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

   1 /*
   2  * Fixed function pipeline replacement using GL_ATI_fragment_shader
   3  *
   4  * Copyright 2008 Stefan Dösinger(for CodeWeavers)
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  19  */
  20
  21 #include "config.h"
  22
  23 #include <math.h>
  24 #include <stdio.h>
  25
  26 #include "wined3d_private.h"
  27
  28 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  29 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 #define GLINFO_LOCATION (*gl_info)
 194
 195 static void wrap_op1(const struct wined3d_gl_info *gl_info, GLuint op, GLuint dst, GLuint dstMask, GLuint dstMod,
 196         GLuint arg1, GLuint arg1Rep, GLuint arg1Mod)
 197 {
 198     if(dstMask == GL_ALPHA) {
 199         TRACE("glAlphaFragmentOp1ATI(%s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst), debug_dstmod(dstMod),
 200               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod));
 201         GL_EXTCALL(glAlphaFragmentOp1ATI(op, dst, dstMod, arg1, arg1Rep, arg1Mod));
 202     } else {
 203         TRACE("glColorFragmentOp1ATI(%s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst),
 204               debug_mask(dstMask), debug_dstmod(dstMod),
 205               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod));
 206         GL_EXTCALL(glColorFragmentOp1ATI(op, dst, dstMask, dstMod, arg1, arg1Rep, arg1Mod));
 207     }
 208 }
 209
 210 static void wrap_op2(const struct wined3d_gl_info *gl_info, GLuint op, GLuint dst, GLuint dstMask, GLuint dstMod,
 211         GLuint arg1, GLuint arg1Rep, GLuint arg1Mod, GLuint arg2, GLuint arg2Rep, GLuint arg2Mod)
 212 {
 213     if(dstMask == GL_ALPHA) {
 214         TRACE("glAlphaFragmentOp2ATI(%s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst), debug_dstmod(dstMod),
 215               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 216               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod));
 217         GL_EXTCALL(glAlphaFragmentOp2ATI(op, dst, dstMod, arg1, arg1Rep, arg1Mod, arg2, arg2Rep, arg2Mod));
 218     } else {
 219         TRACE("glColorFragmentOp2ATI(%s, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst),
 220               debug_mask(dstMask), debug_dstmod(dstMod),
 221               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 222               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod));
 223         GL_EXTCALL(glColorFragmentOp2ATI(op, dst, dstMask, dstMod, arg1, arg1Rep, arg1Mod, arg2, arg2Rep, arg2Mod));
 224     }
 225 }
 226
 227 static void wrap_op3(const struct wined3d_gl_info *gl_info, GLuint op, GLuint dst, GLuint dstMask, GLuint dstMod,
 228         GLuint arg1, GLuint arg1Rep, GLuint arg1Mod, GLuint arg2, GLuint arg2Rep, GLuint arg2Mod,
 229         GLuint arg3, GLuint arg3Rep, GLuint arg3Mod)
 230 {
 231     if(dstMask == GL_ALPHA) {
 232         /* Leave some free space to fit "GL_NONE, " in to align most alpha and color op lines */
 233         TRACE("glAlphaFragmentOp3ATI(%s, %s,          %s, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst), debug_dstmod(dstMod),
 234               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 235               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod),
 236               debug_register(arg3), debug_rep(arg3Rep), debug_argmod(arg3Mod));
 237         GL_EXTCALL(glAlphaFragmentOp3ATI(op, dst, dstMod,
 238                                          arg1, arg1Rep, arg1Mod,
 239                                          arg2, arg2Rep, arg2Mod,
 240                                          arg3, arg3Rep, arg3Mod));
 241     } else {
 242         TRACE("glColorFragmentOp3ATI(%s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst),
 243               debug_mask(dstMask), debug_dstmod(dstMod),
 244               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 245               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod),
 246               debug_register(arg3), debug_rep(arg3Rep), debug_argmod(arg3Mod));
 247         GL_EXTCALL(glColorFragmentOp3ATI(op, dst, dstMask, dstMod,
 248                                          arg1, arg1Rep, arg1Mod,
 249                                          arg2, arg2Rep, arg2Mod,
 250                                          arg3, arg3Rep, arg3Mod));
 251     }
 252 }
 253
 254 static GLuint register_for_arg(DWORD arg, const struct wined3d_gl_info *gl_info,
 255         unsigned int stage, GLuint *mod, GLuint *rep, GLuint tmparg)
 256 {
 257     GLenum ret;
 258
 259     if(mod) *mod = GL_NONE;
 260     if(arg == ARG_UNUSED)
 261     {
 262         if (rep) *rep = GL_NONE;
 263         return -1; /* This is the marker for unused registers */
 264     }
 265
 266     switch(arg & WINED3DTA_SELECTMASK) {
 267         case WINED3DTA_DIFFUSE:
 268             ret = GL_PRIMARY_COLOR;
 269             break;
 270
 271         case WINED3DTA_CURRENT:
 272             /* Note that using GL_REG_0_ATI for the passed on register is safe because
 273              * texture0 is read at stage0, so in the worst case it is read in the
 274              * instruction writing to reg0. Afterwards texture0 is not used any longer.
