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 /* GL locking for state handlers is done by the caller. */
  32
  33 /* Some private defines, Constant associations, etc.
  34  * Env bump matrix and per stage constant should be independent,
  35  * a stage that bump maps can't read the per state constant
  36  */
  37 #define ATI_FFP_CONST_BUMPMAT(i) (GL_CON_0_ATI + i)
  38 #define ATI_FFP_CONST_CONSTANT0 GL_CON_0_ATI
  39 #define ATI_FFP_CONST_CONSTANT1 GL_CON_1_ATI
  40 #define ATI_FFP_CONST_CONSTANT2 GL_CON_2_ATI
  41 #define ATI_FFP_CONST_CONSTANT3 GL_CON_3_ATI
  42 #define ATI_FFP_CONST_CONSTANT4 GL_CON_4_ATI
  43 #define ATI_FFP_CONST_CONSTANT5 GL_CON_5_ATI
  44 #define ATI_FFP_CONST_TFACTOR   GL_CON_6_ATI
  45
  46 /* GL_ATI_fragment_shader specific fixed function pipeline description. "Inherits" from the common one */
  47 struct atifs_ffp_desc
  48 {
  49     struct ffp_frag_desc parent;
  50     GLuint shader;
  51     unsigned int num_textures_used;
  52 };
  53
  54 struct atifs_private_data
  55 {
  56     struct wine_rb_tree fragment_shaders; /* A rb-tree to track fragment pipeline replacement shaders */
  57 };
  58
  59 static const char *debug_dstmod(GLuint mod) {
  60     switch(mod) {
  61         case GL_NONE:               return "GL_NONE";
  62         case GL_2X_BIT_ATI:         return "GL_2X_BIT_ATI";
  63         case GL_4X_BIT_ATI:         return "GL_4X_BIT_ATI";
  64         case GL_8X_BIT_ATI:         return "GL_8X_BIT_ATI";
  65         case GL_HALF_BIT_ATI:       return "GL_HALF_BIT_ATI";
  66         case GL_QUARTER_BIT_ATI:    return "GL_QUARTER_BIT_ATI";
  67         case GL_EIGHTH_BIT_ATI:     return "GL_EIGHTH_BIT_ATI";
  68         case GL_SATURATE_BIT_ATI:   return "GL_SATURATE_BIT_ATI";
  69         default:                    return "Unexpected modifier\n";
  70     }
  71 }
  72
  73 static const char *debug_argmod(GLuint mod) {
  74     switch(mod) {
  75         case GL_NONE:
  76             return "GL_NONE";
  77
  78         case GL_2X_BIT_ATI:
  79             return "GL_2X_BIT_ATI";
  80         case GL_COMP_BIT_ATI:
  81             return "GL_COMP_BIT_ATI";
  82         case GL_NEGATE_BIT_ATI:
  83             return "GL_NEGATE_BIT_ATI";
  84         case GL_BIAS_BIT_ATI:
  85             return "GL_BIAS_BIT_ATI";
  86
  87         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI:
  88             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI";
  89         case GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI:
  90             return "GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI";
  91         case GL_2X_BIT_ATI | GL_BIAS_BIT_ATI:
  92             return "GL_2X_BIT_ATI | GL_BIAS_BIT_ATI";
  93         case GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI:
  94             return "GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI";
  95         case GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI:
  96             return "GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI";
  97         case GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
  98             return "GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
  99
 100         case GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 101             return "GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 102         case GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 103             return "GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 104         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI:
 105             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI";
 106         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI:
 107             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI";
 108
 109         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 110             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 111
 112         default:
 113             return "Unexpected argmod combination\n";
 114     }
 115 }
 116 static const char *debug_register(GLuint reg) {
 117     switch(reg) {
 118         case GL_REG_0_ATI:                  return "GL_REG_0_ATI";
 119         case GL_REG_1_ATI:                  return "GL_REG_1_ATI";
 120         case GL_REG_2_ATI:                  return "GL_REG_2_ATI";
 121         case GL_REG_3_ATI:                  return "GL_REG_3_ATI";
 122         case GL_REG_4_ATI:                  return "GL_REG_4_ATI";
 123         case GL_REG_5_ATI:                  return "GL_REG_5_ATI";
 124
 125         case GL_CON_0_ATI:                  return "GL_CON_0_ATI";
 126         case GL_CON_1_ATI:                  return "GL_CON_1_ATI";
 127         case GL_CON_2_ATI:                  return "GL_CON_2_ATI";
 128         case GL_CON_3_ATI:                  return "GL_CON_3_ATI";
 129         case GL_CON_4_ATI:                  return "GL_CON_4_ATI";
 130         case GL_CON_5_ATI:                  return "GL_CON_5_ATI";
 131         case GL_CON_6_ATI:                  return "GL_CON_6_ATI";
 132         case GL_CON_7_ATI:                  return "GL_CON_7_ATI";
 133
 134         case GL_ZERO:                       return "GL_ZERO";
 135         case GL_ONE:                        return "GL_ONE";
 136         case GL_PRIMARY_COLOR:              return "GL_PRIMARY_COLOR";
 137         case GL_SECONDARY_INTERPOLATOR_ATI: return "GL_SECONDARY_INTERPOLATOR_ATI";
 138
 139         default:                            return "Unknown register\n";
 140     }
 141 }
 142
 143 static const char *debug_swizzle(GLuint swizzle) {
 144     switch(swizzle) {
 145         case GL_SWIZZLE_STR_ATI:        return "GL_SWIZZLE_STR_ATI";
 146         case GL_SWIZZLE_STQ_ATI:        return "GL_SWIZZLE_STQ_ATI";
 147         case GL_SWIZZLE_STR_DR_ATI:     return "GL_SWIZZLE_STR_DR_ATI";
 148         case GL_SWIZZLE_STQ_DQ_ATI:     return "GL_SWIZZLE_STQ_DQ_ATI";
 149         default:                        return "unknown swizzle";
 150     }
 151 }
 152
 153 static const char *debug_rep(GLuint rep) {
 154     switch(rep) {
 155         case GL_NONE:                   return "GL_NONE";
 156         case GL_RED:                    return "GL_RED";
 157         case GL_GREEN:                  return "GL_GREEN";
 158         case GL_BLUE:                   return "GL_BLUE";
 159         default:                        return "unknown argrep";
 160     }
 161 }
 162
 163 static const char *debug_op(GLuint op) {
