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 WINE_DECLARE_DEBUG_CHANNEL(d3d);
  30
  31 /* Some private defines, Constant associations, etc.
  32  * Env bump matrix and per stage constant should be independent,
  33  * a stage that bump maps can't read the per state constant
  34  */
  35 #define ATI_FFP_CONST_BUMPMAT(i) (GL_CON_0_ATI + i)
  36 #define ATI_FFP_CONST_CONSTANT0 GL_CON_0_ATI
  37 #define ATI_FFP_CONST_CONSTANT1 GL_CON_1_ATI
  38 #define ATI_FFP_CONST_CONSTANT2 GL_CON_2_ATI
  39 #define ATI_FFP_CONST_CONSTANT3 GL_CON_3_ATI
  40 #define ATI_FFP_CONST_CONSTANT4 GL_CON_4_ATI
  41 #define ATI_FFP_CONST_CONSTANT5 GL_CON_5_ATI
  42 #define ATI_FFP_CONST_TFACTOR   GL_CON_6_ATI
  43
  44 /* GL_ATI_fragment_shader specific fixed function pipeline description. "Inherits" from the common one */
  45 struct atifs_ffp_desc
  46 {
  47     struct ffp_frag_desc parent;
  48     GLuint shader;
  49     unsigned int num_textures_used;
  50 };
  51
  52 struct atifs_private_data
  53 {
  54     struct hash_table_t *fragment_shaders; /* A hashtable to track fragment pipeline replacement shaders */
  55
  56 };
  57
  58 static const char *debug_dstmod(GLuint mod) {
  59     switch(mod) {
  60         case GL_NONE:               return "GL_NONE";
  61         case GL_2X_BIT_ATI:         return "GL_2X_BIT_ATI";
  62         case GL_4X_BIT_ATI:         return "GL_4X_BIT_ATI";
  63         case GL_8X_BIT_ATI:         return "GL_8X_BIT_ATI";
  64         case GL_HALF_BIT_ATI:       return "GL_HALF_BIT_ATI";
  65         case GL_QUARTER_BIT_ATI:    return "GL_QUARTER_BIT_ATI";
  66         case GL_EIGHTH_BIT_ATI:     return "GL_EIGHTH_BIT_ATI";
  67         case GL_SATURATE_BIT_ATI:   return "GL_SATURATE_BIT_ATI";
  68         default:                    return "Unexpected modifier\n";
  69     }
  70 }
  71
  72 static const char *debug_argmod(GLuint mod) {
  73     switch(mod) {
  74         case GL_NONE:
  75             return "GL_NONE";
  76
  77         case GL_2X_BIT_ATI:
  78             return "GL_2X_BIT_ATI";
  79         case GL_COMP_BIT_ATI:
  80             return "GL_COMP_BIT_ATI";
  81         case GL_NEGATE_BIT_ATI:
  82             return "GL_NEGATE_BIT_ATI";
  83         case GL_BIAS_BIT_ATI:
  84             return "GL_BIAS_BIT_ATI";
  85
  86         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI:
  87             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI";
  88         case GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI:
  89             return "GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI";
  90         case GL_2X_BIT_ATI | GL_BIAS_BIT_ATI:
  91             return "GL_2X_BIT_ATI | GL_BIAS_BIT_ATI";
  92         case GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI:
  93             return "GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI";
  94         case GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI:
  95             return "GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI";
  96         case GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
  97             return "GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
  98
  99         case GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 100             return "GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 101         case GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 102             return "GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 103         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI:
 104             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI";
 105         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI:
 106             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI";
 107
 108         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 109             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 110
 111         default:
 112             return "Unexpected argmod combination\n";
 113     }
 114 }
 115 static const char *debug_register(GLuint reg) {
 116     switch(reg) {
 117         case GL_REG_0_ATI:                  return "GL_REG_0_ATI";
 118         case GL_REG_1_ATI:                  return "GL_REG_1_ATI";
 119         case GL_REG_2_ATI:                  return "GL_REG_2_ATI";
 120         case GL_REG_3_ATI:                  return "GL_REG_3_ATI";
 121         case GL_REG_4_ATI:                  return "GL_REG_4_ATI";
 122         case GL_REG_5_ATI:                  return "GL_REG_5_ATI";
 123
 124         case GL_CON_0_ATI:                  return "GL_CON_0_ATI";
 125         case GL_CON_1_ATI:                  return "GL_CON_1_ATI";
 126         case GL_CON_2_ATI:                  return "GL_CON_2_ATI";
 127         case GL_CON_3_ATI:                  return "GL_CON_3_ATI";
 128         case GL_CON_4_ATI:                  return "GL_CON_4_ATI";
 129         case GL_CON_5_ATI:                  return "GL_CON_5_ATI";
 130         case GL_CON_6_ATI:                  return "GL_CON_6_ATI";
 131         case GL_CON_7_ATI:                  return "GL_CON_7_ATI";
 132
 133         case GL_ZERO:                       return "GL_ZERO";
 134         case GL_ONE:                        return "GL_ONE";
 135         case GL_PRIMARY_COLOR:              return "GL_PRIMARY_COLOR";
 136         case GL_SECONDARY_INTERPOLATOR_ATI: return "GL_SECONDARY_INTERPOLATOR_ATI";
 137
 138         default:                            return "Unknown register\n";
 139     }
 140 }
 141
 142 static const char *debug_swizzle(GLuint swizzle) {
 143     switch(swizzle) {
 144         case GL_SWIZZLE_STR_ATI:        return "GL_SWIZZLE_STR_ATI";
 145         case GL_SWIZZLE_STQ_ATI:        return "GL_SWIZZLE_STQ_ATI";
 146         case GL_SWIZZLE_STR_DR_ATI:     return "GL_SWIZZLE_STR_DR_ATI";
 147         case GL_SWIZZLE_STQ_DQ_ATI:     return "GL_SWIZZLE_STQ_DQ_ATI";
 148         default:                        return "unknown swizzle";
 149     }
 150 }
 151
 152 static const char *debug_rep(GLuint rep) {
 153     switch(rep) {
 154         case GL_NONE:                   return "GL_NONE";
 155         case GL_RED:                    return "GL_RED";
 156         case GL_GREEN:                  return "GL_GREEN";
 157         case GL_BLUE:                   return "GL_BLUE";
 158         default:                        return "unknown argrep";
 159     }
 160 }
 161
 162 #define GLINFO_LOCATION (*gl_info)
 163 static GLuint register_for_arg(DWORD arg, const WineD3D_GL_Info *gl_info,
 164         unsigned int stage, GLuint *mod, GLuint *rep, GLuint tmparg)
 165 {
 166     GLenum ret;
 167
 168     if(mod) *mod = GL_NONE;
 169     if(arg == ARG_UNUSED)
 170     {
 171         if (rep) *rep = GL_NONE;
 172         return -1; /* This is the marker for unused registers */
 173     }
 174
 175     switch(arg & WINED3DTA_SELECTMASK) {
 176         case WINED3DTA_DIFFUSE:
 177             ret = GL_PRIMARY_COLOR;
 178             break;
 179
 180         case WINED3DTA_CURRENT:
 181             /* Note that using GL_REG_0_ATI for the passed on register is safe because
 182              * texture0 is read at stage0, so in the worst case it is read in the
 183              * instruction writing to reg0. Afterwards texture0 is not used any longer.
