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

   1 /*
   2  * Fixed function pipeline replacement using GL_ATI_fragment_shader
   3  *
   4  * Copyright 2008 Stefan Dösinger(for CodeWeavers)
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  19  */
  20
  21 #include "config.h"
  22 #include "wine/port.h"
  23
  24 #include <math.h>
  25 #include <stdio.h>
  26
  27 #include "wined3d_private.h"
  28
  29 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  30 WINE_DECLARE_DEBUG_CHANNEL(d3d);
  31
  32 /* Context activation for state handlers is done by the caller. */
  33
  34 /* Some private defines, Constant associations, etc.
  35  * Env bump matrix and per stage constant should be independent,
  36  * a stage that bump maps can't read the per state constant
  37  */
  38 #define ATI_FFP_CONST_BUMPMAT(i) (GL_CON_0_ATI + i)
  39 #define ATI_FFP_CONST_CONSTANT0 GL_CON_0_ATI
  40 #define ATI_FFP_CONST_CONSTANT1 GL_CON_1_ATI
  41 #define ATI_FFP_CONST_CONSTANT2 GL_CON_2_ATI
  42 #define ATI_FFP_CONST_CONSTANT3 GL_CON_3_ATI
  43 #define ATI_FFP_CONST_CONSTANT4 GL_CON_4_ATI
  44 #define ATI_FFP_CONST_CONSTANT5 GL_CON_5_ATI
  45 #define ATI_FFP_CONST_TFACTOR   GL_CON_6_ATI
  46
  47 /* GL_ATI_fragment_shader specific fixed function pipeline description. "Inherits" from the common one */
  48 struct atifs_ffp_desc
  49 {
  50     struct ffp_frag_desc parent;
  51     GLuint shader;
  52     unsigned int num_textures_used;
  53 };
  54
  55 struct atifs_private_data
  56 {
  57     struct wine_rb_tree fragment_shaders; /* A rb-tree to track fragment pipeline replacement shaders */
  58 };
  59
  60 static const char *debug_dstmod(GLuint mod) {
  61     switch(mod) {
  62         case GL_NONE:               return "GL_NONE";
  63         case GL_2X_BIT_ATI:         return "GL_2X_BIT_ATI";
  64         case GL_4X_BIT_ATI:         return "GL_4X_BIT_ATI";
  65         case GL_8X_BIT_ATI:         return "GL_8X_BIT_ATI";
  66         case GL_HALF_BIT_ATI:       return "GL_HALF_BIT_ATI";
  67         case GL_QUARTER_BIT_ATI:    return "GL_QUARTER_BIT_ATI";
  68         case GL_EIGHTH_BIT_ATI:     return "GL_EIGHTH_BIT_ATI";
  69         case GL_SATURATE_BIT_ATI:   return "GL_SATURATE_BIT_ATI";
  70         default:                    return "Unexpected modifier\n";
  71     }
  72 }
  73
  74 static const char *debug_argmod(GLuint mod) {
  75     switch(mod) {
  76         case GL_NONE:
  77             return "GL_NONE";
  78
  79         case GL_2X_BIT_ATI:
  80             return "GL_2X_BIT_ATI";
  81         case GL_COMP_BIT_ATI:
  82             return "GL_COMP_BIT_ATI";
  83         case GL_NEGATE_BIT_ATI:
  84             return "GL_NEGATE_BIT_ATI";
  85         case GL_BIAS_BIT_ATI:
  86             return "GL_BIAS_BIT_ATI";
  87
  88         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI:
  89             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI";
  90         case GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI:
  91             return "GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI";
  92         case GL_2X_BIT_ATI | GL_BIAS_BIT_ATI:
  93             return "GL_2X_BIT_ATI | GL_BIAS_BIT_ATI";
  94         case GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI:
  95             return "GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI";
  96         case GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI:
  97             return "GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI";
  98         case GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
  99             return "GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 100
 101         case GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 102             return "GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 103         case GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 104             return "GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 105         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI:
 106             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI";
 107         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI:
 108             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI";
 109
 110         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 111             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 112
 113         default:
 114             return "Unexpected argmod combination\n";
 115     }
 116 }
 117 static const char *debug_register(GLuint reg) {
 118     switch(reg) {
 119         case GL_REG_0_ATI:                  return "GL_REG_0_ATI";
 120         case GL_REG_1_ATI:                  return "GL_REG_1_ATI";
 121         case GL_REG_2_ATI:                  return "GL_REG_2_ATI";
 122         case GL_REG_3_ATI:                  return "GL_REG_3_ATI";
 123         case GL_REG_4_ATI:                  return "GL_REG_4_ATI";
 124         case GL_REG_5_ATI:                  return "GL_REG_5_ATI";
 125
 126         case GL_CON_0_ATI:                  return "GL_CON_0_ATI";
 127         case GL_CON_1_ATI:                  return "GL_CON_1_ATI";
 128         case GL_CON_2_ATI:                  return "GL_CON_2_ATI";
 129         case GL_CON_3_ATI:                  return "GL_CON_3_ATI";
 130         case GL_CON_4_ATI:                  return "GL_CON_4_ATI";
 131         case GL_CON_5_ATI:                  return "GL_CON_5_ATI";
 132         case GL_CON_6_ATI:                  return "GL_CON_6_ATI";
 133         case GL_CON_7_ATI:                  return "GL_CON_7_ATI";
 134
 135         case GL_ZERO:                       return "GL_ZERO";
 136         case GL_ONE:                        return "GL_ONE";
 137         case GL_PRIMARY_COLOR:              return "GL_PRIMARY_COLOR";
 138         case GL_SECONDARY_INTERPOLATOR_ATI: return "GL_SECONDARY_INTERPOLATOR_ATI";
 139
 140         default:                            return "Unknown register\n";
 141     }
 142 }
 143
 144 static const char *debug_swizzle(GLuint swizzle) {
 145     switch(swizzle) {
 146         case GL_SWIZZLE_STR_ATI:        return "GL_SWIZZLE_STR_ATI";
 147         case GL_SWIZZLE_STQ_ATI:        return "GL_SWIZZLE_STQ_ATI";
 148         case GL_SWIZZLE_STR_DR_ATI:     return "GL_SWIZZLE_STR_DR_ATI";
 149         case GL_SWIZZLE_STQ_DQ_ATI:     return "GL_SWIZZLE_STQ_DQ_ATI";
 150         default:                        return "unknown swizzle";
 151     }
 152 }
 153
 154 static const char *debug_rep(GLuint rep) {
 155     switch(rep) {
 156         case GL_NONE:                   return "GL_NONE";
 157         case GL_RED:                    return "GL_RED";