 275              * If we're reading from current
 276              */
 277             if(stage == 0) {
 278                 ret = GL_PRIMARY_COLOR;
 279             } else {
 280                 ret = GL_REG_0_ATI;
 281             }
 282             break;
 283
 284         case WINED3DTA_TEXTURE:
 285             ret = GL_REG_0_ATI + stage;
 286             break;
 287
 288         case WINED3DTA_TFACTOR:
 289             ret = ATI_FFP_CONST_TFACTOR;
 290             break;
 291
 292         case WINED3DTA_SPECULAR:
 293             ret = GL_SECONDARY_INTERPOLATOR_ATI;
 294             break;
 295
 296         case WINED3DTA_TEMP:
 297             ret = tmparg;
 298             break;
 299
 300         case WINED3DTA_CONSTANT:
 301             FIXME("Unhandled source argument WINED3DTA_TEMP\n");
 302             ret = GL_CON_0_ATI;
 303             break;
 304
 305         default:
 306             FIXME("Unknown source argument %d\n", arg);
 307             ret = GL_ZERO;
 308     }
 309
 310     if(arg & WINED3DTA_COMPLEMENT) {
 311         if(mod) *mod |= GL_COMP_BIT_ATI;
 312     }
 313     if(arg & WINED3DTA_ALPHAREPLICATE) {
 314         if(rep) *rep = GL_ALPHA;
 315     } else {
 316         if(rep) *rep = GL_NONE;
 317     }
 318     return ret;
 319 }
 320
 321 static GLuint find_tmpreg(const struct texture_stage_op op[MAX_TEXTURES])
 322 {
 323     int lowest_read = -1;
 324     int lowest_write = -1;
 325     int i;
 326     BOOL tex_used[MAX_TEXTURES];
 327
 328     memset(tex_used, 0, sizeof(tex_used));
 329     for(i = 0; i < MAX_TEXTURES; i++) {
 330         if(op[i].cop == WINED3DTOP_DISABLE) {
 331             break;
 332         }
 333
 334         if(lowest_read == -1 &&
 335           (op[i].carg1 == WINED3DTA_TEMP || op[i].carg2 == WINED3DTA_TEMP || op[i].carg0 == WINED3DTA_TEMP ||
 336            op[i].aarg1 == WINED3DTA_TEMP || op[i].aarg2 == WINED3DTA_TEMP || op[i].aarg0 == WINED3DTA_TEMP)) {
 337             lowest_read = i;
 338         }
 339
 340         if(lowest_write == -1 && op[i].dst == tempreg) {
 341             lowest_write = i;
 342         }
 343
 344         if(op[i].carg1 == WINED3DTA_TEXTURE || op[i].carg2 == WINED3DTA_TEXTURE || op[i].carg0 == WINED3DTA_TEXTURE ||
 345            op[i].aarg1 == WINED3DTA_TEXTURE || op[i].aarg2 == WINED3DTA_TEXTURE || op[i].aarg0 == WINED3DTA_TEXTURE) {
 346             tex_used[i] = TRUE;
 347         }
 348     }
 349
 350     /* Temp reg not read? We don't need it, return GL_NONE */
 351     if(lowest_read == -1) return GL_NONE;
 352
 353     if(lowest_write >= lowest_read) {
 354         FIXME("Temp register read before being written\n");
 355     }
 356
 357     if(lowest_write == -1) {
 358         /* This needs a test. Maybe we are supposed to return 0.0/0.0/0.0/0.0, or fail drawprim, or whatever */
 359         FIXME("Temp register read without being written\n");
 360         return GL_REG_1_ATI;
 361     } else if(lowest_write >= 1) {
 362         /* If we're writing to the temp reg at earliest in stage 1, we can use register 1 for the temp result.
 363          * there may be texture data stored in reg 1, but we do not need it any longer since stage 1 already
 364          * read it
 365          */
 366         return GL_REG_1_ATI;
 367     } else {
 368         /* Search for a free texture register. We have 6 registers available. GL_REG_0_ATI is already used
 369          * for the regular result
 370          */
 371         for(i = 1; i < 6; i++) {
 372             if(!tex_used[i]) {
 373                 return GL_REG_0_ATI + i;
 374             }
 375         }
 376         /* What to do here? Report it in ValidateDevice? */
 377         FIXME("Could not find a register for the temporary register\n");
 378         return 0;
 379     }
 380 }
 381
 382 static GLuint gen_ati_shader(const struct texture_stage_op op[MAX_TEXTURES], const struct wined3d_gl_info *gl_info)
 383 {
 384     GLuint ret = GL_EXTCALL(glGenFragmentShadersATI(1));
 385     unsigned int stage;
 386     GLuint arg0, arg1, arg2, extrarg;
 387     GLuint dstmod, argmod0, argmod1, argmod2, argmodextra;
 388     GLuint rep0, rep1, rep2;
 389     GLuint swizzle;
 390     GLuint tmparg = find_tmpreg(op);
 391     GLuint dstreg;
 392
 393     if(!ret) {
 394         ERR("Failed to generate a GL_ATI_fragment_shader shader id\n");
 395         return 0;
 396     }
 397     GL_EXTCALL(glBindFragmentShaderATI(ret));
 398     checkGLcall("GL_EXTCALL(glBindFragmentShaderATI(ret))");
 399
 400     TRACE("glBeginFragmentShaderATI()\n");
 401     GL_EXTCALL(glBeginFragmentShaderATI());
 402     checkGLcall("GL_EXTCALL(glBeginFragmentShaderATI())");
 403
 404     /* Pass 1: Generate sampling instructions for perturbation maps */
 405     for (stage = 0; stage < gl_info->limits.textures; ++stage)
 406     {
 407         if(op[stage].cop == WINED3DTOP_DISABLE) break;
 408         if(op[stage].cop != WINED3DTOP_BUMPENVMAP &&
 409            op[stage].cop != WINED3DTOP_BUMPENVMAPLUMINANCE) continue;
 410
 411         TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, GL_SWIZZLE_STR_ATI)\n",
 412               stage, stage);
 413         GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 414                    GL_TEXTURE0_ARB + stage,
 415                    GL_SWIZZLE_STR_ATI));
