         case GL_HALF_BIT_ATI:       return "GL_HALF_BIT_ATI";
  64         case GL_QUARTER_BIT_ATI:    return "GL_QUARTER_BIT_ATI";
  65         case GL_EIGHTH_BIT_ATI:     return "GL_EIGHTH_BIT_ATI";
  66         case GL_SATURATE_BIT_ATI:   return "GL_SATURATE_BIT_ATI";
  67         default:                    return "Unexpected modifier\n";
  68     }
  69 }
  70
  71 static const char *debug_argmod(GLuint mod) {
  72     switch(mod) {
  73         case GL_NONE:
  74             return "GL_NONE";
  75
  76         case GL_2X_BIT_ATI:
  77             return "GL_2X_BIT_ATI";
  78         case GL_COMP_BIT_ATI:
  79             return "GL_COMP_BIT_ATI";
  80         case GL_NEGATE_BIT_ATI:
  81             return "GL_NEGATE_BIT_ATI";
  82         case GL_BIAS_BIT_ATI:
  83             return "GL_BIAS_BIT_ATI";
  84
  85         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI:
  86             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI";
  87         case GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI:
  88             return "GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI";
  89         case GL_2X_BIT_ATI | GL_BIAS_BIT_ATI:
  90             return "GL_2X_BIT_ATI | GL_BIAS_BIT_ATI";
  91         case GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI:
  92             return "GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI";
  93         case GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI:
  94             return "GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI";
  95         case GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
  96             return "GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
  97
  98         case GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
  99             return "GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 100         case GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 101             return "GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 102         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI:
 103             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI";
 104         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI:
 105             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI";
 106
 107         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 108             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 109
 110         default:
 111             return "Unexpected argmod combination\n";
 112     }
 113 }
 114 static const char *debug_register(GLuint reg) {
 115     switch(reg) {
 116         case GL_REG_0_ATI:                  return "GL_REG_0_ATI";
 117         case GL_REG_1_ATI:                  return "GL_REG_1_ATI";
 118         case GL_REG_2_ATI:                  return "GL_REG_2_ATI";
 119         case GL_REG_3_ATI:                  return "GL_REG_3_ATI";
 120         case GL_REG_4_ATI:                  return "GL_REG_4_ATI";
 121         case GL_REG_5_ATI:                  return "GL_REG_5_ATI";
 122
 123         case GL_CON_0_ATI:                  return "GL_CON_0_ATI";
 124         case GL_CON_1_ATI:                  return "GL_CON_1_ATI";
 125         case GL_CON_2_ATI:                  return "GL_CON_2_ATI";
 126         case GL_CON_3_ATI:                  return "GL_CON_3_ATI";
 127         case GL_CON_4_ATI:                  return "GL_CON_4_ATI";
 128         case GL_CON_5_ATI:                  return "GL_CON_5_ATI";
 129         case GL_CON_6_ATI:                  return "GL_CON_6_ATI";
 130         case GL_CON_7_ATI:                  return "GL_CON_7_ATI";
 131
 132         case GL_ZERO:                       return "GL_ZERO";
 133         case GL_ONE:                        return "GL_ONE";
 134         case GL_PRIMARY_COLOR:              return "GL_PRIMARY_COLOR";
 135         case GL_SECONDARY_INTERPOLATOR_ATI: return "GL_SECONDARY_INTERPOLATOR_ATI";
 136
 137         default:                            return "Unknown register\n";
 138     }
 139 }
 140
 141 static const char *debug_swizzle(GLuint swizzle) {
 142     switch(swizzle) {
 143         case GL_SWIZZLE_STR_ATI:        return "GL_SWIZZLE_STR_ATI";
 144         case GL_SWIZZLE_STQ_ATI:        return "GL_SWIZZLE_STQ_ATI";
 145         case GL_SWIZZLE_STR_DR_ATI:     return "GL_SWIZZLE_STR_DR_ATI";
 146         case GL_SWIZZLE_STQ_DQ_ATI:     return "GL_SWIZZLE_STQ_DQ_ATI";
 147         default:                        return "unknown swizzle";
 148     }
 149 }
 150
 151 #define GLINFO_LOCATION (*gl_info)
 152 static GLuint register_for_arg(DWORD arg, WineD3D_GL_Info *gl_info, unsigned int stage, GLuint *mod, GLuint tmparg) {
 153     GLenum ret;
 154
 155     if(mod) *mod = GL_NONE;
 156     if(arg == 0xFFFFFFFF) return -1; /* This is the marker for unused registers */
 157
 158     switch(arg & WINED3DTA_SELECTMASK) {
 159         case WINED3DTA_DIFFUSE:
 160             ret = GL_PRIMARY_COLOR;
 161             break;
 162
 163         case WINED3DTA_CURRENT:
 164             /* Note that using GL_REG_0_ATI for the passed on register is safe because
 165              * texture0 is read at stage0, so in the worst case it is read in the
 166              * instruction writing to reg0. Afterwards texture0 is not used any longer.