 164     switch(op) {
 165         case GL_MOV_ATI:                return "GL_MOV_ATI";
 166         case GL_ADD_ATI:                return "GL_ADD_ATI";
 167         case GL_MUL_ATI:                return "GL_MUL_ATI";
 168         case GL_SUB_ATI:                return "GL_SUB_ATI";
 169         case GL_DOT3_ATI:               return "GL_DOT3_ATI";
 170         case GL_DOT4_ATI:               return "GL_DOT4_ATI";
 171         case GL_MAD_ATI:                return "GL_MAD_ATI";
 172         case GL_LERP_ATI:               return "GL_LERP_ATI";
 173         case GL_CND_ATI:                return "GL_CND_ATI";
 174         case GL_CND0_ATI:               return "GL_CND0_ATI";
 175         case GL_DOT2_ADD_ATI:           return "GL_DOT2_ADD_ATI";
 176         default:                        return "unexpected op";
 177     }
 178 }
 179
 180 static const char *debug_mask(GLuint mask) {
 181     switch(mask) {
 182         case GL_NONE:                           return "GL_NONE";
 183         case GL_RED_BIT_ATI:                    return "GL_RED_BIT_ATI";
 184         case GL_GREEN_BIT_ATI:                  return "GL_GREEN_BIT_ATI";
 185         case GL_BLUE_BIT_ATI:                   return "GL_BLUE_BIT_ATI";
 186         case GL_RED_BIT_ATI | GL_GREEN_BIT_ATI: return "GL_RED_BIT_ATI | GL_GREEN_BIT_ATI";
 187         case GL_RED_BIT_ATI | GL_BLUE_BIT_ATI:  return "GL_RED_BIT_ATI | GL_BLUE_BIT_ATI";
 188         case GL_GREEN_BIT_ATI | GL_BLUE_BIT_ATI:return "GL_GREEN_BIT_ATI | GL_BLUE_BIT_ATI";
 189         case GL_RED_BIT_ATI | GL_GREEN_BIT_ATI | GL_BLUE_BIT_ATI:return "GL_RED_BIT_ATI | GL_GREEN_BIT_ATI | GL_BLUE_BIT_ATI";
 190         default:                                return "Unexpected writemask";
 191     }
 192 }
 193
 194 static void wrap_op1(const struct wined3d_gl_info *gl_info, GLuint op, GLuint dst, GLuint dstMask, GLuint dstMod,
 195         GLuint arg1, GLuint arg1Rep, GLuint arg1Mod)
 196 {
 197     if(dstMask == GL_ALPHA) {
 198         TRACE("glAlphaFragmentOp1ATI(%s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst), debug_dstmod(dstMod),
 199               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod));
 200         GL_EXTCALL(glAlphaFragmentOp1ATI(op, dst, dstMod, arg1, arg1Rep, arg1Mod));
 201     } else {
 202         TRACE("glColorFragmentOp1ATI(%s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst),
 203               debug_mask(dstMask), debug_dstmod(dstMod),
 204               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod));
 205         GL_EXTCALL(glColorFragmentOp1ATI(op, dst, dstMask, dstMod, arg1, arg1Rep, arg1Mod));
 206     }
 207 }
 208
 209 static void wrap_op2(const struct wined3d_gl_info *gl_info, GLuint op, GLuint dst, GLuint dstMask, GLuint dstMod,
 210         GLuint arg1, GLuint arg1Rep, GLuint arg1Mod, GLuint arg2, GLuint arg2Rep, GLuint arg2Mod)
 211 {
 212     if(dstMask == GL_ALPHA) {
 213         TRACE("glAlphaFragmentOp2ATI(%s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst), debug_dstmod(dstMod),
 214               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 215               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod));
 216         GL_EXTCALL(glAlphaFragmentOp2ATI(op, dst, dstMod, arg1, arg1Rep, arg1Mod, arg2, arg2Rep, arg2Mod));
 217     } else {
 218         TRACE("glColorFragmentOp2ATI(%s, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst),
 219               debug_mask(dstMask), debug_dstmod(dstMod),
 220               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 221               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod));
 222         GL_EXTCALL(glColorFragmentOp2ATI(op, dst, dstMask, dstMod, arg1, arg1Rep, arg1Mod, arg2, arg2Rep, arg2Mod));
 223     }
 224 }
 225
 226 static void wrap_op3(const struct wined3d_gl_info *gl_info, GLuint op, GLuint dst, GLuint dstMask, GLuint dstMod,
 227         GLuint arg1, GLuint arg1Rep, GLuint arg1Mod, GLuint arg2, GLuint arg2Rep, GLuint arg2Mod,
 228         GLuint arg3, GLuint arg3Rep, GLuint arg3Mod)
 229 {
 230     if(dstMask == GL_ALPHA) {
 231         /* Leave some free space to fit "GL_NONE, " in to align most alpha and color op lines */
 232         TRACE("glAlphaFragmentOp3ATI(%s, %s,          %s, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst), debug_dstmod(dstMod),
 233               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 234               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod),
 235               debug_register(arg3), debug_rep(arg3Rep), debug_argmod(arg3Mod));
 236         GL_EXTCALL(glAlphaFragmentOp3ATI(op, dst, dstMod,
 237                                          arg1, arg1Rep, arg1Mod,
 238                                          arg2, arg2Rep, arg2Mod,
 239                                          arg3, arg3Rep, arg3Mod));
 240     } else {
 241         TRACE("glColorFragmentOp3ATI(%s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst),
 242               debug_mask(dstMask), debug_dstmod(dstMod),
 243               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 244               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod),
 245               debug_register(arg3), debug_rep(arg3Rep), debug_argmod(arg3Mod));
 246         GL_EXTCALL(glColorFragmentOp3ATI(op, dst, dstMask, dstMod,
 247                                          arg1, arg1Rep, arg1Mod,
 248                                          arg2, arg2Rep, arg2Mod,
 249                                          arg3, arg3Rep, arg3Mod));
 250     }
 251 }
 252
 253 static GLuint register_for_arg(DWORD arg, const struct wined3d_gl_info *gl_info,
 254         unsigned int stage, GLuint *mod, GLuint *rep, GLuint tmparg)
 255 {
 256     GLenum ret;
 257
 258     if(mod) *mod = GL_NONE;
 259     if(arg == ARG_UNUSED)
 260     {
 261         if (rep) *rep = GL_NONE;
 262         return -1; /* This is the marker for unused registers */
 263     }
 264
 265     switch(arg & WINED3DTA_SELECTMASK) {
 266         case WINED3DTA_DIFFUSE:
 267             ret = GL_PRIMARY_COLOR;
 268             break;
 269
 270         case WINED3DTA_CURRENT:
 271             /* Note that using GL_REG_0_ATI for the passed on register is safe because
 272              * texture0 is read at stage0, so in the worst case it is read in the
 273              * instruction writing to reg0. Afterwards texture0 is not used any longer.