 158         case GL_GREEN:                  return "GL_GREEN";
 159         case GL_BLUE:                   return "GL_BLUE";
 160         case GL_ALPHA:                  return "GL_ALPHA";
 161         default:                        return "unknown argrep";
 162     }
 163 }
 164
 165 static const char *debug_op(GLuint op) {
 166     switch(op) {
 167         case GL_MOV_ATI:                return "GL_MOV_ATI";
 168         case GL_ADD_ATI:                return "GL_ADD_ATI";
 169         case GL_MUL_ATI:                return "GL_MUL_ATI";
 170         case GL_SUB_ATI:                return "GL_SUB_ATI";
 171         case GL_DOT3_ATI:               return "GL_DOT3_ATI";
 172         case GL_DOT4_ATI:               return "GL_DOT4_ATI";
 173         case GL_MAD_ATI:                return "GL_MAD_ATI";
 174         case GL_LERP_ATI:               return "GL_LERP_ATI";
 175         case GL_CND_ATI:                return "GL_CND_ATI";
 176         case GL_CND0_ATI:               return "GL_CND0_ATI";
 177         case GL_DOT2_ADD_ATI:           return "GL_DOT2_ADD_ATI";
 178         default:                        return "unexpected op";
 179     }
 180 }
 181
 182 static const char *debug_mask(GLuint mask) {
 183     switch(mask) {
 184         case GL_NONE:                           return "GL_NONE";
 185         case GL_RED_BIT_ATI:                    return "GL_RED_BIT_ATI";
 186         case GL_GREEN_BIT_ATI:                  return "GL_GREEN_BIT_ATI";
 187         case GL_BLUE_BIT_ATI:                   return "GL_BLUE_BIT_ATI";
 188         case GL_RED_BIT_ATI | GL_GREEN_BIT_ATI: return "GL_RED_BIT_ATI | GL_GREEN_BIT_ATI";
 189         case GL_RED_BIT_ATI | GL_BLUE_BIT_ATI:  return "GL_RED_BIT_ATI | GL_BLUE_BIT_ATI";
 190         case GL_GREEN_BIT_ATI | GL_BLUE_BIT_ATI:return "GL_GREEN_BIT_ATI | GL_BLUE_BIT_ATI";
 191         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";
 192         default:                                return "Unexpected writemask";
 193     }
 194 }
 195
 196 static void wrap_op1(const struct wined3d_gl_info *gl_info, GLuint op, GLuint dst, GLuint dstMask, GLuint dstMod,
 197         GLuint arg1, GLuint arg1Rep, GLuint arg1Mod)
 198 {
 199     if(dstMask == GL_ALPHA) {
 200         TRACE("glAlphaFragmentOp1ATI(%s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst), debug_dstmod(dstMod),
 201               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod));
 202         GL_EXTCALL(glAlphaFragmentOp1ATI(op, dst, dstMod, arg1, arg1Rep, arg1Mod));
 203     } else {
 204         TRACE("glColorFragmentOp1ATI(%s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst),
 205               debug_mask(dstMask), debug_dstmod(dstMod),
 206               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod));
 207         GL_EXTCALL(glColorFragmentOp1ATI(op, dst, dstMask, dstMod, arg1, arg1Rep, arg1Mod));
 208     }
 209 }
 210
 211 static void wrap_op2(const struct wined3d_gl_info *gl_info, GLuint op, GLuint dst, GLuint dstMask, GLuint dstMod,
 212         GLuint arg1, GLuint arg1Rep, GLuint arg1Mod, GLuint arg2, GLuint arg2Rep, GLuint arg2Mod)
 213 {
 214     if(dstMask == GL_ALPHA) {
 215         TRACE("glAlphaFragmentOp2ATI(%s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst), debug_dstmod(dstMod),
 216               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 217               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod));
 218         GL_EXTCALL(glAlphaFragmentOp2ATI(op, dst, dstMod, arg1, arg1Rep, arg1Mod, arg2, arg2Rep, arg2Mod));
 219     } else {
 220         TRACE("glColorFragmentOp2ATI(%s, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst),
 221               debug_mask(dstMask), debug_dstmod(dstMod),
 222               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 223               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod));
 224         GL_EXTCALL(glColorFragmentOp2ATI(op, dst, dstMask, dstMod, arg1, arg1Rep, arg1Mod, arg2, arg2Rep, arg2Mod));
 225     }
 226 }
 227
 228 static void wrap_op3(const struct wined3d_gl_info *gl_info, GLuint op, GLuint dst, GLuint dstMask, GLuint dstMod,
 229         GLuint arg1, GLuint arg1Rep, GLuint arg1Mod, GLuint arg2, GLuint arg2Rep, GLuint arg2Mod,
 230         GLuint arg3, GLuint arg3Rep, GLuint arg3Mod)
 231 {
 232     if(dstMask == GL_ALPHA) {
 233         /* Leave some free space to fit "GL_NONE, " in to align most alpha and color op lines */
 234         TRACE("glAlphaFragmentOp3ATI(%s, %s,          %s, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst), debug_dstmod(dstMod),
 235               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 236               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod),
 237               debug_register(arg3), debug_rep(arg3Rep), debug_argmod(arg3Mod));
 238         GL_EXTCALL(glAlphaFragmentOp3ATI(op, dst, dstMod,
 239                                          arg1, arg1Rep, arg1Mod,
 240                                          arg2, arg2Rep, arg2Mod,
 241                                          arg3, arg3Rep, arg3Mod));
 242     } else {
 243         TRACE("glColorFragmentOp3ATI(%s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst),
 244               debug_mask(dstMask), debug_dstmod(dstMod),
 245               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 246               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod),
 247               debug_register(arg3), debug_rep(arg3Rep), debug_argmod(arg3Mod));
 248         GL_EXTCALL(glColorFragmentOp3ATI(op, dst, dstMask, dstMod,
 249                                          arg1, arg1Rep, arg1Mod,
 250                                          arg2, arg2Rep, arg2Mod,
 251                                          arg3, arg3Rep, arg3Mod));
 252     }
 253 }
 254
 255 static GLuint register_for_arg(DWORD arg, const struct wined3d_gl_info *gl_info,
 256         unsigned int stage, GLuint *mod, GLuint *rep, GLuint tmparg)
 257 {
 258     GLenum ret;
 259
 260     if(mod) *mod = GL_NONE;
 261     if(arg == ARG_UNUSED)
 262     {
 263         if (rep) *rep = GL_NONE;
 264         return -1; /* This is the marker for unused registers */
 265     }
 266
 267     switch(arg & WINED3DTA_SELECTMASK) {
 268         case WINED3DTA_DIFFUSE:
 269             ret = GL_PRIMARY_COLOR;
 270             break;
 271
 272         case WINED3DTA_CURRENT:
 273             /* Note that using GL_REG_0_ATI for the passed on register is safe because
 274              * texture0 is read at stage0, so in the worst case it is read in the
 275              * instruction writing to reg0. Afterwards texture0 is not used any longer.