 184              * If we're reading from current
 185              */
 186             if(stage == 0) {
 187                 ret = GL_PRIMARY_COLOR;
 188             } else {
 189                 ret = GL_REG_0_ATI;
 190             }
 191             break;
 192
 193         case WINED3DTA_TEXTURE:
 194             ret = GL_REG_0_ATI + stage;
 195             break;
 196
 197         case WINED3DTA_TFACTOR:
 198             ret = ATI_FFP_CONST_TFACTOR;
 199             break;
 200
 201         case WINED3DTA_SPECULAR:
 202             ret = GL_SECONDARY_INTERPOLATOR_ATI;
 203             break;
 204
 205         case WINED3DTA_TEMP:
 206             ret = tmparg;
 207             break;
 208
 209         case WINED3DTA_CONSTANT:
 210             FIXME("Unhandled source argument WINED3DTA_TEMP\n");
 211             ret = GL_CON_0_ATI;
 212             break;
 213
 214         default:
 215             FIXME("Unknown source argument %d\n", arg);
 216             ret = GL_ZERO;
 217     }
 218
 219     if(arg & WINED3DTA_COMPLEMENT) {
 220         if(mod) *mod |= GL_COMP_BIT_ATI;
 221     }
 222     if(arg & WINED3DTA_ALPHAREPLICATE) {
 223         if(rep) *rep = GL_ALPHA;
 224     } else {
 225         if(rep) *rep = GL_NONE;
 226     }
 227     return ret;
 228 }
 229
 230 static GLuint find_tmpreg(const struct texture_stage_op op[MAX_TEXTURES])
 231 {
 232     int lowest_read = -1;
 233     int lowest_write = -1;
 234     int i;
 235     BOOL tex_used[MAX_TEXTURES];
 236
 237     memset(tex_used, 0, sizeof(tex_used));
 238     for(i = 0; i < MAX_TEXTURES; i++) {
 239         if(op[i].cop == WINED3DTOP_DISABLE) {
 240             break;
 241         }
 242
 243         if(lowest_read == -1 &&
 244           (op[i].carg1 == WINED3DTA_TEMP || op[i].carg2 == WINED3DTA_TEMP || op[i].carg0 == WINED3DTA_TEMP ||
 245            op[i].aarg1 == WINED3DTA_TEMP || op[i].aarg2 == WINED3DTA_TEMP || op[i].aarg0 == WINED3DTA_TEMP)) {
 246             lowest_read = i;
 247         }
 248
 249         if(lowest_write == -1 && op[i].dst == tempreg) {
 250             lowest_write = i;
 251         }
 252
 253         if(op[i].carg1 == WINED3DTA_TEXTURE || op[i].carg2 == WINED3DTA_TEXTURE || op[i].carg0 == WINED3DTA_TEXTURE ||
 254            op[i].aarg1 == WINED3DTA_TEXTURE || op[i].aarg2 == WINED3DTA_TEXTURE || op[i].aarg0 == WINED3DTA_TEXTURE) {
 255             tex_used[i] = TRUE;
 256         }
 257     }
 258
 259     /* Temp reg not read? We don't need it, return GL_NONE */
 260     if(lowest_read == -1) return GL_NONE;
 261
 262     if(lowest_write >= lowest_read) {
 263         FIXME("Temp register read before being written\n");
 264     }
 265
 266     if(lowest_write == -1) {
 267         /* This needs a test. Maybe we are supposed to return 0.0/0.0/0.0/0.0, or fail drawprim, or whatever */
 268         FIXME("Temp register read without being written\n");
 269         return GL_REG_1_ATI;
 270     } else if(lowest_write >= 1) {
 271         /* If we're writing to the temp reg at earliest in stage 1, we can use register 1 for the temp result.
 272          * there may be texture data stored in reg 1, but we do not need it any longer since stage 1 already
 273          * read it
 274          */
 275         return GL_REG_1_ATI;
 276     } else {
 277         /* Search for a free texture register. We have 6 registers available. GL_REG_0_ATI is already used
 278          * for the regular result
 279          */
 280         for(i = 1; i < 6; i++) {
 281             if(!tex_used[i]) {
 282                 return GL_REG_0_ATI + i;
 283             }
 284         }
 285         /* What to do here? Report it in ValidateDevice? */
 286         FIXME("Could not find a register for the temporary register\n");
 287         return 0;
 288     }
 289 }
 290
 291 static GLuint gen_ati_shader(const struct texture_stage_op op[MAX_TEXTURES], const WineD3D_GL_Info *gl_info)
 292 {
 293     GLuint ret = GL_EXTCALL(glGenFragmentShadersATI(1));
 294     unsigned int stage;
 295     GLuint arg0, arg1, arg2, extrarg;
 296     GLuint dstmod, argmod0, argmod1, argmod2, argmodextra;
 297     GLuint rep0, rep1, rep2;
 298     GLuint swizzle;
 299     GLuint tmparg = find_tmpreg(op);
 300     GLuint dstreg;
 301
 302     if(!ret) {
 303         ERR("Failed to generate a GL_ATI_fragment_shader shader id\n");
 304         return 0;
 305     }
 306     GL_EXTCALL(glBindFragmentShaderATI(ret));
 307     checkGLcall("GL_EXTCALL(glBindFragmentShaderATI(ret))");
 308
 309     TRACE("glBeginFragmentShaderATI()\n");
 310     GL_EXTCALL(glBeginFragmentShaderATI());
 311     checkGLcall("GL_EXTCALL(glBeginFragmentShaderATI())");
 312
 313     /* Pass 1: Generate sampling instructions for perturbation maps */
 314       for(stage = 0; stage < GL_LIMITS(textures); stage++) {
 315         if(op[stage].cop == WINED3DTOP_DISABLE) break;
 316         if(op[stage].cop != WINED3DTOP_BUMPENVMAP &&
 317            op[stage].cop != WINED3DTOP_BUMPENVMAPLUMINANCE) continue;
 318
 319         TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, GL_SWIZZLE_STR_ATI)\n",
 320               stage, stage);
 321         GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 322                    GL_TEXTURE0_ARB + stage,
 323                    GL_SWIZZLE_STR_ATI));
 324         if(op[stage + 1].projected == proj_none) {
 325             swizzle = GL_SWIZZLE_STR_ATI;
 326         } else if(op[stage + 1].projected == proj_count4) {
 327             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 328         } else {
 329             swizzle = GL_SWIZZLE_STR_DR_ATI;
 330         }
 331         TRACE("glPassTexCoordATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, %s)\n",
 332               stage + 1, stage + 1, debug_swizzle(swizzle));
 333         GL_EXTCALL(glPassTexCoordATI(GL_REG_0_ATI + stage + 1,
 334                    GL_TEXTURE0_ARB + stage + 1,
 335                    swizzle));
 336
 337         /* We need GL_REG_5_ATI as a temporary register to swizzle the bump matrix. So we run into
 338          * issues if we're bump mapping on stage 4 or 5
 339          */
 340         if(stage >= 4) {
 341             FIXME("Bump mapping in stage %d\n", stage);
 342         }
 343     }
 344
 345     /* Pass 2: Generate perturbation calculations */
 346     for(stage = 0; stage < GL_LIMITS(textures); stage++) {
 347         GLuint argmodextra_x, argmodextra_y;