 274              * If we're reading from current
 275              */
 276             ret = stage ? GL_REG_0_ATI : GL_PRIMARY_COLOR;
 277             break;
 278
 279         case WINED3DTA_TEXTURE:
 280             ret = GL_REG_0_ATI + stage;
 281             break;
 282
 283         case WINED3DTA_TFACTOR:
 284             ret = ATI_FFP_CONST_TFACTOR;
 285             break;
 286
 287         case WINED3DTA_SPECULAR:
 288             ret = GL_SECONDARY_INTERPOLATOR_ATI;
 289             break;
 290
 291         case WINED3DTA_TEMP:
 292             ret = tmparg;
 293             break;
 294
 295         case WINED3DTA_CONSTANT:
 296             FIXME("Unhandled source argument WINED3DTA_TEMP\n");
 297             ret = GL_CON_0_ATI;
 298             break;
 299
 300         default:
 301             FIXME("Unknown source argument %d\n", arg);
 302             ret = GL_ZERO;
 303     }
 304
 305     if(arg & WINED3DTA_COMPLEMENT) {
 306         if(mod) *mod |= GL_COMP_BIT_ATI;
 307     }
 308     if(arg & WINED3DTA_ALPHAREPLICATE) {
 309         if(rep) *rep = GL_ALPHA;
 310     } else {
 311         if(rep) *rep = GL_NONE;
 312     }
 313     return ret;
 314 }
 315
 316 static GLuint find_tmpreg(const struct texture_stage_op op[MAX_TEXTURES])
 317 {
 318     int lowest_read = -1;
 319     int lowest_write = -1;
 320     int i;
 321     BOOL tex_used[MAX_TEXTURES];
 322
 323     memset(tex_used, 0, sizeof(tex_used));
 324     for (i = 0; i < MAX_TEXTURES; ++i)
 325     {
 326         if (op[i].cop == WINED3D_TOP_DISABLE)
 327             break;
 328
 329         if(lowest_read == -1 &&
 330           (op[i].carg1 == WINED3DTA_TEMP || op[i].carg2 == WINED3DTA_TEMP || op[i].carg0 == WINED3DTA_TEMP ||
 331            op[i].aarg1 == WINED3DTA_TEMP || op[i].aarg2 == WINED3DTA_TEMP || op[i].aarg0 == WINED3DTA_TEMP)) {
 332             lowest_read = i;
 333         }
 334
 335         if(lowest_write == -1 && op[i].dst == tempreg) {
 336             lowest_write = i;
 337         }
 338
 339         if(op[i].carg1 == WINED3DTA_TEXTURE || op[i].carg2 == WINED3DTA_TEXTURE || op[i].carg0 == WINED3DTA_TEXTURE ||
 340            op[i].aarg1 == WINED3DTA_TEXTURE || op[i].aarg2 == WINED3DTA_TEXTURE || op[i].aarg0 == WINED3DTA_TEXTURE) {
 341             tex_used[i] = TRUE;
 342         }
 343     }
 344
 345     /* Temp reg not read? We don't need it, return GL_NONE */
 346     if(lowest_read == -1) return GL_NONE;
 347
 348     if(lowest_write >= lowest_read) {
 349         FIXME("Temp register read before being written\n");
 350     }
 351
 352     if(lowest_write == -1) {
 353         /* This needs a test. Maybe we are supposed to return 0.0/0.0/0.0/0.0, or fail drawprim, or whatever */
 354         FIXME("Temp register read without being written\n");
 355         return GL_REG_1_ATI;
 356     } else if(lowest_write >= 1) {
 357         /* If we're writing to the temp reg at earliest in stage 1, we can use register 1 for the temp result.
 358          * there may be texture data stored in reg 1, but we do not need it any longer since stage 1 already
 359          * read it
 360          */
 361         return GL_REG_1_ATI;
 362     } else {
 363         /* Search for a free texture register. We have 6 registers available. GL_REG_0_ATI is already used
 364          * for the regular result
 365          */
 366         for(i = 1; i < 6; i++) {
 367             if(!tex_used[i]) {
 368                 return GL_REG_0_ATI + i;
 369             }
 370         }
 371         /* What to do here? Report it in ValidateDevice? */
 372         FIXME("Could not find a register for the temporary register\n");
 373         return 0;
 374     }
 375 }
 376
 377 static GLuint gen_ati_shader(const struct texture_stage_op op[MAX_TEXTURES], const struct wined3d_gl_info *gl_info)
 378 {
 379     GLuint ret = GL_EXTCALL(glGenFragmentShadersATI(1));
 380     unsigned int stage;
 381     GLuint arg0, arg1, arg2, extrarg;
 382     GLuint dstmod, argmod0, argmod1, argmod2, argmodextra;
 383     GLuint rep0, rep1, rep2;
 384     GLuint swizzle;
 385     GLuint tmparg = find_tmpreg(op);
 386     GLuint dstreg;
 387
 388     if(!ret) {
 389         ERR("Failed to generate a GL_ATI_fragment_shader shader id\n");
 390         return 0;
 391     }
 392     GL_EXTCALL(glBindFragmentShaderATI(ret));
 393     checkGLcall("GL_EXTCALL(glBindFragmentShaderATI(ret))");
 394
 395     TRACE("glBeginFragmentShaderATI()\n");
 396     GL_EXTCALL(glBeginFragmentShaderATI());
 397     checkGLcall("GL_EXTCALL(glBeginFragmentShaderATI())");
 398
 399     /* Pass 1: Generate sampling instructions for perturbation maps */
 400     for (stage = 0; stage < gl_info->limits.textures; ++stage)
 401     {
 402         if (op[stage].cop == WINED3D_TOP_DISABLE)
 403             break;
 404         if (op[stage].cop != WINED3D_TOP_BUMPENVMAP
 405                 && op[stage].cop != WINED3D_TOP_BUMPENVMAP_LUMINANCE)
 406             continue;
 407
 408         TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, GL_SWIZZLE_STR_ATI)\n",
 409               stage, stage);
 410         GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 411                    GL_TEXTURE0_ARB + stage,
 412                    GL_SWIZZLE_STR_ATI));
 413         if(op[stage + 1].projected == proj_none) {
 414             swizzle = GL_SWIZZLE_STR_ATI;
 415         } else if(op[stage + 1].projected == proj_count4) {
 416             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 417         } else {
 418             swizzle = GL_SWIZZLE_STR_DR_ATI;
 419         }
 420         TRACE("glPassTexCoordATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, %s)\n",
 421               stage + 1, stage + 1, debug_swizzle(swizzle));
 422         GL_EXTCALL(glPassTexCoordATI(GL_REG_0_ATI + stage + 1,
 423                    GL_TEXTURE0_ARB + stage + 1,
 424                    swizzle));
 425     }
 426
 427     /* Pass 2: Generate perturbation calculations */
 428     for (stage = 0; stage < gl_info->limits.textures; ++stage)
 429     {
 430         GLuint argmodextra_x, argmodextra_y;
 431         struct color_fixup_desc fixup;
 432
 433         if (op[stage].cop == WINED3D_TOP_DISABLE)
 434             break;
 435         if (op[stage].cop != WINED3D_TOP_BUMPENVMAP
 436                 && op[stage].cop != WINED3D_TOP_BUMPENVMAP_LUMINANCE)
 437             continue;
 438
 439         fixup = op[stage].color_fixup;
 440         if (fixup.x_source != CHANNEL_SOURCE_X || fixup.y_source != CHANNEL_SOURCE_Y)
 441         {
 442             FIXME("Swizzles not implemented\n");
 443             argmodextra_x = GL_NONE;
 444             argmodextra_y = GL_NONE;
 445         }
 446         else
 447         {
 448             /* Nice thing, we get the color correction for free :-) */
 449             argmodextra_x = fixup.x_sign_fixup ? GL_2X_BIT_ATI | GL_BIAS_BIT_ATI : GL_NONE;
 450             argmodextra_y = fixup.y_sign_fixup ? GL_2X_BIT_ATI | GL_BIAS_BIT_ATI : GL_NONE;
 451         }
 452
 453         wrap_op3(gl_info, GL_DOT2_ADD_ATI, GL_REG_0_ATI + stage + 1, GL_RED_BIT_ATI, GL_NONE,
 454                  GL_REG_0_ATI + stage, GL_NONE, argmodextra_x,
 455                  ATI_FFP_CONST_BUMPMAT(stage), GL_NONE, GL_2X_BIT_ATI | GL_BIAS_BIT_ATI,
 456                  GL_REG_0_ATI + stage + 1, GL_RED, GL_NONE);
 457
 458         /* Don't use GL_DOT2_ADD_ATI here because we cannot configure it to read the blue and alpha
 459          * component of the bump matrix. Instead do this with two MADs:
 460          *
 461          * coord.a = tex.r * bump.b + coord.g
 462          * coord.g = tex.g * bump.a + coord.a
 463          *
 464          * The first instruction writes to alpha so it can be coissued with the above DOT2_ADD.
 465          * coord.a is unused. If the perturbed texture is projected, this was already handled
 466          * in the glPassTexCoordATI above.