 276              * If we're reading from current
 277              */
 278             ret = stage ? GL_REG_0_ATI : GL_PRIMARY_COLOR;
 279             break;
 280
 281         case WINED3DTA_TEXTURE:
 282             ret = GL_REG_0_ATI + stage;
 283             break;
 284
 285         case WINED3DTA_TFACTOR:
 286             ret = ATI_FFP_CONST_TFACTOR;
 287             break;
 288
 289         case WINED3DTA_SPECULAR:
 290             ret = GL_SECONDARY_INTERPOLATOR_ATI;
 291             break;
 292
 293         case WINED3DTA_TEMP:
 294             ret = tmparg;
 295             break;
 296
 297         case WINED3DTA_CONSTANT:
 298             FIXME("Unhandled source argument WINED3DTA_TEMP\n");
 299             ret = GL_CON_0_ATI;
 300             break;
 301
 302         default:
 303             FIXME("Unknown source argument %d\n", arg);
 304             ret = GL_ZERO;
 305     }
 306
 307     if(arg & WINED3DTA_COMPLEMENT) {
 308         if(mod) *mod |= GL_COMP_BIT_ATI;
 309     }
 310     if(arg & WINED3DTA_ALPHAREPLICATE) {
 311         if(rep) *rep = GL_ALPHA;
 312     } else {
 313         if(rep) *rep = GL_NONE;
 314     }
 315     return ret;
 316 }
 317
 318 static GLuint find_tmpreg(const struct texture_stage_op op[MAX_TEXTURES])
 319 {
 320     int lowest_read = -1;
 321     int lowest_write = -1;
 322     int i;
 323     BOOL tex_used[MAX_TEXTURES];
 324
 325     memset(tex_used, 0, sizeof(tex_used));
 326     for (i = 0; i < MAX_TEXTURES; ++i)
 327     {
 328         if (op[i].cop == WINED3D_TOP_DISABLE)
 329             break;
 330
 331         if(lowest_read == -1 &&
 332           (op[i].carg1 == WINED3DTA_TEMP || op[i].carg2 == WINED3DTA_TEMP || op[i].carg0 == WINED3DTA_TEMP ||
 333            op[i].aarg1 == WINED3DTA_TEMP || op[i].aarg2 == WINED3DTA_TEMP || op[i].aarg0 == WINED3DTA_TEMP)) {
 334             lowest_read = i;
 335         }
 336
 337         if(lowest_write == -1 && op[i].dst == tempreg) {
 338             lowest_write = i;
 339         }
 340
 341         if(op[i].carg1 == WINED3DTA_TEXTURE || op[i].carg2 == WINED3DTA_TEXTURE || op[i].carg0 == WINED3DTA_TEXTURE ||
 342            op[i].aarg1 == WINED3DTA_TEXTURE || op[i].aarg2 == WINED3DTA_TEXTURE || op[i].aarg0 == WINED3DTA_TEXTURE) {
 343             tex_used[i] = TRUE;
 344         }
 345     }
 346
 347     /* Temp reg not read? We don't need it, return GL_NONE */
 348     if(lowest_read == -1) return GL_NONE;
 349
 350     if(lowest_write >= lowest_read) {
 351         FIXME("Temp register read before being written\n");
 352     }
 353
 354     if(lowest_write == -1) {
 355         /* This needs a test. Maybe we are supposed to return 0.0/0.0/0.0/0.0, or fail drawprim, or whatever */
 356         FIXME("Temp register read without being written\n");
 357         return GL_REG_1_ATI;
 358     } else if(lowest_write >= 1) {
 359         /* If we're writing to the temp reg at earliest in stage 1, we can use register 1 for the temp result.
 360          * there may be texture data stored in reg 1, but we do not need it any longer since stage 1 already
 361          * read it
 362          */
 363         return GL_REG_1_ATI;
 364     } else {
 365         /* Search for a free texture register. We have 6 registers available. GL_REG_0_ATI is already used
 366          * for the regular result
 367          */
 368         for(i = 1; i < 6; i++) {
 369             if(!tex_used[i]) {
 370                 return GL_REG_0_ATI + i;
 371             }
 372         }
 373         /* What to do here? Report it in ValidateDevice? */
 374         FIXME("Could not find a register for the temporary register\n");
 375         return 0;
 376     }
 377 }
 378
 379 static GLuint gen_ati_shader(const struct texture_stage_op op[MAX_TEXTURES], const struct wined3d_gl_info *gl_info)
 380 {
 381     GLuint ret = GL_EXTCALL(glGenFragmentShadersATI(1));
 382     unsigned int stage;
 383     GLuint arg0, arg1, arg2, extrarg;
 384     GLuint dstmod, argmod0, argmod1, argmod2, argmodextra;
 385     GLuint rep0, rep1, rep2;
 386     GLuint swizzle;
 387     GLuint tmparg = find_tmpreg(op);
 388     GLuint dstreg;
 389
 390     if(!ret) {
 391         ERR("Failed to generate a GL_ATI_fragment_shader shader id\n");
 392         return 0;
 393     }
 394     GL_EXTCALL(glBindFragmentShaderATI(ret));
 395     checkGLcall("GL_EXTCALL(glBindFragmentShaderATI(ret))");
 396
 397     TRACE("glBeginFragmentShaderATI()\n");
 398     GL_EXTCALL(glBeginFragmentShaderATI());
 399     checkGLcall("GL_EXTCALL(glBeginFragmentShaderATI())");
 400
 401     /* Pass 1: Generate sampling instructions for perturbation maps */
 402     for (stage = 0; stage < gl_info->limits.textures; ++stage)
 403     {
 404         if (op[stage].cop == WINED3D_TOP_DISABLE)
 405             break;
 406         if (op[stage].cop != WINED3D_TOP_BUMPENVMAP
 407                 && op[stage].cop != WINED3D_TOP_BUMPENVMAP_LUMINANCE)
 408             continue;
 409
 410         TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, GL_SWIZZLE_STR_ATI)\n",
 411               stage, stage);
 412         GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 413                    GL_TEXTURE0_ARB + stage,
 414                    GL_SWIZZLE_STR_ATI));
 415         if(op[stage + 1].projected == proj_none) {
 416             swizzle = GL_SWIZZLE_STR_ATI;
 417         } else if(op[stage + 1].projected == proj_count4) {
 418             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 419         } else {
 420             swizzle = GL_SWIZZLE_STR_DR_ATI;
 421         }
 422         TRACE("glPassTexCoordATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, %s)\n",
 423               stage + 1, stage + 1, debug_swizzle(swizzle));
 424         GL_EXTCALL(glPassTexCoordATI(GL_REG_0_ATI + stage + 1,
 425                    GL_TEXTURE0_ARB + stage + 1,
 426                    swizzle));
 427     }
 428
 429     /* Pass 2: Generate perturbation calculations */
 430     for (stage = 0; stage < gl_info->limits.textures; ++stage)
 431     {
 432         GLuint argmodextra_x, argmodextra_y;
 433         struct color_fixup_desc fixup;
 434
 435         if (op[stage].cop == WINED3D_TOP_DISABLE)
 436             break;
 437         if (op[stage].cop != WINED3D_TOP_BUMPENVMAP
 438                 && op[stage].cop != WINED3D_TOP_BUMPENVMAP_LUMINANCE)
 439             continue;
 440
 441         fixup = op[stage].color_fixup;
 442         if (fixup.x_source != CHANNEL_SOURCE_X || fixup.y_source != CHANNEL_SOURCE_Y)
 443         {
 444             FIXME("Swizzles not implemented\n");
 445             argmodextra_x = GL_NONE;
 446             argmodextra_y = GL_NONE;
 447         }
 448         else
 449         {
 450             /* Nice thing, we get the color correction for free :-) */
 451             argmodextra_x = fixup.x_sign_fixup ? GL_2X_BIT_ATI | GL_BIAS_BIT_ATI : GL_NONE;
 452             argmodextra_y = fixup.y_sign_fixup ? GL_2X_BIT_ATI | GL_BIAS_BIT_ATI : GL_NONE;
 453         }
 454
 455         wrap_op3(gl_info, GL_DOT2_ADD_ATI, GL_REG_0_ATI + stage + 1, GL_RED_BIT_ATI, GL_NONE,
 456                  GL_REG_0_ATI + stage, GL_NONE, argmodextra_x,
 457                  ATI_FFP_CONST_BUMPMAT(stage), GL_NONE, GL_2X_BIT_ATI | GL_BIAS_BIT_ATI,
 458                  GL_REG_0_ATI + stage + 1, GL_RED, GL_NONE);
 459
 460         /* Don't use GL_DOT2_ADD_ATI here because we cannot configure it to read the blue and alpha
 461          * component of the bump matrix. Instead do this with two MADs:
 462          *
 463          * coord.a = tex.r * bump.b + coord.g
 464          * coord.g = tex.g * bump.a + coord.a
 465          *
 466          * The first instruction writes to alpha so it can be coissued with the above DOT2_ADD.
 467          * coord.a is unused. If the perturbed texture is projected, this was already handled
 468          * in the glPassTexCoordATI above.
