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

   1 /*
   2  * Pixel and vertex shaders implementation using ARB_vertex_program
   3  * and ARB_fragment_program GL extensions.
   4  *
   5  * Copyright 2002-2003 Jason Edmeades
   6  * Copyright 2002-2003 Raphael Junqueira
   7  * Copyright 2004 Christian Costa
   8  * Copyright 2005 Oliver Stieber
   9  * Copyright 2006 Ivan Gyurdiev
  10  * Copyright 2006 Jason Green
  11  * Copyright 2006 Henri Verbeet
  12  * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
  13  *
  14  * This library is free software; you can redistribute it and/or
  15  * modify it under the terms of the GNU Lesser General Public
  16  * License as published by the Free Software Foundation; either
  17  * version 2.1 of the License, or (at your option) any later version.
  18  *
  19  * This library is distributed in the hope that it will be useful,
  20  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  21  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22  * Lesser General Public License for more details.
  23  *
  24  * You should have received a copy of the GNU Lesser General Public
  25  * License along with this library; if not, write to the Free Software
  26  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  27  */
  28
  29 #include "config.h"
  30
  31 #include <math.h>
  32 #include <stdio.h>
  33
  34 #include "wined3d_private.h"
  35
  36 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  37 WINE_DECLARE_DEBUG_CHANNEL(d3d_constants);
  38 WINE_DECLARE_DEBUG_CHANNEL(d3d_caps);
  39 WINE_DECLARE_DEBUG_CHANNEL(d3d);
  40
  41 #define GLINFO_LOCATION      (*gl_info)
  42
  43 /* Internally used shader constants. Applications can use constants 0 to GL_LIMITS(vshader_constantsF) - 1,
  44  * so upload them above that
  45  */
  46 #define ARB_SHADER_PRIVCONST_BASE GL_LIMITS(vshader_constantsF)
  47 #define ARB_SHADER_PRIVCONST_POS ARB_SHADER_PRIVCONST_BASE + 0
  48
  49 /* ARB_program_shader private data */
  50 struct shader_arb_priv {
  51     GLuint                  current_vprogram_id;
  52     GLuint                  current_fprogram_id;
  53     GLuint                  depth_blt_vprogram_id;
  54     GLuint                  depth_blt_fprogram_id[tex_type_count];
  55     BOOL                    use_arbfp_fixed_func;
  56     struct hash_table_t     *fragment_shaders;
  57 };
  58
  59 /********************************************************
  60  * ARB_[vertex/fragment]_program helper functions follow
  61  ********************************************************/
  62
  63 /**
  64  * Loads floating point constants into the currently set ARB_vertex/fragment_program.
  65  * When constant_list == NULL, it will load all the constants.
  66  *
  67  * @target_type should be either GL_VERTEX_PROGRAM_ARB (for vertex shaders)
  68  *  or GL_FRAGMENT_PROGRAM_ARB (for pixel shaders)
  69  */
  70 static unsigned int shader_arb_load_constantsF(IWineD3DBaseShaderImpl* This, const WineD3D_GL_Info *gl_info,
  71         GLuint target_type, unsigned int max_constants, const float *constants, char *dirty_consts)
  72 {
  73     DWORD shader_version = This->baseShader.reg_maps.shader_version;
  74     local_constant* lconst;
  75     DWORD i, j;
  76     unsigned int ret;
  77
  78     if (TRACE_ON(d3d_shader)) {
  79         for(i = 0; i < max_constants; i++) {
  80             if(!dirty_consts[i]) continue;
  81             TRACE_(d3d_constants)("Loading constants %i: %f, %f, %f, %f\n", i,
  82                         constants[i * 4 + 0], constants[i * 4 + 1],
  83                         constants[i * 4 + 2], constants[i * 4 + 3]);
  84         }
  85     }
  86     /* In 1.X pixel shaders constants are implicitly clamped in the range [-1;1] */
  87     if (target_type == GL_FRAGMENT_PROGRAM_ARB && WINED3DSHADER_VERSION_MAJOR(shader_version) == 1)
  88     {
  89         float lcl_const[4];
  90         for(i = 0; i < max_constants; i++) {
  91             if(!dirty_consts[i]) continue;
  92             dirty_consts[i] = 0;
  93
  94             j = 4 * i;
  95             if(constants[j + 0] > 1.0) lcl_const[0] = 1.0;
  96             else if(constants[j + 0] < -1.0) lcl_const[0] = -1.0;
  97             else lcl_const[0] = constants[j + 0];
  98
  99             if(constants[j + 1] > 1.0) lcl_const[1] = 1.0;
 100             else if(constants[j + 1] < -1.0) lcl_const[1] = -1.0;
 101             else lcl_const[1] = constants[j + 1];
 102
 103             if(constants[j + 2] > 1.0) lcl_const[2] = 1.0;
 104             else if(constants[j + 2] < -1.0) lcl_const[2] = -1.0;
 105             else lcl_const[2] = constants[j + 2];
 106
 107             if(constants[j + 3] > 1.0) lcl_const[3] = 1.0;
 108             else if(constants[j + 3] < -1.0) lcl_const[3] = -1.0;
 109             else lcl_const[3] = constants[j + 3];
 110
 111             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, lcl_const));
 112         }
 113     } else {
 114         if(GL_SUPPORT(EXT_GPU_PROGRAM_PARAMETERS)) {
 115             /* TODO: Benchmark if we're better of with finding the dirty constants ourselves,
 116              * or just reloading *all* constants at once
 117              *
 118             GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, 0, max_constants, constants));
 119              */
 120             for(i = 0; i < max_constants; i++) {
 121                 if(!dirty_consts[i]) continue;
 122
 123                 /* Find the next block of dirty constants */
 124                 dirty_consts[i] = 0;
 125                 j = i;
 126                 for(i++; (i < max_constants) && dirty_consts[i]; i++) {
 127                     dirty_consts[i] = 0;
 128                 }
 129
 130                 GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, j, i - j, constants + (j * 4)));
 131             }
 132         } else {
 133             for(i = 0; i < max_constants; i++) {
 134                 if(dirty_consts[i]) {
 135                     dirty_consts[i] = 0;
 136                     GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, constants + (i * 4)));
 137                 }
 138             }
 139         }
 140     }
 141     checkGLcall("glProgramEnvParameter4fvARB()");
 142
 143     /* Load immediate constants */
 144     if(This->baseShader.load_local_constsF) {
 145         if (TRACE_ON(d3d_shader)) {
 146             LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 147                 GLfloat* values = (GLfloat*)lconst->value;
 148                 TRACE_(d3d_constants)("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
 149                         values[0], values[1], values[2], values[3]);
 150             }
 151         }
 152         /* Immediate constants are clamped for 1.X shaders at loading times */
 153         ret = 0;
 154         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 155             dirty_consts[lconst->idx] = 1; /* Dirtify so the non-immediate constant overwrites it next time */
 156             ret = max(ret, lconst->idx + 1);
 157             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, lconst->idx, (GLfloat*)lconst->value));
 158         }
 159         checkGLcall("glProgramEnvParameter4fvARB()");
 160         return ret; /* The loaded immediate constants need reloading for the next shader */
 161     } else {
 162         return 0; /* No constants are dirty now */
 163     }
 164 }
 165
 166 /**
 167  * Loads the app-supplied constants into the currently set ARB_[vertex/fragment]_programs.