 467          */
 468         wrap_op3(gl_info, GL_MAD_ATI, GL_REG_0_ATI + stage + 1, GL_ALPHA, GL_NONE,
 469                  GL_REG_0_ATI + stage, GL_RED, argmodextra_y,
 470                  ATI_FFP_CONST_BUMPMAT(stage), GL_BLUE, GL_NONE,
 471                  GL_REG_0_ATI + stage + 1, GL_GREEN, GL_NONE);
 472         wrap_op3(gl_info, GL_MAD_ATI, GL_REG_0_ATI + stage + 1, GL_GREEN_BIT_ATI, GL_NONE,
 473                  GL_REG_0_ATI + stage, GL_GREEN, argmodextra_y,
 474                  ATI_FFP_CONST_BUMPMAT(stage), GL_ALPHA, GL_NONE,
 475                  GL_REG_0_ATI + stage + 1, GL_ALPHA, GL_NONE);
 476     }
 477
 478     /* Pass 3: Generate sampling instructions for regular textures */
 479     for (stage = 0; stage < gl_info->limits.textures; ++stage)
 480     {
 481         if (op[stage].cop == WINED3D_TOP_DISABLE)
 482             break;
 483
 484         if(op[stage].projected == proj_none) {
 485             swizzle = GL_SWIZZLE_STR_ATI;
 486         } else if(op[stage].projected == proj_count3) {
 487             swizzle = GL_SWIZZLE_STR_DR_ATI;
 488         } else {
 489             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 490         }
 491
 492         if ((op[stage].carg0 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE
 493                 || (op[stage].carg1 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE
 494                 || (op[stage].carg2 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE
 495                 || (op[stage].aarg0 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE
 496                 || (op[stage].aarg1 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE
 497                 || (op[stage].aarg2 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE
 498                 || op[stage].cop == WINED3D_TOP_BLEND_TEXTURE_ALPHA)
 499         {
 500             if (stage > 0
 501                     && (op[stage - 1].cop == WINED3D_TOP_BUMPENVMAP
 502                     || op[stage - 1].cop == WINED3D_TOP_BUMPENVMAP_LUMINANCE))
 503             {
 504                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_REG_%d_ATI, GL_SWIZZLE_STR_ATI)\n",
 505                       stage, stage);
 506                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 507                            GL_REG_0_ATI + stage,
 508                            GL_SWIZZLE_STR_ATI));
 509             } else {
 510                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, %s)\n",
 511                     stage, stage, debug_swizzle(swizzle));
 512                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 513                                         GL_TEXTURE0_ARB + stage,
 514                                         swizzle));
 515             }
 516         }
 517     }
 518
 519     /* Pass 4: Generate the arithmetic instructions */
 520     for (stage = 0; stage < MAX_TEXTURES; ++stage)
 521     {
 522         if (op[stage].cop == WINED3D_TOP_DISABLE)
 523         {
 524             if (!stage)
 525             {
 526                 /* Handle complete texture disabling gracefully */
 527                 wrap_op1(gl_info, GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 528                          GL_PRIMARY_COLOR, GL_NONE, GL_NONE);
 529                 wrap_op1(gl_info, GL_MOV_ATI, GL_REG_0_ATI, GL_ALPHA, GL_NONE,
 530                          GL_PRIMARY_COLOR, GL_NONE, GL_NONE);
 531             }
 532             break;
 533         }
 534
 535         if(op[stage].dst == tempreg) {
 536             /* If we're writing to D3DTA_TEMP, but never reading from it we don't have to write there in the first place.
 537              * skip the entire stage, this saves some GPU time
 538              */
 539             if(tmparg == GL_NONE) continue;
 540
 541             dstreg = tmparg;
 542         } else {
 543             dstreg = GL_REG_0_ATI;
 544         }
 545
 546         arg0 = register_for_arg(op[stage].carg0, gl_info, stage, &argmod0, &rep0, tmparg);
 547         arg1 = register_for_arg(op[stage].carg1, gl_info, stage, &argmod1, &rep1, tmparg);
 548         arg2 = register_for_arg(op[stage].carg2, gl_info, stage, &argmod2, &rep2, tmparg);
 549         dstmod = GL_NONE;
 550         argmodextra = GL_NONE;
 551         extrarg = GL_NONE;
 552
 553         switch (op[stage].cop)
 554         {
 555             case WINED3D_TOP_SELECT_ARG2:
 556                 arg1 = arg2;
 557                 argmod1 = argmod2;
 558                 rep1 = rep2;
 559                 /* fall through */
 560             case WINED3D_TOP_SELECT_ARG1:
 561                 wrap_op1(gl_info, GL_MOV_ATI, dstreg, GL_NONE, GL_NONE,
 562                          arg1, rep1, argmod1);
 563                 break;
 564
 565             case WINED3D_TOP_MODULATE_4X:
 566                 if(dstmod == GL_NONE) dstmod = GL_4X_BIT_ATI;
 567                 /* fall through */
 568             case WINED3D_TOP_MODULATE_2X:
 569                 if(dstmod == GL_NONE) dstmod = GL_2X_BIT_ATI;
 570                 dstmod |= GL_SATURATE_BIT_ATI;
 571                 /* fall through */
 572             case WINED3D_TOP_MODULATE:
 573                 wrap_op2(gl_info, GL_MUL_ATI, dstreg, GL_NONE, dstmod,
 574                          arg1, rep1, argmod1,
 575                          arg2, rep2, argmod2);
 576                 break;
 577
 578             case WINED3D_TOP_ADD_SIGNED_2X:
 579                 dstmod = GL_2X_BIT_ATI;
 580                 /* fall through */
 581             case WINED3D_TOP_ADD_SIGNED:
 582                 argmodextra = GL_BIAS_BIT_ATI;
 583                 /* fall through */
 584             case WINED3D_TOP_ADD:
 585                 dstmod |= GL_SATURATE_BIT_ATI;
 586                 wrap_op2(gl_info, GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, dstmod,
 587                          arg1, rep1, argmod1,
 588                          arg2, rep2, argmodextra | argmod2);
 589                 break;
 590
 591             case WINED3D_TOP_SUBTRACT:
 592                 dstmod |= GL_SATURATE_BIT_ATI;
 593                 wrap_op2(gl_info, GL_SUB_ATI, dstreg, GL_NONE, dstmod,
 594                          arg1, rep1, argmod1,
 595                          arg2, rep2, argmod2);
 596                 break;
 597
 598             case WINED3D_TOP_ADD_SMOOTH:
 599                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 600                 /* Dst = arg1 + * arg2(1 -arg 1)
 601                  *     = arg2 * (1 - arg1) + arg1
 602                  */
 603                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 604                          arg2, rep2, argmod2,
 605                          arg1, rep1, argmodextra,
 606                          arg1, rep1, argmod1);
 607                 break;
 608
 609             case WINED3D_TOP_BLEND_CURRENT_ALPHA:
 610                 if (extrarg == GL_NONE)
 611                     extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, NULL, -1);
 612                 /* fall through */
 613             case WINED3D_TOP_BLEND_FACTOR_ALPHA:
 614                 if (extrarg == GL_NONE)
 615                     extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, NULL, -1);
 616                 /* fall through */
 617             case WINED3D_TOP_BLEND_TEXTURE_ALPHA:
 618                 if (extrarg == GL_NONE)
 619                     extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 620                 /* fall through */
 621             case WINED3D_TOP_BLEND_DIFFUSE_ALPHA:
