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

   1 /*
   2  * Fixed function pipeline replacement using GL_ATI_fragment_shader
   3  *
   4  * Copyright 2008 Stefan Dösinger(for CodeWeavers)
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  19  */
  20
  21 #include "config.h"
  22
  23 #include <math.h>
  24 #include <stdio.h>
  25
  26 #include "wined3d_private.h"
  27
  28 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  29
  30 /* Some private defines, Constant associations, etc.
  31  * Env bump matrix and per stage constant should be independent,
  32  * a stage that bump maps can't read the per state constant
  33  */
  34 #define ATI_FFP_CONST_BUMPMAT(i) (GL_CON_0_ATI + i)
  35 #define ATI_FFP_CONST_CONSTANT0 GL_CON_0_ATI
  36 #define ATI_FFP_CONST_CONSTANT1 GL_CON_1_ATI
  37 #define ATI_FFP_CONST_CONSTANT2 GL_CON_2_ATI
  38 #define ATI_FFP_CONST_CONSTANT3 GL_CON_3_ATI
  39 #define ATI_FFP_CONST_CONSTANT4 GL_CON_4_ATI
  40 #define ATI_FFP_CONST_CONSTANT5 GL_CON_5_ATI
  41 #define ATI_FFP_CONST_TFACTOR   GL_CON_6_ATI
  42
  43 /* GL_ATI_fragment_shader specific fixed function pipeline description. "Inherits" from the common one */
  44 struct atifs_ffp_desc
  45 {
  46     struct ffp_desc parent;
  47     GLuint shader;
  48     unsigned int num_textures_used;
  49 };
  50
  51 struct atifs_private_data
  52 {
  53     struct shader_arb_priv parent;
  54     struct list fragment_shaders; /* A linked list to track fragment pipeline replacement shaders */
  55
  56 };
  57
  58 static const char *debug_dstmod(GLuint mod) {
  59     switch(mod) {
  60         case GL_NONE:               return "GL_NONE";
  61         case GL_2X_BIT_ATI:         return "GL_2X_BIT_ATI";
  62         case GL_4X_BIT_ATI:         return "GL_4X_BIT_ATI";
  63         case GL_8X_BIT_ATI:         return "GL_8X_BIT_ATI";
  64         case GL_HALF_BIT_ATI:       return "GL_HALF_BIT_ATI";
  65         case GL_QUARTER_BIT_ATI:    return "GL_QUARTER_BIT_ATI";
  66         case GL_EIGHTH_BIT_ATI:     return "GL_EIGHTH_BIT_ATI";
  67         case GL_SATURATE_BIT_ATI:   return "GL_SATURATE_BIT_ATI";
  68         default:                    return "Unexpected modifier\n";
  69     }
  70 }
  71
  72 static const char *debug_argmod(GLuint mod) {
  73     switch(mod) {
  74         case GL_NONE:
  75             return "GL_NONE";
  76
  77         case GL_2X_BIT_ATI:
  78             return "GL_2X_BIT_ATI";
  79         case GL_COMP_BIT_ATI:
  80             return "GL_COMP_BIT_ATI";
  81         case GL_NEGATE_BIT_ATI:
  82             return "GL_NEGATE_BIT_ATI";
  83         case GL_BIAS_BIT_ATI:
  84             return "GL_BIAS_BIT_ATI";
  85
  86         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI:
  87             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI";
  88         case GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI:
  89             return "GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI";
  90         case GL_2X_BIT_ATI | GL_BIAS_BIT_ATI:
  91             return "GL_2X_BIT_ATI | GL_BIAS_BIT_ATI";
  92         case GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI:
  93             return "GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI";
  94         case GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI:
  95             return "GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI";
  96         case GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
  97             return "GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
  98
  99         case GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 100             return "GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 101         case GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 102             return "GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 103         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI:
 104             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI";
 105         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI:
 106             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI";
 107
 108         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 109             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 110
 111         default:
 112             return "Unexpected argmod combination\n";
 113     }
 114 }
 115 static const char *debug_register(GLuint reg) {
 116     switch(reg) {
 117         case GL_REG_0_ATI:                  return "GL_REG_0_ATI";
 118         case GL_REG_1_ATI:                  return "GL_REG_1_ATI";
 119         case GL_REG_2_ATI:                  return "GL_REG_2_ATI";
 120         case GL_REG_3_ATI:                  return "GL_REG_3_ATI";
 121         case GL_REG_4_ATI:                  return "GL_REG_4_ATI";
 122         case GL_REG_5_ATI:                  return "GL_REG_5_ATI";
 123
 124         case GL_CON_0_ATI:                  return "GL_CON_0_ATI";
 125         case GL_CON_1_ATI:                  return "GL_CON_1_ATI";
 126         case GL_CON_2_ATI:                  return "GL_CON_2_ATI";
 127         case GL_CON_3_ATI:                  return "GL_CON_3_ATI";
 128         case GL_CON_4_ATI:                  return "GL_CON_4_ATI";
 129         case GL_CON_5_ATI:                  return "GL_CON_5_ATI";
 130         case GL_CON_6_ATI:                  return "GL_CON_6_ATI";
 131         case GL_CON_7_ATI:                  return "GL_CON_7_ATI";
 132
 133         case GL_ZERO:                       return "GL_ZERO";
 134         case GL_ONE:                        return "GL_ONE";
 135         case GL_PRIMARY_COLOR:              return "GL_PRIMARY_COLOR";
 136         case GL_SECONDARY_INTERPOLATOR_ATI: return "GL_SECONDARY_INTERPOLATOR_ATI";
 137
 138         default:                            return "Unknown register\n";
 139     }
 140 }
 141
 142 static const char *debug_swizzle(GLuint swizzle) {
 143     switch(swizzle) {
 144         case GL_SWIZZLE_STR_ATI:        return "GL_SWIZZLE_STR_ATI";
 145         case GL_SWIZZLE_STQ_ATI:        return "GL_SWIZZLE_STQ_ATI";
 146         case GL_SWIZZLE_STR_DR_ATI:     return "GL_SWIZZLE_STR_DR_ATI";
 147         case GL_SWIZZLE_STQ_DQ_ATI:     return "GL_SWIZZLE_STQ_DQ_ATI";
 148         default:                        return "unknown swizzle";
 149     }
 150 }
 151
 152 #define GLINFO_LOCATION (*gl_info)
 153 static GLuint register_for_arg(DWORD arg, WineD3D_GL_Info *gl_info, unsigned int stage, GLuint *mod, GLuint tmparg) {
 154     GLenum ret;
 155
 156     if(mod) *mod = GL_NONE;
 157     if(arg == 0xFFFFFFFF) return -1; /* This is the marker for unused registers */
 158
 159     switch(arg & WINED3DTA_SELECTMASK) {
 160         case WINED3DTA_DIFFUSE:
 161             ret = GL_PRIMARY_COLOR;
 162             break;
 163
 164         case WINED3DTA_CURRENT:
 165             /* Note that using GL_REG_0_ATI for the passed on register is safe because
 166              * texture0 is read at stage0, so in the worst case it is read in the
 167              * instruction writing to reg0. Afterwards texture0 is not used any longer.