 168  *
 169  * We only support float constants in ARB at the moment, so don't
 170  * worry about the Integers or Booleans
 171  */
 172 static void shader_arb_load_constants(
 173     IWineD3DDevice* device,
 174     char usePixelShader,
 175     char useVertexShader) {
 176
 177     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
 178     IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
 179     const WineD3D_GL_Info *gl_info = &deviceImpl->adapter->gl_info;
 180     unsigned char i;
 181
 182     if (useVertexShader) {
 183         IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
 184
 185         /* Load DirectX 9 float constants for vertex shader */
 186         deviceImpl->highest_dirty_vs_const = shader_arb_load_constantsF(
 187                 vshader, gl_info, GL_VERTEX_PROGRAM_ARB,
 188                 deviceImpl->highest_dirty_vs_const,
 189                 stateBlock->vertexShaderConstantF,
 190                 deviceImpl->activeContext->vshader_const_dirty);
 191
 192         /* Upload the position fixup */
 193         GL_EXTCALL(glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, ARB_SHADER_PRIVCONST_POS, deviceImpl->posFixup));
 194     }
 195
 196     if (usePixelShader) {
 197
 198         IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
 199         IWineD3DPixelShaderImpl *psi = (IWineD3DPixelShaderImpl *) pshader;
 200
 201         /* Load DirectX 9 float constants for pixel shader */
 202         deviceImpl->highest_dirty_ps_const = shader_arb_load_constantsF(
 203                 pshader, gl_info, GL_FRAGMENT_PROGRAM_ARB,
 204                 deviceImpl->highest_dirty_ps_const,
 205                 stateBlock->pixelShaderConstantF,
 206                 deviceImpl->activeContext->pshader_const_dirty);
 207
 208         for(i = 0; i < psi->numbumpenvmatconsts; i++) {
 209             /* The state manager takes care that this function is always called if the bump env matrix changes
 210              */
 211             const float *data = (const float *)&stateBlock->textureState[(int) psi->bumpenvmatconst[i].texunit][WINED3DTSS_BUMPENVMAT00];
 212             GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->bumpenvmatconst[i].const_num, data));
 213             deviceImpl->activeContext->pshader_const_dirty[psi->bumpenvmatconst[i].const_num] = 1;
 214
 215             if(psi->luminanceconst[i].const_num != -1) {
 216                 /* WINED3DTSS_BUMPENVLSCALE and WINED3DTSS_BUMPENVLOFFSET are next to each other.
 217                  * point gl to the scale, and load 4 floats. x = scale, y = offset, z and w are junk, we
 218                  * don't care about them. The pointers are valid for sure because the stateblock is bigger.
 219                  * (they're WINED3DTSS_TEXTURETRANSFORMFLAGS and WINED3DTSS_ADDRESSW, so most likely 0 or NaN
 220                  */
 221                 const float *scale = (const float *)&stateBlock->textureState[(int) psi->luminanceconst[i].texunit][WINED3DTSS_BUMPENVLSCALE];
 222                 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->luminanceconst[i].const_num, scale));
 223                 deviceImpl->activeContext->pshader_const_dirty[psi->luminanceconst[i].const_num] = 1;
 224             }
 225         }
 226     }
 227 }
 228
 229 static void shader_arb_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count)
 230 {
 231     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 232
 233     /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
 234      * context. On a context switch the old context will be fully dirtified */
 235     memset(This->activeContext->vshader_const_dirty + start, 1,
 236             sizeof(*This->activeContext->vshader_const_dirty) * count);
 237     This->highest_dirty_vs_const = max(This->highest_dirty_vs_const, start + count + 1);
 238 }
 239
 240 static void shader_arb_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count)
 241 {
 242     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 243
 244     /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
 245      * context. On a context switch the old context will be fully dirtified */
 246     memset(This->activeContext->pshader_const_dirty + start, 1,
 247             sizeof(*This->activeContext->pshader_const_dirty) * count);
 248     This->highest_dirty_ps_const = max(This->highest_dirty_ps_const, start + count + 1);
 249 }
 250
 251 /* Generate the variable & register declarations for the ARB_vertex_program output target */
 252 static void shader_generate_arb_declarations(IWineD3DBaseShader *iface, const shader_reg_maps *reg_maps,
 253         SHADER_BUFFER *buffer, const WineD3D_GL_Info *gl_info)
 254 {
 255     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 256     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
 257     DWORD i, cur;
 258     char pshader = shader_is_pshader_version(reg_maps->shader_version);
 259     unsigned max_constantsF = min(This->baseShader.limits.constant_float,
 260             (pshader ? GL_LIMITS(pshader_constantsF) : GL_LIMITS(vshader_constantsF)));
 261     UINT extra_constants_needed = 0;
 262     const local_constant *lconst;
 263
 264     /* Temporary Output register */
 265     shader_addline(buffer, "TEMP TMP_OUT;\n");
 266
 267     for(i = 0; i < This->baseShader.limits.temporary; i++) {
 268         if (reg_maps->temporary[i])
 269             shader_addline(buffer, "TEMP R%u;\n", i);
 270     }
 271
 272     for (i = 0; i < This->baseShader.limits.address; i++) {
 273         if (reg_maps->address[i])
 274             shader_addline(buffer, "ADDRESS A%d;\n", i);
 275     }
 276
 277     for(i = 0; i < This->baseShader.limits.texcoord; i++) {
 278         if (reg_maps->texcoord[i])
 279             shader_addline(buffer,"TEMP T%u;\n", i);
 280     }
 281
 282     /* Texture coordinate registers must be pre-loaded */
 283     for (i = 0; i < This->baseShader.limits.texcoord; i++) {
 284         if (reg_maps->texcoord[i])
 285             shader_addline(buffer, "MOV T%u, fragment.texcoord[%u];\n", i, i);
 286     }
 287
 288     for(i = 0; i < (sizeof(reg_maps->bumpmat) / sizeof(reg_maps->bumpmat[0])); i++) {
 289         IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *) This;
 290         if(!reg_maps->bumpmat[i]) continue;
 291
 292         cur = ps->numbumpenvmatconsts;
 293         ps->bumpenvmatconst[cur].const_num = -1;
 294         ps->bumpenvmatconst[cur].texunit = i;
 295         ps->luminanceconst[cur].const_num = -1;
 296         ps->luminanceconst[cur].texunit = i;
 297
 298         /* If the shader does not use all available constants, use the next free constant to load the bump mapping environment matrix from
 299          * the stateblock into the shader. If no constant is available don't load, texbem will then just sample the texture without applying
 300          * bump mapping.
 301          */
 302         if(max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF)) {
 303             ps->bumpenvmatconst[cur].const_num = max_constantsF + extra_constants_needed;
 304             shader_addline(buffer, "PARAM bumpenvmat%d = program.env[%d];\n",
 305                            i, ps->bumpenvmatconst[cur].const_num);
 306             extra_constants_needed++;
 307
 308             if(reg_maps->luminanceparams && max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF)) {
 309                 ((IWineD3DPixelShaderImpl *)This)->luminanceconst[cur].const_num = max_constantsF + extra_constants_needed;
 310                 shader_addline(buffer, "PARAM luminance%d = program.env[%d];\n",
 311                                i, ps->luminanceconst[cur].const_num);
 312                 extra_constants_needed++;
 313             } else if(reg_maps->luminanceparams) {
 314                 FIXME("No free constant to load the luminance parameters\n");
 315             }
 316         } else {
 317             FIXME("No free constant found to load environment bump mapping matrix into the shader. texbem instruction will not apply bump mapping\n");
 318         }
 319
 320         ps->numbumpenvmatconsts = cur + 1;
 321     }
 322
 323     if(device->stateBlock->renderState[WINED3DRS_SRGBWRITEENABLE] && pshader) {