 167              * If we're reading from current
 168              */
 169             if(stage == 0) {
 170                 ret = GL_PRIMARY_COLOR;
 171             } else {
 172                 ret = GL_REG_0_ATI;
 173             }
 174             break;
 175
 176         case WINED3DTA_TEXTURE:
 177             ret = GL_REG_0_ATI + stage;
 178             break;
 179
 180         case WINED3DTA_TFACTOR:
 181             ret = ATI_FFP_CONST_TFACTOR;
 182             break;
 183
 184         case WINED3DTA_SPECULAR:
 185             ret = GL_SECONDARY_INTERPOLATOR_ATI;
 186             break;
 187
 188         case WINED3DTA_TEMP:
 189             ret = tmparg;
 190             break;
 191
 192         case WINED3DTA_CONSTANT:
 193             FIXME("Unhandled source argument WINED3DTA_TEMP\n");
 194             ret = GL_CON_0_ATI;
 195             break;
 196
 197         default:
 198             FIXME("Unknown source argument %d\n", arg);
 199             ret = GL_ZERO;
 200     }
 201
 202     if(arg & WINED3DTA_COMPLEMENT) {
 203         if(mod) *mod |= GL_COMP_BIT_ATI;
 204     }
 205     if(arg & WINED3DTA_ALPHAREPLICATE) {
 206         FIXME("Unhandled read modifier WINED3DTA_ALPHAREPLICATE\n");
 207     }
 208     return ret;
 209 }
 210
 211 static GLuint find_tmpreg(struct texture_stage_op op[MAX_TEXTURES]) {
 212     int lowest_read = -1;
 213     int lowest_write = -1;
 214     int i;
 215     BOOL tex_used[MAX_TEXTURES];
 216
 217     memset(tex_used, 0, sizeof(tex_used));
 218     for(i = 0; i < MAX_TEXTURES; i++) {
 219         if(op[i].cop == WINED3DTOP_DISABLE) {
 220             break;
 221         }
 222
 223         if(lowest_read == -1 &&
 224           (op[i].carg1 == WINED3DTA_TEMP || op[i].carg2 == WINED3DTA_TEMP || op[i].carg0 == WINED3DTA_TEMP ||
 225            op[i].aarg1 == WINED3DTA_TEMP || op[i].aarg2 == WINED3DTA_TEMP || op[i].aarg0 == WINED3DTA_TEMP)) {
 226             lowest_read = i;
 227         }
 228
 229         if(lowest_write == -1 && op[i].dst == WINED3DTA_TEMP) {
 230             lowest_write = i;
 231         }
 232
 233         if(op[i].carg1 == WINED3DTA_TEXTURE || op[i].carg2 == WINED3DTA_TEXTURE || op[i].carg0 == WINED3DTA_TEXTURE ||
 234            op[i].aarg1 == WINED3DTA_TEXTURE || op[i].aarg2 == WINED3DTA_TEXTURE || op[i].aarg0 == WINED3DTA_TEXTURE) {
 235             tex_used[i] = TRUE;
 236         }
 237     }
 238
 239     /* Temp reg not read? We don't need it, return GL_NONE */
 240     if(lowest_read == -1) return GL_NONE;
 241
 242     if(lowest_write >= lowest_read) {
 243         FIXME("Temp register read before being written\n");
 244     }
 245
 246     if(lowest_write == -1) {
 247         /* This needs a test. Maybe we are supposed to return 0.0/0.0/0.0/0.0, or fail drawprim, or whatever */
 248         FIXME("Temp register read without being written\n");
 249         return GL_REG_1_ATI;
 250     } else if(lowest_write >= 1) {
 251         /* If we're writing to the temp reg at earliest in stage 1, we can use register 1 for the temp result.
 252          * there may be texture data stored in reg 1, but we do not need it any longer since stage 1 already
 253          * read it
 254          */
 255         return GL_REG_1_ATI;
 256     } else {
 257         /* Search for a free texture register. We have 6 registers available. GL_REG_0_ATI is already used
 258          * for the regular result
 259          */
 260         for(i = 1; i < 6; i++) {
 261             if(!tex_used[i]) {
 262                 return GL_REG_0_ATI + i;
 263             }
 264         }
 265         /* What to do here? Report it in ValidateDevice? */
 266         FIXME("Could not find a register for the temporary register\n");
 267         return 0;
 268     }
 269 }
 270
 271 static GLuint gen_ati_shader(struct texture_stage_op op[MAX_TEXTURES], WineD3D_GL_Info *gl_info) {
 272     GLuint ret = GL_EXTCALL(glGenFragmentShadersATI(1));
 273     unsigned int stage;
 274     GLuint arg0, arg1, arg2, extrarg;
 275     GLuint dstmod, argmod0, argmod1, argmod2, argmodextra;
 276     GLuint swizzle;
 277     GLuint tmparg = find_tmpreg(op);
 278     GLuint dstreg;
 279
 280     if(!ret) {
 281         ERR("Failed to generate a GL_ATI_fragment_shader shader id\n");
 282         return 0;
 283     }
 284     GL_EXTCALL(glBindFragmentShaderATI(ret));
 285     checkGLcall("GL_EXTCALL(glBindFragmentShaderATI(ret))");
 286
 287     TRACE("glBeginFragmentShaderATI()\n");
 288     GL_EXTCALL(glBeginFragmentShaderATI());
 289     checkGLcall("GL_EXTCALL(glBeginFragmentShaderATI())");
 290
 291     /* Pass 1: Generate sampling instructions for perturbation maps */
 292       for(stage = 0; stage < GL_LIMITS(textures); stage++) {
 293         if(op[stage].cop == WINED3DTOP_DISABLE) break;
 294         if(op[stage].cop != WINED3DTOP_BUMPENVMAP &&
 295            op[stage].cop != WINED3DTOP_BUMPENVMAPLUMINANCE) continue;
 296
 297         TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, GL_SWIZZLE_STR_ATI)\n",
 298               stage, stage);
 299         GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 300                    GL_TEXTURE0_ARB + stage,
 301                    GL_SWIZZLE_STR_ATI));
 302         TRACE("glPassTexCoordATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, GL_SWIZZLE_STR_ATI)\n",
 303               stage + 1, stage + 1);
 304         GL_EXTCALL(glPassTexCoordATI(GL_REG_0_ATI + stage + 1,
 305                    GL_TEXTURE0_ARB + stage + 1,
 306                    GL_SWIZZLE_STR_ATI));
 307
 308         /* We need GL_REG_5_ATI as a temporary register to swizzle the bump matrix. So we run into
 309          * issues if we're bump mapping on stage 4 or 5
 310          */
 311         if(stage >= 4) {
 312             FIXME("Bump mapping in stage %d\n", stage);
 313         }
 314     }
 315
 316     /* Pass 2: Generate perturbation calculations */
 317     for(stage = 0; stage < GL_LIMITS(textures); stage++) {
 318         if(op[stage].cop == WINED3DTOP_DISABLE) break;
 319         if(op[stage].cop != WINED3DTOP_BUMPENVMAP &&
 320            op[stage].cop != WINED3DTOP_BUMPENVMAPLUMINANCE) continue;
 321
 322         /* Nice thing, we get the color correction for free :-) */
 323         if(op[stage].color_correction == WINED3DFMT_V8U8) {
 324             argmodextra = GL_2X_BIT_ATI | GL_BIAS_BIT_ATI;
 325         } else {
 326             argmodextra = 0;
 327         }
 328
 329         TRACE("glColorFragmentOp3ATI(GL_DOT2_ADD_ATI, GL_REG_%d_ATI, GL_RED_BIT_ATI, GL_NONE, GL_REG_%d_ATI, GL_NONE, %s, ATI_FFP_CONST_BUMPMAT(%d), GL_NONE, GL_NONE, GL_REG_%d_ATI, GL_RED, GL_NONE)\n",
 330               stage + 1, stage, debug_argmod(argmodextra), stage, stage + 1);
 331         GL_EXTCALL(glColorFragmentOp3ATI(GL_DOT2_ADD_ATI, GL_REG_0_ATI + stage + 1, GL_RED_BIT_ATI, GL_NONE,
 332                                          GL_REG_0_ATI + stage, GL_NONE, argmodextra,
 333                                          ATI_FFP_CONST_BUMPMAT(stage), GL_NONE, GL_2X_BIT_ATI | GL_BIAS_BIT_ATI,
 334                                          GL_REG_0_ATI + stage + 1, GL_RED, GL_NONE));
 335
 336         /* FIXME: How can I make GL_DOT2_ADD_ATI read the factors from blue and alpha? It defaults to red and green,
 337          * and it is fairly easy to make it read GL_BLUE or BL_ALPHA, but I can't get an R * B + G * A. So we're wasting
 338          * one register and two instructions in this pass for a simple swizzling operation.