 469          */
 470         wrap_op3(gl_info, GL_MAD_ATI, GL_REG_0_ATI + stage + 1, GL_ALPHA, GL_NONE,
 471                  GL_REG_0_ATI + stage, GL_RED, argmodextra_y,
 472                  ATI_FFP_CONST_BUMPMAT(stage), GL_BLUE, GL_2X_BIT_ATI | GL_BIAS_BIT_ATI,
 473                  GL_REG_0_ATI + stage + 1, GL_GREEN, GL_NONE);
 474         wrap_op3(gl_info, GL_MAD_ATI, GL_REG_0_ATI + stage + 1, GL_GREEN_BIT_ATI, GL_NONE,
 475                  GL_REG_0_ATI + stage, GL_GREEN, argmodextra_y,
 476                  ATI_FFP_CONST_BUMPMAT(stage), GL_ALPHA, GL_2X_BIT_ATI | GL_BIAS_BIT_ATI,
 477                  GL_REG_0_ATI + stage + 1, GL_ALPHA, GL_NONE);
 478     }
 479
 480     /* Pass 3: Generate sampling instructions for regular textures */
 481     for (stage = 0; stage < gl_info->limits.textures; ++stage)
 482     {
 483         if (op[stage].cop == WINED3D_TOP_DISABLE)
 484             break;
 485
 486         if(op[stage].projected == proj_none) {
 487             swizzle = GL_SWIZZLE_STR_ATI;
 488         } else if(op[stage].projected == proj_count3) {
 489             swizzle = GL_SWIZZLE_STR_DR_ATI;
 490         } else {
 491             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 492         }
 493
 494         if ((op[stage].carg0 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE
 495                 || (op[stage].carg1 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE
 496                 || (op[stage].carg2 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE
 497                 || (op[stage].aarg0 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE
 498                 || (op[stage].aarg1 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE
 499                 || (op[stage].aarg2 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE
 500                 || op[stage].cop == WINED3D_TOP_BLEND_TEXTURE_ALPHA)
 501         {
 502             if (stage > 0
 503                     && (op[stage - 1].cop == WINED3D_TOP_BUMPENVMAP
 504                     || op[stage - 1].cop == WINED3D_TOP_BUMPENVMAP_LUMINANCE))
 505             {
 506                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_REG_%d_ATI, GL_SWIZZLE_STR_ATI)\n",
 507                       stage, stage);
 508                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 509                            GL_REG_0_ATI + stage,
 510                            GL_SWIZZLE_STR_ATI));
 511             } else {
 512                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, %s)\n",
 513                     stage, stage, debug_swizzle(swizzle));
 514                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 515                                         GL_TEXTURE0_ARB + stage,
 516                                         swizzle));
 517             }
 518         }
 519     }
 520
 521     /* Pass 4: Generate the arithmetic instructions */
 522     for (stage = 0; stage < MAX_TEXTURES; ++stage)
 523     {
 524         if (op[stage].cop == WINED3D_TOP_DISABLE)
 525         {
 526             if (!stage)
 527             {
 528                 /* Handle complete texture disabling gracefully */
 529                 wrap_op1(gl_info, GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 530                          GL_PRIMARY_COLOR, GL_NONE, GL_NONE);
 531                 wrap_op1(gl_info, GL_MOV_ATI, GL_REG_0_ATI, GL_ALPHA, GL_NONE,
 532                          GL_PRIMARY_COLOR, GL_NONE, GL_NONE);
 533             }
 534             break;
 535         }
 536
 537         if(op[stage].dst == tempreg) {
 538             /* If we're writing to D3DTA_TEMP, but never reading from it we don't have to write there in the first place.
 539              * skip the entire stage, this saves some GPU time
 540              */
 541             if(tmparg == GL_NONE) continue;
 542
 543             dstreg = tmparg;
 544         } else {
 545             dstreg = GL_REG_0_ATI;
 546         }
 547
 548         arg0 = register_for_arg(op[stage].carg0, gl_info, stage, &argmod0, &rep0, tmparg);
 549         arg1 = register_for_arg(op[stage].carg1, gl_info, stage, &argmod1, &rep1, tmparg);
 550         arg2 = register_for_arg(op[stage].carg2, gl_info, stage, &argmod2, &rep2, tmparg);
 551         dstmod = GL_NONE;
 552         argmodextra = GL_NONE;
 553         extrarg = GL_NONE;
 554
 555         switch (op[stage].cop)
 556         {
 557             case WINED3D_TOP_SELECT_ARG2:
 558                 arg1 = arg2;
 559                 argmod1 = argmod2;
 560                 rep1 = rep2;
 561                 /* fall through */
 562             case WINED3D_TOP_SELECT_ARG1:
 563                 wrap_op1(gl_info, GL_MOV_ATI, dstreg, GL_NONE, GL_NONE,
 564                          arg1, rep1, argmod1);
 565                 break;
 566
 567             case WINED3D_TOP_MODULATE_4X:
 568                 if(dstmod == GL_NONE) dstmod = GL_4X_BIT_ATI;
 569                 /* fall through */
 570             case WINED3D_TOP_MODULATE_2X:
 571                 if(dstmod == GL_NONE) dstmod = GL_2X_BIT_ATI;
 572                 dstmod |= GL_SATURATE_BIT_ATI;
 573                 /* fall through */
 574             case WINED3D_TOP_MODULATE:
 575                 wrap_op2(gl_info, GL_MUL_ATI, dstreg, GL_NONE, dstmod,
 576                          arg1, rep1, argmod1,
 577                          arg2, rep2, argmod2);
 578                 break;
 579
 580             case WINED3D_TOP_ADD_SIGNED_2X:
 581                 dstmod = GL_2X_BIT_ATI;
 582                 /* fall through */
 583             case WINED3D_TOP_ADD_SIGNED:
 584                 argmodextra = GL_BIAS_BIT_ATI;
 585                 /* fall through */
 586             case WINED3D_TOP_ADD:
 587                 dstmod |= GL_SATURATE_BIT_ATI;
 588                 wrap_op2(gl_info, GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, dstmod,
 589                          arg1, rep1, argmod1,
 590                          arg2, rep2, argmodextra | argmod2);
 591                 break;
 592
 593             case WINED3D_TOP_SUBTRACT:
 594                 dstmod |= GL_SATURATE_BIT_ATI;
 595                 wrap_op2(gl_info, GL_SUB_ATI, dstreg, GL_NONE, dstmod,
 596                          arg1, rep1, argmod1,
 597                          arg2, rep2, argmod2);
 598                 break;
 599
 600             case WINED3D_TOP_ADD_SMOOTH:
 601                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 602                 /* Dst = arg1 + * arg2(1 -arg 1)
 603                  *     = arg2 * (1 - arg1) + arg1
 604                  */
 605                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 606                          arg2, rep2, argmod2,
 607                          arg1, rep1, argmodextra,
 608                          arg1, rep1, argmod1);
 609                 break;
 610
 611             case WINED3D_TOP_BLEND_CURRENT_ALPHA:
 612                 if (extrarg == GL_NONE)
 613                     extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, NULL, -1);
 614                 /* fall through */
 615             case WINED3D_TOP_BLEND_FACTOR_ALPHA:
 616                 if (extrarg == GL_NONE)
 617                     extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, NULL, -1);
 618                 /* fall through */
 619             case WINED3D_TOP_BLEND_TEXTURE_ALPHA:
 620                 if (extrarg == GL_NONE)
 621                     extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 622                 /* fall through */