 348         struct color_fixup_desc fixup;
 349
 350         if(op[stage].cop == WINED3DTOP_DISABLE) break;
 351         if(op[stage].cop != WINED3DTOP_BUMPENVMAP &&
 352            op[stage].cop != WINED3DTOP_BUMPENVMAPLUMINANCE) continue;
 353
 354         fixup = op[stage].color_fixup;
 355         if (fixup.x_source != CHANNEL_SOURCE_X || fixup.y_source != CHANNEL_SOURCE_Y)
 356         {
 357             FIXME("Swizzles not implemented\n");
 358             argmodextra_x = GL_NONE;
 359             argmodextra_y = GL_NONE;
 360         }
 361         else
 362         {
 363             /* Nice thing, we get the color correction for free :-) */
 364             argmodextra_x = fixup.x_sign_fixup ? GL_2X_BIT_ATI | GL_BIAS_BIT_ATI : GL_NONE;
 365             argmodextra_y = fixup.y_sign_fixup ? GL_2X_BIT_ATI | GL_BIAS_BIT_ATI : GL_NONE;
 366         }
 367
 368         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",
 369               stage + 1, stage, debug_argmod(argmodextra_x), stage, stage + 1);
 370         GL_EXTCALL(glColorFragmentOp3ATI(GL_DOT2_ADD_ATI, GL_REG_0_ATI + stage + 1, GL_RED_BIT_ATI, GL_NONE,
 371                                          GL_REG_0_ATI + stage, GL_NONE, argmodextra_x,
 372                                          ATI_FFP_CONST_BUMPMAT(stage), GL_NONE, GL_2X_BIT_ATI | GL_BIAS_BIT_ATI,
 373                                          GL_REG_0_ATI + stage + 1, GL_RED, GL_NONE));
 374
 375         /* FIXME: How can I make GL_DOT2_ADD_ATI read the factors from blue and alpha? It defaults to red and green,
 376          * 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
 377          * one register and two instructions in this pass for a simple swizzling operation.
 378          * For starters it might be good enough to merge the two movs into one, but even that isn't possible :-(
 379          *
 380          * NOTE: GL_BLUE | GL_ALPHA is not possible. It doesn't throw a compilation error, but an OR operation on the
 381          * constants doesn't make sense, considering their values.
 382          */
 383         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);
 384         GL_EXTCALL(glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_5_ATI, GL_RED_BIT_ATI, GL_NONE,
 385                                          ATI_FFP_CONST_BUMPMAT(stage), GL_BLUE, GL_NONE));
 386         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);
 387         GL_EXTCALL(glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_5_ATI, GL_GREEN_BIT_ATI, GL_NONE,
 388                                         ATI_FFP_CONST_BUMPMAT(stage), GL_ALPHA, GL_NONE));
 389         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",
 390               stage + 1, stage, debug_argmod(argmodextra_y), stage + 1);
 391         GL_EXTCALL(glColorFragmentOp3ATI(GL_DOT2_ADD_ATI, GL_REG_0_ATI + stage + 1, GL_GREEN_BIT_ATI, GL_NONE,
 392                                          GL_REG_0_ATI + stage, GL_NONE, argmodextra_y,
 393                                          GL_REG_5_ATI, GL_NONE, GL_2X_BIT_ATI | GL_BIAS_BIT_ATI,
 394                                          GL_REG_0_ATI + stage + 1, GL_GREEN, GL_NONE));
 395     }
 396
 397     /* Pass 3: Generate sampling instructions for regular textures */
 398     for(stage = 0; stage < GL_LIMITS(textures); stage++) {
 399         if(op[stage].cop == WINED3DTOP_DISABLE) {
 400             break;
 401         }
 402
 403         if(op[stage].projected == proj_none) {
 404             swizzle = GL_SWIZZLE_STR_ATI;
 405         } else if(op[stage].projected == proj_count3) {
 406             swizzle = GL_SWIZZLE_STR_DR_ATI;
 407         } else {
 408             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 409         }
 410
 411         if((op[stage].carg0 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 412            (op[stage].carg1 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 413            (op[stage].carg2 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 414            (op[stage].aarg0 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 415            (op[stage].aarg1 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 416            (op[stage].aarg2 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 417             op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) {
 418
 419             if(stage > 0 &&
 420                (op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
 421                 op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
 422                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_REG_%d_ATI, GL_SWIZZLE_STR_ATI)\n",
 423                       stage, stage);
 424                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 425                            GL_REG_0_ATI + stage,
 426                            GL_SWIZZLE_STR_ATI));
 427             } else {
 428                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, %s)\n",
 429                     stage, stage, debug_swizzle(swizzle));
 430                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 431                                         GL_TEXTURE0_ARB + stage,
 432                                         swizzle));
 433             }
 434         }
 435     }
 436
 437     /* Pass 4: Generate the arithmetic instructions */
 438     for(stage = 0; stage < MAX_TEXTURES; stage++) {
 439         if(op[stage].cop == WINED3DTOP_DISABLE) {
 440             if(stage == 0) {
 441                 /* Handle complete texture disabling gracefully */
 442                 TRACE("glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE, GL_PRIMARY_COLOR, GL_NONE, GL_NONE)\n");
 443                 GL_EXTCALL(glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 444                                                  GL_PRIMARY_COLOR, GL_NONE, GL_NONE));
 445                 TRACE("glAlphaFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_PRIMARY_COLOR, GL_NONE, GL_NONE)\n");
 446                 GL_EXTCALL(glAlphaFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE,
 447                                                  GL_PRIMARY_COLOR, GL_NONE, GL_NONE));
 448             }
 449             break;
 450         }
 451
 452         if(op[stage].dst == tempreg) {
 453             /* If we're writing to D3DTA_TEMP, but never reading from it we don't have to write there in the first place.