 339          * For starters it might be good enough to merge the two movs into one, but even that isn't possible :-(
 340          *
 341          * NOTE: GL_BLUE | GL_ALPHA is not possible. It doesn't throw a compilation error, but an OR operation on the
 342          * constants doesn't make sense, considering their values.
 343          */
 344         TRACE("glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_5_ATI, GL_RED_BIT_ATI, GL_NONE, ATI_FFP_CONST_BUMPMAT(%d), GL_BLUE, GL_NONE)\n", stage);
 345         GL_EXTCALL(glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_5_ATI, GL_RED_BIT_ATI, GL_NONE,
 346                                          ATI_FFP_CONST_BUMPMAT(stage), GL_BLUE, GL_NONE));
 347         TRACE("glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_5_ATI, GL_GREEN_BIT_ATI, GL_NONE, ATI_FFP_CONST_BUMPMAT(%d), GL_ALPHA, GL_NONE)\n", stage);
 348         GL_EXTCALL(glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_5_ATI, GL_GREEN_BIT_ATI, GL_NONE,
 349                                         ATI_FFP_CONST_BUMPMAT(stage), GL_ALPHA, GL_NONE));
 350         TRACE("glColorFragmentOp3ATI(GL_DOT2_ADD_ATI, GL_REG_%d_ATI, GL_GREEN_BIT_ATI, GL_NONE, GL_REG_%d_ATI, GL_NONE, %s, GL_REG_5_ATI, GL_NONE, GL_NONE, GL_REG_%d_ATI, GL_GREEN, GL_NONE)\n",
 351               stage + 1, stage, debug_argmod(argmodextra), stage + 1);
 352         GL_EXTCALL(glColorFragmentOp3ATI(GL_DOT2_ADD_ATI, GL_REG_0_ATI + stage + 1, GL_GREEN_BIT_ATI, GL_NONE,
 353                                          GL_REG_0_ATI + stage, GL_NONE, argmodextra,
 354                                          GL_REG_5_ATI, GL_NONE, GL_2X_BIT_ATI | GL_BIAS_BIT_ATI,
 355                                          GL_REG_0_ATI + stage + 1, GL_GREEN, GL_NONE));
 356     }
 357
 358     /* Pass 3: Generate sampling instructions for regular textures */
 359     for(stage = 0; stage < GL_LIMITS(textures); stage++) {
 360         if(op[stage].cop == WINED3DTOP_DISABLE) {
 361             break;
 362         }
 363
 364         if(op[stage].projected == proj_none) {
 365             swizzle = GL_SWIZZLE_STR_ATI;
 366         } else if(op[stage].projected == proj_count3) {
 367             /* TODO: D3DTTFF_COUNT3 | D3DTTFF_PROJECTED would be GL_SWIZZLE_STR_DR_ATI.
 368              * However, the FFP vertex processing texture transform matrix handler does
 369              * some transformations in the texture matrix which makes the 3rd coordinate
 370              * arrive in Q, not R in that case. This is needed for opengl fixed function
 371              * fragment processing which always divides by Q. In this backend we can
 372              * handle that properly and be compatible with vertex shader output and avoid
 373              * side effects of the texture matrix games
 374              */
 375             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 376         } else {
 377             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 378         }
 379
 380         if((op[stage].carg0 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 381            (op[stage].carg1 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 382            (op[stage].carg2 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 383            (op[stage].aarg0 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 384            (op[stage].aarg1 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 385            (op[stage].aarg2 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 386             op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) {
 387
 388             if(stage > 0 &&
 389                (op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
 390                 op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
 391                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_REG_%d_ATI, GL_SWIZZLE_STR_ATI)\n",
 392                       stage, stage);
 393                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 394                            GL_REG_0_ATI + stage,
 395                            GL_SWIZZLE_STR_ATI));
 396             } else {
 397                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, %s)\n",
 398                     stage, stage, debug_swizzle(swizzle));
 399                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 400                                         GL_TEXTURE0_ARB + stage,
 401                                         swizzle));
 402             }
 403         }
 404     }
 405
 406     /* Pass 4: Generate the arithmetic instructions */
 407     for(stage = 0; stage < MAX_TEXTURES; stage++) {
 408         if(op[stage].cop == WINED3DTOP_DISABLE) {
 409             if(stage == 0) {
 410                 /* Handle complete texture disabling gracefully */
 411                 TRACE("glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE, GL_PRIMARY_COLOR, GL_NONE, GL_NONE)\n");
 412                 GL_EXTCALL(glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 413                                                  GL_PRIMARY_COLOR, GL_NONE, GL_NONE));
 414                 TRACE("glAlphaFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_PRIMARY_COLOR, GL_NONE, GL_NONE)\n");
 415                 GL_EXTCALL(glAlphaFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE,
 416                                                  GL_PRIMARY_COLOR, GL_NONE, GL_NONE));
 417             }
 418             break;
 419         }
 420
 421         if(op[stage].dst == WINED3DTA_TEMP) {
 422             /* If we're writing to D3DTA_TEMP, but never reading from it we don't have to write there in the first place.