 324         shader_addline(buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n",
 325                         srgb_mul_low, srgb_mul_low, srgb_mul_low);
 326         shader_addline(buffer, "PARAM srgb_comparison =  {%f, %f, %f, %f};\n",
 327                         srgb_cmp, srgb_cmp, srgb_cmp, srgb_cmp);
 328         shader_addline(buffer, "PARAM srgb_pow =  {%f, %f, %f, 1.0};\n",
 329                        srgb_pow, srgb_pow, srgb_pow);
 330         shader_addline(buffer, "PARAM srgb_mul_hi =  {%f, %f, %f, 1.0};\n",
 331                        srgb_mul_high, srgb_mul_high, srgb_mul_high);
 332         shader_addline(buffer, "PARAM srgb_sub_hi =  {%f, %f, %f, 0.0};\n",
 333                        srgb_sub_high, srgb_sub_high, srgb_sub_high);
 334     }
 335
 336     /* Load local constants using the program-local space,
 337      * this avoids reloading them each time the shader is used
 338      */
 339     if(!This->baseShader.load_local_constsF) {
 340         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 341             shader_addline(buffer, "PARAM C%u = program.local[%u];\n", lconst->idx,
 342                            lconst->idx);
 343         }
 344     }
 345
 346     /* we use the array-based constants array if the local constants are marked for loading,
 347      * because then we use indirect addressing, or when the local constant list is empty,
 348      * because then we don't know if we're using indirect addressing or not. If we're hardcoding
 349      * local constants do not declare the loaded constants as an array because ARB compilers usually
 350      * do not optimize unused constants away
 351      */
 352     if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) {
 353         /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
 354         shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
 355                     max_constantsF, max_constantsF - 1);
 356     } else {
 357         for(i = 0; i < max_constantsF; i++) {
 358             if(!shader_constant_is_local(This, i)) {
 359                 shader_addline(buffer, "PARAM C%d = program.env[%d];\n",i, i);
 360             }
 361         }
 362     }
 363 }
 364
 365 static const char * const shift_tab[] = {
 366     "dummy",     /*  0 (none) */
 367     "coefmul.x", /*  1 (x2)   */
 368     "coefmul.y", /*  2 (x4)   */
 369     "coefmul.z", /*  3 (x8)   */
 370     "coefmul.w", /*  4 (x16)  */
 371     "dummy",     /*  5 (x32)  */
 372     "dummy",     /*  6 (x64)  */
 373     "dummy",     /*  7 (x128) */
 374     "dummy",     /*  8 (d256) */
 375     "dummy",     /*  9 (d128) */
 376     "dummy",     /* 10 (d64)  */
 377     "dummy",     /* 11 (d32)  */
 378     "coefdiv.w", /* 12 (d16)  */
 379     "coefdiv.z", /* 13 (d8)   */
 380     "coefdiv.y", /* 14 (d4)   */
 381     "coefdiv.x"  /* 15 (d2)   */
 382 };
 383
 384 static void shader_arb_get_write_mask(const SHADER_OPCODE_ARG *arg, const DWORD param, char *write_mask)
 385 {
 386     char *ptr = write_mask;
 387     char vshader = shader_is_vshader_version(arg->reg_maps->shader_version);
 388
 389     if(vshader && shader_get_regtype(param) == WINED3DSPR_ADDR) {
 390         *ptr++ = '.';
 391         *ptr++ = 'x';
 392     } else if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
 393         *ptr++ = '.';
 394         if (param & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
 395         if (param & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
 396         if (param & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
 397         if (param & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
 398     }
 399
 400     *ptr = '\0';
 401 }
 402
 403 static void shader_arb_get_swizzle(const DWORD param, BOOL fixup, char *swizzle_str) {
 404     /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
 405      * but addressed as "rgba". To fix this we need to swap the register's x
 406      * and z components. */
 407     const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
 408     char *ptr = swizzle_str;
 409
 410     /* swizzle bits fields: wwzzyyxx */
 411     DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
 412     DWORD swizzle_x = swizzle & 0x03;
 413     DWORD swizzle_y = (swizzle >> 2) & 0x03;
 414     DWORD swizzle_z = (swizzle >> 4) & 0x03;
 415     DWORD swizzle_w = (swizzle >> 6) & 0x03;
 416
 417     /* If the swizzle is the default swizzle (ie, "xyzw"), we don't need to
 418      * generate a swizzle string. Unless we need to our own swizzling. */
 419     if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) != swizzle || fixup) {
 420         *ptr++ = '.';
 421         if (swizzle_x == swizzle_y && swizzle_x == swizzle_z && swizzle_x == swizzle_w) {
 422             *ptr++ = swizzle_chars[swizzle_x];
 423         } else {
 424             *ptr++ = swizzle_chars[swizzle_x];
 425             *ptr++ = swizzle_chars[swizzle_y];
 426             *ptr++ = swizzle_chars[swizzle_z];
 427             *ptr++ = swizzle_chars[swizzle_w];
 428         }
 429     }
 430
 431     *ptr = '\0';
 432 }
 433
 434 static void pshader_get_register_name(IWineD3DBaseShader* iface,
 435     const DWORD param, char* regstr) {
 436
 437     DWORD reg = param & WINED3DSP_REGNUM_MASK;
 438     DWORD regtype = shader_get_regtype(param);
 439     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *) iface;
 440
 441     switch (regtype) {
 442     case WINED3DSPR_TEMP:
 443         sprintf(regstr, "R%u", reg);
 444     break;
 445     case WINED3DSPR_INPUT:
 446         if (reg==0) {
 447             strcpy(regstr, "fragment.color.primary");
 448         } else {
 449             strcpy(regstr, "fragment.color.secondary");
 450         }
 451     break;
 452     case WINED3DSPR_CONST:
 453         if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) {
 454             sprintf(regstr, "C[%u]", reg);
 455         } else {
 456             sprintf(regstr, "C%u", reg);
 457         }
 458     break;
 459     case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
 460         sprintf(regstr,"T%u", reg);
 461     break;
 462     case WINED3DSPR_COLOROUT:
 463         if (reg == 0)
 464             sprintf(regstr, "TMP_COLOR");
 465         else {
 466             /* TODO: See GL_ARB_draw_buffers */
 467             FIXME("Unsupported write to render target %u\n", reg);
 468             sprintf(regstr, "unsupported_register");
 469         }
 470     break;
 471     case WINED3DSPR_DEPTHOUT:
 472         sprintf(regstr, "result.depth");
 473     break;
 474     case WINED3DSPR_ATTROUT:
 475         sprintf(regstr, "oD[%u]", reg);
 476     break;
 477     case WINED3DSPR_TEXCRDOUT:
 478         sprintf(regstr, "oT[%u]", reg);
 479     break;
 480     default:
 481         FIXME("Unhandled register name Type(%d)\n", regtype);
 482         sprintf(regstr, "unrecognized_register");
 483     break;
 484     }
 485 }
 486
 487 /* TODO: merge with pixel shader */
 488 static void vshader_program_add_param(const SHADER_OPCODE_ARG *arg, const DWORD param, BOOL is_input, char *hwLine)
 489 {
 490   IWineD3DVertexShaderImpl* This = (IWineD3DVertexShaderImpl*) arg->shader;
 491
 492   /* oPos, oFog and oPts in D3D */
 493   static const char * const hwrastout_reg_names[] = { "TMP_OUT", "result.fogcoord", "result.pointsize" };
 494
 495   DWORD reg = param & WINED3DSP_REGNUM_MASK;
 496   DWORD regtype = shader_get_regtype(param);
 497   char  tmpReg[255];
 498   BOOL is_color = FALSE;
 499
 500   if ((param & WINED3DSP_SRCMOD_MASK) == WINED3DSPSM_NEG) {
 501       strcat(hwLine, " -");
 502   } else {
 503       strcat(hwLine, " ");
 504   }
 505
 506   switch (regtype) {
 507   case WINED3DSPR_TEMP:
 508     sprintf(tmpReg, "R%u", reg);
 509     strcat(hwLine, tmpReg);
 510     break;
 511   case WINED3DSPR_INPUT:
 512
 513     if (vshader_input_is_color((IWineD3DVertexShader*) This, reg))
 514         is_color = TRUE;
 515
 516     sprintf(tmpReg, "vertex.attrib[%u]", reg);
 517     strcat(hwLine, tmpReg);
 518     break;
 519   case WINED3DSPR_CONST:
 520       if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
 521           if(reg >= This->rel_offset) {