 454              * skip the entire stage, this saves some GPU time
 455              */
 456             if(tmparg == GL_NONE) continue;
 457
 458             dstreg = tmparg;
 459         } else {
 460             dstreg = GL_REG_0_ATI;
 461         }
 462
 463         arg0 = register_for_arg(op[stage].carg0, gl_info, stage, &argmod0, &rep0, tmparg);
 464         arg1 = register_for_arg(op[stage].carg1, gl_info, stage, &argmod1, &rep1, tmparg);
 465         arg2 = register_for_arg(op[stage].carg2, gl_info, stage, &argmod2, &rep2, tmparg);
 466         dstmod = GL_NONE;
 467         argmodextra = GL_NONE;
 468         extrarg = GL_NONE;
 469
 470         switch(op[stage].cop) {
 471             case WINED3DTOP_SELECTARG2:
 472                 arg1 = arg2;
 473                 argmod1 = argmod2;
 474                 rep1 = rep2;
 475             case WINED3DTOP_SELECTARG1:
 476                 TRACE("glColorFragmentOp1ATI(GL_MOV_ATI, %s, GL_NONE, GL_NONE, %s, %s, %s)\n",
 477                       debug_register(dstreg), debug_register(arg1), debug_rep(rep1), debug_argmod(argmod1));
 478                 GL_EXTCALL(glColorFragmentOp1ATI(GL_MOV_ATI, dstreg, GL_NONE, GL_NONE,
 479                                                  arg1, rep1, argmod1));
 480                 break;
 481
 482             case WINED3DTOP_MODULATE4X:
 483                 if(dstmod == GL_NONE) dstmod = GL_4X_BIT_ATI;
 484             case WINED3DTOP_MODULATE2X:
 485                 if(dstmod == GL_NONE) dstmod = GL_2X_BIT_ATI;
 486                 dstmod |= GL_SATURATE_BIT_ATI;
 487             case WINED3DTOP_MODULATE:
 488                 TRACE("glColorFragmentOp2ATI(GL_MUL_ATI, %s, GL_NONE, %s, %s, %s, %s, %s, %s, %s)\n",
 489                       debug_register(dstreg), debug_dstmod(dstmod),
 490                       debug_register(arg1), debug_rep(rep1), debug_argmod(argmod1),
 491                       debug_register(arg2), debug_rep(rep2), debug_argmod(argmod2));
 492                 GL_EXTCALL(glColorFragmentOp2ATI(GL_MUL_ATI, dstreg, GL_NONE, dstmod,
 493                                                  arg1, rep1, argmod1,
 494                                                  arg2, rep2, argmod2));
 495                 break;
 496
 497             case WINED3DTOP_ADDSIGNED2X:
 498                 dstmod = GL_2X_BIT_ATI;
 499             case WINED3DTOP_ADDSIGNED:
 500                 argmodextra = GL_BIAS_BIT_ATI;
 501             case WINED3DTOP_ADD:
 502                 dstmod |= GL_SATURATE_BIT_ATI;
 503                 TRACE("glColorFragmentOp2ATI(GL_ADD_ATI, %s, GL_NONE, %s, %s, %s, %s, %s, %s, %s)\n",
 504                       debug_register(dstreg), debug_dstmod(dstmod),
 505                       debug_register(arg1), debug_rep(rep1), debug_argmod(argmod1),
 506                       debug_register(arg2), debug_rep(rep2), debug_argmod(argmodextra | argmod2));
 507                 GL_EXTCALL(glColorFragmentOp2ATI(GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, dstmod,
 508                                                  arg1, rep1, argmod1,
 509                                                  arg2, rep2, argmodextra | argmod2));
 510                 break;
 511
 512             case WINED3DTOP_SUBTRACT:
 513                 dstmod |= GL_SATURATE_BIT_ATI;
 514                 TRACE("glColorFragmentOp2ATI(GL_SUB_ATI, %s, GL_NONE, %s, %s, %s, %s, %s, %s, %s)\n",
 515                       debug_register(dstreg), debug_dstmod(dstmod),
 516                       debug_register(arg1), debug_rep(rep1), debug_argmod(argmod1),
 517                       debug_register(arg2), debug_rep(rep2), debug_argmod(argmod2));
 518                 GL_EXTCALL(glColorFragmentOp2ATI(GL_SUB_ATI, dstreg, GL_NONE, dstmod,
 519                                                  arg1, rep1, argmod1,
 520                                                  arg2, rep2, argmod2));
 521                 break;
 522
 523             case WINED3DTOP_ADDSMOOTH:
 524                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 525                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_SATURATE_BIT_ATI, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n",
 526                       debug_register(dstreg),
 527                       debug_register(arg2), debug_rep(rep2), debug_argmod(argmod2),
 528                       debug_register(arg1), debug_rep(rep1), debug_argmod(argmodextra),
 529                       debug_register(arg1), debug_rep(rep1), debug_argmod(argmod1));
 530                 /* Dst = arg1 + * arg2(1 -arg 1)
 531                  *     = arg2 * (1 - arg1) + arg1
 532                  */
 533                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 534                                                  arg2, rep2, argmod2,
 535                                                  arg1, rep1, argmodextra,
 536                                                  arg1, rep1, argmod1));
 537                 break;
 538
 539             case WINED3DTOP_BLENDCURRENTALPHA:
 540                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, NULL, -1);
 541             case WINED3DTOP_BLENDFACTORALPHA:
 542                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, NULL, -1);
 543             case WINED3DTOP_BLENDTEXTUREALPHA:
 544                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 545             case WINED3DTOP_BLENDDIFFUSEALPHA:
 546                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, NULL, -1);
 547                 TRACE("glColorFragmentOp3ATI(GL_LERP_ATI, %s, GL_NONE, GL_NONE, %s, GL_ALPHA, GL_NONE, %s, %s, %s, %s, %s, %s)\n",
 548                       debug_register(dstreg),
 549                       debug_register(extrarg),
 550                       debug_register(arg1), debug_rep(rep1), debug_argmod(argmod1),
 551                       debug_register(arg2), debug_rep(rep2), debug_argmod(argmod2));
 552                 GL_EXTCALL(glColorFragmentOp3ATI(GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 553                                                  extrarg, GL_ALPHA, GL_NONE,
 554                                                  arg1, rep1, argmod1,
 555                                                  arg2, rep2, argmod2));
 556                 break;
 557
 558             case WINED3DTOP_BLENDTEXTUREALPHAPM:
 559                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 560                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_NONE, %s, %s, %s, %s, GL_ALPHA, GL_COMP_BIT_ATI, %s, %s, %s)\n",
 561                       debug_register(dstreg),
 562                       debug_register(arg2), debug_rep(rep2), debug_argmod(argmod2),
 563                       debug_register(arg0),
 564                       debug_register(arg1), debug_rep(rep1), debug_argmod(argmod1));
 565                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 566                                                  arg2, rep2,  argmod2,
 567                                                  arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 568                                                  arg1, rep1,  argmod1));
 569                 break;
 570
 571             /* D3DTOP_PREMODULATE ???? */
 572
 573             case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
 574                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 575             case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
 576                 if(!argmodextra) argmodextra = argmod1;
 577                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_SATURATE_BIT_ATI, %s, %s, %s, %s, GL_ALPHA, %s, %s, %s, %s)\n",
 578                       debug_register(dstreg),
 579                       debug_register(arg2), debug_rep(rep2), debug_argmod(argmod2),
 580                       debug_register(arg1), debug_argmod(argmodextra), debug_register(arg1), debug_rep(rep1), debug_argmod(arg1));
 581                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 582                                                  arg2, rep2,  argmod2,
 583                                                  arg1, GL_ALPHA, argmodextra,
 584                                                  arg1, rep1,  argmod1));
 585                 break;
 586
 587             case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
 588                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 589             case WINED3DTOP_MODULATECOLOR_ADDALPHA:
 590                 if(!argmodextra) argmodextra = argmod1;
 591                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_SATURATE_BIT_ATI, %s, %s, %s, %s, %s, %s, %s, GL_ALPHA, %s)\n",
 592                       debug_register(dstreg),
 593                       debug_register(arg2), debug_rep(rep2), debug_argmod(argmod2),
 594                       debug_register(arg1), debug_rep(rep1), debug_argmod(argmodextra),
 595                       debug_register(arg1), debug_argmod(argmod1));
 596                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 597                                                  arg2, rep2,  argmod2,
 598                                                  arg1, rep1,  argmodextra,
 599                                                  arg1, GL_ALPHA, argmod1));
 600                 break;
 601
 602             case WINED3DTOP_DOTPRODUCT3:
 603                 TRACE("glColorFragmentOp2ATI(GL_DOT3_ATI, %s, GL_NONE, GL_4X_BIT_ATI | GL_SATURATE_BIT_ATI, %s, %s, %s, %s, %s, %s)\n",
 604                       debug_register(dstreg),
 605                       debug_register(arg1), debug_rep(rep1), debug_argmod(argmod1 | GL_BIAS_BIT_ATI),
 606                       debug_register(arg2), debug_rep(rep2), debug_argmod(argmod2 | GL_BIAS_BIT_ATI));
 607                 GL_EXTCALL(glColorFragmentOp2ATI(GL_DOT3_ATI, dstreg, GL_NONE, GL_4X_BIT_ATI | GL_SATURATE_BIT_ATI,
 608                                                  arg1, rep1, argmod1 | GL_BIAS_BIT_ATI,
 609                                                  arg2, rep2, argmod2 | GL_BIAS_BIT_ATI));
 610                 break;
 611
 612             case WINED3DTOP_MULTIPLYADD:
 613                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_SATURATE_BIT_ATI, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n",
 614                       debug_register(dstreg),
 615                       debug_register(arg1), debug_rep(rep1), debug_argmod(argmod1),
 616                       debug_register(arg2), debug_rep(rep2), debug_argmod(argmod2),
 617                       debug_register(arg0), debug_rep(rep0), debug_argmod(argmod0));
 618                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 619                                                  arg1, rep1, argmod1,
 620                                                  arg2, rep2, argmod2,
 621                                                  arg0, rep0, argmod0));
 622                 break;
 623
 624             case WINED3DTOP_LERP:
 625                 TRACE("glColorFragmentOp3ATI(GL_LERP_ATI, %s, GL_NONE, GL_NONE, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n",
 626                       debug_register(dstreg),
 627                       debug_register(arg1), debug_rep(rep1), debug_argmod(argmod1),
 628                       debug_register(arg2), debug_rep(rep2), debug_argmod(argmod2),
 629                       debug_register(arg0), debug_rep(rep0), debug_argmod(argmod0));
 630                 GL_EXTCALL(glColorFragmentOp3ATI(GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 631                                                  arg0, rep0, argmod0,
 632                                                  arg1, rep1, argmod1,
 633                                                  arg2, rep2, argmod2));
 634                 break;
 635
 636             case WINED3DTOP_BUMPENVMAP:
 637             case WINED3DTOP_BUMPENVMAPLUMINANCE:
 638                 /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
 639                 break;
 640
 641             default: FIXME("Unhandled color operation %d on stage %d\n", op[stage].cop, stage);
 642         }
 643
 644         arg0 = register_for_arg(op[stage].aarg0, gl_info, stage, &argmod0, NULL, tmparg);
 645         arg1 = register_for_arg(op[stage].aarg1, gl_info, stage, &argmod1, NULL, tmparg);
 646         arg2 = register_for_arg(op[stage].aarg2, gl_info, stage, &argmod2, NULL, tmparg);
 647         dstmod = GL_NONE;
 648         argmodextra = GL_NONE;
 649         extrarg = GL_NONE;
 650
 651         switch(op[stage].aop) {
 652             case WINED3DTOP_DISABLE:
 653                 /* Get the primary color to the output if on stage 0, otherwise leave register 0 untouched */
 654                 if(stage == 0) {
 655                     TRACE("glAlphaFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_PRIMARY_COLOR, GL_NONE, GL_NONE)\n");
 656                     GL_EXTCALL(glAlphaFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE,
 657                                GL_PRIMARY_COLOR, GL_NONE, GL_NONE));
 658                 }
 659                 break;
 660
 661             case WINED3DTOP_SELECTARG2:
 662                 arg1 = arg2;
 663                 argmod1 = argmod2;
 664             case WINED3DTOP_SELECTARG1:
 665                 TRACE("glAlphaFragmentOp1ATI(GL_MOV_ATI, %s,          GL_NONE, %s, GL_NONE, %s)\n",
 666                       debug_register(dstreg),
 667                       debug_register(arg1), debug_argmod(argmod1));
 668                 GL_EXTCALL(glAlphaFragmentOp1ATI(GL_MOV_ATI, dstreg, GL_NONE,
 669                                                  arg1, GL_NONE, argmod1));
 670                 break;
 671
 672             case WINED3DTOP_MODULATE4X:
 673                 if(dstmod == GL_NONE) dstmod = GL_4X_BIT_ATI;
 674             case WINED3DTOP_MODULATE2X:
 675                 if(dstmod == GL_NONE) dstmod = GL_2X_BIT_ATI;
 676                 dstmod |= GL_SATURATE_BIT_ATI;
 677             case WINED3DTOP_MODULATE:
 678                 TRACE("glAlphaFragmentOp2ATI(GL_MUL_ATI, %s,          %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 679                       debug_register(dstreg), debug_dstmod(dstmod),
 680                       debug_register(arg1), debug_argmod(argmod1),
 681                       debug_register(arg2), debug_argmod(argmod2));
 682                 GL_EXTCALL(glAlphaFragmentOp2ATI(GL_MUL_ATI, dstreg, dstmod,
 683                                                  arg1, GL_NONE, argmod1,
 684                                                  arg2, GL_NONE, argmod2));
 685                 break;
 686
 687             case WINED3DTOP_ADDSIGNED2X:
 688                 dstmod = GL_2X_BIT_ATI;
 689             case WINED3DTOP_ADDSIGNED:
 690                 argmodextra = GL_BIAS_BIT_ATI;
 691             case WINED3DTOP_ADD:
 692                 dstmod |= GL_SATURATE_BIT_ATI;
 693                 TRACE("glAlphaFragmentOp2ATI(GL_ADD_ATI, %s,          %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 694                       debug_register(dstreg), debug_dstmod(dstmod),
 695                       debug_register(arg1), debug_argmod(argmod1),
 696                       debug_register(arg2), debug_argmod(argmodextra | argmod2));
 697                 GL_EXTCALL(glAlphaFragmentOp2ATI(GL_ADD_ATI, dstreg, dstmod,
 698                                                  arg1, GL_NONE, argmod1,
 699                                                  arg2, GL_NONE, argmodextra | argmod2));
 700                 break;
 701
 702             case WINED3DTOP_SUBTRACT:
 703                 dstmod |= GL_SATURATE_BIT_ATI;
 704                 TRACE("glAlphaFragmentOp2ATI(GL_SUB_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 705                       debug_register(dstreg), debug_dstmod(dstmod),
 706                       debug_register(arg1), debug_argmod(argmod1),
 707                       debug_register(arg2), debug_argmod(argmod2));
 708                 GL_EXTCALL(glAlphaFragmentOp2ATI(GL_SUB_ATI, dstreg, dstmod,
 709                                                  arg1, GL_NONE, argmod1,
 710                                                  arg2, GL_NONE, argmod2));
 711                 break;
 712
 713             case WINED3DTOP_ADDSMOOTH:
 714                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 715                 TRACE("glAlphaFragmentOp3ATI(GL_MAD_ATI, %s,          GL_SATURATE_BIT_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 716                       debug_register(dstreg),
 717                       debug_register(arg2), debug_argmod(argmod2),
 718                       debug_register(arg1), debug_argmod(argmodextra),
 719                       debug_register(arg1), debug_argmod(argmod1));
 720                 /* Dst = arg1 + * arg2(1 -arg 1)
 721                  *     = arg2 * (1 - arg1) + arg1
 722                  */
 723                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_SATURATE_BIT_ATI,