 168              * If we're reading from current
 169              */
 170             if(stage == 0) {
 171                 ret = GL_PRIMARY_COLOR;
 172             } else {
 173                 ret = GL_REG_0_ATI;
 174             }
 175             break;
 176
 177         case WINED3DTA_TEXTURE:
 178             ret = GL_REG_0_ATI + stage;
 179             break;
 180
 181         case WINED3DTA_TFACTOR:
 182             ret = ATI_FFP_CONST_TFACTOR;
 183             break;
 184
 185         case WINED3DTA_SPECULAR:
 186             ret = GL_SECONDARY_INTERPOLATOR_ATI;
 187             break;
 188
 189         case WINED3DTA_TEMP:
 190             ret = tmparg;
 191             break;
 192
 193         case WINED3DTA_CONSTANT:
 194             FIXME("Unhandled source argument WINED3DTA_TEMP\n");
 195             ret = GL_CON_0_ATI;
 196             break;
 197
 198         default:
 199             FIXME("Unknown source argument %d\n", arg);
 200             ret = GL_ZERO;
 201     }
 202
 203     if(arg & WINED3DTA_COMPLEMENT) {
 204         if(mod) *mod |= GL_COMP_BIT_ATI;
 205     }
 206     if(arg & WINED3DTA_ALPHAREPLICATE) {
 207         FIXME("Unhandled read modifier WINED3DTA_ALPHAREPLICATE\n");
 208     }
 209     return ret;
 210 }
 211
 212 static GLuint find_tmpreg(struct texture_stage_op op[MAX_TEXTURES]) {
 213     int lowest_read = -1;
 214     int lowest_write = -1;
 215     int i;
 216     BOOL tex_used[MAX_TEXTURES];
 217
 218     memset(tex_used, 0, sizeof(tex_used));
 219     for(i = 0; i < MAX_TEXTURES; i++) {
 220         if(op[i].cop == WINED3DTOP_DISABLE) {
 221             break;
 222         }
 223
 224         if(lowest_read == -1 &&
 225           (op[i].carg1 == WINED3DTA_TEMP || op[i].carg2 == WINED3DTA_TEMP || op[i].carg0 == WINED3DTA_TEMP ||
 226            op[i].aarg1 == WINED3DTA_TEMP || op[i].aarg2 == WINED3DTA_TEMP || op[i].aarg0 == WINED3DTA_TEMP)) {
 227             lowest_read = i;
 228         }
 229
 230         if(lowest_write == -1 && op[i].dst == WINED3DTA_TEMP) {
 231             lowest_write = i;
 232         }
 233
 234         if(op[i].carg1 == WINED3DTA_TEXTURE || op[i].carg2 == WINED3DTA_TEXTURE || op[i].carg0 == WINED3DTA_TEXTURE ||
 235            op[i].aarg1 == WINED3DTA_TEXTURE || op[i].aarg2 == WINED3DTA_TEXTURE || op[i].aarg0 == WINED3DTA_TEXTURE) {
 236             tex_used[i] = TRUE;
 237         }
 238     }
 239
 240     /* Temp reg not read? We don't need it, return GL_NONE */
 241     if(lowest_read == -1) return GL_NONE;
 242
 243     if(lowest_write >= lowest_read) {
 244         FIXME("Temp register read before being written\n");
 245     }
 246
 247     if(lowest_write == -1) {
 248         /* This needs a test. Maybe we are supposed to return 0.0/0.0/0.0/0.0, or fail drawprim, or whatever */
 249         FIXME("Temp register read without being written\n");
 250         return GL_REG_1_ATI;
 251     } else if(lowest_write >= 1) {
 252         /* If we're writing to the temp reg at earliest in stage 1, we can use register 1 for the temp result.
 253          * there may be texture data stored in reg 1, but we do not need it any longer since stage 1 already
 254          * read it
 255          */
 256         return GL_REG_1_ATI;
 257     } else {
 258         /* Search for a free texture register. We have 6 registers available. GL_REG_0_ATI is already used
 259          * for the regular result
 260          */
 261         for(i = 1; i < 6; i++) {
 262             if(!tex_used[i]) {
 263                 return GL_REG_0_ATI + i;
 264             }
 265         }
 266         /* What to do here? Report it in ValidateDevice? */
 267         FIXME("Could not find a register for the temporary register\n");
 268         return 0;
 269     }
 270 }
 271
 272 static GLuint gen_ati_shader(struct texture_stage_op op[MAX_TEXTURES], WineD3D_GL_Info *gl_info) {
 273     GLuint ret = GL_EXTCALL(glGenFragmentShadersATI(1));
 274     unsigned int stage;
 275     GLuint arg0, arg1, arg2, extrarg;
 276     GLuint dstmod, argmod0, argmod1, argmod2, argmodextra;
 277     GLuint swizzle;
 278     GLuint tmparg = find_tmpreg(op);
 279     GLuint dstreg;
 280
 281     if(!ret) {
 282         ERR("Failed to generate a GL_ATI_fragment_shader shader id\n");
 283         return 0;
 284     }
 285     GL_EXTCALL(glBindFragmentShaderATI(ret));
 286     checkGLcall("GL_EXTCALL(glBindFragmentShaderATI(ret))");
 287
 288     TRACE("glBeginFragmentShaderATI()\n");
 289     GL_EXTCALL(glBeginFragmentShaderATI());
 290     checkGLcall("GL_EXTCALL(glBeginFragmentShaderATI())");
 291
 292     /* Pass 1: Generate sampling instructions for perturbation maps */
 293       for(stage = 0; stage < GL_LIMITS(textures); stage++) {
 294         if(op[stage].cop == WINED3DTOP_DISABLE) break;
 295         if(op[stage].cop != WINED3DTOP_BUMPENVMAP &&
 296            op[stage].cop != WINED3DTOP_BUMPENVMAPLUMINANCE) continue;
 297
 298         TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, GL_SWIZZLE_STR_ATI)\n",
 299               stage, stage);
 300         GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 301                    GL_TEXTURE0_ARB + stage,
 302                    GL_SWIZZLE_STR_ATI));
 303         TRACE("glPassTexCoordATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, GL_SWIZZLE_STR_ATI)\n",
 304               stage + 1, stage + 1);
 305         GL_EXTCALL(glPassTexCoordATI(GL_REG_0_ATI + stage + 1,
 306                    GL_TEXTURE0_ARB + stage + 1,
 307                    GL_SWIZZLE_STR_ATI));
 308
 309         /* We need GL_REG_5_ATI as a temporary register to swizzle the bump matrix. So we run into
 310          * issues if we're bump mapping on stage 4 or 5
 311          */
 312         if(stage >= 4) {
 313             FIXME("Bump mapping in stage %d\n", stage);
 314         }
 315     }
 316
 317     /* Pass 2: Generate perturbation calculations */
 318     for(stage = 0; stage < GL_LIMITS(textures); stage++) {
 319         if(op[stage].cop == WINED3DTOP_DISABLE) break;
 320         if(op[stage].cop != WINED3DTOP_BUMPENVMAP &&
 321            op[stage].cop != WINED3DTOP_BUMPENVMAPLUMINANCE) continue;
 322
 323         /* Nice thing, we get the color correction for free :-) */
 324         if(op[stage].color_correction == WINED3DFMT_V8U8) {
 325             argmodextra = GL_2X_BIT_ATI | GL_BIAS_BIT_ATI;
 326         } else {
 327             argmodextra = 0;
 328         }
 329
 330         TRACE("glColorFragmentOp3ATI(GL_DOT2_ADD_ATI, GL_REG_%d_ATI, GL_RED_BIT_ATI, GL_NONE, GL_REG_%d_ATI, GL_NONE, %s, ATI_FFP_CONST_BUMPMAT(%d), GL_NONE, GL_NONE, GL_REG_%d_ATI, GL_RED, GL_NONE)\n",