 423              * skip the entire stage, this saves some GPU time
 424              */
 425             if(tmparg == GL_NONE) continue;
 426
 427             dstreg = tmparg;
 428         } else {
 429             dstreg = GL_REG_0_ATI;
 430         }
 431
 432         arg0 = register_for_arg(op[stage].carg0, gl_info, stage, &argmod0, tmparg);
 433         arg1 = register_for_arg(op[stage].carg1, gl_info, stage, &argmod1, tmparg);
 434         arg2 = register_for_arg(op[stage].carg2, gl_info, stage, &argmod2, tmparg);
 435         dstmod = GL_NONE;
 436         argmodextra = GL_NONE;
 437         extrarg = GL_NONE;
 438
 439         switch(op[stage].cop) {
 440             case WINED3DTOP_SELECTARG2:
 441                 arg1 = arg2;
 442                 argmod1 = argmod2;
 443             case WINED3DTOP_SELECTARG1:
 444                 TRACE("glColorFragmentOp1ATI(GL_MOV_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s)\n",
 445                       debug_register(dstreg), debug_register(arg1), debug_argmod(argmod1));
 446                 GL_EXTCALL(glColorFragmentOp1ATI(GL_MOV_ATI, dstreg, GL_NONE, GL_NONE,
 447                                                  arg1, GL_NONE, argmod1));
 448                 break;
 449
 450             case WINED3DTOP_MODULATE4X:
 451                 if(dstmod == GL_NONE) dstmod = GL_4X_BIT_ATI;
 452             case WINED3DTOP_MODULATE2X:
 453                 if(dstmod == GL_NONE) dstmod = GL_2X_BIT_ATI;
 454             case WINED3DTOP_MODULATE:
 455                 TRACE("glColorFragmentOp2ATI(GL_MUL_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 456                       debug_register(dstreg), debug_dstmod(dstmod),
 457                       debug_register(arg1), debug_argmod(argmod1),
 458                       debug_register(arg2), debug_argmod(argmod2));
 459                 GL_EXTCALL(glColorFragmentOp2ATI(GL_MUL_ATI, dstreg, GL_NONE, dstmod,
 460                                                  arg1, GL_NONE, argmod1,
 461                                                  arg2, GL_NONE, argmod2));
 462                 break;
 463
 464             case WINED3DTOP_ADDSIGNED2X:
 465                 dstmod = GL_2X_BIT_ATI;
 466             case WINED3DTOP_ADDSIGNED:
 467                 argmodextra = GL_BIAS_BIT_ATI;
 468             case WINED3DTOP_ADD:
 469                 TRACE("glColorFragmentOp2ATI(GL_ADD_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 470                       debug_register(dstreg), debug_dstmod(dstmod),
 471                       debug_register(arg1), debug_argmod(argmod1),
 472                       debug_register(arg2), debug_argmod(argmodextra | argmod2));
 473                 GL_EXTCALL(glColorFragmentOp2ATI(GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, dstmod,
 474                                                  arg1, GL_NONE, argmod1,
 475                                                  arg2, GL_NONE, argmodextra | argmod2));
 476                 break;
 477
 478             case WINED3DTOP_SUBTRACT:
 479                 TRACE("glColorFragmentOp2ATI(GL_SUB_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 480                       debug_register(dstreg), debug_dstmod(dstmod),
 481                       debug_register(arg1), debug_argmod(argmod1),
 482                       debug_register(arg2), debug_argmod(argmod2));
 483                 GL_EXTCALL(glColorFragmentOp2ATI(GL_SUB_ATI, dstreg, GL_NONE, dstmod,
 484                                                  arg1, GL_NONE, argmod1,
 485                                                  arg2, GL_NONE, argmod2));
 486                 break;
 487
 488             case WINED3DTOP_ADDSMOOTH:
 489                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 490                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 491                       debug_register(dstreg),
 492                       debug_register(arg2), debug_argmod(argmod2),
 493                       debug_register(arg1), debug_argmod(argmodextra),
 494                       debug_register(arg1), debug_argmod(argmod1));
 495                 /* Dst = arg1 + * arg2(1 -arg 1)
 496                  *     = arg2 * (1 - arg1) + arg1
 497                  */
 498                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 499                                                  arg2, GL_NONE, argmod2,
 500                                                  arg1, GL_NONE, argmodextra,
 501                                                  arg1, GL_NONE, argmod1));
 502                 break;
 503
 504             case WINED3DTOP_BLENDCURRENTALPHA:
 505                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, -1);
 506             case WINED3DTOP_BLENDFACTORALPHA:
 507                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, -1);
 508             case WINED3DTOP_BLENDTEXTUREALPHA:
 509                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, -1);
 510             case WINED3DTOP_BLENDDIFFUSEALPHA:
 511                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, -1);
 512                 TRACE("glColorFragmentOp3ATI(GL_LERP_ATI, %s, GL_NONE, GL_NONE, %s, GL_ALPHA, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 513                       debug_register(dstreg),
 514                       debug_register(extrarg),
 515                       debug_register(arg1), debug_argmod(argmod1),
 516                       debug_register(arg2), debug_argmod(argmod2));
 517                 GL_EXTCALL(glColorFragmentOp3ATI(GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 518                                                  extrarg, GL_ALPHA, GL_NONE,
 519                                                  arg1, GL_NONE, argmod1,
 520                                                  arg2, GL_NONE, argmod2));
 521                 break;
 522
 523             case WINED3DTOP_BLENDTEXTUREALPHAPM:
 524                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, -1);
 525                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s, %s, GL_ALPHA, GL_COMP_BIT_ATI, %s, GL_NONE, %s)\n",
 526                       debug_register(dstreg),
 527                       debug_register(arg2), debug_argmod(argmod2),
 528                       debug_register(arg0),
 529                       debug_register(arg1), debug_argmod(argmod1));
 530                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 531                                                  arg2, GL_NONE,  argmod2,
 532                                                  arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 533                                                  arg1, GL_NONE,  argmod1));
 534                 break;
 535
 536             /* D3DTOP_PREMODULATE ???? */
 537
 538             case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
 539                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 540             case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
 541                 if(!argmodextra) argmodextra = argmod1;