 522               sprintf(tmpReg, "C[A0.x + %u]", reg - This->rel_offset);
 523           } else {
 524               sprintf(tmpReg, "C[A0.x - %u]", -reg + This->rel_offset);
 525           }
 526       } else {
 527           if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) {
 528               sprintf(tmpReg, "C[%u]", reg);
 529           } else {
 530               sprintf(tmpReg, "C%u", reg);
 531           }
 532       }
 533     strcat(hwLine, tmpReg);
 534     break;
 535   case WINED3DSPR_ADDR: /*case D3DSPR_TEXTURE:*/
 536     sprintf(tmpReg, "A%u", reg);
 537     strcat(hwLine, tmpReg);
 538     break;
 539   case WINED3DSPR_RASTOUT:
 540     sprintf(tmpReg, "%s", hwrastout_reg_names[reg]);
 541     strcat(hwLine, tmpReg);
 542     break;
 543   case WINED3DSPR_ATTROUT:
 544     if (reg==0) {
 545        strcat(hwLine, "result.color.primary");
 546     } else {
 547        strcat(hwLine, "result.color.secondary");
 548     }
 549     break;
 550   case WINED3DSPR_TEXCRDOUT:
 551     sprintf(tmpReg, "result.texcoord[%u]", reg);
 552     strcat(hwLine, tmpReg);
 553     break;
 554   default:
 555     FIXME("Unknown reg type %d %d\n", regtype, reg);
 556     strcat(hwLine, "unrecognized_register");
 557     break;
 558   }
 559
 560   if (!is_input) {
 561     char write_mask[6];
 562     shader_arb_get_write_mask(arg, param, write_mask);
 563     strcat(hwLine, write_mask);
 564   } else {
 565     char swizzle[6];
 566     shader_arb_get_swizzle(param, is_color, swizzle);
 567     strcat(hwLine, swizzle);
 568   }
 569 }
 570
 571 static void shader_hw_sample(const SHADER_OPCODE_ARG *arg, DWORD sampler_idx, const char *dst_str,
 572         const char *coord_reg, BOOL projected, BOOL bias)
 573 {
 574     SHADER_BUFFER* buffer = arg->buffer;
 575     DWORD sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
 576     const char *tex_type;
 577
 578     switch(sampler_type) {
 579         case WINED3DSTT_1D:
 580             tex_type = "1D";
 581             break;
 582
 583         case WINED3DSTT_2D:
 584         {
 585             IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *) arg->shader;
 586             IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
 587             if(device->stateBlock->textures[sampler_idx] &&
 588                IWineD3DBaseTexture_GetTextureDimensions(device->stateBlock->textures[sampler_idx]) == GL_TEXTURE_RECTANGLE_ARB) {
 589                 tex_type = "RECT";
 590             } else {
 591                 tex_type = "2D";
 592             }
 593             break;
 594         }
 595
 596         case WINED3DSTT_VOLUME:
 597             tex_type = "3D";
 598             break;
 599
 600         case WINED3DSTT_CUBE:
 601             tex_type = "CUBE";
 602             break;
 603
 604         default:
 605             ERR("Unexpected texture type %d\n", sampler_type);
 606             tex_type = "";
 607     }
 608
 609     if (bias) {
 610         /* Shouldn't be possible, but let's check for it */
 611         if(projected) FIXME("Biased and Projected texture sampling\n");
 612         /* TXB takes the 4th component of the source vector automatically, as d3d. Nothing more to do */
 613         shader_addline(buffer, "TXB %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 614     } else if (projected) {
 615         shader_addline(buffer, "TXP %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 616     } else {
 617         shader_addline(buffer, "TEX %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 618     }
 619 }
 620
 621 static const char *shader_arb_get_fixup_swizzle(enum fixup_channel_source channel_source)
 622 {
 623     switch(channel_source)
 624     {
 625         case CHANNEL_SOURCE_ZERO: return "0";
 626         case CHANNEL_SOURCE_ONE: return "1";
 627         case CHANNEL_SOURCE_X: return "x";
 628         case CHANNEL_SOURCE_Y: return "y";
 629         case CHANNEL_SOURCE_Z: return "z";
 630         case CHANNEL_SOURCE_W: return "w";
 631         default:
 632             FIXME("Unhandled channel source %#x\n", channel_source);
 633             return "undefined";
 634     }
 635 }
 636
 637 static void gen_color_correction(SHADER_BUFFER *buffer, const char *reg, DWORD dst_mask,
 638         const char *one, const char *two, struct color_fixup_desc fixup)
 639 {
 640     DWORD mask;
 641
 642     if (is_yuv_fixup(fixup))
 643     {
 644         enum yuv_fixup yuv_fixup = get_yuv_fixup(fixup);
 645         FIXME("YUV fixup (%#x) not supported\n", yuv_fixup);
 646         return;
 647     }
 648
 649     mask = 0;
 650     if (fixup.x_source != CHANNEL_SOURCE_X) mask |= WINED3DSP_WRITEMASK_0;
 651     if (fixup.y_source != CHANNEL_SOURCE_Y) mask |= WINED3DSP_WRITEMASK_1;
 652     if (fixup.z_source != CHANNEL_SOURCE_Z) mask |= WINED3DSP_WRITEMASK_2;
 653     if (fixup.w_source != CHANNEL_SOURCE_W) mask |= WINED3DSP_WRITEMASK_3;
 654     mask &= dst_mask;
 655
 656     if (mask)
 657     {
 658         shader_addline(buffer, "SWZ %s, %s, %s, %s, %s, %s;\n", reg, reg,
 659                 shader_arb_get_fixup_swizzle(fixup.x_source), shader_arb_get_fixup_swizzle(fixup.y_source),
 660                 shader_arb_get_fixup_swizzle(fixup.z_source), shader_arb_get_fixup_swizzle(fixup.w_source));
 661     }
 662
 663     mask = 0;
 664     if (fixup.x_sign_fixup) mask |= WINED3DSP_WRITEMASK_0;
 665     if (fixup.y_sign_fixup) mask |= WINED3DSP_WRITEMASK_1;
 666     if (fixup.z_sign_fixup) mask |= WINED3DSP_WRITEMASK_2;
 667     if (fixup.w_sign_fixup) mask |= WINED3DSP_WRITEMASK_3;
 668     mask &= dst_mask;
 669
 670     if (mask)
 671     {
 672         char reg_mask[6];
 673         char *ptr = reg_mask;
 674
 675         if (mask != WINED3DSP_WRITEMASK_ALL)
 676         {
 677             *ptr++ = '.';
 678             if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
 679             if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
 680             if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
 681             if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
 682         }
 683         *ptr = '\0';
 684
 685         shader_addline(buffer, "MAD %s%s, %s, %s, -%s;\n", reg, reg_mask, reg, two, one);
 686     }
 687 }
 688
 689 static void shader_arb_color_correction(const struct SHADER_OPCODE_ARG* arg, struct color_fixup_desc fixup)
 690 {
 691     char reg[256];
 692     pshader_get_register_name(arg->shader, arg->dst, reg);
 693     gen_color_correction(arg->buffer, reg, arg->dst & WINED3DSP_WRITEMASK_ALL, "one", "coefmul.x", fixup);
 694 }
 695
 696 static void pshader_gen_input_modifier_line (
 697     IWineD3DBaseShader *iface,
 698     SHADER_BUFFER* buffer,
 699     const DWORD instr,
 700     int tmpreg,
 701     char *outregstr) {
 702
 703     /* Generate a line that does the input modifier computation and return the input register to use */
 704     char regstr[256];
 705     char swzstr[20];
 706     int insert_line;
 707
 708     /* Assume a new line will be added */
 709     insert_line = 1;
 710
 711     /* Get register name */
 712     pshader_get_register_name(iface, instr, regstr);
 713     shader_arb_get_swizzle(instr, FALSE, swzstr);
 714
 715     switch (instr & WINED3DSP_SRCMOD_MASK) {
 716     case WINED3DSPSM_NONE:
 717         sprintf(outregstr, "%s%s", regstr, swzstr);
 718         insert_line = 0;
 719         break;
 720     case WINED3DSPSM_NEG:
 721         sprintf(outregstr, "-%s%s", regstr, swzstr);
 722         insert_line = 0;
 723         break;
 724     case WINED3DSPSM_BIAS:
 725         shader_addline(buffer, "ADD T%c, %s, -coefdiv.x;\n", 'A' + tmpreg, regstr);
 726         break;
 727     case WINED3DSPSM_BIASNEG:
 728         shader_addline(buffer, "ADD T%c, -%s, coefdiv.x;\n", 'A' + tmpreg, regstr);
 729         break;
 730     case WINED3DSPSM_SIGN:
 731         shader_addline(buffer, "MAD T%c, %s, coefmul.x, -one.x;\n", 'A' + tmpreg, regstr);
 732         break;
 733     case WINED3DSPSM_SIGNNEG:
 734         shader_addline(buffer, "MAD T%c, %s, -coefmul.x, one.x;\n", 'A' + tmpreg, regstr);
 735         break;
 736     case WINED3DSPSM_COMP:
 737         shader_addline(buffer, "SUB T%c, one.x, %s;\n", 'A' + tmpreg, regstr);