 622                 if (extrarg == GL_NONE)
 623                     extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, NULL, -1);
 624                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 625                          extrarg, GL_ALPHA, GL_NONE,
 626                          arg1, rep1, argmod1,
 627                          arg2, rep2, argmod2);
 628                 break;
 629
 630             case WINED3D_TOP_BLEND_TEXTURE_ALPHA_PM:
 631                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 632                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 633                          arg2, rep2,  argmod2,
 634                          arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 635                          arg1, rep1,  argmod1);
 636                 break;
 637
 638             /* D3DTOP_PREMODULATE ???? */
 639
 640             case WINED3D_TOP_MODULATE_INVALPHA_ADD_COLOR:
 641                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 642                 /* fall through */
 643             case WINED3D_TOP_MODULATE_ALPHA_ADD_COLOR:
 644                 if (!argmodextra)
 645                     argmodextra = argmod1;
 646                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 647                          arg2, rep2,  argmod2,
 648                          arg1, GL_ALPHA, argmodextra,
 649                          arg1, rep1,  argmod1);
 650                 break;
 651
 652             case WINED3D_TOP_MODULATE_INVCOLOR_ADD_ALPHA:
 653                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 654                 /* fall through */
 655             case WINED3D_TOP_MODULATE_COLOR_ADD_ALPHA:
 656                 if (!argmodextra)
 657                     argmodextra = argmod1;
 658                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 659                          arg2, rep2,  argmod2,
 660                          arg1, rep1,  argmodextra,
 661                          arg1, GL_ALPHA, argmod1);
 662                 break;
 663
 664             case WINED3D_TOP_DOTPRODUCT3:
 665                 wrap_op2(gl_info, GL_DOT3_ATI, dstreg, GL_NONE, GL_4X_BIT_ATI | GL_SATURATE_BIT_ATI,
 666                          arg1, rep1, argmod1 | GL_BIAS_BIT_ATI,
 667                          arg2, rep2, argmod2 | GL_BIAS_BIT_ATI);
 668                 break;
 669
 670             case WINED3D_TOP_MULTIPLY_ADD:
 671                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 672                          arg1, rep1, argmod1,
 673                          arg2, rep2, argmod2,
 674                          arg0, rep0, argmod0);
 675                 break;
 676
 677             case WINED3D_TOP_LERP:
 678                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 679                          arg0, rep0, argmod0,
 680                          arg1, rep1, argmod1,
 681                          arg2, rep2, argmod2);
 682                 break;
 683
 684             case WINED3D_TOP_BUMPENVMAP:
 685             case WINED3D_TOP_BUMPENVMAP_LUMINANCE:
 686                 /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
 687                 break;
 688
 689             default: FIXME("Unhandled color operation %d on stage %d\n", op[stage].cop, stage);
 690         }
 691
 692         arg0 = register_for_arg(op[stage].aarg0, gl_info, stage, &argmod0, NULL, tmparg);
 693         arg1 = register_for_arg(op[stage].aarg1, gl_info, stage, &argmod1, NULL, tmparg);
 694         arg2 = register_for_arg(op[stage].aarg2, gl_info, stage, &argmod2, NULL, tmparg);
 695         dstmod = GL_NONE;
 696         argmodextra = GL_NONE;
 697         extrarg = GL_NONE;
 698
 699         switch (op[stage].aop)
 700         {
 701             case WINED3D_TOP_DISABLE:
 702                 /* Get the primary color to the output if on stage 0, otherwise leave register 0 untouched */
 703                 if (!stage)
 704                 {
 705                     wrap_op1(gl_info, GL_MOV_ATI, GL_REG_0_ATI, GL_ALPHA, GL_NONE,
 706                              GL_PRIMARY_COLOR, GL_NONE, GL_NONE);
 707                 }
 708                 break;
 709
 710             case WINED3D_TOP_SELECT_ARG2:
 711                 arg1 = arg2;
 712                 argmod1 = argmod2;
 713                 /* fall through */
 714             case WINED3D_TOP_SELECT_ARG1:
 715                 wrap_op1(gl_info, GL_MOV_ATI, dstreg, GL_ALPHA, GL_NONE,
 716                          arg1, GL_NONE, argmod1);
 717                 break;
 718
 719             case WINED3D_TOP_MODULATE_4X:
 720                 if (dstmod == GL_NONE)
 721                     dstmod = GL_4X_BIT_ATI;
 722                 /* fall through */
 723             case WINED3D_TOP_MODULATE_2X:
 724                 if (dstmod == GL_NONE)
 725                     dstmod = GL_2X_BIT_ATI;
 726                 dstmod |= GL_SATURATE_BIT_ATI;
 727                 /* fall through */
 728             case WINED3D_TOP_MODULATE:
 729                 wrap_op2(gl_info, GL_MUL_ATI, dstreg, GL_ALPHA, dstmod,
 730                          arg1, GL_NONE, argmod1,
 731                          arg2, GL_NONE, argmod2);
 732                 break;
 733
 734             case WINED3D_TOP_ADD_SIGNED_2X:
 735                 dstmod = GL_2X_BIT_ATI;
 736                 /* fall through */
 737             case WINED3D_TOP_ADD_SIGNED:
 738                 argmodextra = GL_BIAS_BIT_ATI;
 739                 /* fall through */
 740             case WINED3D_TOP_ADD:
 741                 dstmod |= GL_SATURATE_BIT_ATI;
 742                 wrap_op2(gl_info, GL_ADD_ATI, dstreg, GL_ALPHA, dstmod,
 743                          arg1, GL_NONE, argmod1,
 744                          arg2, GL_NONE, argmodextra | argmod2);
 745                 break;
 746
 747             case WINED3D_TOP_SUBTRACT:
 748                 dstmod |= GL_SATURATE_BIT_ATI;
 749                 wrap_op2(gl_info, GL_SUB_ATI, dstreg, GL_ALPHA, dstmod,
 750                          arg1, GL_NONE, argmod1,
 751                          arg2, GL_NONE, argmod2);
 752                 break;
 753
 754             case WINED3D_TOP_ADD_SMOOTH:
 755                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 756                 /* Dst = arg1 + * arg2(1 -arg 1)
 757                  *     = arg2 * (1 - arg1) + arg1
 758                  */
 759                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_ALPHA, GL_SATURATE_BIT_ATI,
 760                          arg2, GL_NONE, argmod2,
 761                          arg1, GL_NONE, argmodextra,
 762                          arg1, GL_NONE, argmod1);
 763                 break;
 764
 765             case WINED3D_TOP_BLEND_CURRENT_ALPHA:
 766                 if (extrarg == GL_NONE)
 767                     extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, NULL, -1);
 768                 /* fall through */
 769             case WINED3D_TOP_BLEND_FACTOR_ALPHA:
 770                 if (extrarg == GL_NONE)
 771                     extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, NULL, -1);
 772                 /* fall through */
 773             case WINED3D_TOP_BLEND_TEXTURE_ALPHA:
 774                 if (extrarg == GL_NONE)
 775                     extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 776                 /* fall through */
 777             case WINED3D_TOP_BLEND_DIFFUSE_ALPHA:
 778                 if (extrarg == GL_NONE)