 724                                                  arg2, GL_NONE, argmod2,
 725                                                  arg1, GL_NONE, argmodextra,
 726                                                  arg1, GL_NONE, argmod1));
 727                 break;
 728
 729             case WINED3DTOP_BLENDCURRENTALPHA:
 730                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, NULL, -1);
 731             case WINED3DTOP_BLENDFACTORALPHA:
 732                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, NULL, -1);
 733             case WINED3DTOP_BLENDTEXTUREALPHA:
 734                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 735             case WINED3DTOP_BLENDDIFFUSEALPHA:
 736                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, NULL, -1);
 737                 TRACE("glAlphaFragmentOp3ATI(GL_LERP_ATI, %s,          GL_NONE, %s, GL_ALPHA, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 738                       debug_register(dstreg),
 739                       debug_register(extrarg),
 740                       debug_register(arg1), debug_argmod(argmod1),
 741                       debug_register(arg2), debug_argmod(argmod2));
 742                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_LERP_ATI, dstreg, GL_NONE,
 743                                                  extrarg, GL_ALPHA, GL_NONE,
 744                                                  arg1, GL_NONE, argmod1,
 745                                                  arg2, GL_NONE, argmod2));
 746                 break;
 747
 748             case WINED3DTOP_BLENDTEXTUREALPHAPM:
 749                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 750                 TRACE("glAlphaFragmentOp3ATI(GL_MAD_ATI, %s,          GL_NONE, %s, GL_NONE, %s, %s, GL_ALPHA, GL_COMP_BIT_ATI, %s, GL_NONE, %s)\n",
 751                       debug_register(dstreg),
 752                       debug_register(arg2), debug_argmod(argmod2),
 753                       debug_register(arg0),
 754                       debug_register(arg1), debug_argmod(argmod1));
 755                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE,
 756                                                  arg2, GL_NONE,  argmod2,
 757                                                  arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 758                                                  arg1, GL_NONE,  argmod1));
 759                 break;
 760
 761             /* D3DTOP_PREMODULATE ???? */
 762
 763             case WINED3DTOP_DOTPRODUCT3:
 764                 TRACE("glAlphaFragmentOp2ATI(GL_DOT3_ATI, %s,          GL_4X_BIT_ATI | GL_SATURATE_BIT_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 765                       debug_register(dstreg),
 766                       debug_register(arg1), debug_argmod(argmod1 | GL_BIAS_BIT_ATI),
 767                       debug_register(arg2), debug_argmod(argmod2 | GL_BIAS_BIT_ATI));
 768                 GL_EXTCALL(glAlphaFragmentOp2ATI(GL_DOT3_ATI, dstreg, GL_4X_BIT_ATI | GL_SATURATE_BIT_ATI,
 769                                                  arg1, GL_NONE, argmod1 | GL_BIAS_BIT_ATI,
 770                                                  arg2, GL_NONE, argmod2 | GL_BIAS_BIT_ATI));
 771                 break;
 772
 773             case WINED3DTOP_MULTIPLYADD:
 774                 TRACE("glAlphaFragmentOp3ATI(GL_MAD_ATI, %s,          GL_SATURATE_BIT_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 775                       debug_register(dstreg),
 776                       debug_register(arg1), debug_argmod(argmod1),
 777                       debug_register(arg2), debug_argmod(argmod2),
 778                       debug_register(arg0), debug_argmod(argmod0));
 779                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_MAD_ATI, dstreg,          GL_SATURATE_BIT_ATI,
 780                            arg1, GL_NONE, argmod1,
 781                            arg2, GL_NONE, argmod2,
 782                            arg0, GL_NONE, argmod0));
 783                 break;
 784
 785             case WINED3DTOP_LERP:
 786                 TRACE("glAlphaFragmentOp3ATI(GL_LERP_ATI, %s,          GL_SATURATE_BIT_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 787                       debug_register(dstreg),
 788                       debug_register(arg1), debug_argmod(argmod1),
 789                       debug_register(arg2), debug_argmod(argmod2),
 790                       debug_register(arg0), debug_argmod(argmod0));
 791                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_LERP_ATI, dstreg, GL_SATURATE_BIT_ATI,
 792                                                  arg1, GL_NONE, argmod1,
 793                                                  arg2, GL_NONE, argmod2,
 794                                                  arg0, GL_NONE, argmod0));
 795                 break;
 796
 797             case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
 798             case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
 799             case WINED3DTOP_MODULATECOLOR_ADDALPHA:
 800             case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
 801             case WINED3DTOP_BUMPENVMAP:
 802             case WINED3DTOP_BUMPENVMAPLUMINANCE:
 803                 ERR("Application uses an invalid alpha operation\n");
 804                 break;
 805
 806             default: FIXME("Unhandled alpha operation %d on stage %d\n", op[stage].aop, stage);
 807         }
 808     }
 809
 810     TRACE("glEndFragmentShaderATI()\n");
 811     GL_EXTCALL(glEndFragmentShaderATI());
 812     checkGLcall("GL_EXTCALL(glEndFragmentShaderATI())");
 813     return ret;
 814 }
 815 #undef GLINFO_LOCATION
 816
 817 #define GLINFO_LOCATION stateblock->wineD3DDevice->adapter->gl_info
 818 static void set_tex_op_atifs(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
 819     IWineD3DDeviceImpl          *This = stateblock->wineD3DDevice;
 820     const struct atifs_ffp_desc *desc;
 821     struct ffp_frag_settings     settings;
 822     struct atifs_private_data   *priv = (struct atifs_private_data *) This->fragment_priv;
 823     DWORD mapped_stage;
 824     unsigned int i;
 825
 826     gen_ffp_frag_op(stateblock, &settings, TRUE);
 827     desc = (const struct atifs_ffp_desc *)find_ffp_frag_shader(priv->fragment_shaders, &settings);
 828     if(!desc) {
 829         struct atifs_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
 830         if (!new_desc)
 831         {
 832             ERR("Out of memory\n");
 833             return;
 834         }
 835         new_desc->num_textures_used = 0;
 836         for(i = 0; i < GL_LIMITS(texture_stages); i++) {
 837             if(settings.op[i].cop == WINED3DTOP_DISABLE) break;
 838             new_desc->num_textures_used = i;
 839         }
 840
 841         memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
 842         new_desc->shader = gen_ati_shader(settings.op, &GLINFO_LOCATION);
 843         add_ffp_frag_shader(priv->fragment_shaders, &new_desc->parent);
 844         TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
 845         desc = new_desc;
 846     }
 847
 848     /* GL_ATI_fragment_shader depends on the GL_TEXTURE_xD enable settings. Update the texture stages
 849      * used by this shader
 850      */
 851     for(i = 0; i < desc->num_textures_used; i++) {
 852         mapped_stage = This->texUnitMap[i];
 853         if(mapped_stage != -1) {
 854             GL_EXTCALL(glActiveTextureARB(GL_TEXTURE0_ARB + mapped_stage));
 855             checkGLcall("glActiveTextureARB");
 856             texture_activate_dimensions(i, stateblock, context);
 857         }
 858     }
 859
 860     GL_EXTCALL(glBindFragmentShaderATI(desc->shader));
 861 }
 862
 863 static void state_texfactor_atifs(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
 864     float col[4];
 865     D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
 866
 867     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col));
 868     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col)");
 869 }
 870
 871 static void set_bumpmat(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
 872     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / WINED3D_HIGHEST_TEXTURE_STATE;
 873     float mat[2][2];
 874
 875     mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
 876     mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
 877     mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