 738         break;
 739     case WINED3DSPSM_X2:
 740         shader_addline(buffer, "ADD T%c, %s, %s;\n", 'A' + tmpreg, regstr, regstr);
 741         break;
 742     case WINED3DSPSM_X2NEG:
 743         shader_addline(buffer, "ADD T%c, -%s, -%s;\n", 'A' + tmpreg, regstr, regstr);
 744         break;
 745     case WINED3DSPSM_DZ:
 746         shader_addline(buffer, "RCP T%c, %s.z;\n", 'A' + tmpreg, regstr);
 747         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 748         break;
 749     case WINED3DSPSM_DW:
 750         shader_addline(buffer, "RCP T%c, %s.w;\n", 'A' + tmpreg, regstr);
 751         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 752         break;
 753     default:
 754         sprintf(outregstr, "%s%s", regstr, swzstr);
 755         insert_line = 0;
 756     }
 757
 758     /* Return modified or original register, with swizzle */
 759     if (insert_line)
 760         sprintf(outregstr, "T%c%s", 'A' + tmpreg, swzstr);
 761 }
 762
 763 static inline void pshader_gen_output_modifier_line(SHADER_BUFFER *buffer, int saturate, const char *write_mask,
 764         int shift, const char *regstr)
 765 {
 766     /* Generate a line that does the output modifier computation */
 767     shader_addline(buffer, "MUL%s %s%s, %s, %s;\n", saturate ? "_SAT" : "",
 768         regstr, write_mask, regstr, shift_tab[shift]);
 769 }
 770
 771 static void pshader_hw_bem(const SHADER_OPCODE_ARG *arg)
 772 {
 773     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
 774
 775     SHADER_BUFFER* buffer = arg->buffer;
 776     char dst_name[50];
 777     char src_name[2][50];
 778     char dst_wmask[20];
 779     DWORD sampler_code = arg->dst & WINED3DSP_REGNUM_MASK;
 780     BOOL has_bumpmat = FALSE;
 781     int i;
 782
 783     for(i = 0; i < This->numbumpenvmatconsts; i++) {
 784         if(This->bumpenvmatconst[i].const_num != -1 && This->bumpenvmatconst[i].texunit == sampler_code) {
 785             has_bumpmat = TRUE;
 786             break;
 787         }
 788     }
 789
 790     pshader_get_register_name(arg->shader, arg->dst, dst_name);
 791     shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
 792     strcat(dst_name, dst_wmask);
 793
 794     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_name[0]);
 795     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[1], 1, src_name[1]);
 796
 797     if(has_bumpmat) {
 798         /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
 799         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", sampler_code);
 800         shader_addline(buffer, "DP3 TMP.r, TMP2, %s;\n", src_name[1]);
 801         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", sampler_code);
 802         shader_addline(buffer, "DP3 TMP.g, TMP2, %s;\n", src_name[1]);
 803
 804         shader_addline(buffer, "ADD %s, %s, TMP;\n", dst_name, src_name[0]);
 805     } else {
 806         shader_addline(buffer, "MOV %s, %s;\n", dst_name, src_name[0]);
 807     }
 808 }
 809
 810 static void pshader_hw_cnd(const SHADER_OPCODE_ARG *arg)
 811 {
 812     SHADER_BUFFER* buffer = arg->buffer;
 813     char dst_wmask[20];
 814     char dst_name[50];
 815     char src_name[3][50];
 816     BOOL sat = (arg->dst & WINED3DSP_DSTMOD_MASK) & WINED3DSPDM_SATURATE;
 817     DWORD shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
 818
 819     /* FIXME: support output modifiers */
 820
 821     /* Handle output register */
 822     pshader_get_register_name(arg->shader, arg->dst, dst_name);
 823     shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
 824
 825     /* Generate input register names (with modifiers) */
 826     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_name[0]);
 827     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[1], 1, src_name[1]);
 828     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[2], 2, src_name[2]);
 829
 830     /* The coissue flag changes the semantic of the cnd instruction in <= 1.3 shaders */
 831     if (arg->reg_maps->shader_version <= WINED3DPS_VERSION(1, 3) && arg->opcode_token & WINED3DSI_COISSUE)
 832     {
 833         shader_addline(buffer, "MOV%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1]);
 834     } else {
 835         shader_addline(buffer, "ADD TMP, -%s, coefdiv.x;\n", src_name[0]);
 836         shader_addline(buffer, "CMP%s %s%s, TMP, %s, %s;\n",
 837                                 sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1], src_name[2]);
 838     }
 839     if (shift != 0)
 840         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
 841 }
 842
 843 static void pshader_hw_cmp(const SHADER_OPCODE_ARG *arg)
 844 {
 845     SHADER_BUFFER* buffer = arg->buffer;
 846     char dst_wmask[20];
 847     char dst_name[50];
 848     char src_name[3][50];
 849     DWORD shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
 850     BOOL sat = (arg->dst & WINED3DSP_DSTMOD_MASK) & WINED3DSPDM_SATURATE;
 851
 852     /* FIXME: support output modifiers */
 853
 854     /* Handle output register */
 855     pshader_get_register_name(arg->shader, arg->dst, dst_name);
 856     shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
 857
 858     /* Generate input register names (with modifiers) */
 859     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_name[0]);
 860     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[1], 1, src_name[1]);
 861     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[2], 2, src_name[2]);
 862
 863     shader_addline(buffer, "CMP%s %s%s, %s, %s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
 864                    src_name[0], src_name[2], src_name[1]);
 865
 866     if (shift != 0)
 867         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
 868 }
 869
 870 /** Process the WINED3DSIO_DP2ADD instruction in ARB.
 871  * dst = dot2(src0, src1) + src2 */
 872 static void pshader_hw_dp2add(const SHADER_OPCODE_ARG *arg)
 873 {
 874     SHADER_BUFFER* buffer = arg->buffer;
 875     char dst_wmask[20];
 876     char dst_name[50];
 877     char src_name[3][50];
 878     DWORD shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
 879     BOOL sat = (arg->dst & WINED3DSP_DSTMOD_MASK) & WINED3DSPDM_SATURATE;
 880
 881     pshader_get_register_name(arg->shader, arg->dst, dst_name);
 882     shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
 883
 884     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_name[0]);
 885     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[1], 1, src_name[1]);
 886     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[2], 2, src_name[2]);
 887
 888     /* Emulate a DP2 with a DP3 and 0.0 */
 889     shader_addline(buffer, "MOV TMP, %s;\n", src_name[0]);
 890     shader_addline(buffer, "MOV TMP.z, 0.0;\n");
 891     shader_addline(buffer, "DP3 TMP2, TMP, %s;\n", src_name[1]);
 892     shader_addline(buffer, "ADD%s %s%s, TMP2, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[2]);
 893
 894     if (shift != 0)
 895         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
 896 }
 897
 898 /* Map the opcode 1-to-1 to the GL code */
 899 static void shader_hw_map2gl(const SHADER_OPCODE_ARG *arg)
 900 {
 901     CONST SHADER_OPCODE* curOpcode = arg->opcode;
 902     SHADER_BUFFER* buffer = arg->buffer;
 903     DWORD dst = arg->dst;
 904     const DWORD *src = arg->src;
 905     const char *instruction;
 906     char arguments[256];
 907     unsigned int i;
 908
 909     switch (curOpcode->opcode)
 910     {
 911         case WINED3DSIO_ABS: instruction = "ABS"; break;
 912         case WINED3DSIO_ADD: instruction = "ADD"; break;
 913         case WINED3DSIO_CRS: instruction = "XPD"; break;
 914         case WINED3DSIO_DP3: instruction = "DP3"; break;
 915         case WINED3DSIO_DP4: instruction = "DP4"; break;
 916         case WINED3DSIO_DST: instruction = "DST"; break;
 917         case WINED3DSIO_EXP: instruction = "EX2"; break;
 918         case WINED3DSIO_EXPP: instruction = "EXP"; break;
 919         case WINED3DSIO_FRC: instruction = "FRC"; break;
 920         case WINED3DSIO_LIT: instruction = "LIT"; break;
 921         case WINED3DSIO_LOG: instruction = "LG2"; break;