 542                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s, %s, GL_ALPHA, %s, %s, GL_NONE, %s)\n",
 543                       debug_register(dstreg),
 544                       debug_register(arg2), debug_argmod(argmod2),
 545                       debug_register(arg1), debug_argmod(argmodextra), debug_register(arg1), debug_argmod(arg1));
 546                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 547                                                  arg2, GL_NONE,  argmod2,
 548                                                  arg1, GL_ALPHA, argmodextra,
 549                                                  arg1, GL_NONE,  argmod1));
 550                 break;
 551
 552             case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
 553                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 554             case WINED3DTOP_MODULATECOLOR_ADDALPHA:
 555                 if(!argmodextra) argmodextra = argmod1;
 556                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_ALPHA, %s)\n",
 557                       debug_register(dstreg),
 558                       debug_register(arg2), debug_argmod(argmod2),
 559                       debug_register(arg1), debug_argmod(argmodextra),
 560                       debug_register(arg1), debug_argmod(argmod1));
 561                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 562                                                  arg2, GL_NONE,  argmod2,
 563                                                  arg1, GL_NONE,  argmodextra,
 564                                                  arg1, GL_ALPHA, argmod1));
 565                 break;
 566
 567             case WINED3DTOP_DOTPRODUCT3:
 568                 TRACE("glColorFragmentOp2ATI(GL_DOT3_ATI, %s, GL_NONE, GL_4X_BIT_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 569                       debug_register(dstreg),
 570                       debug_register(arg1), debug_argmod(argmod1 | GL_BIAS_BIT_ATI),
 571                       debug_register(arg2), debug_argmod(argmod2 | GL_BIAS_BIT_ATI));
 572                 GL_EXTCALL(glColorFragmentOp2ATI(GL_DOT3_ATI, dstreg, GL_NONE, GL_4X_BIT_ATI,
 573                                                  arg1, GL_NONE, argmod1 | GL_BIAS_BIT_ATI,
 574                                                  arg2, GL_NONE, argmod2 | GL_BIAS_BIT_ATI));
 575                 break;
 576
 577             case WINED3DTOP_MULTIPLYADD:
 578                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 579                       debug_register(dstreg),
 580                       debug_register(arg0), debug_argmod(argmod0),
 581                       debug_register(arg2), debug_argmod(argmod2),
 582                       debug_register(arg1), debug_argmod(argmod1));
 583                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 584                                                  arg0, GL_NONE, argmod0,
 585                                                  arg2, GL_NONE, argmod2,
 586                                                  arg1, GL_NONE, argmod1));
 587                 break;
 588
 589             case WINED3DTOP_LERP:
 590                 TRACE("glColorFragmentOp3ATI(GL_LERP_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 591                       debug_register(dstreg),
 592                       debug_register(arg1), debug_argmod(argmod1),
 593                       debug_register(arg2), debug_argmod(argmod2),
 594                       debug_register(arg0), debug_argmod(argmod0));
 595                 GL_EXTCALL(glColorFragmentOp3ATI(GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 596                                                  arg1, GL_NONE, argmod1,
 597                                                  arg2, GL_NONE, argmod2,
 598                                                  arg0, GL_NONE, argmod0));
 599                 break;
 600
 601             case WINED3DTOP_BUMPENVMAP:
 602             case WINED3DTOP_BUMPENVMAPLUMINANCE:
 603                 /* Those are handled in the first pass of the shader(generation pass 1 and 2) alraedy */
 604                 break;
 605
 606             default: FIXME("Unhandled color operation %d on stage %d\n", op[stage].cop, stage);
 607         }
 608
 609         arg0 = register_for_arg(op[stage].aarg0, gl_info, stage, &argmod0, tmparg);
 610         arg1 = register_for_arg(op[stage].aarg1, gl_info, stage, &argmod1, tmparg);
 611         arg2 = register_for_arg(op[stage].aarg2, gl_info, stage, &argmod2, tmparg);
 612         dstmod = GL_NONE;
 613         argmodextra = GL_NONE;
 614         extrarg = GL_NONE;
 615
 616         switch(op[stage].aop) {
 617             case WINED3DTOP_DISABLE:
 618                 /* Get the primary color to the output if on stage 0, otherwise leave register 0 untouched */
 619                 if(stage == 0) {
 620                     TRACE("glAlphaFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_PRIMARY_COLOR, GL_NONE, GL_NONE)\n");
 621                     GL_EXTCALL(glAlphaFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE,
 622                                GL_PRIMARY_COLOR, GL_NONE, GL_NONE));
 623                 }
 624                 break;
 625
 626             case WINED3DTOP_SELECTARG2:
 627                 arg1 = arg2;
 628                 argmod1 = argmod2;
 629             case WINED3DTOP_SELECTARG1:
 630                 TRACE("glAlphaFragmentOp1ATI(GL_MOV_ATI, %s,          GL_NONE, %s, GL_NONE, %s)\n",
 631                       debug_register(dstreg),
 632                       debug_register(arg1), debug_argmod(argmod1));
 633                 GL_EXTCALL(glAlphaFragmentOp1ATI(GL_MOV_ATI, dstreg, GL_NONE,
 634                                                  arg1, GL_NONE, argmod1));
 635                 break;
 636
 637             case WINED3DTOP_MODULATE4X:
 638                 if(dstmod == GL_NONE) dstmod = GL_4X_BIT_ATI;
 639             case WINED3DTOP_MODULATE2X:
 640                 if(dstmod == GL_NONE) dstmod = GL_2X_BIT_ATI;
 641             case WINED3DTOP_MODULATE:
 642                 TRACE("glAlphaFragmentOp2ATI(GL_MUL_ATI, %s,          %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 643                       debug_register(dstreg), debug_dstmod(dstmod),
 644                       debug_register(arg1), debug_argmod(argmod1),
 645                       debug_register(arg2), debug_argmod(argmod2));
 646                 GL_EXTCALL(glAlphaFragmentOp2ATI(GL_MUL_ATI, dstreg, dstmod,
 647                                                  arg1, GL_NONE, argmod1,
 648                                                  arg2, GL_NONE, argmod2));
 649                 break;
 650
 651             case WINED3DTOP_ADDSIGNED2X:
 652                 dstmod = GL_2X_BIT_ATI;
 653             case WINED3DTOP_ADDSIGNED:
 654                 argmodextra = GL_BIAS_BIT_ATI;
 655             case WINED3DTOP_ADD:
 656                 TRACE("glAlphaFragmentOp2ATI(GL_ADD_ATI, %s,          %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 657                       debug_register(dstreg), debug_dstmod(dstmod),
 658                       debug_register(arg1), debug_argmod(argmod1),