 779                     extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, NULL, -1);
 780                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_ALPHA, GL_NONE,
 781                          extrarg, GL_ALPHA, GL_NONE,
 782                          arg1, GL_NONE, argmod1,
 783                          arg2, GL_NONE, argmod2);
 784                 break;
 785
 786             case WINED3D_TOP_BLEND_TEXTURE_ALPHA_PM:
 787                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 788                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_ALPHA, GL_NONE,
 789                          arg2, GL_NONE,  argmod2,
 790                          arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 791                          arg1, GL_NONE,  argmod1);
 792                 break;
 793
 794             /* D3DTOP_PREMODULATE ???? */
 795
 796             case WINED3D_TOP_DOTPRODUCT3:
 797                 wrap_op2(gl_info, GL_DOT3_ATI, dstreg, GL_ALPHA, GL_4X_BIT_ATI | GL_SATURATE_BIT_ATI,
 798                          arg1, GL_NONE, argmod1 | GL_BIAS_BIT_ATI,
 799                          arg2, GL_NONE, argmod2 | GL_BIAS_BIT_ATI);
 800                 break;
 801
 802             case WINED3D_TOP_MULTIPLY_ADD:
 803                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_ALPHA, GL_SATURATE_BIT_ATI,
 804                          arg1, GL_NONE, argmod1,
 805                          arg2, GL_NONE, argmod2,
 806                          arg0, GL_NONE, argmod0);
 807                 break;
 808
 809             case WINED3D_TOP_LERP:
 810                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_ALPHA, GL_SATURATE_BIT_ATI,
 811                          arg1, GL_NONE, argmod1,
 812                          arg2, GL_NONE, argmod2,
 813                          arg0, GL_NONE, argmod0);
 814                 break;
 815
 816             case WINED3D_TOP_MODULATE_INVALPHA_ADD_COLOR:
 817             case WINED3D_TOP_MODULATE_ALPHA_ADD_COLOR:
 818             case WINED3D_TOP_MODULATE_COLOR_ADD_ALPHA:
 819             case WINED3D_TOP_MODULATE_INVCOLOR_ADD_ALPHA:
 820             case WINED3D_TOP_BUMPENVMAP:
 821             case WINED3D_TOP_BUMPENVMAP_LUMINANCE:
 822                 ERR("Application uses an invalid alpha operation\n");
 823                 break;
 824
 825             default: FIXME("Unhandled alpha operation %d on stage %d\n", op[stage].aop, stage);
 826         }
 827     }
 828
 829     TRACE("glEndFragmentShaderATI()\n");
 830     GL_EXTCALL(glEndFragmentShaderATI());
 831     checkGLcall("GL_EXTCALL(glEndFragmentShaderATI())");
 832     return ret;
 833 }
 834
 835 static void set_tex_op_atifs(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
 836 {
 837     const struct wined3d_device *device = context->swapchain->device;
 838     const struct wined3d_gl_info *gl_info = context->gl_info;
 839     const struct atifs_ffp_desc *desc;
 840     struct ffp_frag_settings settings;
 841     struct atifs_private_data *priv = device->fragment_priv;
 842     DWORD mapped_stage;
 843     unsigned int i;
 844
 845     gen_ffp_frag_op(device, state, &settings, TRUE);
 846     desc = (const struct atifs_ffp_desc *)find_ffp_frag_shader(&priv->fragment_shaders, &settings);
 847     if(!desc) {
 848         struct atifs_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
 849         if (!new_desc)
 850         {
 851             ERR("Out of memory\n");
 852             return;
 853         }
 854         new_desc->num_textures_used = 0;
 855         for (i = 0; i < gl_info->limits.texture_stages; ++i)
 856         {
 857             if (settings.op[i].cop == WINED3D_TOP_DISABLE)
 858                 break;
 859             new_desc->num_textures_used = i;
 860         }
 861
 862         memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
 863         new_desc->shader = gen_ati_shader(settings.op, gl_info);
 864         add_ffp_frag_shader(&priv->fragment_shaders, &new_desc->parent);
 865         TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
 866         desc = new_desc;
 867     }
 868
 869     /* GL_ATI_fragment_shader depends on the GL_TEXTURE_xD enable settings. Update the texture stages
 870      * used by this shader
 871      */
 872     for (i = 0; i < desc->num_textures_used; ++i)
 873     {
 874         mapped_stage = device->texUnitMap[i];
 875         if (mapped_stage != WINED3D_UNMAPPED_STAGE)
 876         {
 877             context_active_texture(context, gl_info, mapped_stage);
 878             texture_activate_dimensions(state->textures[i], gl_info);
 879         }
 880     }
 881
 882     GL_EXTCALL(glBindFragmentShaderATI(desc->shader));
 883 }
 884
 885 static void state_texfactor_atifs(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
 886 {
 887     const struct wined3d_gl_info *gl_info = context->gl_info;
 888     float col[4];
 889
 890     D3DCOLORTOGLFLOAT4(state->render_states[WINED3D_RS_TEXTUREFACTOR], col);
 891     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col));
 892     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col)");
 893 }
 894
 895 static void set_bumpmat(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
 896 {
 897     DWORD stage = (state_id - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
 898     const struct wined3d_gl_info *gl_info = context->gl_info;
 899     float mat[2][2];
 900
 901     mat[0][0] = *((float *)&state->texture_states[stage][WINED3D_TSS_BUMPENV_MAT00]);
 902     mat[1][0] = *((float *)&state->texture_states[stage][WINED3D_TSS_BUMPENV_MAT01]);
 903     mat[0][1] = *((float *)&state->texture_states[stage][WINED3D_TSS_BUMPENV_MAT10]);
 904     mat[1][1] = *((float *)&state->texture_states[stage][WINED3D_TSS_BUMPENV_MAT11]);
 905     /* GL_ATI_fragment_shader allows only constants from 0.0 to 1.0, but the bumpmat
 906      * constants can be in any range. While they should stay between [-1.0 and 1.0] because
 907      * Shader Model 1.x pixel shaders are clamped to that range negative values are used occasionally,
 908      * for example by our d3d9 test. So to get negative values scale -1;1 to 0;1 and undo that in the
 909      * shader(it is free). This might potentially reduce precision. However, if the hardware does
 910      * support proper floats it shouldn't, and if it doesn't we can't get anything better anyway
 911      */
 912     mat[0][0] = (mat[0][0] + 1.0f) * 0.5f;
 913     mat[1][0] = (mat[1][0] + 1.0f) * 0.5f;
 914     mat[0][1] = (mat[0][1] + 1.0f) * 0.5f;
 915     mat[1][1] = (mat[1][1] + 1.0f) * 0.5f;
 916     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), (float *) mat));
 917     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), mat)");
 918 }
 919
 920 static void textransform(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
 921 {
 922     if (!isStateDirty(context, STATE_PIXELSHADER))
 923         set_tex_op_atifs(context, state, state_id);
 924 }
 925
 926 static void atifs_apply_pixelshader(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
 927 {
 928     const struct wined3d_device *device = context->swapchain->device;
 929     BOOL use_vshader = use_vs(state);
 930
 931     context->last_was_pshader = use_ps(state);
 932     /* The ATIFS code does not support pixel shaders currently, but we have to
 933      * provide a state handler to call shader_select to select a vertex shader
 934      * if one is applied because the vertex shader state may defer calling the
 935      * shader backend if the pshader state is dirty.