 922         case WINED3DSIO_LOGP: instruction = "LOG"; break;
 923         case WINED3DSIO_LRP: instruction = "LRP"; break;
 924         case WINED3DSIO_MAD: instruction = "MAD"; break;
 925         case WINED3DSIO_MAX: instruction = "MAX"; break;
 926         case WINED3DSIO_MIN: instruction = "MIN"; break;
 927         case WINED3DSIO_MUL: instruction = "MUL"; break;
 928         case WINED3DSIO_NOP: instruction = "NOP"; break;
 929         case WINED3DSIO_POW: instruction = "POW"; break;
 930         case WINED3DSIO_SGE: instruction = "SGE"; break;
 931         case WINED3DSIO_SLT: instruction = "SLT"; break;
 932         case WINED3DSIO_SUB: instruction = "SUB"; break;
 933         default: instruction = "";
 934             FIXME("Unhandled opcode %s\n", curOpcode->name);
 935             break;
 936     }
 937
 938     if (shader_is_pshader_version(arg->reg_maps->shader_version))
 939     {
 940         /* Output token related */
 941         char output_rname[256];
 942         char output_wmask[20];
 943         char operands[4][100];
 944         BOOL saturate = FALSE;
 945         BOOL centroid = FALSE;
 946         BOOL partialprecision = FALSE;
 947         const char *modifier;
 948         DWORD shift;
 949
 950         if (!curOpcode->num_params)
 951         {
 952             ERR("Opcode \"%s\" has no parameters\n", curOpcode->name);
 953             return;
 954         }
 955
 956         /* Process modifiers */
 957         if (dst & WINED3DSP_DSTMOD_MASK)
 958         {
 959             DWORD mask = dst & WINED3DSP_DSTMOD_MASK;
 960
 961             saturate = mask & WINED3DSPDM_SATURATE;
 962             centroid = mask & WINED3DSPDM_MSAMPCENTROID;
 963             partialprecision = mask & WINED3DSPDM_PARTIALPRECISION;
 964             mask &= ~(WINED3DSPDM_MSAMPCENTROID | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_SATURATE);
 965             if (mask)
 966                 FIXME("Unrecognized modifier(%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
 967
 968             if (centroid)
 969                 FIXME("Unhandled modifier(%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
 970         }
 971         shift = (dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
 972         modifier = (saturate && !shift) ? "_SAT" : "";
 973
 974         /* Generate input register names (with modifiers) */
 975         for (i = 1; i < curOpcode->num_params; ++i)
 976             pshader_gen_input_modifier_line(arg->shader, buffer, src[i-1], i-1, operands[i]);
 977
 978         /* Handle output register */
 979         pshader_get_register_name(arg->shader, dst, output_rname);
 980         strcpy(operands[0], output_rname);
 981         shader_arb_get_write_mask(arg, dst, output_wmask);
 982         strcat(operands[0], output_wmask);
 983
 984         arguments[0] = '\0';
 985         strcat(arguments, operands[0]);
 986         for (i = 1; i < curOpcode->num_params; i++)
 987         {
 988             strcat(arguments, ", ");
 989             strcat(arguments, operands[i]);
 990         }
 991         shader_addline(buffer, "%s%s %s;\n", instruction, modifier, arguments);
 992
 993         /* A shift requires another line. */
 994         if (shift) pshader_gen_output_modifier_line(buffer, saturate, output_wmask, shift, output_rname);
 995     } else {
 996         /* Note that vshader_program_add_param() adds spaces. */
 997
 998         arguments[0] = '\0';
 999         if (curOpcode->num_params > 0)
1000         {
1001             vshader_program_add_param(arg, dst, FALSE, arguments);
1002             for (i = 1; i < curOpcode->num_params; ++i)
1003             {
1004                 strcat(arguments, ",");
1005                 vshader_program_add_param(arg, src[i-1], TRUE, arguments);
1006             }
1007         }
1008         shader_addline(buffer, "%s%s;\n", instruction, arguments);
1009     }
1010 }
1011
1012 static void shader_hw_mov(const SHADER_OPCODE_ARG *arg)
1013 {
1014     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl*)arg->shader;
1015
1016     if ((WINED3DSHADER_VERSION_MAJOR(arg->reg_maps->shader_version) == 1
1017             && !shader_is_pshader_version(arg->reg_maps->shader_version)
1018             && shader_get_regtype(arg->dst) == WINED3DSPR_ADDR)
1019             || arg->opcode->opcode == WINED3DSIO_MOVA)
1020     {
1021         SHADER_BUFFER *buffer = arg->buffer;
1022         char src0_param[256];
1023
1024         if (arg->opcode->opcode == WINED3DSIO_MOVA)
1025             FIXME("mova should round\n");
1026
1027         src0_param[0] = '\0';
1028         if (((IWineD3DVertexShaderImpl *)shader)->rel_offset)
1029         {
1030             vshader_program_add_param(arg, arg->src[0], TRUE, src0_param);
1031             shader_addline(buffer, "ADD TMP.x, %s, helper_const.z;\n", src0_param);
1032             shader_addline(buffer, "ARL A0.x, TMP.x;\n");
1033         }
1034         else
1035         {
1036             /* Apple's ARB_vertex_program implementation does not accept an ARL source argument
1037              * with more than one component. Thus replicate the first source argument over all
1038              * 4 components. For example, .xyzw -> .x (or better: .xxxx), .zwxy -> .z, etc)
1039              */
1040             DWORD parm = arg->src[0] & ~(WINED3DVS_SWIZZLE_MASK);
1041             if((arg->src[0] & WINED3DVS_X_W) == WINED3DVS_X_W)
1042                 parm |= WINED3DVS_X_W | WINED3DVS_Y_W | WINED3DVS_Z_W | WINED3DVS_W_W;
1043             else if((arg->src[0] & WINED3DVS_X_Z) == WINED3DVS_X_Z)
1044                 parm |= WINED3DVS_X_Z | WINED3DVS_Y_Z | WINED3DVS_Z_Z | WINED3DVS_W_Z;
1045             else if((arg->src[0] & WINED3DVS_X_Y) == WINED3DVS_X_Y)
1046                 parm |= WINED3DVS_X_Y | WINED3DVS_Y_Y | WINED3DVS_Z_Y | WINED3DVS_W_Y;
1047             else if((arg->src[0] & WINED3DVS_X_X) == WINED3DVS_X_X)
1048                 parm |= WINED3DVS_X_X | WINED3DVS_Y_X | WINED3DVS_Z_X | WINED3DVS_W_X;
1049             vshader_program_add_param(arg, parm, TRUE, src0_param);
1050             shader_addline(buffer, "ARL A0.x, %s;\n", src0_param);
1051         }
1052     }
1053     else
1054     {
1055         shader_hw_map2gl(arg);
1056     }
1057 }
1058
1059 static void pshader_hw_texkill(const SHADER_OPCODE_ARG *arg)
1060 {
1061     DWORD shader_version = arg->reg_maps->shader_version;
1062     SHADER_BUFFER* buffer = arg->buffer;
1063     char reg_dest[40];
1064
1065     /* No swizzles are allowed in d3d's texkill. PS 1.x ignores the 4th component as documented,
1066      * but >= 2.0 honors it(undocumented, but tested by the d3d9 testsuit)
1067      */
1068     pshader_get_register_name(arg->shader, arg->dst, reg_dest);
1069
1070     if (shader_version >= WINED3DPS_VERSION(2,0))
1071     {
1072         /* The arb backend doesn't claim ps 2.0 support, but try to eat what the app feeds to us */
1073         shader_addline(buffer, "KIL %s;\n", reg_dest);
1074     } else {
1075         /* ARB fp doesn't like swizzles on the parameter of the KIL instruction. To mask the 4th component,
1076          * copy the register into our general purpose TMP variable, overwrite .w and pass TMP to KIL
1077          */
1078         shader_addline(buffer, "MOV TMP, %s;\n", reg_dest);
1079         shader_addline(buffer, "MOV TMP.w, one.w;\n");
1080         shader_addline(buffer, "KIL TMP;\n");
1081     }
1082 }
1083
1084 static void pshader_hw_tex(const SHADER_OPCODE_ARG *arg)
1085 {
1086     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1087     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1088
1089     DWORD dst = arg->dst;
1090     const DWORD *src = arg->src;
1091     SHADER_BUFFER* buffer = arg->buffer;
1092     DWORD shader_version = arg->reg_maps->shader_version;
1093     BOOL projected = FALSE, bias = FALSE;
1094
1095     char reg_dest[40];
1096     char reg_coord[40];
1097     DWORD reg_dest_code;
1098     DWORD reg_sampler_code;
1099
1100     /* All versions have a destination register */
1101     reg_dest_code = dst & WINED3DSP_REGNUM_MASK;
1102     pshader_get_register_name(arg->shader, dst, reg_dest);
1103
1104     /* 1.0-1.3: Use destination register as coordinate source.