 878     mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
 879     /* GL_ATI_fragment_shader allows only constants from 0.0 to 1.0, but the bumpmat
 880      * constants can be in any range. While they should stay between [-1.0 and 1.0] because
 881      * Shader Model 1.x pixel shaders are clamped to that range negative values are used occasionally,
 882      * for example by our d3d9 test. So to get negative values scale -1;1 to 0;1 and undo that in the
 883      * shader(it is free). This might potentially reduce precision. However, if the hardware does
 884      * support proper floats it shouldn't, and if it doesn't we can't get anything better anyway
 885      */
 886     mat[0][0] = (mat[0][0] + 1.0) * 0.5;
 887     mat[1][0] = (mat[1][0] + 1.0) * 0.5;
 888     mat[0][1] = (mat[0][1] + 1.0) * 0.5;
 889     mat[1][1] = (mat[1][1] + 1.0) * 0.5;
 890     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), (float *) mat));
 891     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), mat)");
 892 }
 893
 894 static void textransform(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
 895     if(!isStateDirty(context, STATE_PIXELSHADER)) {
 896         set_tex_op_atifs(state, stateblock, context);
 897     }
 898 }
 899
 900 #undef GLINFO_LOCATION
 901
 902 static const struct StateEntryTemplate atifs_fragmentstate_template[] = {
 903     {STATE_RENDER(WINED3DRS_TEXTUREFACTOR),               { STATE_RENDER(WINED3DRS_TEXTUREFACTOR),              state_texfactor_atifs   }, 0                               },
 904     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 905     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 906     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 907     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 908     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 909     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 910     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 911     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 912     {STATE_TEXTURESTAGE(0, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 913     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 914     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 915     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 916     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 917     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 918     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 919     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 920     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 921     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 922     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 923     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 924     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 925     {STATE_TEXTURESTAGE(1, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 926     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 927     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 928     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 929     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 930     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 931     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 932     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 933     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 934     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 935     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 936     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 937     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 938     {STATE_TEXTURESTAGE(2, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 939     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 940     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 941     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 942     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 943     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 944     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 945     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 946     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 947     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 948     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 949     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 950     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 951     {STATE_TEXTURESTAGE(3, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 952     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 953     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 954     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 955     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 956     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 957     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 958     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 959     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 960     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 961     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 962     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 963     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 964     {STATE_TEXTURESTAGE(4, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 965     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 966     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 967     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 968     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 969     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 970     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 971     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 972     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 973     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 974     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 975     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 976     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 977     {STATE_TEXTURESTAGE(5, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 978     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 979     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 980     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 981     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 982     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 983     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 984     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 985     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 986     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 987     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 988     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 989     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 990     {STATE_TEXTURESTAGE(6, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 991     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 992     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 993     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 994     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
 995     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 996     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 997     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 998     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
 999     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
1000     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
1001     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
1002     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
1003     {STATE_TEXTURESTAGE(7, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, 0                               },
1004     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
1005     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
1006     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
1007     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, 0                               },
1008     { STATE_SAMPLER(0),                                   { STATE_SAMPLER(0),                                   sampler_texdim          }, 0                               },
1009     { STATE_SAMPLER(1),                                   { STATE_SAMPLER(1),                                   sampler_texdim          }, 0                               },