 331               stage + 1, stage, debug_argmod(argmodextra), stage, stage + 1);
 332         GL_EXTCALL(glColorFragmentOp3ATI(GL_DOT2_ADD_ATI, GL_REG_0_ATI + stage + 1, GL_RED_BIT_ATI, GL_NONE,
 333                                          GL_REG_0_ATI + stage, GL_NONE, argmodextra,
 334                                          ATI_FFP_CONST_BUMPMAT(stage), GL_NONE, GL_2X_BIT_ATI | GL_BIAS_BIT_ATI,
 335                                          GL_REG_0_ATI + stage + 1, GL_RED, GL_NONE));
 336
 337         /* FIXME: How can I make GL_DOT2_ADD_ATI read the factors from blue and alpha? It defaults to red and green,
 338          * and it is fairly easy to make it read GL_BLUE or BL_ALPHA, but I can't get an R * B + G * A. So we're wasting
 339          * one register and two instructions in this pass for a simple swizzling operation.
 340          * For starters it might be good enough to merge the two movs into one, but even that isn't possible :-(
 341          *
 342          * NOTE: GL_BLUE | GL_ALPHA is not possible. It doesn't throw a compilation error, but an OR operation on the
 343          * constants doesn't make sense, considering their values.
 344          */
 345         TRACE("glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_5_ATI, GL_RED_BIT_ATI, GL_NONE, ATI_FFP_CONST_BUMPMAT(%d), GL_BLUE, GL_NONE)\n", stage);
 346         GL_EXTCALL(glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_5_ATI, GL_RED_BIT_ATI, GL_NONE,
 347                                          ATI_FFP_CONST_BUMPMAT(stage), GL_BLUE, GL_NONE));
 348         TRACE("glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_5_ATI, GL_GREEN_BIT_ATI, GL_NONE, ATI_FFP_CONST_BUMPMAT(%d), GL_ALPHA, GL_NONE)\n", stage);
 349         GL_EXTCALL(glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_5_ATI, GL_GREEN_BIT_ATI, GL_NONE,
 350                                         ATI_FFP_CONST_BUMPMAT(stage), GL_ALPHA, GL_NONE));
 351         TRACE("glColorFragmentOp3ATI(GL_DOT2_ADD_ATI, GL_REG_%d_ATI, GL_GREEN_BIT_ATI, GL_NONE, GL_REG_%d_ATI, GL_NONE, %s, GL_REG_5_ATI, GL_NONE, GL_NONE, GL_REG_%d_ATI, GL_GREEN, GL_NONE)\n",
 352               stage + 1, stage, debug_argmod(argmodextra), stage + 1);
 353         GL_EXTCALL(glColorFragmentOp3ATI(GL_DOT2_ADD_ATI, GL_REG_0_ATI + stage + 1, GL_GREEN_BIT_ATI, GL_NONE,
 354                                          GL_REG_0_ATI + stage, GL_NONE, argmodextra,
 355                                          GL_REG_5_ATI, GL_NONE, GL_2X_BIT_ATI | GL_BIAS_BIT_ATI,
 356                                          GL_REG_0_ATI + stage + 1, GL_GREEN, GL_NONE));
 357     }
 358
 359     /* Pass 3: Generate sampling instructions for regular textures */
 360     for(stage = 0; stage < GL_LIMITS(textures); stage++) {
 361         if(op[stage].cop == WINED3DTOP_DISABLE) {
 362             break;
 363         }
 364
 365         if(op[stage].projected == proj_none) {
 366             swizzle = GL_SWIZZLE_STR_ATI;
 367         } else if(op[stage].projected == proj_count3) {
 368             /* TODO: D3DTTFF_COUNT3 | D3DTTFF_PROJECTED would be GL_SWIZZLE_STR_DR_ATI.
 369              * However, the FFP vertex processing texture transform matrix handler does
 370              * some transformations in the texture matrix which makes the 3rd coordinate
 371              * arrive in Q, not R in that case. This is needed for opengl fixed function
 372              * fragment processing which always divides by Q. In this backend we can
 373              * handle that properly and be compatible with vertex shader output and avoid
 374              * side effects of the texture matrix games
 375              */
 376             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 377         } else {
 378             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 379         }
 380
 381         if((op[stage].carg0 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 382            (op[stage].carg1 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 383            (op[stage].carg2 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 384            (op[stage].aarg0 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 385            (op[stage].aarg1 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 386            (op[stage].aarg2 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 387             op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) {
 388
 389             if(stage > 0 &&
 390                (op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
 391                 op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
 392                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_REG_%d_ATI, GL_SWIZZLE_STR_ATI)\n",
 393                       stage, stage);
 394                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 395                            GL_REG_0_ATI + stage,
 396                            GL_SWIZZLE_STR_ATI));
 397             } else {
 398                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, %s)\n",
 399                     stage, stage, debug_swizzle(swizzle));
 400                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 401                                         GL_TEXTURE0_ARB + stage,
 402                                         swizzle));
 403             }
 404         }
 405     }
 406
 407     /* Pass 4: Generate the arithmetic instructions */
 408     for(stage = 0; stage < MAX_TEXTURES; stage++) {
 409         if(op[stage].cop == WINED3DTOP_DISABLE) {
 410             if(stage == 0) {
 411                 /* Handle complete texture disabling gracefully */
 412                 TRACE("glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE, GL_PRIMARY_COLOR, GL_NONE, GL_NONE)\n");
 413                 GL_EXTCALL(glColorFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 414                                                  GL_PRIMARY_COLOR, GL_NONE, GL_NONE));
 415                 TRACE("glAlphaFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_PRIMARY_COLOR, GL_NONE, GL_NONE)\n");
 416                 GL_EXTCALL(glAlphaFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE,
 417                                                  GL_PRIMARY_COLOR, GL_NONE, GL_NONE));
 418             }
 419             break;
 420         }
 421
 422         if(op[stage].dst == WINED3DTA_TEMP) {
 423             /* If we're writing to D3DTA_TEMP, but never reading from it we don't have to write there in the first place.