 623             case WINED3D_TOP_BLEND_DIFFUSE_ALPHA:
 624                 if (extrarg == GL_NONE)
 625                     extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, NULL, -1);
 626                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 627                          extrarg, GL_ALPHA, GL_NONE,
 628                          arg1, rep1, argmod1,
 629                          arg2, rep2, argmod2);
 630                 break;
 631
 632             case WINED3D_TOP_BLEND_TEXTURE_ALPHA_PM:
 633                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 634                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 635                          arg2, rep2,  argmod2,
 636                          arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 637                          arg1, rep1,  argmod1);
 638                 break;
 639
 640             /* D3DTOP_PREMODULATE ???? */
 641
 642             case WINED3D_TOP_MODULATE_INVALPHA_ADD_COLOR:
 643                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 644                 /* fall through */
 645             case WINED3D_TOP_MODULATE_ALPHA_ADD_COLOR:
 646                 if (!argmodextra)
 647                     argmodextra = argmod1;
 648                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 649                          arg2, rep2,  argmod2,
 650                          arg1, GL_ALPHA, argmodextra,
 651                          arg1, rep1,  argmod1);
 652                 break;
 653
 654             case WINED3D_TOP_MODULATE_INVCOLOR_ADD_ALPHA:
 655                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 656                 /* fall through */
 657             case WINED3D_TOP_MODULATE_COLOR_ADD_ALPHA:
 658                 if (!argmodextra)
 659                     argmodextra = argmod1;
 660                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 661                          arg2, rep2,  argmod2,
 662                          arg1, rep1,  argmodextra,
 663                          arg1, GL_ALPHA, argmod1);
 664                 break;
 665
 666             case WINED3D_TOP_DOTPRODUCT3:
 667                 wrap_op2(gl_info, GL_DOT3_ATI, dstreg, GL_NONE, GL_4X_BIT_ATI | GL_SATURATE_BIT_ATI,
 668                          arg1, rep1, argmod1 | GL_BIAS_BIT_ATI,
 669                          arg2, rep2, argmod2 | GL_BIAS_BIT_ATI);
 670                 break;
 671
 672             case WINED3D_TOP_MULTIPLY_ADD:
 673                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 674                          arg1, rep1, argmod1,
 675                          arg2, rep2, argmod2,
 676                          arg0, rep0, argmod0);
 677                 break;
 678
 679             case WINED3D_TOP_LERP:
 680                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 681                          arg0, rep0, argmod0,
 682                          arg1, rep1, argmod1,
 683                          arg2, rep2, argmod2);
 684                 break;
 685
 686             case WINED3D_TOP_BUMPENVMAP:
 687             case WINED3D_TOP_BUMPENVMAP_LUMINANCE:
 688                 /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
 689                 break;
 690
 691             default: FIXME("Unhandled color operation %d on stage %d\n", op[stage].cop, stage);
 692         }
 693
 694         arg0 = register_for_arg(op[stage].aarg0, gl_info, stage, &argmod0, NULL, tmparg);
 695         arg1 = register_for_arg(op[stage].aarg1, gl_info, stage, &argmod1, NULL, tmparg);
 696         arg2 = register_for_arg(op[stage].aarg2, gl_info, stage, &argmod2, NULL, tmparg);
 697         dstmod = GL_NONE;
 698         argmodextra = GL_NONE;
 699         extrarg = GL_NONE;
 700
 701         switch (op[stage].aop)
 702         {
 703             case WINED3D_TOP_DISABLE:
 704                 /* Get the primary color to the output if on stage 0, otherwise leave register 0 untouched */
 705                 if (!stage)
 706                 {
 707                     wrap_op1(gl_info, GL_MOV_ATI, GL_REG_0_ATI, GL_ALPHA, GL_NONE,
 708                              GL_PRIMARY_COLOR, GL_NONE, GL_NONE);
 709                 }
 710                 break;
 711
 712             case WINED3D_TOP_SELECT_ARG2:
 713                 arg1 = arg2;
 714                 argmod1 = argmod2;
 715                 /* fall through */
 716             case WINED3D_TOP_SELECT_ARG1:
 717                 wrap_op1(gl_info, GL_MOV_ATI, dstreg, GL_ALPHA, GL_NONE,
 718                          arg1, GL_NONE, argmod1);
 719                 break;
 720
 721             case WINED3D_TOP_MODULATE_4X:
 722                 if (dstmod == GL_NONE)
 723                     dstmod = GL_4X_BIT_ATI;
 724                 /* fall through */
 725             case WINED3D_TOP_MODULATE_2X:
 726                 if (dstmod == GL_NONE)
 727                     dstmod = GL_2X_BIT_ATI;
 728                 dstmod |= GL_SATURATE_BIT_ATI;
 729                 /* fall through */
 730             case WINED3D_TOP_MODULATE:
 731                 wrap_op2(gl_info, GL_MUL_ATI, dstreg, GL_ALPHA, dstmod,
 732                          arg1, GL_NONE, argmod1,
 733                          arg2, GL_NONE, argmod2);
 734                 break;
 735
 736             case WINED3D_TOP_ADD_SIGNED_2X:
 737                 dstmod = GL_2X_BIT_ATI;
 738                 /* fall through */
 739             case WINED3D_TOP_ADD_SIGNED:
 740                 argmodextra = GL_BIAS_BIT_ATI;
 741                 /* fall through */
 742             case WINED3D_TOP_ADD:
 743                 dstmod |= GL_SATURATE_BIT_ATI;
 744                 wrap_op2(gl_info, GL_ADD_ATI, dstreg, GL_ALPHA, dstmod,
 745                          arg1, GL_NONE, argmod1,
 746                          arg2, GL_NONE, argmodextra | argmod2);
 747                 break;
 748
 749             case WINED3D_TOP_SUBTRACT:
 750                 dstmod |= GL_SATURATE_BIT_ATI;
 751                 wrap_op2(gl_info, GL_SUB_ATI, dstreg, GL_ALPHA, dstmod,
 752                          arg1, GL_NONE, argmod1,
 753                          arg2, GL_NONE, argmod2);
 754                 break;
 755
 756             case WINED3D_TOP_ADD_SMOOTH:
 757                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 758                 /* Dst = arg1 + * arg2(1 -arg 1)
 759                  *     = arg2 * (1 - arg1) + arg1
 760                  */
 761                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_ALPHA, GL_SATURATE_BIT_ATI,
 762                          arg2, GL_NONE, argmod2,
 763                          arg1, GL_NONE, argmodextra,
 764                          arg1, GL_NONE, argmod1);
 765                 break;
 766
 767             case WINED3D_TOP_BLEND_CURRENT_ALPHA:
 768                 if (extrarg == GL_NONE)
 769                     extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, NULL, -1);
 770                 /* fall through */
 771             case WINED3D_TOP_BLEND_FACTOR_ALPHA:
 772                 if (extrarg == GL_NONE)
 773                     extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, NULL, -1);
 774                 /* fall through */
 775             case WINED3D_TOP_BLEND_TEXTURE_ALPHA:
 776                 if (extrarg == GL_NONE)
 777                     extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 778                 /* fall through */
 779             case WINED3D_TOP_BLEND_DIFFUSE_ALPHA:
 780                 if (extrarg == GL_NONE)
 781                     extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, NULL, -1);
 782                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_ALPHA, GL_NONE,
 783                          extrarg, GL_ALPHA, GL_NONE,
 784                          arg1, GL_NONE, argmod1,
 785                          arg2, GL_NONE, argmod2);
 786                 break;
 787
 788             case WINED3D_TOP_BLEND_TEXTURE_ALPHA_PM:
 789                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 790                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_ALPHA, GL_NONE,
 791                          arg2, GL_NONE,  argmod2,
 792                          arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 793                          arg1, GL_NONE,  argmod1);
 794                 break;
 795
 796             /* D3DTOP_PREMODULATE ???? */
 797
 798             case WINED3D_TOP_DOTPRODUCT3:
 799                 wrap_op2(gl_info, GL_DOT3_ATI, dstreg, GL_ALPHA, GL_4X_BIT_ATI | GL_SATURATE_BIT_ATI,
 800                          arg1, GL_NONE, argmod1 | GL_BIAS_BIT_ATI,
 801                          arg2, GL_NONE, argmod2 | GL_BIAS_BIT_ATI);
 802                 break;
 803
 804             case WINED3D_TOP_MULTIPLY_ADD:
 805                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_ALPHA, GL_SATURATE_BIT_ATI,
 806                          arg1, GL_NONE, argmod1,
 807                          arg2, GL_NONE, argmod2,
 808                          arg0, GL_NONE, argmod0);
 809                 break;
 810
 811             case WINED3D_TOP_LERP:
 812                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_ALPHA, GL_SATURATE_BIT_ATI,
 813                          arg1, GL_NONE, argmod1,
 814                          arg2, GL_NONE, argmod2,
 815                          arg0, GL_NONE, argmod0);
 816                 break;
 817
 818             case WINED3D_TOP_MODULATE_INVALPHA_ADD_COLOR:
 819             case WINED3D_TOP_MODULATE_ALPHA_ADD_COLOR:
 820             case WINED3D_TOP_MODULATE_COLOR_ADD_ALPHA:
 821             case WINED3D_TOP_MODULATE_INVCOLOR_ADD_ALPHA:
 822             case WINED3D_TOP_BUMPENVMAP:
 823             case WINED3D_TOP_BUMPENVMAP_LUMINANCE:
 824                 ERR("Application uses an invalid alpha operation\n");
 825                 break;
 826
 827             default: FIXME("Unhandled alpha operation %d on stage %d\n", op[stage].aop, stage);
 828         }
 829     }
 830
 831     TRACE("glEndFragmentShaderATI()\n");
 832     GL_EXTCALL(glEndFragmentShaderATI());
 833     checkGLcall("GL_EXTCALL(glEndFragmentShaderATI())");
 834     return ret;
 835 }
 836
 837 static void set_tex_op_atifs(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
 838 {
 839     const struct wined3d_device *device = context->swapchain->device;
 840     const struct wined3d_gl_info *gl_info = context->gl_info;
 841     const struct atifs_ffp_desc *desc;
 842     struct ffp_frag_settings settings;
 843     struct atifs_private_data *priv = device->fragment_priv;
 844     DWORD mapped_stage;
 845     unsigned int i;
 846
 847     gen_ffp_frag_op(device, state, &settings, TRUE);
 848     desc = (const struct atifs_ffp_desc *)find_ffp_frag_shader(&priv->fragment_shaders, &settings);
 849     if(!desc) {
 850         struct atifs_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
 851         if (!new_desc)
 852         {
 853             ERR("Out of memory\n");
 854             return;
 855         }
 856         new_desc->num_textures_used = 0;
 857         for (i = 0; i < gl_info->limits.texture_stages; ++i)
 858         {
 859             if (settings.op[i].cop == WINED3D_TOP_DISABLE)
 860                 break;
 861             new_desc->num_textures_used = i + 1;
 862         }
 863
 864         memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
 865         new_desc->shader = gen_ati_shader(settings.op, gl_info);
 866         add_ffp_frag_shader(&priv->fragment_shaders, &new_desc->parent);
 867         TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
 868         desc = new_desc;
 869     }
 870
 871     /* GL_ATI_fragment_shader depends on the GL_TEXTURE_xD enable settings. Update the texture stages
 872      * used by this shader
 873      */
 874     for (i = 0; i < desc->num_textures_used; ++i)
 875     {
 876         mapped_stage = device->texUnitMap[i];
 877         if (mapped_stage != WINED3D_UNMAPPED_STAGE)
 878         {
 879             context_active_texture(context, gl_info, mapped_stage);
 880             texture_activate_dimensions(state->textures[i], gl_info);
 881         }
 882     }
 883
 884     GL_EXTCALL(glBindFragmentShaderATI(desc->shader));
 885 }
 886
 887 static void state_texfactor_atifs(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
 888 {
 889     const struct wined3d_gl_info *gl_info = context->gl_info;
 890     float col[4];
 891
 892     D3DCOLORTOGLFLOAT4(state->render_states[WINED3D_RS_TEXTUREFACTOR], col);
 893     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col));
 894     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col)");
 895 }
 896
 897 static void set_bumpmat(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
 898 {
 899     DWORD stage = (state_id - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
 900     const struct wined3d_gl_info *gl_info = context->gl_info;
 901     float mat[2][2];
 902
 903     mat[0][0] = *((float *)&state->texture_states[stage][WINED3D_TSS_BUMPENV_MAT00]);
 904     mat[1][0] = *((float *)&state->texture_states[stage][WINED3D_TSS_BUMPENV_MAT01]);
 905     mat[0][1] = *((float *)&state->texture_states[stage][WINED3D_TSS_BUMPENV_MAT10]);
 906     mat[1][1] = *((float *)&state->texture_states[stage][WINED3D_TSS_BUMPENV_MAT11]);
 907     /* GL_ATI_fragment_shader allows only constants from 0.0 to 1.0, but the bumpmat
 908      * constants can be in any range. While they should stay between [-1.0 and 1.0] because
 909      * Shader Model 1.x pixel shaders are clamped to that range negative values are used occasionally,
 910      * for example by our d3d9 test. So to get negative values scale -1;1 to 0;1 and undo that in the
 911      * shader(it is free). This might potentially reduce precision. However, if the hardware does
 912      * support proper floats it shouldn't, and if it doesn't we can't get anything better anyway
 913      */
 914     mat[0][0] = (mat[0][0] + 1.0f) * 0.5f;
 915     mat[1][0] = (mat[1][0] + 1.0f) * 0.5f;
 916     mat[0][1] = (mat[0][1] + 1.0f) * 0.5f;
 917     mat[1][1] = (mat[1][1] + 1.0f) * 0.5f;
 918     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), (float *) mat));
 919     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), mat)");
 920 }
 921
 922 static void textransform(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
 923 {
 924     if (!isStateDirty(context, STATE_PIXELSHADER))
 925         set_tex_op_atifs(context, state, state_id);
 926 }
 927
 928 static void atifs_apply_pixelshader(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
 929 {
 930     context->last_was_pshader = use_ps(state);
 931     /* The ATIFS code does not support pixel shaders currently, but we have to
 932      * provide a state handler to call shader_select to select a vertex shader
 933      * if one is applied because the vertex shader state may defer calling the
 934      * shader backend if the pshader state is dirty.