 416         if(op[stage + 1].projected == proj_none) {
 417             swizzle = GL_SWIZZLE_STR_ATI;
 418         } else if(op[stage + 1].projected == proj_count4) {
 419             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 420         } else {
 421             swizzle = GL_SWIZZLE_STR_DR_ATI;
 422         }
 423         TRACE("glPassTexCoordATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, %s)\n",
 424               stage + 1, stage + 1, debug_swizzle(swizzle));
 425         GL_EXTCALL(glPassTexCoordATI(GL_REG_0_ATI + stage + 1,
 426                    GL_TEXTURE0_ARB + stage + 1,
 427                    swizzle));
 428     }
 429
 430     /* Pass 2: Generate perturbation calculations */
 431     for (stage = 0; stage < gl_info->limits.textures; ++stage)
 432     {
 433         GLuint argmodextra_x, argmodextra_y;
 434         struct color_fixup_desc fixup;
 435
 436         if(op[stage].cop == WINED3DTOP_DISABLE) break;
 437         if(op[stage].cop != WINED3DTOP_BUMPENVMAP &&
 438            op[stage].cop != WINED3DTOP_BUMPENVMAPLUMINANCE) continue;
 439
 440         fixup = op[stage].color_fixup;
 441         if (fixup.x_source != CHANNEL_SOURCE_X || fixup.y_source != CHANNEL_SOURCE_Y)
 442         {
 443             FIXME("Swizzles not implemented\n");
 444             argmodextra_x = GL_NONE;
 445             argmodextra_y = GL_NONE;
 446         }
 447         else
 448         {
 449             /* Nice thing, we get the color correction for free :-) */
 450             argmodextra_x = fixup.x_sign_fixup ? GL_2X_BIT_ATI | GL_BIAS_BIT_ATI : GL_NONE;
 451             argmodextra_y = fixup.y_sign_fixup ? GL_2X_BIT_ATI | GL_BIAS_BIT_ATI : GL_NONE;
 452         }
 453
 454         wrap_op3(gl_info, GL_DOT2_ADD_ATI, GL_REG_0_ATI + stage + 1, GL_RED_BIT_ATI, GL_NONE,
 455                  GL_REG_0_ATI + stage, GL_NONE, argmodextra_x,
 456                  ATI_FFP_CONST_BUMPMAT(stage), GL_NONE, GL_2X_BIT_ATI | GL_BIAS_BIT_ATI,
 457                  GL_REG_0_ATI + stage + 1, GL_RED, GL_NONE);
 458
 459         /* Don't use GL_DOT2_ADD_ATI here because we cannot configure it to read the blue and alpha
 460          * component of the bump matrix. Instead do this with two MADs:
 461          *
 462          * coord.a = tex.r * bump.b + coord.g
 463          * coord.g = tex.g * bump.a + coord.a
 464          *
 465          * The first instruction writes to alpha so it can be coissued with the above DOT2_ADD.
 466          * coord.a is unused. If the perturbed texture is projected, this was already handled
 467          * in the glPassTexCoordATI above.
 468          */
 469         wrap_op3(gl_info, GL_MAD_ATI, GL_REG_0_ATI + stage + 1, GL_ALPHA, GL_NONE,
 470                  GL_REG_0_ATI + stage, GL_RED, argmodextra_y,
 471                  ATI_FFP_CONST_BUMPMAT(stage), GL_BLUE, GL_NONE,
 472                  GL_REG_0_ATI + stage + 1, GL_GREEN, GL_NONE);
 473         wrap_op3(gl_info, GL_MAD_ATI, GL_REG_0_ATI + stage + 1, GL_GREEN_BIT_ATI, GL_NONE,
 474                  GL_REG_0_ATI + stage, GL_GREEN, argmodextra_y,
 475                  ATI_FFP_CONST_BUMPMAT(stage), GL_ALPHA, GL_NONE,
 476                  GL_REG_0_ATI + stage + 1, GL_ALPHA, GL_NONE);
 477     }
 478
 479     /* Pass 3: Generate sampling instructions for regular textures */
 480     for (stage = 0; stage < gl_info->limits.textures; ++stage)
 481     {
 482         if(op[stage].cop == WINED3DTOP_DISABLE) {
 483             break;
 484         }
 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 == WINED3DTOP_BLENDTEXTUREALPHA) {
 501
 502             if(stage > 0 &&
 503                (op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
 504                 op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
 505                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_REG_%d_ATI, GL_SWIZZLE_STR_ATI)\n",
 506                       stage, stage);
 507                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 508                            GL_REG_0_ATI + stage,
 509                            GL_SWIZZLE_STR_ATI));
 510             } else {
 511                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, %s)\n",
 512                     stage, stage, debug_swizzle(swizzle));
 513                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 514                                         GL_TEXTURE0_ARB + stage,
 515                                         swizzle));
 516             }
 517         }
 518     }
 519
 520     /* Pass 4: Generate the arithmetic instructions */
 521     for(stage = 0; stage < MAX_TEXTURES; stage++) {
 522         if(op[stage].cop == WINED3DTOP_DISABLE) {
 523             if(stage == 0) {
 524                 /* Handle complete texture disabling gracefully */
 525                 wrap_op1(gl_info, GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 526                          GL_PRIMARY_COLOR, GL_NONE, GL_NONE);
 527                 wrap_op1(gl_info, GL_MOV_ATI, GL_REG_0_ATI, GL_ALPHA, GL_NONE,
 528                          GL_PRIMARY_COLOR, GL_NONE, GL_NONE);
 529             }
 530             break;
 531         }
 532
 533         if(op[stage].dst == tempreg) {
 534             /* If we're writing to D3DTA_TEMP, but never reading from it we don't have to write there in the first place.