 424              * skip the entire stage, this saves some GPU time
 425              */
 426             if(tmparg == GL_NONE) continue;
 427
 428             dstreg = tmparg;
 429         } else {
 430             dstreg = GL_REG_0_ATI;
 431         }
 432
 433         arg0 = register_for_arg(op[stage].carg0, gl_info, stage, &argmod0, tmparg);
 434         arg1 = register_for_arg(op[stage].carg1, gl_info, stage, &argmod1, tmparg);
 435         arg2 = register_for_arg(op[stage].carg2, gl_info, stage, &argmod2, tmparg);
 436         dstmod = GL_NONE;
 437         argmodextra = GL_NONE;
 438         extrarg = GL_NONE;
 439
 440         switch(op[stage].cop) {
 441             case WINED3DTOP_SELECTARG2:
 442                 arg1 = arg2;
 443                 argmod1 = argmod2;
 444             case WINED3DTOP_SELECTARG1:
 445                 TRACE("glColorFragmentOp1ATI(GL_MOV_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s)\n",
 446                       debug_register(dstreg), debug_register(arg1), debug_argmod(argmod1));
 447                 GL_EXTCALL(glColorFragmentOp1ATI(GL_MOV_ATI, dstreg, GL_NONE, GL_NONE,
 448                                                  arg1, GL_NONE, argmod1));
 449                 break;
 450
 451             case WINED3DTOP_MODULATE4X:
 452                 if(dstmod == GL_NONE) dstmod = GL_4X_BIT_ATI;
 453             case WINED3DTOP_MODULATE2X:
 454                 if(dstmod == GL_NONE) dstmod = GL_2X_BIT_ATI;
 455             case WINED3DTOP_MODULATE:
 456                 TRACE("glColorFragmentOp2ATI(GL_MUL_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 457                       debug_register(dstreg), debug_dstmod(dstmod),
 458                       debug_register(arg1), debug_argmod(argmod1),
 459                       debug_register(arg2), debug_argmod(argmod2));
 460                 GL_EXTCALL(glColorFragmentOp2ATI(GL_MUL_ATI, dstreg, GL_NONE, dstmod,
 461                                                  arg1, GL_NONE, argmod1,
 462                                                  arg2, GL_NONE, argmod2));
 463                 break;
 464
 465             case WINED3DTOP_ADDSIGNED2X:
 466                 dstmod = GL_2X_BIT_ATI;
 467             case WINED3DTOP_ADDSIGNED:
 468                 argmodextra = GL_BIAS_BIT_ATI;
 469             case WINED3DTOP_ADD:
 470                 TRACE("glColorFragmentOp2ATI(GL_ADD_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 471                       debug_register(dstreg), debug_dstmod(dstmod),
 472                       debug_register(arg1), debug_argmod(argmod1),
 473                       debug_register(arg2), debug_argmod(argmodextra | argmod2));
 474                 GL_EXTCALL(glColorFragmentOp2ATI(GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, dstmod,
 475                                                  arg1, GL_NONE, argmod1,
 476                                                  arg2, GL_NONE, argmodextra | argmod2));
 477                 break;
 478
 479             case WINED3DTOP_SUBTRACT:
 480                 TRACE("glColorFragmentOp2ATI(GL_SUB_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 481                       debug_register(dstreg), debug_dstmod(dstmod),
 482                       debug_register(arg1), debug_argmod(argmod1),
 483                       debug_register(arg2), debug_argmod(argmod2));
 484                 GL_EXTCALL(glColorFragmentOp2ATI(GL_SUB_ATI, dstreg, GL_NONE, dstmod,
 485                                                  arg1, GL_NONE, argmod1,
 486                                                  arg2, GL_NONE, argmod2));
 487                 break;
 488
 489             case WINED3DTOP_ADDSMOOTH:
 490                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 491                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 492                       debug_register(dstreg),
 493                       debug_register(arg2), debug_argmod(argmod2),
 494                       debug_register(arg1), debug_argmod(argmodextra),
 495                       debug_register(arg1), debug_argmod(argmod1));
 496                 /* Dst = arg1 + * arg2(1 -arg 1)
 497                  *     = arg2 * (1 - arg1) + arg1
 498                  */
 499                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 500                                                  arg2, GL_NONE, argmod2,
 501                                                  arg1, GL_NONE, argmodextra,
 502                                                  arg1, GL_NONE, argmod1));
 503                 break;
 504
 505             case WINED3DTOP_BLENDCURRENTALPHA:
 506                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, -1);
 507             case WINED3DTOP_BLENDFACTORALPHA:
 508                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, -1);
 509             case WINED3DTOP_BLENDTEXTUREALPHA:
 510                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, -1);
 511             case WINED3DTOP_BLENDDIFFUSEALPHA:
 512                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, -1);
 513                 TRACE("glColorFragmentOp3ATI(GL_LERP_ATI, %s, GL_NONE, GL_NONE, %s, GL_ALPHA, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 514                       debug_register(dstreg),
 515                       debug_register(extrarg),
 516                       debug_register(arg1), debug_argmod(argmod1),
 517                       debug_register(arg2), debug_argmod(argmod2));
 518                 GL_EXTCALL(glColorFragmentOp3ATI(GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 519                                                  extrarg, GL_ALPHA, GL_NONE,
 520                                                  arg1, GL_NONE, argmod1,
 521                                                  arg2, GL_NONE, argmod2));
 522                 break;
 523
 524             case WINED3DTOP_BLENDTEXTUREALPHAPM:
 525                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, -1);
 526                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s, %s, GL_ALPHA, GL_COMP_BIT_ATI, %s, GL_NONE, %s)\n",
 527                       debug_register(dstreg),
 528                       debug_register(arg2), debug_argmod(argmod2),
 529                       debug_register(arg0),
 530                       debug_register(arg1), debug_argmod(argmod1));
 531                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 532                                                  arg2, GL_NONE,  argmod2,
 533                                                  arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 534                                                  arg1, GL_NONE,  argmod1));
 535                 break;
 536
 537             /* D3DTOP_PREMODULATE ???? */
 538
 539             case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
 540                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 541             case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
 542                 if(!argmodextra) argmodextra = argmod1;
 543                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s, %s, GL_ALPHA, %s, %s, GL_NONE, %s)\n",
 544                       debug_register(dstreg),
 545                       debug_register(arg2), debug_argmod(argmod2),