 936      *
 937      * In theory the application should not be able to mark the pixel shader
 938      * dirty because it cannot create a shader, and thus has no way to set the
 939      * state to something != NULL. However, a different pipeline part may link
 940      * a different state to its pixelshader handler, thus a pshader state
 941      * exists and can be dirtified. Also the pshader is always dirtified at
 942      * startup, and blitting disables all shaders and dirtifies all shader
 943      * states. If atifs can deal with this it keeps the rest of the code
 944      * simpler. */
 945     if (!isStateDirty(context, context->state_table[STATE_VSHADER].representative))
 946     {
 947         device->shader_backend->shader_select(context, FALSE, use_vshader);
 948
 949         if (!isStateDirty(context, STATE_VERTEXSHADERCONSTANT) && use_vshader)
 950             context_apply_state(context, state, STATE_VERTEXSHADERCONSTANT);
 951     }
 952 }
 953
 954 static void atifs_srgbwriteenable(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
 955 {
 956     if (state->render_states[WINED3D_RS_SRGBWRITEENABLE])
 957         WARN("sRGB writes are not supported by this fragment pipe.\n");
 958 }
 959
 960 static const struct StateEntryTemplate atifs_fragmentstate_template[] = {
 961     {STATE_RENDER(WINED3D_RS_TEXTUREFACTOR),              { STATE_RENDER(WINED3D_RS_TEXTUREFACTOR),             state_texfactor_atifs   }, WINED3D_GL_EXT_NONE             },
 962     {STATE_RENDER(WINED3D_RS_FOGCOLOR),                   { STATE_RENDER(WINED3D_RS_FOGCOLOR),                  state_fogcolor          }, WINED3D_GL_EXT_NONE             },
 963     {STATE_RENDER(WINED3D_RS_FOGDENSITY),                 { STATE_RENDER(WINED3D_RS_FOGDENSITY),                state_fogdensity        }, WINED3D_GL_EXT_NONE             },
 964     {STATE_RENDER(WINED3D_RS_FOGENABLE),                  { STATE_RENDER(WINED3D_RS_FOGENABLE),                 state_fog_fragpart      }, WINED3D_GL_EXT_NONE             },
 965     {STATE_RENDER(WINED3D_RS_FOGTABLEMODE),               { STATE_RENDER(WINED3D_RS_FOGENABLE),                 NULL                    }, WINED3D_GL_EXT_NONE             },
 966     {STATE_RENDER(WINED3D_RS_FOGVERTEXMODE),              { STATE_RENDER(WINED3D_RS_FOGENABLE),                 NULL                    }, WINED3D_GL_EXT_NONE             },
 967     {STATE_RENDER(WINED3D_RS_FOGSTART),                   { STATE_RENDER(WINED3D_RS_FOGSTART),                  state_fogstartend       }, WINED3D_GL_EXT_NONE             },
 968     {STATE_RENDER(WINED3D_RS_FOGEND),                     { STATE_RENDER(WINED3D_RS_FOGSTART),                  NULL                    }, WINED3D_GL_EXT_NONE             },
 969     {STATE_RENDER(WINED3D_RS_SRGBWRITEENABLE),            { STATE_RENDER(WINED3D_RS_SRGBWRITEENABLE),           atifs_srgbwriteenable   }, WINED3D_GL_EXT_NONE             },
 970     {STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        set_tex_op_atifs        }, WINED3D_GL_EXT_NONE             },
 971     {STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 972     {STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 973     {STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 974     {STATE_TEXTURESTAGE(0, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 975     {STATE_TEXTURESTAGE(0, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 976     {STATE_TEXTURESTAGE(0, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 977     {STATE_TEXTURESTAGE(0, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 978     {STATE_TEXTURESTAGE(0, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 979     {STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 980     {STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
 981     {STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
 982     {STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
 983     {STATE_TEXTURESTAGE(1, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 984     {STATE_TEXTURESTAGE(1, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 985     {STATE_TEXTURESTAGE(1, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 986     {STATE_TEXTURESTAGE(1, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 987     {STATE_TEXTURESTAGE(1, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 988     {STATE_TEXTURESTAGE(1, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 989     {STATE_TEXTURESTAGE(1, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 990     {STATE_TEXTURESTAGE(1, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 991     {STATE_TEXTURESTAGE(1, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 992     {STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 993     {STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
 994     {STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
 995     {STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
 996     {STATE_TEXTURESTAGE(2, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 997     {STATE_TEXTURESTAGE(2, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 998     {STATE_TEXTURESTAGE(2, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 999     {STATE_TEXTURESTAGE(2, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1000     {STATE_TEXTURESTAGE(2, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1001     {STATE_TEXTURESTAGE(2, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1002     {STATE_TEXTURESTAGE(2, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1003     {STATE_TEXTURESTAGE(2, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1004     {STATE_TEXTURESTAGE(2, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1005     {STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1006     {STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1007     {STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1008     {STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1009     {STATE_TEXTURESTAGE(3, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1010     {STATE_TEXTURESTAGE(3, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1011     {STATE_TEXTURESTAGE(3, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1012     {STATE_TEXTURESTAGE(3, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1013     {STATE_TEXTURESTAGE(3, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1014     {STATE_TEXTURESTAGE(3, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1015     {STATE_TEXTURESTAGE(3, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1016     {STATE_TEXTURESTAGE(3, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1017     {STATE_TEXTURESTAGE(3, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1018     {STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1019     {STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1020     {STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1021     {STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1022     {STATE_TEXTURESTAGE(4, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1023     {STATE_TEXTURESTAGE(4, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1024     {STATE_TEXTURESTAGE(4, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1025     {STATE_TEXTURESTAGE(4, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1026     {STATE_TEXTURESTAGE(4, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1027     {STATE_TEXTURESTAGE(4, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1028     {STATE_TEXTURESTAGE(4, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1029     {STATE_TEXTURESTAGE(4, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1030     {STATE_TEXTURESTAGE(4, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1031     {STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1032     {STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1033     {STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1034     {STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1035     {STATE_TEXTURESTAGE(5, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1036     {STATE_TEXTURESTAGE(5, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1037     {STATE_TEXTURESTAGE(5, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1038     {STATE_TEXTURESTAGE(5, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1039     {STATE_TEXTURESTAGE(5, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1040     {STATE_TEXTURESTAGE(5, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1041     {STATE_TEXTURESTAGE(5, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1042     {STATE_TEXTURESTAGE(5, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1043     {STATE_TEXTURESTAGE(5, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1044     {STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1045     {STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1046     {STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1047     {STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1048     {STATE_TEXTURESTAGE(6, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1049     {STATE_TEXTURESTAGE(6, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1050     {STATE_TEXTURESTAGE(6, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1051     {STATE_TEXTURESTAGE(6, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1052     {STATE_TEXTURESTAGE(6, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1053     {STATE_TEXTURESTAGE(6, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1054     {STATE_TEXTURESTAGE(6, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1055     {STATE_TEXTURESTAGE(6, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1056     {STATE_TEXTURESTAGE(6, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1057     {STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1058     {STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1059     {STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1060     {STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1061     {STATE_TEXTURESTAGE(7, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1062     {STATE_TEXTURESTAGE(7, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1063     {STATE_TEXTURESTAGE(7, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1064     {STATE_TEXTURESTAGE(7, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1065     {STATE_TEXTURESTAGE(7, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1066     {STATE_TEXTURESTAGE(7, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1067     {STATE_TEXTURESTAGE(7, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1068     {STATE_TEXTURESTAGE(7, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1069     {STATE_TEXTURESTAGE(7, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1070     {STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1071     {STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1072     {STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1073     {STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1074     { STATE_SAMPLER(0),                                   { STATE_SAMPLER(0),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1075     { STATE_SAMPLER(1),                                   { STATE_SAMPLER(1),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1076     { STATE_SAMPLER(2),                                   { STATE_SAMPLER(2),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1077     { STATE_SAMPLER(3),                                   { STATE_SAMPLER(3),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1078     { STATE_SAMPLER(4),                                   { STATE_SAMPLER(4),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1079     { STATE_SAMPLER(5),                                   { STATE_SAMPLER(5),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1080     { STATE_SAMPLER(6),                                   { STATE_SAMPLER(6),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1081     { STATE_SAMPLER(7),                                   { STATE_SAMPLER(7),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1082     {STATE_TEXTURESTAGE(0,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(0, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1083     {STATE_TEXTURESTAGE(1,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(1, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1084     {STATE_TEXTURESTAGE(2,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(2, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1085     {STATE_TEXTURESTAGE(3,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(3, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1086     {STATE_TEXTURESTAGE(4,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(4, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1087     {STATE_TEXTURESTAGE(5,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(5, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1088     {STATE_TEXTURESTAGE(6,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(6, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1089     {STATE_TEXTURESTAGE(7,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(7, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1090     {STATE_PIXELSHADER,                                   { STATE_PIXELSHADER,                                  atifs_apply_pixelshader }, WINED3D_GL_EXT_NONE             },
1091     {0 /* Terminate */,                                   { 0,                                                  0                       }, WINED3D_GL_EXT_NONE             },
1092 };
1093
1094 /* Context activation and GL locking are done by the caller. */
1095 static void atifs_enable(BOOL enable)
1096 {
1097     if(enable) {
1098         glEnable(GL_FRAGMENT_SHADER_ATI);
1099         checkGLcall("glEnable(GL_FRAGMENT_SHADER_ATI)");
1100     } else {
1101         glDisable(GL_FRAGMENT_SHADER_ATI);
1102         checkGLcall("glDisable(GL_FRAGMENT_SHADER_ATI)");
1103     }
1104 }
1105
1106 static void atifs_get_caps(const struct wined3d_gl_info *gl_info, struct fragment_caps *caps)
1107 {
1108     caps->PrimitiveMiscCaps = WINED3DPMISCCAPS_TSSARGTEMP;
1109     caps->TextureOpCaps =  WINED3DTEXOPCAPS_DISABLE                     |
1110                            WINED3DTEXOPCAPS_SELECTARG1                  |
1111                            WINED3DTEXOPCAPS_SELECTARG2                  |
1112                            WINED3DTEXOPCAPS_MODULATE4X                  |
1113                            WINED3DTEXOPCAPS_MODULATE2X                  |
1114                            WINED3DTEXOPCAPS_MODULATE                    |
1115                            WINED3DTEXOPCAPS_ADDSIGNED2X                 |
1116                            WINED3DTEXOPCAPS_ADDSIGNED                   |
1117                            WINED3DTEXOPCAPS_ADD                         |
1118                            WINED3DTEXOPCAPS_SUBTRACT                    |
1119                            WINED3DTEXOPCAPS_ADDSMOOTH                   |
1120                            WINED3DTEXOPCAPS_BLENDCURRENTALPHA           |
1121                            WINED3DTEXOPCAPS_BLENDFACTORALPHA            |
1122                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHA           |
1123                            WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA           |
1124                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM         |
1125                            WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR      |
1126                            WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA      |
1127                            WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA   |
1128                            WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR   |
1129                            WINED3DTEXOPCAPS_DOTPRODUCT3                 |
1130                            WINED3DTEXOPCAPS_MULTIPLYADD                 |
1131                            WINED3DTEXOPCAPS_LERP                        |
1132                            WINED3DTEXOPCAPS_BUMPENVMAP;
1133
1134     /* TODO: Implement WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE
1135     and WINED3DTEXOPCAPS_PREMODULATE */
1136
1137     /* GL_ATI_fragment_shader always supports 6 textures, which was the limit on r200 cards
1138      * which this extension is exclusively focused on(later cards have GL_ARB_fragment_program).
1139      * If the current card has more than 8 fixed function textures in OpenGL's regular fixed
1140      * function pipeline then the ATI_fragment_shader backend imposes a stricter limit. This
1141      * shouldn't be too hard since Nvidia cards have a limit of 4 textures with the default ffp
1142      * pipeline, and almost all games are happy with that. We can however support up to 8
1143      * texture stages because we have a 2nd pass limit of 8 instructions, and per stage we use
1144      * only 1 instruction.
1145      *
1146      * The proper fix for this is not to use GL_ATI_fragment_shader on cards newer than the
1147      * r200 series and use an ARB or GLSL shader instead
1148      */
1149     caps->MaxTextureBlendStages   = 8;
1150     caps->MaxSimultaneousTextures = 6;
1151 }
1152
1153 static HRESULT atifs_alloc(struct wined3d_device *device)
1154 {
1155     struct atifs_private_data *priv;
1156
1157     device->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct atifs_private_data));
1158     if (!device->fragment_priv)
1159     {
1160         ERR("Out of memory\n");
1161         return E_OUTOFMEMORY;
1162     }
1163     priv = device->fragment_priv;
1164     if (wine_rb_init(&priv->fragment_shaders, &wined3d_ffp_frag_program_rb_functions) == -1)
1165     {
1166         ERR("Failed to initialize rbtree.\n");
1167         HeapFree(GetProcessHeap(), 0, device->fragment_priv);
1168         return E_OUTOFMEMORY;
1169     }
1170     return WINED3D_OK;
1171 }
1172
1173 /* Context activation is done by the caller. */
1174 static void atifs_free_ffpshader(struct wine_rb_entry *entry, void *context)
1175 {
1176     struct wined3d_device *device = context;
1177     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1178     struct atifs_ffp_desc *entry_ati = WINE_RB_ENTRY_VALUE(entry, struct atifs_ffp_desc, parent.entry);
1179
1180     ENTER_GL();
1181     GL_EXTCALL(glDeleteFragmentShaderATI(entry_ati->shader));
1182     checkGLcall("glDeleteFragmentShaderATI(entry->shader)");
1183     HeapFree(GetProcessHeap(), 0, entry_ati);
1184     LEAVE_GL();
1185 }
1186
1187 /* Context activation is done by the caller. */
1188 static void atifs_free(struct wined3d_device *device)
1189 {
1190     struct atifs_private_data *priv = device->fragment_priv;
1191
1192     wine_rb_destroy(&priv->fragment_shaders, atifs_free_ffpshader, device);
1193
1194     HeapFree(GetProcessHeap(), 0, priv);
1195     device->fragment_priv = NULL;
1196 }
1197
1198 static BOOL atifs_color_fixup_supported(struct color_fixup_desc fixup)
1199 {
1200     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1201     {
1202         TRACE("Checking support for fixup:\n");
1203         dump_color_fixup_desc(fixup);
1204     }
1205
1206     /* We only support sign fixup of the first two channels. */
1207     if (fixup.x_source == CHANNEL_SOURCE_X && fixup.y_source == CHANNEL_SOURCE_Y
1208             && fixup.z_source == CHANNEL_SOURCE_Z && fixup.w_source == CHANNEL_SOURCE_W
1209             && !fixup.z_sign_fixup && !fixup.w_sign_fixup)
1210     {
1211         TRACE("[OK]\n");
1212         return TRUE;
1213     }
1214
1215     TRACE("[FAILED]\n");
1216     return FALSE;
1217 }
1218
1219 const struct fragment_pipeline atifs_fragment_pipeline = {
1220     atifs_enable,
1221     atifs_get_caps,
1222     atifs_alloc,
1223     atifs_free,
1224     atifs_color_fixup_supported,
1225     atifs_fragmentstate_template,
1226     TRUE /* We can disable projected textures */
1227 };