 935      *
 936      * In theory the application should not be able to mark the pixel shader
 937      * dirty because it cannot create a shader, and thus has no way to set the
 938      * state to something != NULL. However, a different pipeline part may link
 939      * a different state to its pixelshader handler, thus a pshader state
 940      * exists and can be dirtified. Also the pshader is always dirtified at
 941      * startup, and blitting disables all shaders and dirtifies all shader
 942      * states. If atifs can deal with this it keeps the rest of the code
 943      * simpler. */
 944     context->select_shader = 1;
 945     context->load_constants = 1;
 946 }
 947
 948 static void atifs_srgbwriteenable(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
 949 {
 950     if (state->render_states[WINED3D_RS_SRGBWRITEENABLE])
 951         WARN("sRGB writes are not supported by this fragment pipe.\n");
 952 }
 953
 954 static const struct StateEntryTemplate atifs_fragmentstate_template[] = {
 955     {STATE_RENDER(WINED3D_RS_TEXTUREFACTOR),              { STATE_RENDER(WINED3D_RS_TEXTUREFACTOR),             state_texfactor_atifs   }, WINED3D_GL_EXT_NONE             },
 956     {STATE_RENDER(WINED3D_RS_FOGCOLOR),                   { STATE_RENDER(WINED3D_RS_FOGCOLOR),                  state_fogcolor          }, WINED3D_GL_EXT_NONE             },
 957     {STATE_RENDER(WINED3D_RS_FOGDENSITY),                 { STATE_RENDER(WINED3D_RS_FOGDENSITY),                state_fogdensity        }, WINED3D_GL_EXT_NONE             },
 958     {STATE_RENDER(WINED3D_RS_FOGENABLE),                  { STATE_RENDER(WINED3D_RS_FOGENABLE),                 state_fog_fragpart      }, WINED3D_GL_EXT_NONE             },
 959     {STATE_RENDER(WINED3D_RS_FOGTABLEMODE),               { STATE_RENDER(WINED3D_RS_FOGENABLE),                 NULL                    }, WINED3D_GL_EXT_NONE             },
 960     {STATE_RENDER(WINED3D_RS_FOGVERTEXMODE),              { STATE_RENDER(WINED3D_RS_FOGENABLE),                 NULL                    }, WINED3D_GL_EXT_NONE             },
 961     {STATE_RENDER(WINED3D_RS_FOGSTART),                   { STATE_RENDER(WINED3D_RS_FOGSTART),                  state_fogstartend       }, WINED3D_GL_EXT_NONE             },
 962     {STATE_RENDER(WINED3D_RS_FOGEND),                     { STATE_RENDER(WINED3D_RS_FOGSTART),                  NULL                    }, WINED3D_GL_EXT_NONE             },
 963     {STATE_RENDER(WINED3D_RS_SRGBWRITEENABLE),            { STATE_RENDER(WINED3D_RS_SRGBWRITEENABLE),           atifs_srgbwriteenable   }, WINED3D_GL_EXT_NONE             },
 964     {STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        set_tex_op_atifs        }, WINED3D_GL_EXT_NONE             },
 965     {STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 966     {STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 967     {STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 968     {STATE_TEXTURESTAGE(0, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 969     {STATE_TEXTURESTAGE(0, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 970     {STATE_TEXTURESTAGE(0, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 971     {STATE_TEXTURESTAGE(0, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 972     {STATE_TEXTURESTAGE(0, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 973     {STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 974     {STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
 975     {STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
 976     {STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(0, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
 977     {STATE_TEXTURESTAGE(1, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 978     {STATE_TEXTURESTAGE(1, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 979     {STATE_TEXTURESTAGE(1, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 980     {STATE_TEXTURESTAGE(1, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 981     {STATE_TEXTURESTAGE(1, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 982     {STATE_TEXTURESTAGE(1, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 983     {STATE_TEXTURESTAGE(1, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 984     {STATE_TEXTURESTAGE(1, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 985     {STATE_TEXTURESTAGE(1, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 986     {STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 987     {STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
 988     {STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
 989     {STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(1, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
 990     {STATE_TEXTURESTAGE(2, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 991     {STATE_TEXTURESTAGE(2, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 992     {STATE_TEXTURESTAGE(2, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 993     {STATE_TEXTURESTAGE(2, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 994     {STATE_TEXTURESTAGE(2, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 995     {STATE_TEXTURESTAGE(2, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 996     {STATE_TEXTURESTAGE(2, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 997     {STATE_TEXTURESTAGE(2, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 998     {STATE_TEXTURESTAGE(2, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
 999     {STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1000     {STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1001     {STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1002     {STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(2, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1003     {STATE_TEXTURESTAGE(3, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1004     {STATE_TEXTURESTAGE(3, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1005     {STATE_TEXTURESTAGE(3, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1006     {STATE_TEXTURESTAGE(3, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1007     {STATE_TEXTURESTAGE(3, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1008     {STATE_TEXTURESTAGE(3, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1009     {STATE_TEXTURESTAGE(3, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1010     {STATE_TEXTURESTAGE(3, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1011     {STATE_TEXTURESTAGE(3, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1012     {STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1013     {STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1014     {STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1015     {STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(3, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1016     {STATE_TEXTURESTAGE(4, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1017     {STATE_TEXTURESTAGE(4, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1018     {STATE_TEXTURESTAGE(4, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1019     {STATE_TEXTURESTAGE(4, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1020     {STATE_TEXTURESTAGE(4, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1021     {STATE_TEXTURESTAGE(4, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1022     {STATE_TEXTURESTAGE(4, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1023     {STATE_TEXTURESTAGE(4, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1024     {STATE_TEXTURESTAGE(4, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1025     {STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1026     {STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1027     {STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1028     {STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(4, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1029     {STATE_TEXTURESTAGE(5, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1030     {STATE_TEXTURESTAGE(5, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1031     {STATE_TEXTURESTAGE(5, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1032     {STATE_TEXTURESTAGE(5, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1033     {STATE_TEXTURESTAGE(5, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1034     {STATE_TEXTURESTAGE(5, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1035     {STATE_TEXTURESTAGE(5, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1036     {STATE_TEXTURESTAGE(5, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1037     {STATE_TEXTURESTAGE(5, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1038     {STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1039     {STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1040     {STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1041     {STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(5, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1042     {STATE_TEXTURESTAGE(6, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1043     {STATE_TEXTURESTAGE(6, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1044     {STATE_TEXTURESTAGE(6, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1045     {STATE_TEXTURESTAGE(6, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1046     {STATE_TEXTURESTAGE(6, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1047     {STATE_TEXTURESTAGE(6, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1048     {STATE_TEXTURESTAGE(6, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1049     {STATE_TEXTURESTAGE(6, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1050     {STATE_TEXTURESTAGE(6, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1051     {STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1052     {STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1053     {STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1054     {STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(6, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1055     {STATE_TEXTURESTAGE(7, WINED3D_TSS_COLOR_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1056     {STATE_TEXTURESTAGE(7, WINED3D_TSS_COLOR_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1057     {STATE_TEXTURESTAGE(7, WINED3D_TSS_COLOR_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1058     {STATE_TEXTURESTAGE(7, WINED3D_TSS_COLOR_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1059     {STATE_TEXTURESTAGE(7, WINED3D_TSS_ALPHA_OP),         { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1060     {STATE_TEXTURESTAGE(7, WINED3D_TSS_ALPHA_ARG1),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1061     {STATE_TEXTURESTAGE(7, WINED3D_TSS_ALPHA_ARG2),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1062     {STATE_TEXTURESTAGE(7, WINED3D_TSS_ALPHA_ARG0),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1063     {STATE_TEXTURESTAGE(7, WINED3D_TSS_RESULT_ARG),       { STATE_TEXTURESTAGE(0, WINED3D_TSS_COLOR_OP),        NULL                    }, WINED3D_GL_EXT_NONE             },