 546                       debug_register(arg1), debug_argmod(argmodextra), debug_register(arg1), debug_argmod(arg1));
 547                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 548                                                  arg2, GL_NONE,  argmod2,
 549                                                  arg1, GL_ALPHA, argmodextra,
 550                                                  arg1, GL_NONE,  argmod1));
 551                 break;
 552
 553             case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
 554                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 555             case WINED3DTOP_MODULATECOLOR_ADDALPHA:
 556                 if(!argmodextra) argmodextra = argmod1;
 557                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_ALPHA, %s)\n",
 558                       debug_register(dstreg),
 559                       debug_register(arg2), debug_argmod(argmod2),
 560                       debug_register(arg1), debug_argmod(argmodextra),
 561                       debug_register(arg1), debug_argmod(argmod1));
 562                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 563                                                  arg2, GL_NONE,  argmod2,
 564                                                  arg1, GL_NONE,  argmodextra,
 565                                                  arg1, GL_ALPHA, argmod1));
 566                 break;
 567
 568             case WINED3DTOP_DOTPRODUCT3:
 569                 TRACE("glColorFragmentOp2ATI(GL_DOT3_ATI, %s, GL_NONE, GL_4X_BIT_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 570                       debug_register(dstreg),
 571                       debug_register(arg1), debug_argmod(argmod1 | GL_BIAS_BIT_ATI),
 572                       debug_register(arg2), debug_argmod(argmod2 | GL_BIAS_BIT_ATI));
 573                 GL_EXTCALL(glColorFragmentOp2ATI(GL_DOT3_ATI, dstreg, GL_NONE, GL_4X_BIT_ATI,
 574                                                  arg1, GL_NONE, argmod1 | GL_BIAS_BIT_ATI,
 575                                                  arg2, GL_NONE, argmod2 | GL_BIAS_BIT_ATI));
 576                 break;
 577
 578             case WINED3DTOP_MULTIPLYADD:
 579                 TRACE("glColorFragmentOp3ATI(GL_MAD_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 580                       debug_register(dstreg),
 581                       debug_register(arg0), debug_argmod(argmod0),
 582                       debug_register(arg2), debug_argmod(argmod2),
 583                       debug_register(arg1), debug_argmod(argmod1));
 584                 GL_EXTCALL(glColorFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 585                                                  arg0, GL_NONE, argmod0,
 586                                                  arg2, GL_NONE, argmod2,
 587                                                  arg1, GL_NONE, argmod1));
 588                 break;
 589
 590             case WINED3DTOP_LERP:
 591                 TRACE("glColorFragmentOp3ATI(GL_LERP_ATI, %s, GL_NONE, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 592                       debug_register(dstreg),
 593                       debug_register(arg1), debug_argmod(argmod1),
 594                       debug_register(arg2), debug_argmod(argmod2),
 595                       debug_register(arg0), debug_argmod(argmod0));
 596                 GL_EXTCALL(glColorFragmentOp3ATI(GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 597                                                  arg1, GL_NONE, argmod1,
 598                                                  arg2, GL_NONE, argmod2,
 599                                                  arg0, GL_NONE, argmod0));
 600                 break;
 601
 602             case WINED3DTOP_BUMPENVMAP:
 603             case WINED3DTOP_BUMPENVMAPLUMINANCE:
 604                 /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
 605                 break;
 606
 607             default: FIXME("Unhandled color operation %d on stage %d\n", op[stage].cop, stage);
 608         }
 609
 610         arg0 = register_for_arg(op[stage].aarg0, gl_info, stage, &argmod0, tmparg);
 611         arg1 = register_for_arg(op[stage].aarg1, gl_info, stage, &argmod1, tmparg);
 612         arg2 = register_for_arg(op[stage].aarg2, gl_info, stage, &argmod2, tmparg);
 613         dstmod = GL_NONE;
 614         argmodextra = GL_NONE;
 615         extrarg = GL_NONE;
 616
 617         switch(op[stage].aop) {
 618             case WINED3DTOP_DISABLE:
 619                 /* Get the primary color to the output if on stage 0, otherwise leave register 0 untouched */
 620                 if(stage == 0) {
 621                     TRACE("glAlphaFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_PRIMARY_COLOR, GL_NONE, GL_NONE)\n");
 622                     GL_EXTCALL(glAlphaFragmentOp1ATI(GL_MOV_ATI, GL_REG_0_ATI, GL_NONE,
 623                                GL_PRIMARY_COLOR, GL_NONE, GL_NONE));
 624                 }
 625                 break;
 626
 627             case WINED3DTOP_SELECTARG2:
 628                 arg1 = arg2;
 629                 argmod1 = argmod2;
 630             case WINED3DTOP_SELECTARG1:
 631                 TRACE("glAlphaFragmentOp1ATI(GL_MOV_ATI, %s,          GL_NONE, %s, GL_NONE, %s)\n",
 632                       debug_register(dstreg),
 633                       debug_register(arg1), debug_argmod(argmod1));
 634                 GL_EXTCALL(glAlphaFragmentOp1ATI(GL_MOV_ATI, dstreg, GL_NONE,
 635                                                  arg1, GL_NONE, argmod1));
 636                 break;
 637
 638             case WINED3DTOP_MODULATE4X:
 639                 if(dstmod == GL_NONE) dstmod = GL_4X_BIT_ATI;
 640             case WINED3DTOP_MODULATE2X:
 641                 if(dstmod == GL_NONE) dstmod = GL_2X_BIT_ATI;
 642             case WINED3DTOP_MODULATE:
 643                 TRACE("glAlphaFragmentOp2ATI(GL_MUL_ATI, %s,          %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 644                       debug_register(dstreg), debug_dstmod(dstmod),
 645                       debug_register(arg1), debug_argmod(argmod1),
 646                       debug_register(arg2), debug_argmod(argmod2));
 647                 GL_EXTCALL(glAlphaFragmentOp2ATI(GL_MUL_ATI, dstreg, dstmod,
 648                                                  arg1, GL_NONE, argmod1,
 649                                                  arg2, GL_NONE, argmod2));
 650                 break;
 651
 652             case WINED3DTOP_ADDSIGNED2X:
 653                 dstmod = GL_2X_BIT_ATI;
 654             case WINED3DTOP_ADDSIGNED:
 655                 argmodextra = GL_BIAS_BIT_ATI;
 656             case WINED3DTOP_ADD:
 657                 TRACE("glAlphaFragmentOp2ATI(GL_ADD_ATI, %s,          %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 658                       debug_register(dstreg), debug_dstmod(dstmod),
 659                       debug_register(arg1), debug_argmod(argmod1),
 660                       debug_register(arg2), debug_argmod(argmodextra | argmod2));
 661                 GL_EXTCALL(glAlphaFragmentOp2ATI(GL_ADD_ATI, dstreg, dstmod,
 662                                                  arg1, GL_NONE, argmod1,
 663                                                  arg2, GL_NONE, argmodextra | argmod2));
 664                 break;
 665