1010     { STATE_SAMPLER(2),                                   { STATE_SAMPLER(2),                                   sampler_texdim          }, 0                               },
1011     { STATE_SAMPLER(3),                                   { STATE_SAMPLER(3),                                   sampler_texdim          }, 0                               },
1012     { STATE_SAMPLER(4),                                   { STATE_SAMPLER(4),                                   sampler_texdim          }, 0                               },
1013     { STATE_SAMPLER(5),                                   { STATE_SAMPLER(5),                                   sampler_texdim          }, 0                               },
1014     { STATE_SAMPLER(6),                                   { STATE_SAMPLER(6),                                   sampler_texdim          }, 0                               },
1015     { STATE_SAMPLER(7),                                   { STATE_SAMPLER(7),                                   sampler_texdim          }, 0                               },
1016     {STATE_TEXTURESTAGE(0,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(0, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
1017     {STATE_TEXTURESTAGE(1,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(1, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
1018     {STATE_TEXTURESTAGE(2,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(2, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
1019     {STATE_TEXTURESTAGE(3,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(3, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
1020     {STATE_TEXTURESTAGE(4,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(4, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
1021     {STATE_TEXTURESTAGE(5,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(5, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
1022     {STATE_TEXTURESTAGE(6,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(6, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
1023     {STATE_TEXTURESTAGE(7,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(7, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
1024     {0 /* Terminate */,                                   { 0,                                                  0                       }, 0                               },
1025 };
1026
1027 static void atifs_enable(IWineD3DDevice *iface, BOOL enable) {
1028     if(enable) {
1029         glEnable(GL_FRAGMENT_SHADER_ATI);
1030         checkGLcall("glEnable(GL_FRAGMENT_SHADER_ATI)");
1031     } else {
1032         glDisable(GL_FRAGMENT_SHADER_ATI);
1033         checkGLcall("glDisable(GL_FRAGMENT_SHADER_ATI)");
1034     }
1035 }
1036
1037 static void atifs_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct fragment_caps *caps)
1038 {
1039     caps->TextureOpCaps =  WINED3DTEXOPCAPS_DISABLE                     |
1040                            WINED3DTEXOPCAPS_SELECTARG1                  |
1041                            WINED3DTEXOPCAPS_SELECTARG2                  |
1042                            WINED3DTEXOPCAPS_MODULATE4X                  |
1043                            WINED3DTEXOPCAPS_MODULATE2X                  |
1044                            WINED3DTEXOPCAPS_MODULATE                    |
1045                            WINED3DTEXOPCAPS_ADDSIGNED2X                 |
1046                            WINED3DTEXOPCAPS_ADDSIGNED                   |
1047                            WINED3DTEXOPCAPS_ADD                         |
1048                            WINED3DTEXOPCAPS_SUBTRACT                    |
1049                            WINED3DTEXOPCAPS_ADDSMOOTH                   |
1050                            WINED3DTEXOPCAPS_BLENDCURRENTALPHA           |
1051                            WINED3DTEXOPCAPS_BLENDFACTORALPHA            |
1052                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHA           |
1053                            WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA           |
1054                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM         |
1055                            WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR      |
1056                            WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA      |
1057                            WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA   |
1058                            WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR   |
1059                            WINED3DTEXOPCAPS_DOTPRODUCT3                 |
1060                            WINED3DTEXOPCAPS_MULTIPLYADD                 |
1061                            WINED3DTEXOPCAPS_LERP                        |
1062                            WINED3DTEXOPCAPS_BUMPENVMAP;
1063
1064     /* TODO: Implement WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE
1065     and WINED3DTEXOPCAPS_PREMODULATE */
1066
1067     /* GL_ATI_fragment_shader always supports 6 textures, which was the limit on r200 cards
1068      * which this extension is exclusively focused on(later cards have GL_ARB_fragment_program).
1069      * If the current card has more than 8 fixed function textures in OpenGL's regular fixed
1070      * function pipeline then the ATI_fragment_shader backend imposes a stricter limit. This
1071      * shouldn't be too hard since Nvidia cards have a limit of 4 textures with the default ffp
1072      * pipeline, and almost all games are happy with that. We can however support up to 8
1073      * texture stages because we have a 2nd pass limit of 8 instructions, and per stage we use
1074      * only 1 instruction.
1075      *
1076      * The proper fix for this is not to use GL_ATI_fragment_shader on cards newer than the
1077      * r200 series and use an ARB or GLSL shader instead
1078      */
1079     caps->MaxTextureBlendStages   = 8;
1080     caps->MaxSimultaneousTextures = 6;
1081
1082     caps->PrimitiveMiscCaps |= WINED3DPMISCCAPS_TSSARGTEMP;
1083 }
1084
1085 static HRESULT atifs_alloc(IWineD3DDevice *iface) {
1086     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
1087     struct atifs_private_data *priv;
1088
1089     This->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct atifs_private_data));
1090     if(!This->fragment_priv) {
1091         ERR("Out of memory\n");
1092         return E_OUTOFMEMORY;
1093     }
1094     priv = (struct atifs_private_data *) This->fragment_priv;
1095     priv->fragment_shaders = hash_table_create(ffp_frag_program_key_hash, ffp_frag_program_key_compare);
1096     return WINED3D_OK;
1097 }
1098
1099 #define GLINFO_LOCATION This->adapter->gl_info
1100 static void atifs_free_ffpshader(void *value, void *device) {
1101     IWineD3DDeviceImpl *This = device;
1102     struct atifs_ffp_desc *entry_ati = value;
1103
1104     ENTER_GL();
1105     GL_EXTCALL(glDeleteFragmentShaderATI(entry_ati->shader));
1106     checkGLcall("glDeleteFragmentShaderATI(entry->shader)");
1107     HeapFree(GetProcessHeap(), 0, entry_ati);
1108     LEAVE_GL();
1109 }
1110
1111 static void atifs_free(IWineD3DDevice *iface) {
1112     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
1113     struct atifs_private_data *priv = (struct atifs_private_data *) This->fragment_priv;
1114
1115     hash_table_destroy(priv->fragment_shaders, atifs_free_ffpshader, This);
1116
1117     HeapFree(GetProcessHeap(), 0, priv);
1118     This->fragment_priv = NULL;
1119 }
1120 #undef GLINFO_LOCATION
1121
1122 static BOOL atifs_color_fixup_supported(struct color_fixup_desc fixup)
1123 {
1124     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1125     {
1126         TRACE("Checking support for fixup:\n");
1127         dump_color_fixup_desc(fixup);
1128     }
1129
1130     /* We only support sign fixup of the first two channels. */
1131     if (fixup.x_source == CHANNEL_SOURCE_X && fixup.y_source == CHANNEL_SOURCE_Y
1132             && fixup.z_source == CHANNEL_SOURCE_Z && fixup.w_source == CHANNEL_SOURCE_W
1133             && !fixup.z_sign_fixup && !fixup.w_sign_fixup)
1134     {
1135         TRACE("[OK]\n");
1136         return TRUE;
1137     }
1138
1139     TRACE("[FAILED]\n");
1140     return FALSE;
1141 }
1142
1143 const struct fragment_pipeline atifs_fragment_pipeline = {
1144     atifs_enable,
1145     atifs_get_caps,
1146     atifs_alloc,
1147     atifs_free,
1148     atifs_color_fixup_supported,
1149     atifs_fragmentstate_template,
1150     TRUE /* We can disable projected textures */
1151 };