1105        1.4+: Use provided coordinate source register. */
1106    if (shader_version < WINED3DPS_VERSION(1,4))
1107       strcpy(reg_coord, reg_dest);
1108    else
1109       pshader_gen_input_modifier_line(arg->shader, buffer, src[0], 0, reg_coord);
1110
1111   /* 1.0-1.4: Use destination register number as texture code.
1112      2.0+: Use provided sampler number as texure code. */
1113   if (shader_version < WINED3DPS_VERSION(2,0))
1114      reg_sampler_code = reg_dest_code;
1115   else
1116      reg_sampler_code = src[1] & WINED3DSP_REGNUM_MASK;
1117
1118   /* projection flag:
1119    * 1.1, 1.2, 1.3: Use WINED3DTSS_TEXTURETRANSFORMFLAGS
1120    * 1.4: Use WINED3DSPSM_DZ or WINED3DSPSM_DW on src[0]
1121    * 2.0+: Use WINED3DSI_TEXLD_PROJECT on the opcode
1122    */
1123   if (shader_version < WINED3DPS_VERSION(1,4))
1124   {
1125       DWORD flags = 0;
1126       if(reg_sampler_code < MAX_TEXTURES) {
1127         flags = deviceImpl->stateBlock->textureState[reg_sampler_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1128       }
1129       if (flags & WINED3DTTFF_PROJECTED) {
1130           projected = TRUE;
1131       }
1132   }
1133   else if (shader_version < WINED3DPS_VERSION(2,0))
1134   {
1135       DWORD src_mod = arg->src[0] & WINED3DSP_SRCMOD_MASK;
1136       if (src_mod == WINED3DSPSM_DZ) {
1137           projected = TRUE;
1138       } else if(src_mod == WINED3DSPSM_DW) {
1139           projected = TRUE;
1140       }
1141   } else {
1142       if(arg->opcode_token & WINED3DSI_TEXLD_PROJECT) {
1143           projected = TRUE;
1144       }
1145       if(arg->opcode_token & WINED3DSI_TEXLD_BIAS) {
1146           bias = TRUE;
1147       }
1148   }
1149   shader_hw_sample(arg, reg_sampler_code, reg_dest, reg_coord, projected, bias);
1150 }
1151
1152 static void pshader_hw_texcoord(const SHADER_OPCODE_ARG *arg)
1153 {
1154     DWORD dst = arg->dst;
1155     SHADER_BUFFER* buffer = arg->buffer;
1156
1157     char tmp[20];
1158     shader_arb_get_write_mask(arg, dst, tmp);
1159     if (arg->reg_maps->shader_version != WINED3DPS_VERSION(1,4))
1160     {
1161         DWORD reg = dst & WINED3DSP_REGNUM_MASK;
1162         shader_addline(buffer, "MOV_SAT T%u%s, fragment.texcoord[%u];\n", reg, tmp, reg);
1163     } else {
1164         DWORD reg1 = dst & WINED3DSP_REGNUM_MASK;
1165         char reg_src[40];
1166
1167         pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, reg_src);
1168         shader_addline(buffer, "MOV R%u%s, %s;\n", reg1, tmp, reg_src);
1169    }
1170 }
1171
1172 static void pshader_hw_texreg2ar(const SHADER_OPCODE_ARG *arg)
1173 {
1174      SHADER_BUFFER* buffer = arg->buffer;
1175      IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1176      IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1177      DWORD flags;
1178
1179      DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
1180      char dst_str[8];
1181      char src_str[50];
1182
1183      sprintf(dst_str, "T%u", reg1);
1184      pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_str);
1185      shader_addline(buffer, "MOV TMP.x, %s.w;\n", src_str);
1186      shader_addline(buffer, "MOV TMP.y, %s.x;\n", src_str);
1187      flags = reg1 < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg1][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1188      shader_hw_sample(arg, reg1, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1189 }
1190
1191 static void pshader_hw_texreg2gb(const SHADER_OPCODE_ARG *arg)
1192 {
1193      SHADER_BUFFER* buffer = arg->buffer;
1194
1195      DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
1196      char dst_str[8];
1197      char src_str[50];
1198
1199      sprintf(dst_str, "T%u", reg1);
1200      pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_str);
1201      shader_addline(buffer, "MOV TMP.x, %s.y;\n", src_str);
1202      shader_addline(buffer, "MOV TMP.y, %s.z;\n", src_str);
1203      shader_hw_sample(arg, reg1, dst_str, "TMP", FALSE, FALSE);
1204 }
1205
1206 static void pshader_hw_texreg2rgb(const SHADER_OPCODE_ARG *arg)
1207 {
1208     SHADER_BUFFER* buffer = arg->buffer;
1209     DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
1210     char dst_str[8];
1211     char src_str[50];
1212
1213     sprintf(dst_str, "T%u", reg1);
1214     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_str);
1215     shader_hw_sample(arg, reg1, dst_str, src_str, FALSE, FALSE);
1216 }
1217
1218 static void pshader_hw_texbem(const SHADER_OPCODE_ARG *arg)
1219 {
1220     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1221     BOOL has_bumpmat = FALSE;
1222     BOOL has_luminance = FALSE;
1223     int i;
1224
1225     DWORD dst = arg->dst;
1226     DWORD src = arg->src[0] & WINED3DSP_REGNUM_MASK;
1227     SHADER_BUFFER* buffer = arg->buffer;
1228
1229     char reg_coord[40];
1230     DWORD reg_dest_code;
1231
1232     /* All versions have a destination register */
1233     reg_dest_code = dst & WINED3DSP_REGNUM_MASK;
1234     /* Can directly use the name because texbem is only valid for <= 1.3 shaders */
1235     pshader_get_register_name(arg->shader, dst, reg_coord);
1236
1237     for(i = 0; i < This->numbumpenvmatconsts; i++) {
1238         if(This->bumpenvmatconst[i].const_num != -1 && reg_dest_code == This->bumpenvmatconst[i].texunit) {
1239             has_bumpmat = TRUE;
1240             break;
1241         }
1242     }
1243     for(i = 0; i < This->numbumpenvmatconsts; i++) {
1244         if(This->luminanceconst[i].const_num != -1 && reg_dest_code == This->luminanceconst[i].texunit) {
1245             has_luminance = TRUE;
1246             break;
1247         }
1248     }
1249
1250     if(has_bumpmat) {
1251         /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
1252
1253         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", reg_dest_code);
1254         shader_addline(buffer, "DP3 TMP.x, TMP2, T%u;\n", src);
1255         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", reg_dest_code);
1256         shader_addline(buffer, "DP3 TMP.y, TMP2, T%u;\n", src);
1257
1258         /* with projective textures, texbem only divides the static texture coord, not the displacement,
1259          * so we can't let the GL handle this.