 666             case WINED3DTOP_SUBTRACT:
 667                 TRACE("glAlphaFragmentOp2ATI(GL_SUB_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 668                       debug_register(dstreg), debug_dstmod(dstmod),
 669                       debug_register(arg1), debug_argmod(argmod1),
 670                       debug_register(arg2), debug_argmod(argmod2));
 671                 GL_EXTCALL(glAlphaFragmentOp2ATI(GL_SUB_ATI, dstreg, dstmod,
 672                                                  arg1, GL_NONE, argmod1,
 673                                                  arg2, GL_NONE, argmod2));
 674                 break;
 675
 676             case WINED3DTOP_ADDSMOOTH:
 677                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 678                 TRACE("glAlphaFragmentOp3ATI(GL_MAD_ATI, %s,          GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 679                       debug_register(dstreg),
 680                       debug_register(arg2), debug_argmod(argmod2),
 681                       debug_register(arg1), debug_argmod(argmodextra),
 682                       debug_register(arg1), debug_argmod(argmod1));
 683                 /* Dst = arg1 + * arg2(1 -arg 1)
 684                  *     = arg2 * (1 - arg1) + arg1
 685                  */
 686                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE,
 687                                                  arg2, GL_NONE, argmod2,
 688                                                  arg1, GL_NONE, argmodextra,
 689                                                  arg1, GL_NONE, argmod1));
 690                 break;
 691
 692             case WINED3DTOP_BLENDCURRENTALPHA:
 693                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, -1);
 694             case WINED3DTOP_BLENDFACTORALPHA:
 695                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, -1);
 696             case WINED3DTOP_BLENDTEXTUREALPHA:
 697                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, -1);
 698             case WINED3DTOP_BLENDDIFFUSEALPHA:
 699                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, -1);
 700                 TRACE("glAlphaFragmentOp3ATI(GL_LERP_ATI, %s,          GL_NONE, %s, GL_ALPHA, GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 701                       debug_register(dstreg),
 702                       debug_register(extrarg),
 703                       debug_register(arg1), debug_argmod(argmod1),
 704                       debug_register(arg2), debug_argmod(argmod2));
 705                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_LERP_ATI, dstreg, GL_NONE,
 706                                                  extrarg, GL_ALPHA, GL_NONE,
 707                                                  arg1, GL_NONE, argmod1,
 708                                                  arg2, GL_NONE, argmod2));
 709                 break;
 710
 711             case WINED3DTOP_BLENDTEXTUREALPHAPM:
 712                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, -1);
 713                 TRACE("glAlphaFragmentOp3ATI(GL_MAD_ATI, %s,          GL_NONE, %s, GL_NONE, %s, %s, GL_ALPHA, GL_COMP_BIT_ATI, %s, GL_NONE, %s)\n",
 714                       debug_register(dstreg),
 715                       debug_register(arg2), debug_argmod(argmod2),
 716                       debug_register(arg0),
 717                       debug_register(arg1), debug_argmod(argmod1));
 718                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_MAD_ATI, dstreg, GL_NONE,
 719                                                  arg2, GL_NONE,  argmod2,
 720                                                  arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 721                                                  arg1, GL_NONE,  argmod1));
 722                 break;
 723
 724             /* D3DTOP_PREMODULATE ???? */
 725
 726             case WINED3DTOP_DOTPRODUCT3:
 727                 TRACE("glAlphaFragmentOp2ATI(GL_DOT3_ATI, %s, GL_NONE, GL_4X_BIT_ATI, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 728                       debug_register(dstreg),
 729                       debug_register(arg1), debug_argmod(argmod1 | GL_BIAS_BIT_ATI),
 730                       debug_register(arg2), debug_argmod(argmod2 | GL_BIAS_BIT_ATI));
 731                 GL_EXTCALL(glAlphaFragmentOp2ATI(GL_DOT3_ATI, dstreg, GL_4X_BIT_ATI,
 732                                                  arg1, GL_NONE, argmod1 | GL_BIAS_BIT_ATI,
 733                                                  arg2, GL_NONE, argmod2 | GL_BIAS_BIT_ATI));
 734                 break;
 735
 736             case WINED3DTOP_MULTIPLYADD:
 737                 TRACE("glAlphaFragmentOp3ATI(GL_MAD_ATI, %s,          GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 738                       debug_register(dstreg),
 739                       debug_register(arg0), debug_argmod(argmod0),
 740                       debug_register(arg2), debug_argmod(argmod2),
 741                       debug_register(arg1), debug_argmod(argmod1));
 742                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_MAD_ATI, dstreg,          GL_NONE,
 743                            arg0, GL_NONE, argmod0,
 744                            arg2, GL_NONE, argmod2,
 745                            arg1, GL_NONE, argmod1));
 746                 break;
 747
 748             case WINED3DTOP_LERP:
 749                 TRACE("glAlphaFragmentOp3ATI(GL_LERP_ATI, %s,          GL_NONE, %s, GL_NONE, %s, %s, GL_NONE, %s, %s, GL_NONE, %s)\n",
 750                       debug_register(dstreg),
 751                       debug_register(arg1), debug_argmod(argmod1),
 752                       debug_register(arg2), debug_argmod(argmod2),
 753                       debug_register(arg0), debug_argmod(argmod0));
 754                 GL_EXTCALL(glAlphaFragmentOp3ATI(GL_LERP_ATI, dstreg, GL_NONE,
 755                                                  arg1, GL_NONE, argmod1,
 756                                                  arg2, GL_NONE, argmod2,
 757                                                  arg0, GL_NONE, argmod0));
 758                 break;
 759
 760             case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
 761             case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
 762             case WINED3DTOP_MODULATECOLOR_ADDALPHA:
 763             case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
 764             case WINED3DTOP_BUMPENVMAP:
 765             case WINED3DTOP_BUMPENVMAPLUMINANCE:
 766                 ERR("Application uses an invalid alpha operation\n");
 767                 break;
 768
 769             default: FIXME("Unhandled alpha operation %d on stage %d\n", op[stage].aop, stage);
 770         }
 771     }
 772
 773     TRACE("glEndFragmentShaderATI()\n");
 774     GL_EXTCALL(glEndFragmentShaderATI());
 775     checkGLcall("GL_EXTCALL(glEndFragmentShaderATI())");
 776     return ret;
 777 }
 778 #undef GLINFO_LOCATION
 779
 780 #define GLINFO_LOCATION stateblock->wineD3DDevice->adapter->gl_info
 781 static void set_tex_op_atifs(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
 782     IWineD3DDeviceImpl          *This = stateblock->wineD3DDevice;
 783     struct atifs_ffp_desc       *desc;
 784     struct texture_stage_op     op[MAX_TEXTURES];