 535              * skip the entire stage, this saves some GPU time
 536              */
 537             if(tmparg == GL_NONE) continue;
 538
 539             dstreg = tmparg;
 540         } else {
 541             dstreg = GL_REG_0_ATI;
 542         }
 543
 544         arg0 = register_for_arg(op[stage].carg0, gl_info, stage, &argmod0, &rep0, tmparg);
 545         arg1 = register_for_arg(op[stage].carg1, gl_info, stage, &argmod1, &rep1, tmparg);
 546         arg2 = register_for_arg(op[stage].carg2, gl_info, stage, &argmod2, &rep2, tmparg);
 547         dstmod = GL_NONE;
 548         argmodextra = GL_NONE;
 549         extrarg = GL_NONE;
 550
 551         switch(op[stage].cop) {
 552             case WINED3DTOP_SELECTARG2:
 553                 arg1 = arg2;
 554                 argmod1 = argmod2;
 555                 rep1 = rep2;
 556             case WINED3DTOP_SELECTARG1:
 557                 wrap_op1(gl_info, GL_MOV_ATI, dstreg, GL_NONE, GL_NONE,
 558                          arg1, rep1, argmod1);
 559                 break;
 560
 561             case WINED3DTOP_MODULATE4X:
 562                 if(dstmod == GL_NONE) dstmod = GL_4X_BIT_ATI;
 563             case WINED3DTOP_MODULATE2X:
 564                 if(dstmod == GL_NONE) dstmod = GL_2X_BIT_ATI;
 565                 dstmod |= GL_SATURATE_BIT_ATI;
 566             case WINED3DTOP_MODULATE:
 567                 wrap_op2(gl_info, GL_MUL_ATI, dstreg, GL_NONE, dstmod,
 568                          arg1, rep1, argmod1,
 569                          arg2, rep2, argmod2);
 570                 break;
 571
 572             case WINED3DTOP_ADDSIGNED2X:
 573                 dstmod = GL_2X_BIT_ATI;
 574             case WINED3DTOP_ADDSIGNED:
 575                 argmodextra = GL_BIAS_BIT_ATI;
 576             case WINED3DTOP_ADD:
 577                 dstmod |= GL_SATURATE_BIT_ATI;
 578                 wrap_op2(gl_info, GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, dstmod,
 579                          arg1, rep1, argmod1,
 580                          arg2, rep2, argmodextra | argmod2);
 581                 break;
 582
 583             case WINED3DTOP_SUBTRACT:
 584                 dstmod |= GL_SATURATE_BIT_ATI;
 585                 wrap_op2(gl_info, GL_SUB_ATI, dstreg, GL_NONE, dstmod,
 586                          arg1, rep1, argmod1,
 587                          arg2, rep2, argmod2);
 588                 break;
 589
 590             case WINED3DTOP_ADDSMOOTH:
 591                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 592                 /* Dst = arg1 + * arg2(1 -arg 1)
 593                  *     = arg2 * (1 - arg1) + arg1
 594                  */
 595                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 596                          arg2, rep2, argmod2,
 597                          arg1, rep1, argmodextra,
 598                          arg1, rep1, argmod1);
 599                 break;
 600
 601             case WINED3DTOP_BLENDCURRENTALPHA:
 602                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, NULL, -1);
 603             case WINED3DTOP_BLENDFACTORALPHA:
 604                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, NULL, -1);
 605             case WINED3DTOP_BLENDTEXTUREALPHA:
 606                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 607             case WINED3DTOP_BLENDDIFFUSEALPHA:
 608                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, NULL, -1);
 609                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 610                          extrarg, GL_ALPHA, GL_NONE,
 611                          arg1, rep1, argmod1,
 612                          arg2, rep2, argmod2);
 613                 break;
 614
 615             case WINED3DTOP_BLENDTEXTUREALPHAPM:
 616                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 617                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 618                          arg2, rep2,  argmod2,
 619                          arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 620                          arg1, rep1,  argmod1);
 621                 break;
 622
 623             /* D3DTOP_PREMODULATE ???? */
 624
 625             case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
 626                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 627             case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
 628                 if(!argmodextra) argmodextra = argmod1;
 629                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 630                          arg2, rep2,  argmod2,
 631                          arg1, GL_ALPHA, argmodextra,
 632                          arg1, rep1,  argmod1);
 633                 break;
 634
 635             case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
 636                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 637             case WINED3DTOP_MODULATECOLOR_ADDALPHA:
 638                 if(!argmodextra) argmodextra = argmod1;
 639                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 640                          arg2, rep2,  argmod2,
 641                          arg1, rep1,  argmodextra,
 642                          arg1, GL_ALPHA, argmod1);
 643                 break;
 644
 645             case WINED3DTOP_DOTPRODUCT3:
 646                 wrap_op2(gl_info, GL_DOT3_ATI, dstreg, GL_NONE, GL_4X_BIT_ATI | GL_SATURATE_BIT_ATI,
 647                          arg1, rep1, argmod1 | GL_BIAS_BIT_ATI,
 648                          arg2, rep2, argmod2 | GL_BIAS_BIT_ATI);
 649                 break;
 650
 651             case WINED3DTOP_MULTIPLYADD:
 652                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 653                          arg1, rep1, argmod1,
 654                          arg2, rep2, argmod2,
 655                          arg0, rep0, argmod0);
 656                 break;
 657
 658             case WINED3DTOP_LERP:
 659                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 660                          arg0, rep0, argmod0,