1064     {STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT00),    { STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT00),   set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1065     {STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT01),    { STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1066     {STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT10),    { STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1067     {STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT11),    { STATE_TEXTURESTAGE(7, WINED3D_TSS_BUMPENV_MAT00),   NULL                    }, WINED3D_GL_EXT_NONE             },
1068     { STATE_SAMPLER(0),                                   { STATE_SAMPLER(0),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1069     { STATE_SAMPLER(1),                                   { STATE_SAMPLER(1),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1070     { STATE_SAMPLER(2),                                   { STATE_SAMPLER(2),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1071     { STATE_SAMPLER(3),                                   { STATE_SAMPLER(3),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1072     { STATE_SAMPLER(4),                                   { STATE_SAMPLER(4),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1073     { STATE_SAMPLER(5),                                   { STATE_SAMPLER(5),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1074     { STATE_SAMPLER(6),                                   { STATE_SAMPLER(6),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1075     { STATE_SAMPLER(7),                                   { STATE_SAMPLER(7),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1076     {STATE_TEXTURESTAGE(0,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(0, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1077     {STATE_TEXTURESTAGE(1,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(1, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1078     {STATE_TEXTURESTAGE(2,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(2, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1079     {STATE_TEXTURESTAGE(3,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(3, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1080     {STATE_TEXTURESTAGE(4,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(4, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1081     {STATE_TEXTURESTAGE(5,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(5, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1082     {STATE_TEXTURESTAGE(6,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(6, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1083     {STATE_TEXTURESTAGE(7,WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), {STATE_TEXTURESTAGE(7, WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1084     {STATE_PIXELSHADER,                                   { STATE_PIXELSHADER,                                  atifs_apply_pixelshader }, WINED3D_GL_EXT_NONE             },
1085     {0 /* Terminate */,                                   { 0,                                                  0                       }, WINED3D_GL_EXT_NONE             },
1086 };
1087
1088 /* Context activation is done by the caller. */
1089 static void atifs_enable(const struct wined3d_gl_info *gl_info, BOOL enable)
1090 {
1091     if (enable)
1092     {
1093         gl_info->gl_ops.gl.p_glEnable(GL_FRAGMENT_SHADER_ATI);
1094         checkGLcall("glEnable(GL_FRAGMENT_SHADER_ATI)");
1095     }
1096     else
1097     {
1098         gl_info->gl_ops.gl.p_glDisable(GL_FRAGMENT_SHADER_ATI);
1099         checkGLcall("glDisable(GL_FRAGMENT_SHADER_ATI)");
1100     }
1101 }
1102
1103 static void atifs_get_caps(const struct wined3d_gl_info *gl_info, struct fragment_caps *caps)
1104 {
1105     caps->PrimitiveMiscCaps = WINED3DPMISCCAPS_TSSARGTEMP;
1106     caps->TextureOpCaps =  WINED3DTEXOPCAPS_DISABLE                     |
1107                            WINED3DTEXOPCAPS_SELECTARG1                  |
1108                            WINED3DTEXOPCAPS_SELECTARG2                  |
1109                            WINED3DTEXOPCAPS_MODULATE4X                  |
1110                            WINED3DTEXOPCAPS_MODULATE2X                  |
1111                            WINED3DTEXOPCAPS_MODULATE                    |
1112                            WINED3DTEXOPCAPS_ADDSIGNED2X                 |
1113                            WINED3DTEXOPCAPS_ADDSIGNED                   |
1114                            WINED3DTEXOPCAPS_ADD                         |
1115                            WINED3DTEXOPCAPS_SUBTRACT                    |
1116                            WINED3DTEXOPCAPS_ADDSMOOTH                   |
1117                            WINED3DTEXOPCAPS_BLENDCURRENTALPHA           |
1118                            WINED3DTEXOPCAPS_BLENDFACTORALPHA            |
1119                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHA           |
1120                            WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA           |
1121                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM         |
1122                            WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR      |
1123                            WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA      |
1124                            WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA   |
1125                            WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR   |
1126                            WINED3DTEXOPCAPS_DOTPRODUCT3                 |
1127                            WINED3DTEXOPCAPS_MULTIPLYADD                 |
1128                            WINED3DTEXOPCAPS_LERP                        |
1129                            WINED3DTEXOPCAPS_BUMPENVMAP;
1130
1131     /* TODO: Implement WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE
1132     and WINED3DTEXOPCAPS_PREMODULATE */
1133
1134     /* GL_ATI_fragment_shader always supports 6 textures, which was the limit on r200 cards
1135      * which this extension is exclusively focused on(later cards have GL_ARB_fragment_program).
1136      * If the current card has more than 8 fixed function textures in OpenGL's regular fixed
1137      * function pipeline then the ATI_fragment_shader backend imposes a stricter limit. This
1138      * shouldn't be too hard since Nvidia cards have a limit of 4 textures with the default ffp
1139      * pipeline, and almost all games are happy with that. We can however support up to 8
1140      * texture stages because we have a 2nd pass limit of 8 instructions, and per stage we use
1141      * only 1 instruction.
1142      *
1143      * The proper fix for this is not to use GL_ATI_fragment_shader on cards newer than the
1144      * r200 series and use an ARB or GLSL shader instead
1145      */
1146     caps->MaxTextureBlendStages   = 8;
1147     caps->MaxSimultaneousTextures = 6;
1148 }
1149
1150 static void *atifs_alloc(const struct wined3d_shader_backend_ops *shader_backend, void *shader_priv)
1151 {
1152     struct atifs_private_data *priv;
1153
1154     if (!(priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*priv))))
1155     {
1156         ERR("Out of memory.\n");
1157         return NULL;
1158     }
1159     if (wine_rb_init(&priv->fragment_shaders, &wined3d_ffp_frag_program_rb_functions) == -1)
1160     {
1161         ERR("Failed to initialize rbtree.\n");
1162         HeapFree(GetProcessHeap(), 0, priv);
1163         return NULL;
1164     }
1165
1166     return priv;
1167 }
1168
1169 /* Context activation is done by the caller. */
1170 static void atifs_free_ffpshader(struct wine_rb_entry *entry, void *context)
1171 {
1172     struct wined3d_device *device = context;
1173     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1174     struct atifs_ffp_desc *entry_ati = WINE_RB_ENTRY_VALUE(entry, struct atifs_ffp_desc, parent.entry);
1175
1176     GL_EXTCALL(glDeleteFragmentShaderATI(entry_ati->shader));
1177     checkGLcall("glDeleteFragmentShaderATI(entry->shader)");
1178     HeapFree(GetProcessHeap(), 0, entry_ati);
1179 }
1180
1181 /* Context activation is done by the caller. */
1182 static void atifs_free(struct wined3d_device *device)
1183 {
1184     struct atifs_private_data *priv = device->fragment_priv;
1185
1186     wine_rb_destroy(&priv->fragment_shaders, atifs_free_ffpshader, device);
1187
1188     HeapFree(GetProcessHeap(), 0, priv);
1189     device->fragment_priv = NULL;
1190 }
1191
1192 static BOOL atifs_color_fixup_supported(struct color_fixup_desc fixup)
1193 {
1194     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1195     {
1196         TRACE("Checking support for fixup:\n");
1197         dump_color_fixup_desc(fixup);
1198     }
1199
1200     /* We only support sign fixup of the first two channels. */
1201     if (fixup.x_source == CHANNEL_SOURCE_X && fixup.y_source == CHANNEL_SOURCE_Y
1202             && fixup.z_source == CHANNEL_SOURCE_Z && fixup.w_source == CHANNEL_SOURCE_W
1203             && !fixup.z_sign_fixup && !fixup.w_sign_fixup)
1204     {
1205         TRACE("[OK]\n");
1206         return TRUE;
1207     }
1208
1209     TRACE("[FAILED]\n");
1210     return FALSE;
1211 }
1212
1213 const struct fragment_pipeline atifs_fragment_pipeline = {
1214     atifs_enable,
1215     atifs_get_caps,
1216     atifs_alloc,
1217     atifs_free,
1218     atifs_color_fixup_supported,
1219     atifs_fragmentstate_template,
1220     TRUE /* We can disable projected textures */
1221 };