 785     struct atifs_private_data   *priv = (struct atifs_private_data *) This->shader_priv;
 786     DWORD mapped_stage;
 787     unsigned int i;
 788
 789     gen_ffp_op(stateblock, op);
 790     desc = (struct atifs_ffp_desc *) find_ffp_shader(&priv->fragment_shaders, op);
 791     if(!desc) {
 792         desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*desc));
 793         if(!desc) {
 794             ERR("Out of memory\n");
 795             return;
 796         }
 797         desc->num_textures_used = 0;
 798         for(i = 0; i < GL_LIMITS(texture_stages); i++) {
 799             if(op[i].cop == WINED3DTOP_DISABLE) break;
 800             desc->num_textures_used = i;
 801         }
 802
 803         memcpy(desc->parent.op, op, sizeof(op));
 804         desc->shader = gen_ati_shader(op, &GLINFO_LOCATION);
 805         add_ffp_shader(&priv->fragment_shaders, &desc->parent);
 806         TRACE("Allocated fixed function replacement shader descriptor %p\n", desc);
 807     }
 808
 809     /* GL_ATI_fragment_shader depends on the GL_TEXTURE_xD enable settings. Update the texture stages
 810      * used by this shader
 811      */
 812     for(i = 0; i < desc->num_textures_used; i++) {
 813         mapped_stage = This->texUnitMap[i];
 814         if(mapped_stage != -1) {
 815             GL_EXTCALL(glActiveTextureARB(GL_TEXTURE0_ARB + mapped_stage));
 816             checkGLcall("glActiveTextureARB");
 817             texture_activate_dimensions(i, stateblock, context);
 818         }
 819     }
 820
 821     GL_EXTCALL(glBindFragmentShaderATI(desc->shader));
 822 }
 823
 824 static void state_texfactor_atifs(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
 825     float col[4];
 826     D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
 827
 828     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col));
 829     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col)");
 830 }
 831
 832 static void set_bumpmat(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
 833     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / WINED3D_HIGHEST_TEXTURE_STATE;
 834     float mat[2][2];
 835
 836     mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
 837     mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
 838     mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
 839     mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
 840     /* GL_ATI_fragment_shader allows only constants from 0.0 to 1.0, but the bumpmat
 841      * constants can be in any range. While they should stay between [-1.0 and 1.0] because
 842      * Shader Model 1.x pixel shaders are clamped to that range negative values are used occasionally,
 843      * for example by our d3d9 test. So to get negative values scale -1;1 to 0;1 and undo that in the
 844      * shader(it is free). This might potentially reduce precision. However, if the hardware does
 845      * support proper floats it shouldn't, and if it doesn't we can't get anything better anyway
 846      */
 847     mat[0][0] = (mat[0][0] + 1.0) * 0.5;
 848     mat[1][0] = (mat[1][0] + 1.0) * 0.5;
 849     mat[0][1] = (mat[0][1] + 1.0) * 0.5;
 850     mat[1][1] = (mat[1][1] + 1.0) * 0.5;
 851     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), (float *) mat));
 852     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), mat)");
 853
 854     /* FIXME: This should go away
 855      * This is currently needed because atifs borrows a pixel shader implementation
 856      * from somewhere else, but consumes bump map matrix change events. The other pixel
 857      * shader implementation may need notification about the change to update the texbem
 858      * constants. Once ATIFS supports real shaders on its own, and GLSL/ARB have a replacement
 859      * pipeline this call can go away
 860      *
 861      * FIXME2: Even considering this workaround calling FFPStateTable directly isn't nice
 862      * as well. Better would be to call the model's table we inherit from, but currently
 863      * it is always the FFP table, and as soon as this changes we can remove the call anyway
 864      */
 865     FFPStateTable[state].apply(state, stateblock, context);
 866 }
 867 #undef GLINFO_LOCATION
 868
 869 /* our state table. Borrows lots of stuff from the base implementation */
 870 struct StateEntry ATIFSStateTable[STATE_HIGHEST + 1];
 871
 872 static void init_state_table() {
 873     unsigned int i;
 874     const DWORD rep = STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP);
 875     memcpy(ATIFSStateTable, arb_program_shader_backend.StateTable, sizeof(ATIFSStateTable));
 876
 877     for(i = 0; i < MAX_TEXTURES; i++) {
 878         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLOROP)].apply = set_tex_op_atifs;
 879         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLOROP)].representative = rep;
 880         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLORARG1)].apply = set_tex_op_atifs;
 881         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLORARG1)].representative = rep;
 882         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLORARG2)].apply = set_tex_op_atifs;
 883         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLORARG2)].representative = rep;
 884         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLORARG0)].apply = set_tex_op_atifs;
 885         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_COLORARG0)].representative = rep;
 886
 887         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAOP)].apply = set_tex_op_atifs;
 888         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAOP)].representative = rep;
 889         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAARG1)].apply = set_tex_op_atifs;
 890         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAARG1)].representative = rep;
 891         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAARG2)].apply = set_tex_op_atifs;
 892         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAARG2)].representative = rep;
 893         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAARG0)].apply = set_tex_op_atifs;
 894         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_ALPHAARG0)].representative = rep;
 895
 896         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_RESULTARG)].apply = set_tex_op_atifs;
 897         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_RESULTARG)].representative = rep;
 898
 899         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00)].apply = set_bumpmat;
 900         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT01)].apply = set_bumpmat;
 901         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT10)].apply = set_bumpmat;
 902         ATIFSStateTable[STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT11)].apply = set_bumpmat;
 903     }
 904
 905     ATIFSStateTable[STATE_RENDER(WINED3DRS_TEXTUREFACTOR)].apply = state_texfactor_atifs;
 906     ATIFSStateTable[STATE_RENDER(WINED3DRS_TEXTUREFACTOR)].representative = STATE_RENDER(WINED3DRS_TEXTUREFACTOR);
 907 }
 908
 909 /* GL_ATI_fragment_shader backend.It borrows a lot from a the
 910  * ARB shader backend, currently the whole vertex processing
 911  * code. This code would also forward pixel shaders, but if
 912  * GL_ARB_fragment_program is supported, the atifs shader backend
 913  * is not used.