 661                          arg1, rep1, argmod1,
 662                          arg2, rep2, argmod2);
 663                 break;
 664
 665             case WINED3DTOP_BUMPENVMAP:
 666             case WINED3DTOP_BUMPENVMAPLUMINANCE:
 667                 /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
 668                 break;
 669
 670             default: FIXME("Unhandled color operation %d on stage %d\n", op[stage].cop, stage);
 671         }
 672
 673         arg0 = register_for_arg(op[stage].aarg0, gl_info, stage, &argmod0, NULL, tmparg);
 674         arg1 = register_for_arg(op[stage].aarg1, gl_info, stage, &argmod1, NULL, tmparg);
 675         arg2 = register_for_arg(op[stage].aarg2, gl_info, stage, &argmod2, NULL, tmparg);
 676         dstmod = GL_NONE;
 677         argmodextra = GL_NONE;
 678         extrarg = GL_NONE;
 679
 680         switch(op[stage].aop) {
 681             case WINED3DTOP_DISABLE:
 682                 /* Get the primary color to the output if on stage 0, otherwise leave register 0 untouched */
 683                 if(stage == 0) {
 684                     wrap_op1(gl_info, GL_MOV_ATI, GL_REG_0_ATI, GL_ALPHA, GL_NONE,
 685                              GL_PRIMARY_COLOR, GL_NONE, GL_NONE);
 686                 }
 687                 break;
 688
 689             case WINED3DTOP_SELECTARG2:
 690                 arg1 = arg2;
 691                 argmod1 = argmod2;
 692             case WINED3DTOP_SELECTARG1:
 693                 wrap_op1(gl_info, GL_MOV_ATI, dstreg, GL_ALPHA, GL_NONE,
 694                          arg1, GL_NONE, argmod1);
 695                 break;
 696
 697             case WINED3DTOP_MODULATE4X:
 698                 if(dstmod == GL_NONE) dstmod = GL_4X_BIT_ATI;
 699             case WINED3DTOP_MODULATE2X:
 700                 if(dstmod == GL_NONE) dstmod = GL_2X_BIT_ATI;
 701                 dstmod |= GL_SATURATE_BIT_ATI;
 702             case WINED3DTOP_MODULATE:
 703                 wrap_op2(gl_info, GL_MUL_ATI, dstreg, GL_ALPHA, dstmod,
 704                          arg1, GL_NONE, argmod1,
 705                          arg2, GL_NONE, argmod2);
 706                 break;
 707
 708             case WINED3DTOP_ADDSIGNED2X:
 709                 dstmod = GL_2X_BIT_ATI;
 710             case WINED3DTOP_ADDSIGNED:
 711                 argmodextra = GL_BIAS_BIT_ATI;
 712             case WINED3DTOP_ADD:
 713                 dstmod |= GL_SATURATE_BIT_ATI;
 714                 wrap_op2(gl_info, GL_ADD_ATI, dstreg, GL_ALPHA, dstmod,
 715                          arg1, GL_NONE, argmod1,
 716                          arg2, GL_NONE, argmodextra | argmod2);
 717                 break;
 718
 719             case WINED3DTOP_SUBTRACT:
 720                 dstmod |= GL_SATURATE_BIT_ATI;
 721                 wrap_op2(gl_info, GL_SUB_ATI, dstreg, GL_ALPHA, dstmod,
 722                          arg1, GL_NONE, argmod1,
 723                          arg2, GL_NONE, argmod2);
 724                 break;
 725
 726             case WINED3DTOP_ADDSMOOTH:
 727                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 728                 /* Dst = arg1 + * arg2(1 -arg 1)
 729                  *     = arg2 * (1 - arg1) + arg1
 730                  */
 731                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_ALPHA, GL_SATURATE_BIT_ATI,
 732                          arg2, GL_NONE, argmod2,
 733                          arg1, GL_NONE, argmodextra,
 734                          arg1, GL_NONE, argmod1);
 735                 break;
 736
 737             case WINED3DTOP_BLENDCURRENTALPHA:
 738                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, NULL, -1);
 739             case WINED3DTOP_BLENDFACTORALPHA:
 740                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, NULL, -1);
 741             case WINED3DTOP_BLENDTEXTUREALPHA:
 742                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 743             case WINED3DTOP_BLENDDIFFUSEALPHA:
 744                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, NULL, -1);
 745                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_ALPHA, GL_NONE,
 746                          extrarg, GL_ALPHA, GL_NONE,
 747                          arg1, GL_NONE, argmod1,
 748                          arg2, GL_NONE, argmod2);
 749                 break;
 750
 751             case WINED3DTOP_BLENDTEXTUREALPHAPM:
 752                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 753                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_ALPHA, GL_NONE,
 754                          arg2, GL_NONE,  argmod2,
 755                          arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 756                          arg1, GL_NONE,  argmod1);
 757                 break;
 758
 759             /* D3DTOP_PREMODULATE ???? */
 760
 761             case WINED3DTOP_DOTPRODUCT3:
 762                 wrap_op2(gl_info, GL_DOT3_ATI, dstreg, GL_ALPHA, GL_4X_BIT_ATI | GL_SATURATE_BIT_ATI,
 763                          arg1, GL_NONE, argmod1 | GL_BIAS_BIT_ATI,
 764                          arg2, GL_NONE, argmod2 | GL_BIAS_BIT_ATI);
 765                 break;
 766
 767             case WINED3DTOP_MULTIPLYADD:
 768                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_ALPHA, GL_SATURATE_BIT_ATI,
 769                          arg1, GL_NONE, argmod1,
 770                          arg2, GL_NONE, argmod2,
 771                          arg0, GL_NONE, argmod0);
 772                 break;
 773
 774             case WINED3DTOP_LERP:
 775                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_ALPHA, GL_SATURATE_BIT_ATI,
 776                          arg1, GL_NONE, argmod1,
 777                          arg2, GL_NONE, argmod2,
 778                          arg0, GL_NONE, argmod0);
 779                 break;
 780
 781             case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
 782             case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