 659                       debug_register(arg2), debug_argmod(argmodextra | argmod2));
 660                 GL_EXTCALL(glAlphaFragmentOp2ATI(GL_ADD_ATI, dstreg, dstmod,
 661                                                  arg1, GL_NONE, argmod1,
 662                                                  arg2, GL_NONE, argmodextra | argmod2));
 663                 break;
 664
 665             case WINED3DTOP_SUBTRACT:
 666                 TRACE("glAlphaFragmentOp2ATI(GL_SUB_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 667                       debug_register(dstreg), debug_dstmod(dstmod),
 668                       debug_register(arg1), debug_argmod(argmod1),
 669                       debug_register(arg2), debug_argmod(argmod2));
 670                 GL_EXTCALL(glAlphaFragmentOp2ATI(GL_SUB_ATI, dstreg, dstmod,
 671                                                  arg1, GL_NONE, argmod1,
 672                                                  arg2, GL_NONE, argmod2));
 673                 break;
 674
 675             case WINED3DTOP_ADDSMOOTH:
 676                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 677                 TRACE("glAlphaFragmentOp3ATI(GL_MAD_ATI, %s,          GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 678                       debug_register(dstreg),
 679                       debug_register(arg2), debug_argmod(argmod2),
 680                       debug_register(arg1), debug_argmod(argmodextra),
 681                       debug_register(arg1), debug_argmod(argmod1));
 682                 /* Dst = arg1 + * arg2(1 -arg 1)
 683                  *     = arg2 * (1 - arg1) + arg1
 684                  */
 685                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE,
 686                                                  arg2, GL_NONE, argmod2,
 687                                                  arg1, GL_NONE, argmodextra,
 688                                                  arg1, GL_NONE, argmod1));
 689                 break;
 690
 691             case WINED3DTOP_BLENDCURRENTALPHA:
 692                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, -1);
 693             case WINED3DTOP_BLENDFACTORALPHA:
 694                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, -1);
 695             case WINED3DTOP_BLENDTEXTUREALPHA:
 696                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, -1);
 697             case WINED3DTOP_BLENDDIFFUSEALPHA:
 698                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, -1);
 699                 TRACE("glAlphaFragmentOp3ATI(GL_LERP_ATI, %s,          GL_NONE, %s, GL_ALPHA, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 700                       debug_register(dstreg),
 701                       debug_register(extrarg),
 702                       debug_register(arg1), debug_argmod(argmod1),
 703                       debug_register(arg2), debug_argmod(argmod2));
 704                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_LERP_ATI, dstreg, GL_NONE,
 705                                                  extrarg, GL_ALPHA, GL_NONE,
 706                                                  arg1, GL_NONE, argmod1,
 707                                                  arg2, GL_NONE, argmod2));
 708                 break;
 709
 710             case WINED3DTOP_BLENDTEXTUREALPHAPM:
 711                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, -1);
 712                 TRACE("glAlphaFragmentOp3ATI(GL_MAD_ATI, %s,          GL_NONE, %s, GL_NONE, %s, %s, GL_ALPHA, GL_COMP_BIT_ATI, %s, GL_NONE, %s)\n",
 713                       debug_register(dstreg),
 714                       debug_register(arg2), debug_argmod(argmod2),
 715                       debug_register(arg0),
 716                       debug_register(arg1), debug_argmod(argmod1));
 717                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE,
 718                                                  arg2, GL_NONE,  argmod2,
 719                                                  arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 720                                                  arg1, GL_NONE,  argmod1));
 721                 break;
 722
 723             /* D3DTOP_PREMODULATE ???? */
 724
 725             case WINED3DTOP_DOTPRODUCT3:
 726                 TRACE("glAlphaFragmentOp2ATI(GL_DOT3_ATI, %s, GL_NONE, GL_4X_BIT_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 727                       debug_register(dstreg),
 728                       debug_register(arg1), debug_argmod(argmod1 | GL_BIAS_BIT_ATI),
 729                       debug_register(arg2), debug_argmod(argmod2 | GL_BIAS_BIT_ATI));
 730                 GL_EXTCALL(glAlphaFragmentOp2ATI(GL_DOT3_ATI, dstreg, GL_4X_BIT_ATI,
 731                                                  arg1, GL_NONE, argmod1 | GL_BIAS_BIT_ATI,
 732                                                  arg2, GL_NONE, argmod2 | GL_BIAS_BIT_ATI));
 733                 break;
 734
 735             case WINED3DTOP_MULTIPLYADD:
 736                 TRACE("glAlphaFragmentOp3ATI(GL_MAD_ATI, %s,          GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 737                       debug_register(dstreg),
 738                       debug_register(arg0), debug_argmod(argmod0),
 739                       debug_register(arg2), debug_argmod(argmod2),
 740                       debug_register(arg1), debug_argmod(argmod1));
 741                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_MAD_ATI, dstreg,          GL_NONE,
 742                            arg0, GL_NONE, argmod0,
 743                            arg2, GL_NONE, argmod2,
 744                            arg1, GL_NONE, argmod1));
 745                 break;
 746
 747             case WINED3DTOP_LERP:
 748                 TRACE("glAlphaFragmentOp3ATI(GL_LERP_ATI, %s,          GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 749                       debug_register(dstreg),
 750                       debug_register(arg1), debug_argmod(argmod1),
 751                       debug_register(arg2), debug_argmod(argmod2),
 752                       debug_register(arg0), debug_argmod(argmod0));
 753                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_LERP_ATI, dstreg, GL_NONE,
 754                                                  arg1, GL_NONE, argmod1,
 755                                                  arg2, GL_NONE, argmod2,
 756                                                  arg0, GL_NONE, argmod0));
 757                 break;
 758
 759             case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
 760             case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
 761             case WINED3DTOP_MODULATECOLOR_ADDALPHA:
 762             case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
 763             case WINED3DTOP_BUMPENVMAP:
 764             case WINED3DTOP_BUMPENVMAPLUMINANCE:
 765                 ERR("Application uses an invalid alpha operation\n");
 766                 break;
 767
 768             default: FIXME("Unhandled alpha operation %d on stage %d\n", op[stage].aop, stage);