 914  */
 915 static void shader_atifs_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
 916     arb_program_shader_backend.shader_select(iface, usePS, useVS);
 917 }
 918
 919 static void shader_atifs_select_depth_blt(IWineD3DDevice *iface) {
 920     arb_program_shader_backend.shader_select_depth_blt(iface);
 921 }
 922
 923 static void shader_atifs_destroy_depth_blt(IWineD3DDevice *iface) {
 924     arb_program_shader_backend.shader_destroy_depth_blt(iface);
 925 }
 926
 927 static void shader_atifs_load_constants(IWineD3DDevice *iface, char usePS, char useVS) {
 928     arb_program_shader_backend.shader_load_constants(iface, usePS, useVS);
 929 }
 930
 931 static void shader_atifs_cleanup(IWineD3DDevice *iface) {
 932     arb_program_shader_backend.shader_cleanup(iface);
 933 }
 934
 935 static void shader_atifs_color_correction(SHADER_OPCODE_ARG* arg) {
 936     arb_program_shader_backend.shader_color_correction(arg);
 937 }
 938
 939 static void shader_atifs_destroy(IWineD3DBaseShader *iface) {
 940     arb_program_shader_backend.shader_destroy(iface);
 941 }
 942
 943 static HRESULT shader_atifs_alloc(IWineD3DDevice *iface) {
 944     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 945     HRESULT hr;
 946     struct atifs_private_data *priv;
 947     hr = arb_program_shader_backend.shader_alloc_private(iface);
 948     if(FAILED(hr)) return hr;
 949
 950     This->shader_priv = HeapReAlloc(GetProcessHeap(), 0, This->shader_priv,
 951                                     sizeof(struct atifs_private_data));
 952     priv = (struct atifs_private_data *) This->shader_priv;
 953     list_init(&priv->fragment_shaders);
 954     return WINED3D_OK;
 955 }
 956
 957 #define GLINFO_LOCATION This->adapter->gl_info
 958 static void shader_atifs_free(IWineD3DDevice *iface) {
 959     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 960     struct atifs_private_data *priv = (struct atifs_private_data *) This->shader_priv;
 961     struct ffp_desc *entry, *entry2;
 962     struct atifs_ffp_desc *entry_ati;
 963
 964     ENTER_GL();
 965     LIST_FOR_EACH_ENTRY_SAFE(entry, entry2, &priv->fragment_shaders, struct ffp_desc, entry) {
 966         entry_ati = (struct atifs_ffp_desc *) entry;
 967         GL_EXTCALL(glDeleteFragmentShaderATI(entry_ati->shader));
 968         checkGLcall("glDeleteFragmentShaderATI(entry->shader)");
 969         list_remove(&entry->entry);
 970         HeapFree(GetProcessHeap(), 0, entry);
 971     }
 972     LEAVE_GL();
 973
 974     /* Not actually needed, but revert what we've done before */
 975     This->shader_priv = HeapReAlloc(GetProcessHeap(), 0, This->shader_priv,
 976                                     sizeof(struct shader_arb_priv));
 977     arb_program_shader_backend.shader_free_private(iface);
 978 }
 979 #undef GLINFO_LOCATION
 980
 981 static BOOL shader_atifs_dirty_const(IWineD3DDevice *iface) {
 982     return arb_program_shader_backend.shader_dirtifyable_constants(iface);
 983 }
 984
 985 static void shader_atifs_load_init(void) {
 986     arb_program_shader_backend.shader_dll_load_init();
 987     init_state_table();
 988 }
 989
 990 static void shader_atifs_get_caps(WINED3DDEVTYPE devtype, WineD3D_GL_Info *gl_info, struct shader_caps *caps) {
 991     arb_program_shader_backend.shader_get_caps(devtype, gl_info, caps);
 992
 993     caps->TextureOpCaps =  WINED3DTEXOPCAPS_DISABLE                     |
 994                            WINED3DTEXOPCAPS_SELECTARG1                  |
 995                            WINED3DTEXOPCAPS_SELECTARG2                  |
 996                            WINED3DTEXOPCAPS_MODULATE4X                  |
 997                            WINED3DTEXOPCAPS_MODULATE2X                  |
 998                            WINED3DTEXOPCAPS_MODULATE                    |
 999                            WINED3DTEXOPCAPS_ADDSIGNED2X                 |
1000                            WINED3DTEXOPCAPS_ADDSIGNED                   |
1001                            WINED3DTEXOPCAPS_ADD                         |
1002                            WINED3DTEXOPCAPS_SUBTRACT                    |
1003                            WINED3DTEXOPCAPS_ADDSMOOTH                   |
1004                            WINED3DTEXOPCAPS_BLENDCURRENTALPHA           |
1005                            WINED3DTEXOPCAPS_BLENDFACTORALPHA            |
1006                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHA           |
1007                            WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA           |
1008                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM         |
1009                            WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR      |
1010                            WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA      |
1011                            WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA   |
1012                            WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR   |
1013                            WINED3DTEXOPCAPS_DOTPRODUCT3                 |
1014                            WINED3DTEXOPCAPS_MULTIPLYADD                 |
1015                            WINED3DTEXOPCAPS_LERP                        |
1016                            WINED3DTEXOPCAPS_BUMPENVMAP;
1017
1018     /* TODO: Implement WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE
1019     and WINED3DTEXOPCAPS_PREMODULATE */
1020
1021     /* GL_ATI_fragment_shader only supports up to 6 textures, which was the limit on r200 cards
1022      * which this extension is exclusively focused on(later cards have GL_ARB_fragment_program).
1023      * If the current card has more than 8 fixed function textures in OpenGL's regular fixed
1024      * function pipeline then the ATI_fragment_shader backend imposes a stricter limit. This
1025      * shouldn't be too hard since Nvidia cards have a limit of 4 textures with the default ffp
1026      * pipeline, and almost all games are happy with that. We can however support up to 8
1027      * texture stages because we have a 2nd pass limit of 8 instructions, and per stage we use
1028      * only 1 instruction.
1029      *
1030      * The proper fix for this is not to use GL_ATI_fragment_shader on cards newer than the
1031      * r200 series and use an ARB or GLSL shader instead
1032      */
1033     if(caps->MaxSimultaneousTextures > 6) {
1034         WARN("OpenGL fixed function supports %d simultaneous textures,\n", caps->MaxSimultaneousTextures);
1035         WARN("but GL_ATI_fragment_shader limits this to 6\n");
1036         caps->MaxSimultaneousTextures = 6;
1037     }
1038
1039     caps->PrimitiveMiscCaps |= WINED3DPMISCCAPS_TSSARGTEMP;
1040 }
1041
1042 static void shader_atifs_generate_pshader(IWineD3DPixelShader *iface, SHADER_BUFFER *buffer) {
1043     ERR("Should not get here\n");
1044 }
1045
1046 static void shader_atifs_generate_vshader(IWineD3DVertexShader *iface, SHADER_BUFFER *buffer) {
1047     arb_program_shader_backend.shader_generate_vshader(iface, buffer);
1048 }
1049
1050 static void shader_atifs_fragment_enable(IWineD3DDevice *iface, BOOL enable) {
1051     if(enable) {
1052         glEnable(GL_FRAGMENT_SHADER_ATI);
1053         checkGLcall("glEnable(GL_FRAGMENT_SHADER_ATI)");
1054     } else {
1055         glDisable(GL_FRAGMENT_SHADER_ATI);
1056         checkGLcall("glDisable(GL_FRAGMENT_SHADER_ATI)");
1057     }
1058 }
1059
1060 const shader_backend_t atifs_shader_backend = {
1061     shader_atifs_select,
1062     shader_atifs_select_depth_blt,
1063     shader_atifs_destroy_depth_blt,
1064     shader_atifs_load_constants,
1065     shader_atifs_cleanup,
1066     shader_atifs_color_correction,
1067     shader_atifs_destroy,
1068     shader_atifs_alloc,
1069     shader_atifs_free,
1070     shader_atifs_dirty_const,
1071     shader_atifs_generate_pshader,
1072     shader_atifs_generate_vshader,
1073     shader_atifs_get_caps,
1074     shader_atifs_load_init,
1075     shader_atifs_fragment_enable,
1076     ATIFSStateTable
1077 };