 783             case WINED3DTOP_MODULATECOLOR_ADDALPHA:
 784             case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
 785             case WINED3DTOP_BUMPENVMAP:
 786             case WINED3DTOP_BUMPENVMAPLUMINANCE:
 787                 ERR("Application uses an invalid alpha operation\n");
 788                 break;
 789
 790             default: FIXME("Unhandled alpha operation %d on stage %d\n", op[stage].aop, stage);
 791         }
 792     }
 793
 794     TRACE("glEndFragmentShaderATI()\n");
 795     GL_EXTCALL(glEndFragmentShaderATI());
 796     checkGLcall("GL_EXTCALL(glEndFragmentShaderATI())");
 797     return ret;
 798 }
 799 #undef GLINFO_LOCATION
 800
 801 #define GLINFO_LOCATION stateblock->device->adapter->gl_info
 802 static void set_tex_op_atifs(DWORD state, IWineD3DStateBlockImpl *stateblock, struct wined3d_context *context)
 803 {
 804     IWineD3DDeviceImpl *This = stateblock->device;
 805     const struct atifs_ffp_desc *desc;
 806     struct ffp_frag_settings     settings;
 807     struct atifs_private_data   *priv = This->fragment_priv;
 808     DWORD mapped_stage;
 809     unsigned int i;
 810
 811     gen_ffp_frag_op(stateblock, &settings, TRUE);
 812     desc = (const struct atifs_ffp_desc *)find_ffp_frag_shader(&priv->fragment_shaders, &settings);
 813     if(!desc) {
 814         struct atifs_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
 815         if (!new_desc)
 816         {
 817             ERR("Out of memory\n");
 818             return;
 819         }
 820         new_desc->num_textures_used = 0;
 821         for (i = 0; i < context->gl_info->limits.texture_stages; ++i)
 822         {
 823             if(settings.op[i].cop == WINED3DTOP_DISABLE) break;
 824             new_desc->num_textures_used = i;
 825         }
 826
 827         memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
 828         new_desc->shader = gen_ati_shader(settings.op, context->gl_info);
 829         add_ffp_frag_shader(&priv->fragment_shaders, &new_desc->parent);
 830         TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
 831         desc = new_desc;
 832     }
 833
 834     /* GL_ATI_fragment_shader depends on the GL_TEXTURE_xD enable settings. Update the texture stages
 835      * used by this shader
 836      */
 837     for(i = 0; i < desc->num_textures_used; i++) {
 838         mapped_stage = This->texUnitMap[i];
 839         if (mapped_stage != WINED3D_UNMAPPED_STAGE)
 840         {
 841             GL_EXTCALL(glActiveTextureARB(GL_TEXTURE0_ARB + mapped_stage));
 842             checkGLcall("glActiveTextureARB");
 843             texture_activate_dimensions(i, stateblock, context);
 844         }
 845     }
 846
 847     GL_EXTCALL(glBindFragmentShaderATI(desc->shader));
 848 }
 849
 850 static void state_texfactor_atifs(DWORD state, IWineD3DStateBlockImpl *stateblock, struct wined3d_context *context)
 851 {
 852     float col[4];
 853     D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
 854
 855     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col));
 856     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col)");
 857 }
 858
 859 static void set_bumpmat(DWORD state, IWineD3DStateBlockImpl *stateblock, struct wined3d_context *context)
 860 {
 861     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
 862     float mat[2][2];
 863
 864     mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
 865     mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
 866     mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
 867     mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
 868     /* GL_ATI_fragment_shader allows only constants from 0.0 to 1.0, but the bumpmat
 869      * constants can be in any range. While they should stay between [-1.0 and 1.0] because
 870      * Shader Model 1.x pixel shaders are clamped to that range negative values are used occasionally,
 871      * for example by our d3d9 test. So to get negative values scale -1;1 to 0;1 and undo that in the
 872      * shader(it is free). This might potentially reduce precision. However, if the hardware does
 873      * support proper floats it shouldn't, and if it doesn't we can't get anything better anyway
 874      */
 875     mat[0][0] = (mat[0][0] + 1.0f) * 0.5f;
 876     mat[1][0] = (mat[1][0] + 1.0f) * 0.5f;
 877     mat[0][1] = (mat[0][1] + 1.0f) * 0.5f;
 878     mat[1][1] = (mat[1][1] + 1.0f) * 0.5f;
 879     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), (float *) mat));
 880     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), mat)");
 881 }
 882
 883 static void textransform(DWORD state, IWineD3DStateBlockImpl *stateblock, struct wined3d_context *context)
 884 {
 885     if(!isStateDirty(context, STATE_PIXELSHADER)) {
 886         set_tex_op_atifs(state, stateblock, context);
 887     }
 888 }
 889
 890 static void atifs_apply_pixelshader(DWORD state, IWineD3DStateBlockImpl *stateblock, struct wined3d_context *context)
 891 {
 892     IWineD3DDeviceImpl *device = stateblock->device;
 893     BOOL use_vshader = use_vs(stateblock);
 894
 895     context->last_was_pshader = use_ps(stateblock);
 896     /* The ATIFS code does not support pixel shaders currently, but we have to provide a state handler
 897      * to call shader_select to select a vertex shader if one is applied because the vertex shader state
 898      * may defer calling the shader backend if the pshader state is dirty.
 899      *
 900      * In theory the application should not be able to mark the pixel shader dirty because it cannot
 901      * create a shader, and thus has no way to set the state to something != NULL. However, a different
 902      * pipeline part may link a different state to its pixelshader handler, thus a pshader state exists
 903      * and can be dirtified. Also the pshader is always dirtified at startup, and blitting disables all
 904      * shaders and dirtifies all shader states. If atifs can deal with this it keeps the rest of the code
 905      * simpler.