 769         }
 770     }
 771
 772     TRACE("glEndFragmentShaderATI()\n");
 773     GL_EXTCALL(glEndFragmentShaderATI());
 774     checkGLcall("GL_EXTCALL(glEndFragmentShaderATI())");
 775     return ret;
 776 }
 777 #undef GLINFO_LOCATION
 778
 779 #define GLINFO_LOCATION stateblock->wineD3DDevice->adapter->gl_info
 780 static void set_tex_op_atifs(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
 781     IWineD3DDeviceImpl          *This = stateblock->wineD3DDevice;
 782     struct atifs_ffp_desc       *desc;
 783     struct texture_stage_op     op[MAX_TEXTURES];
 784     struct atifs_private_data   *priv = (struct atifs_private_data *) This->shader_priv;
 785
 786     gen_ffp_op(stateblock, op);
 787     desc = (struct atifs_ffp_desc *) find_ffp_shader(&priv->fragment_shaders, op);
 788     if(!desc) {
 789         desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*desc));
 790         if(!desc) {
 791             ERR("Out of memory\n");
 792             return;
 793         }
 794         memcpy(desc->parent.op, op, sizeof(op));
 795         desc->shader = gen_ati_shader(op, &GLINFO_LOCATION);
 796         add_ffp_shader(&priv->fragment_shaders, &desc->parent);
 797         TRACE("Allocated fixed function replacement shader descriptor %p\n", desc);
 798     }
 799
 800     GL_EXTCALL(glBindFragmentShaderATI(desc->shader));
 801 }
 802
 803 static void state_texfactor_atifs(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
 804     float col[4];
 805     D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
 806
 807     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col));
 808     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col)");
 809 }
 810
 811 static void set_bumpmat(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
 812     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / WINED3D_HIGHEST_TEXTURE_STATE;
 813     float mat[2][2];
 814
 815     mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
 816     mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
 817     mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
 818     mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
 819     /* GL_ATI_fragment_shader allows only constants from 0.0 to 1.0, but the bumpmat
 820      * constants can be in any range. While they should stay between [-1.0 and 1.0] because
 821      * Shader Model 1.x pixel shaders are clamped to that range negative values are used occasionally,
 822      * for example by our d3d9 test. So to get negative values scale -1;1 to 0;1 and undo that in the
 823      * shader(it is free). This might potentially reduce precision. However, if the hardware does
 824      * support proper floats it shouldn't, and if it doesn't we can't get anything better anyway
 825      */
 826     mat[0][0] = (mat[0][0] + 1.0) * 0.5;
 827     mat[1][0] = (mat[1][0] + 1.0) * 0.5;
 828     mat[0][1] = (mat[0][1] + 1.0) * 0.5;
 829     mat[1][1] = (mat[1][1] + 1.0) * 0.5;
 830     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), (float *) mat));
 831     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), mat)");
 832
 833     /* FIXME: This should go away
 834      * This is currently needed because atifs borrows a pixel shader implementation
 835      * from somewhere else, but consumes bump map matrix change events. The other pixel
 836      * shader implementation may need notification about the change to update the texbem
 837      * constants. Once ATIFS supports real shaders on its own, and GLSL/ARB have a replacement
 838      * pipeline this call can go away
 839      *
 840      * FIXME2: Even considering this workaround calling FFPStateTable directly isn't nice
 841      * as well. Better would be to call the model's table we inherit from, but currently
 842      * it is always the FFP table, and as soon as this changes we can remove the call anyway
 843      */
 844     FFPStateTable[state].apply(state, stateblock, context);
 845 }
 846 #undef GLINFO_LOCATION
 847
 848 /* our state table. Borrows lots of stuff from the base implementation */
 849 struct StateEntry ATIFSStateTable[STATE_HIGHEST + 1];
 850
 851 static void init_state_table() {
 852     unsigned int i;
 853     const DWORD rep = STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP);
 854     memcpy(ATIFSStateTable, arb_program_shader_backend.StateTable, sizeof(ATIFSStateTable));
 855
 856     for(i = 0; i < MAX_TEXTURES; i++) {
 857         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLOROP)].apply = set_tex_op_atifs;
 858         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLOROP)].representative = rep;
 859         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLORARG1)].apply = set_tex_op_atifs;
 860         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLORARG1)].representative = rep;
 861         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLORARG2)].apply = set_tex_op_atifs;
 862         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLORARG2)].representative = rep;
 863         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLORARG0)].apply = set_tex_op_atifs;
 864         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLORARG0)].representative = rep;
 865
 866         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAOP)].apply = set_tex_op_atifs;
 867         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAOP)].representative = rep;
 868         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAARG1)].apply = set_tex_op_atifs;
 869         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAARG1)].representative = rep;
 870         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAARG2)].apply = set_tex_op_atifs;
 871         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAARG2)].representative = rep;
 872         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAARG0)].apply = set_tex_op_atifs;
 873         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAARG0)].representative = rep;
 874
 875         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_RESULTARG)].apply = set_tex_op_atifs;
 876         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_RESULTARG)].representative = rep;
 877
 878         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00)].apply = set_bumpmat;
 879         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT01)].apply = set_bumpmat;
 880         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT10)].apply = set_bumpmat;
 881         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT11)].apply = set_bumpmat;
 882     }
 883
 884     ATIFSStateTable[STATE_RENDER(WINED3DRS_TEXTUREFACTOR)].apply = state_texfactor_atifs;
 885     ATIFSStateTable[STATE_RENDER(WINED3DRS_TEXTUREFACTOR)].representative = STATE_RENDER(WINED3DRS_TEXTUREFACTOR);
 886 }
 887
 888 /* GL_ATI_fragment_shader backend.It borrows a lot from a the
 889  * ARB shader backend, currently the whole vertex processing
 890  * code. This code would also forward pixel shaders, but if
 891  * GL_ARB_fragment_program is supported, the atifs shader backend
 892  * is not used.