 906      */
 907     if(!isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
 908         device->shader_backend->shader_select(context, FALSE, use_vshader);
 909
 910         if (!isStateDirty(context, STATE_VERTEXSHADERCONSTANT) && use_vshader)
 911             stateblock_apply_state(STATE_VERTEXSHADERCONSTANT, stateblock, context);
 912     }
 913 }
 914
 915 #undef GLINFO_LOCATION
 916
 917 static const struct StateEntryTemplate atifs_fragmentstate_template[] = {
 918     {STATE_RENDER(WINED3DRS_TEXTUREFACTOR),               { STATE_RENDER(WINED3DRS_TEXTUREFACTOR),              state_texfactor_atifs   }, WINED3D_GL_EXT_NONE             },
 919     {STATE_RENDER(WINED3DRS_FOGCOLOR),                    { STATE_RENDER(WINED3DRS_FOGCOLOR),                   state_fogcolor          }, WINED3D_GL_EXT_NONE             },
 920     {STATE_RENDER(WINED3DRS_FOGDENSITY),                  { STATE_RENDER(WINED3DRS_FOGDENSITY),                 state_fogdensity        }, WINED3D_GL_EXT_NONE             },
 921     {STATE_RENDER(WINED3DRS_FOGENABLE),                   { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_fog_fragpart      }, WINED3D_GL_EXT_NONE             },
 922     {STATE_RENDER(WINED3DRS_FOGTABLEMODE),                { STATE_RENDER(WINED3DRS_FOGENABLE),                  NULL                    }, WINED3D_GL_EXT_NONE             },
 923     {STATE_RENDER(WINED3DRS_FOGVERTEXMODE),               { STATE_RENDER(WINED3DRS_FOGENABLE),                  NULL                    }, WINED3D_GL_EXT_NONE             },
 924     {STATE_RENDER(WINED3DRS_FOGSTART),                    { STATE_RENDER(WINED3DRS_FOGSTART),                   state_fogstartend       }, WINED3D_GL_EXT_NONE             },
 925     {STATE_RENDER(WINED3DRS_FOGEND),                      { STATE_RENDER(WINED3DRS_FOGSTART),                   NULL                    }, WINED3D_GL_EXT_NONE             },
 926     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, WINED3D_GL_EXT_NONE             },
 927     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 928     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 929     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 930     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 931     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 932     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 933     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 934     {STATE_TEXTURESTAGE(0, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 935     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 936     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 937     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 938     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 939     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 940     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 941     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 942     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 943     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 944     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 945     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 946     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 947     {STATE_TEXTURESTAGE(1, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 948     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 949     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 950     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 951     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 952     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 953     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 954     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 955     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 956     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 957     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 958     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 959     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 960     {STATE_TEXTURESTAGE(2, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 961     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 962     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 963     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 964     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 965     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 966     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 967     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 968     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 969     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 970     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 971     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 972     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 973     {STATE_TEXTURESTAGE(3, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 974     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 975     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 976     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 977     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 978     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 979     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 980     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 981     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 982     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 983     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 984     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 985     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 986     {STATE_TEXTURESTAGE(4, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 987     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 988     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 989     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 990     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 991     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 992     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 993     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 994     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 995     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 996     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 997     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 998     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 999     {STATE_TEXTURESTAGE(5, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1000     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1001     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1002     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1003     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1004     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1005     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1006     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1007     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1008     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1009     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1010     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1011     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1012     {STATE_TEXTURESTAGE(6, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1013     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1014     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1015     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1016     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1017     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1018     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1019     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1020     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1021     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1022     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1023     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1024     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1025     {STATE_TEXTURESTAGE(7, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1026     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1027     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1028     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1029     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1030     { STATE_SAMPLER(0),                                   { STATE_SAMPLER(0),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1031     { STATE_SAMPLER(1),                                   { STATE_SAMPLER(1),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1032     { STATE_SAMPLER(2),                                   { STATE_SAMPLER(2),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1033     { STATE_SAMPLER(3),                                   { STATE_SAMPLER(3),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1034     { STATE_SAMPLER(4),                                   { STATE_SAMPLER(4),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1035     { STATE_SAMPLER(5),                                   { STATE_SAMPLER(5),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1036     { STATE_SAMPLER(6),                                   { STATE_SAMPLER(6),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1037     { STATE_SAMPLER(7),                                   { STATE_SAMPLER(7),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1038     {STATE_TEXTURESTAGE(0,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(0, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1039     {STATE_TEXTURESTAGE(1,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(1, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1040     {STATE_TEXTURESTAGE(2,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(2, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1041     {STATE_TEXTURESTAGE(3,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(3, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1042     {STATE_TEXTURESTAGE(4,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(4, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1043     {STATE_TEXTURESTAGE(5,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(5, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1044     {STATE_TEXTURESTAGE(6,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(6, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1045     {STATE_TEXTURESTAGE(7,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(7, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1046     {STATE_PIXELSHADER,                                   { STATE_PIXELSHADER,                                  atifs_apply_pixelshader }, WINED3D_GL_EXT_NONE             },
1047     {0 /* Terminate */,                                   { 0,                                                  0                       }, WINED3D_GL_EXT_NONE             },
1048 };
1049
1050 /* Context activation is done by the caller. */
1051 static void atifs_enable(IWineD3DDevice *iface, BOOL enable) {
1052     ENTER_GL();
1053     if(enable) {
1054         glEnable(GL_FRAGMENT_SHADER_ATI);
1055         checkGLcall("glEnable(GL_FRAGMENT_SHADER_ATI)");
1056     } else {
1057         glDisable(GL_FRAGMENT_SHADER_ATI);
1058         checkGLcall("glDisable(GL_FRAGMENT_SHADER_ATI)");
1059     }
1060     LEAVE_GL();
1061 }
1062
1063 static void atifs_get_caps(const struct wined3d_gl_info *gl_info, struct fragment_caps *caps)
1064 {
1065     caps->PrimitiveMiscCaps = WINED3DPMISCCAPS_TSSARGTEMP;
1066     caps->TextureOpCaps =  WINED3DTEXOPCAPS_DISABLE                     |
1067                            WINED3DTEXOPCAPS_SELECTARG1                  |
1068                            WINED3DTEXOPCAPS_SELECTARG2                  |
1069                            WINED3DTEXOPCAPS_MODULATE4X                  |
1070                            WINED3DTEXOPCAPS_MODULATE2X                  |
1071                            WINED3DTEXOPCAPS_MODULATE                    |
1072                            WINED3DTEXOPCAPS_ADDSIGNED2X                 |
1073                            WINED3DTEXOPCAPS_ADDSIGNED                   |
1074                            WINED3DTEXOPCAPS_ADD                         |
1075                            WINED3DTEXOPCAPS_SUBTRACT                    |
1076                            WINED3DTEXOPCAPS_ADDSMOOTH                   |
1077                            WINED3DTEXOPCAPS_BLENDCURRENTALPHA           |
1078                            WINED3DTEXOPCAPS_BLENDFACTORALPHA            |
1079                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHA           |
1080                            WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA           |
1081                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM         |
1082                            WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR      |
1083                            WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA      |
1084                            WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA   |
1085                            WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR   |
1086                            WINED3DTEXOPCAPS_DOTPRODUCT3                 |
1087                            WINED3DTEXOPCAPS_MULTIPLYADD                 |
1088                            WINED3DTEXOPCAPS_LERP                        |
1089                            WINED3DTEXOPCAPS_BUMPENVMAP;
1090
1091     /* TODO: Implement WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE
1092     and WINED3DTEXOPCAPS_PREMODULATE */
1093
1094     /* GL_ATI_fragment_shader always supports 6 textures, which was the limit on r200 cards
1095      * which this extension is exclusively focused on(later cards have GL_ARB_fragment_program).
1096      * If the current card has more than 8 fixed function textures in OpenGL's regular fixed
1097      * function pipeline then the ATI_fragment_shader backend imposes a stricter limit. This
1098      * shouldn't be too hard since Nvidia cards have a limit of 4 textures with the default ffp
1099      * pipeline, and almost all games are happy with that. We can however support up to 8
1100      * texture stages because we have a 2nd pass limit of 8 instructions, and per stage we use
1101      * only 1 instruction.
1102      *
1103      * The proper fix for this is not to use GL_ATI_fragment_shader on cards newer than the
1104      * r200 series and use an ARB or GLSL shader instead
1105      */
1106     caps->MaxTextureBlendStages   = 8;
1107     caps->MaxSimultaneousTextures = 6;
1108 }
1109
1110 static HRESULT atifs_alloc(IWineD3DDevice *iface) {
1111     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
1112     struct atifs_private_data *priv;
1113
1114     This->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct atifs_private_data));
1115     if(!This->fragment_priv) {
1116         ERR("Out of memory\n");
1117         return E_OUTOFMEMORY;
1118     }
1119     priv = This->fragment_priv;
1120     if (wine_rb_init(&priv->fragment_shaders, &wined3d_ffp_frag_program_rb_functions) == -1)
1121     {
1122         ERR("Failed to initialize rbtree.\n");
1123         HeapFree(GetProcessHeap(), 0, This->fragment_priv);
1124         return E_OUTOFMEMORY;
1125     }
1126     return WINED3D_OK;
1127 }
1128
1129 #define GLINFO_LOCATION This->adapter->gl_info
1130 /* Context activation is done by the caller. */
1131 static void atifs_free_ffpshader(struct wine_rb_entry *entry, void *context)
1132 {
1133     IWineD3DDeviceImpl *This = context;
1134     struct atifs_ffp_desc *entry_ati = WINE_RB_ENTRY_VALUE(entry, struct atifs_ffp_desc, parent.entry);
1135
1136     ENTER_GL();
1137     GL_EXTCALL(glDeleteFragmentShaderATI(entry_ati->shader));
1138     checkGLcall("glDeleteFragmentShaderATI(entry->shader)");
1139     HeapFree(GetProcessHeap(), 0, entry_ati);
1140     LEAVE_GL();
1141 }
1142
1143 /* Context activation is done by the caller. */
1144 static void atifs_free(IWineD3DDevice *iface) {
1145     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
1146     struct atifs_private_data *priv = This->fragment_priv;
1147
1148     wine_rb_destroy(&priv->fragment_shaders, atifs_free_ffpshader, This);
1149
1150     HeapFree(GetProcessHeap(), 0, priv);
1151     This->fragment_priv = NULL;
1152 }
1153 #undef GLINFO_LOCATION
1154
1155 static BOOL atifs_color_fixup_supported(struct color_fixup_desc fixup)
1156 {
1157     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1158     {
1159         TRACE("Checking support for fixup:\n");
1160         dump_color_fixup_desc(fixup);
1161     }
1162
1163     /* We only support sign fixup of the first two channels. */
1164     if (fixup.x_source == CHANNEL_SOURCE_X && fixup.y_source == CHANNEL_SOURCE_Y
1165             && fixup.z_source == CHANNEL_SOURCE_Z && fixup.w_source == CHANNEL_SOURCE_W
1166             && !fixup.z_sign_fixup && !fixup.w_sign_fixup)
1167     {
1168         TRACE("[OK]\n");
1169         return TRUE;
1170     }
1171
1172     TRACE("[FAILED]\n");
1173     return FALSE;
1174 }
1175
1176 const struct fragment_pipeline atifs_fragment_pipeline = {
1177     atifs_enable,
1178     atifs_get_caps,
1179     atifs_alloc,
1180     atifs_free,
1181     atifs_color_fixup_supported,
1182     atifs_fragmentstate_template,
1183     TRUE /* We can disable projected textures */
1184 };