 893  */
 894 static void shader_atifs_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
 895     arb_program_shader_backend.shader_select(iface, usePS, useVS);
 896 }
 897
 898 static void shader_atifs_select_depth_blt(IWineD3DDevice *iface) {
 899     arb_program_shader_backend.shader_select_depth_blt(iface);
 900 }
 901
 902 static void shader_atifs_destroy_depth_blt(IWineD3DDevice *iface) {
 903     arb_program_shader_backend.shader_destroy_depth_blt(iface);
 904 }
 905
 906 static void shader_atifs_load_constants(IWineD3DDevice *iface, char usePS, char useVS) {
 907     arb_program_shader_backend.shader_load_constants(iface, usePS, useVS);
 908 }
 909
 910 static void shader_atifs_cleanup(IWineD3DDevice *iface) {
 911     arb_program_shader_backend.shader_cleanup(iface);
 912 }
 913
 914 static void shader_atifs_color_correction(SHADER_OPCODE_ARG* arg) {
 915     arb_program_shader_backend.shader_color_correction(arg);
 916 }
 917
 918 static void shader_atifs_destroy(IWineD3DBaseShader *iface) {
 919     arb_program_shader_backend.shader_destroy(iface);
 920 }
 921
 922 static HRESULT shader_atifs_alloc(IWineD3DDevice *iface) {
 923     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 924     HRESULT hr;
 925     struct atifs_private_data *priv;
 926     hr = arb_program_shader_backend.shader_alloc_private(iface);
 927     if(FAILED(hr)) return hr;
 928
 929     This->shader_priv = HeapReAlloc(GetProcessHeap(), 0, This->shader_priv,
 930                                     sizeof(struct atifs_private_data));
 931     priv = (struct atifs_private_data *) This->shader_priv;
 932     list_init(&priv->fragment_shaders);
 933     return WINED3D_OK;
 934 }
 935
 936 #define GLINFO_LOCATION This->adapter->gl_info
 937 static void shader_atifs_free(IWineD3DDevice *iface) {
 938     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 939     struct atifs_private_data *priv = (struct atifs_private_data *) This->shader_priv;
 940     struct ffp_desc *entry, *entry2;
 941     struct atifs_ffp_desc *entry_ati;
 942
 943     ENTER_GL();
 944     LIST_FOR_EACH_ENTRY_SAFE(entry, entry2, &priv->fragment_shaders, struct ffp_desc, entry) {
 945         entry_ati = (struct atifs_ffp_desc *) entry;
 946         GL_EXTCALL(glDeleteFragmentShaderATI(entry_ati->shader));
 947         checkGLcall("glDeleteFragmentShaderATI(entry->shader)");
 948         list_remove(&entry->entry);
 949         HeapFree(GetProcessHeap(), 0, entry);
 950     }
 951     LEAVE_GL();
 952
 953     /* Not actually needed, but revert what we've done before */
 954     This->shader_priv = HeapReAlloc(GetProcessHeap(), 0, This->shader_priv,
 955                                     sizeof(struct shader_arb_priv));
 956     arb_program_shader_backend.shader_free_private(iface);
 957 }
 958 #undef GLINFO_LOCATION
 959
 960 static BOOL shader_atifs_dirty_const(IWineD3DDevice *iface) {
 961     return arb_program_shader_backend.shader_dirtifyable_constants(iface);
 962 }
 963
 964 static void shader_atifs_load_init(void) {
 965     arb_program_shader_backend.shader_dll_load_init();
 966     init_state_table();
 967 }
 968
 969 static void shader_atifs_get_caps(WINED3DDEVTYPE devtype, WineD3D_GL_Info *gl_info, struct shader_caps *caps) {
 970     arb_program_shader_backend.shader_get_caps(devtype, gl_info, caps);
 971
 972     caps->TextureOpCaps =  WINED3DTEXOPCAPS_DISABLE                     |
 973                            WINED3DTEXOPCAPS_SELECTARG1                  |
 974                            WINED3DTEXOPCAPS_SELECTARG2                  |
 975                            WINED3DTEXOPCAPS_MODULATE4X                  |
 976                            WINED3DTEXOPCAPS_MODULATE2X                  |
 977                            WINED3DTEXOPCAPS_MODULATE                    |
 978                            WINED3DTEXOPCAPS_ADDSIGNED2X                 |
 979                            WINED3DTEXOPCAPS_ADDSIGNED                   |
 980                            WINED3DTEXOPCAPS_ADD                         |
 981                            WINED3DTEXOPCAPS_SUBTRACT                    |
 982                            WINED3DTEXOPCAPS_ADDSMOOTH                   |
 983                            WINED3DTEXOPCAPS_BLENDCURRENTALPHA           |
 984                            WINED3DTEXOPCAPS_BLENDFACTORALPHA            |
 985                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHA           |
 986                            WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA           |
 987                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM         |
 988                            WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR      |
 989                            WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA      |
 990                            WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA   |
 991                            WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR   |
 992                            WINED3DTEXOPCAPS_DOTPRODUCT3                 |
 993                            WINED3DTEXOPCAPS_MULTIPLYADD                 |
 994                            WINED3DTEXOPCAPS_LERP                        |
 995                            WINED3DTEXOPCAPS_BUMPENVMAP;
 996
 997     /* TODO: Implement WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE
 998     and WINED3DTEXOPCAPS_PREMODULATE */
 999
1000     /* GL_ATI_fragment_shader only supports up to 6 textures, which was the limit on r200 cards
1001      * which this extension is exclusively focused on(later cards have GL_ARB_fragment_program).
1002      * If the current card has more than 8 fixed function textures in OpenGL's regular fixed
1003      * function pipeline then the ATI_fragment_shader backend imposes a stricter limit. This
1004      * shouldn't be too hard since Nvidia cards have a limit of 4 textures with the default ffp
1005      * pipeline, and almost all games are happy with that. We can however support up to 8
1006      * texture stages because we have a 2nd pass limit of 8 instructions, and per stage we use
1007      * only 1 instruction.
1008      *
1009      * The proper fix for this is not to use GL_ATI_fragment_shader on cards newer than the
1010      * r200 series and use an ARB or GLSL shader instead
1011      */
1012     if(caps->MaxSimultaneousTextures > 6) {
1013         WARN("OpenGL fixed function supports %d simultaneous textures,\n", caps->MaxSimultaneousTextures);
1014         WARN("but GL_ATI_fragment_shader limits this to 6\n");
1015         caps->MaxSimultaneousTextures = 6;
1016     }
1017
1018     caps->PrimitiveMiscCaps |= WINED3DPMISCCAPS_TSSARGTEMP;
1019 }
1020
1021 static void shader_atifs_generate_pshader(IWineD3DPixelShader *iface, SHADER_BUFFER *buffer) {
1022     ERR("Should not get here\n");
1023 }
1024
1025 static void shader_atifs_generate_vshader(IWineD3DVertexShader *iface, SHADER_BUFFER *buffer) {
1026     arb_program_shader_backend.shader_generate_vshader(iface, buffer);
1027 }
1028
1029 const shader_backend_t atifs_shader_backend = {
1030     shader_atifs_select,
1031     shader_atifs_select_depth_blt,
1032     shader_atifs_destroy_depth_blt,
1033     shader_atifs_load_constants,
1034     shader_atifs_cleanup,
1035     shader_atifs_color_correction,
1036     shader_atifs_destroy,
1037     shader_atifs_alloc,
1038     shader_atifs_free,
1039     shader_atifs_dirty_const,
1040     shader_atifs_generate_pshader,
1041     shader_atifs_generate_vshader,
1042     shader_atifs_get_caps,
1043     shader_atifs_load_init,
1044     ATIFSStateTable
1045 };