1260          */
1261         if (((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS]
1262               & WINED3DTTFF_PROJECTED) {
1263             shader_addline(buffer, "RCP TMP2.w, %s.w;\n", reg_coord);
1264             shader_addline(buffer, "MUL TMP2.xy, %s, TMP2.w;\n", reg_coord);
1265             shader_addline(buffer, "ADD TMP.xy, TMP, TMP2;\n");
1266         } else {
1267             shader_addline(buffer, "ADD TMP.xy, TMP, %s;\n", reg_coord);
1268         }
1269
1270         shader_hw_sample(arg, reg_dest_code, reg_coord, "TMP", FALSE, FALSE);
1271
1272         if(arg->opcode->opcode == WINED3DSIO_TEXBEML && has_luminance) {
1273             shader_addline(buffer, "MAD TMP, T%u.z, luminance%d.x, luminance%d.y;\n",
1274                            src, reg_dest_code, reg_dest_code);
1275             shader_addline(buffer, "MUL %s, %s, TMP;\n", reg_coord, reg_coord);
1276         }
1277
1278     } else {
1279         DWORD tf;
1280         if(reg_dest_code < MAX_TEXTURES) {
1281             tf = ((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1282         } else {
1283             tf = 0;
1284         }
1285         /* Without a bump matrix loaded, just sample with the unmodified coordinates */
1286         shader_hw_sample(arg, reg_dest_code, reg_coord, reg_coord, tf & WINED3DTTFF_PROJECTED, FALSE);
1287     }
1288 }
1289
1290 static void pshader_hw_texm3x2pad(const SHADER_OPCODE_ARG *arg)
1291 {
1292     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1293     SHADER_BUFFER* buffer = arg->buffer;
1294     char src0_name[50];
1295
1296     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0_name);
1297     shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", reg, src0_name);
1298 }
1299
1300 static void pshader_hw_texm3x2tex(const SHADER_OPCODE_ARG *arg)
1301 {
1302     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1303     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1304     DWORD flags;
1305     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1306     SHADER_BUFFER* buffer = arg->buffer;
1307     char dst_str[8];
1308     char src0_name[50];
1309
1310     sprintf(dst_str, "T%u", reg);
1311     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0_name);
1312     shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", reg, src0_name);
1313     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1314     shader_hw_sample(arg, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1315 }
1316
1317 static void pshader_hw_texm3x3pad(const SHADER_OPCODE_ARG *arg)
1318 {
1319     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1320     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1321     SHADER_BUFFER* buffer = arg->buffer;
1322     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1323     char src0_name[50];
1324
1325     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0_name);
1326     shader_addline(buffer, "DP3 TMP.%c, T%u, %s;\n", 'x' + current_state->current_row, reg, src0_name);
1327     current_state->texcoord_w[current_state->current_row++] = reg;
1328 }
1329
1330 static void pshader_hw_texm3x3tex(const SHADER_OPCODE_ARG *arg)
1331 {
1332     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1333     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1334     DWORD flags;
1335     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1336     SHADER_BUFFER* buffer = arg->buffer;
1337     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1338     char dst_str[8];
1339     char src0_name[50];
1340
1341     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0_name);
1342     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1343
1344     /* Sample the texture using the calculated coordinates */
1345     sprintf(dst_str, "T%u", reg);
1346     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1347     shader_hw_sample(arg, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1348     current_state->current_row = 0;
1349 }
1350
1351 static void pshader_hw_texm3x3vspec(const SHADER_OPCODE_ARG *arg)
1352 {
1353     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1354     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1355     DWORD flags;
1356     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1357     SHADER_BUFFER* buffer = arg->buffer;
1358     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1359     char dst_str[8];
1360     char src0_name[50];
1361
1362     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0_name);
1363     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1364
1365     /* Construct the eye-ray vector from w coordinates */
1366     shader_addline(buffer, "MOV TMP2.x, fragment.texcoord[%u].w;\n", current_state->texcoord_w[0]);
1367     shader_addline(buffer, "MOV TMP2.y, fragment.texcoord[%u].w;\n", current_state->texcoord_w[1]);
1368     shader_addline(buffer, "MOV TMP2.z, fragment.texcoord[%u].w;\n", reg);
1369
1370     /* Calculate reflection vector
1371      */
1372     shader_addline(buffer, "DP3 TMP.w, TMP, TMP2;\n");
1373     /* The .w is ignored when sampling, so I can use TMP2.w to calculate dot(N, N) */
1374     shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1375     shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1376     shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1377     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1378     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -TMP2;\n");
1379
1380     /* Sample the texture using the calculated coordinates */
1381     sprintf(dst_str, "T%u", reg);
1382     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1383     shader_hw_sample(arg, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1384     current_state->current_row = 0;
1385 }
1386
1387 static void pshader_hw_texm3x3spec(const SHADER_OPCODE_ARG *arg)
1388 {
1389     IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1390     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1391     DWORD flags;
1392     DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1393     DWORD reg3 = arg->src[1] & WINED3DSP_REGNUM_MASK;
1394     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1395     SHADER_BUFFER* buffer = arg->buffer;
1396     char dst_str[8];
1397     char src0_name[50];
1398
1399     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0_name);
1400     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1401
1402     /* Calculate reflection vector.
1403      *
1404      *               dot(N, E)
1405      * TMP.xyz = 2 * --------- * N - E
1406      *               dot(N, N)
1407      *
1408      * Which normalizes the normal vector
1409      */
1410     shader_addline(buffer, "DP3 TMP.w, TMP, C[%u];\n", reg3);
1411     shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1412     shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1413     shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1414     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1415     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -C[%u];\n", reg3);
1416
1417     /* Sample the texture using the calculated coordinates */
1418     sprintf(dst_str, "T%u", reg);
1419     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1420     shader_hw_sample(arg, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1421     current_state->current_row = 0;
1422 }
1423
1424 static void pshader_hw_texdepth(const SHADER_OPCODE_ARG *arg)
1425 {
1426     SHADER_BUFFER* buffer = arg->buffer;
1427     char dst_name[50];
1428
1429     /* texdepth has an implicit destination, the fragment depth value. It's only parameter,
1430      * which is essentially an input, is the destination register because it is the first
1431      * parameter. According to the msdn, this must be register r5, but let's keep it more flexible
1432      * here
1433      */
1434     pshader_get_register_name(arg->shader, arg->dst, dst_name);
1435
1436     /* According to the msdn, the source register(must be r5) is unusable after
1437      * the texdepth instruction, so we're free to modify it
1438      */
1439     shader_addline(buffer, "MIN %s.y, %s.y, one.y;\n", dst_name, dst_name);
1440
1441     /* How to deal with the special case dst_name.g == 0? if r != 0, then
1442      * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1443      * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1444      */
1445     shader_addline(buffer, "RCP %s.y, %s.y;\n", dst_name, dst_name);
1446     shader_addline(buffer, "MUL TMP.x, %s.x, %s.y;\n", dst_name, dst_name);
1447     shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1448     shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1449 }
1450
1451 /** Process the WINED3DSIO_TEXDP3TEX instruction in ARB:
1452  * Take a 3-component dot product of the TexCoord[dstreg] and src,
1453  * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1454 static void pshader_hw_texdp3tex(const SHADER_OPCODE_ARG *arg)
1455 {
1456     SHADER_BUFFER* buffer = arg->buffer;
1457     DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1458     char src0[50];
1459     char dst_str[8];
1460
1461     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0);
1462     shader_addline(buffer, "MOV TMP, 0.0;\n");
1463     shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", sampler_idx, src0);
1464
1465     sprintf(dst_str, "T%u", sampler_idx);
1466     shader_hw_sample(arg, sampler_idx, dst_str, "TMP", FALSE /* Only one coord, can't be projected */, FALSE);
1467 }
1468
1469 /** Process the WINED3DSIO_TEXDP3 instruction in ARB:
1470  * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1471 static void pshader_hw_texdp3(const SHADER_OPCODE_ARG *arg)
1472 {
1473     char src0[50];
1474     char dst_str[50];
1475     char dst_mask[6];
1476     DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1477     SHADER_BUFFER* buffer = arg->buffer;
1478
1479     /* Handle output register */
1480     pshader_get_register_name(arg->shader, arg->dst, dst_str);
1481     shader_arb_get_write_mask(arg, arg->dst, dst_mask);
1482
1483     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0);
1484     shader_addline(buffer, "DP3 %s%s, T%u, %s;\n", dst_str, dst_mask, dstreg, src0);
1485
1486     /* TODO: Handle output modifiers */
1487 }
1488
1489 /** Process the WINED3DSIO_TEXM3X3 instruction in ARB
1490  * Perform the 3rd row of a 3x3 matrix multiply */
1491 static void pshader_hw_texm3x3(const SHADER_OPCODE_ARG *arg)
1492 {
1493     SHADER_BUFFER* buffer = arg->buffer;
1494     char dst_str[50];
1495     char dst_mask[6];
1496     char src0[50];
1497     DWORD dst_reg = arg->dst & WINED3DSP_REGNUM_MASK;
1498
1499     pshader_get_register_name(arg->shader, arg->dst, dst_str);
1500     shader_arb_get_write_mask(arg, arg->dst, dst_mask);
1501
1502     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0);
1503     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", dst_reg, src0);
1504     shader_addline(buffer, "MOV %s%s, TMP;\n", dst_str, dst_mask);
1505
1506     /* TODO: Handle output modifiers */
1507 }
1508
1509 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in ARB:
1510  * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1511  * Calculate tmp0.y = TexCoord[dstreg] . src.xyz;  (tmp0.x has already been calculated)
1512  * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1513  */
1514 static void pshader_hw_texm3x2depth(const SHADER_OPCODE_ARG *arg)
1515 {
1516     SHADER_BUFFER* buffer = arg->buffer;
1517     DWORD dst_reg = arg->dst & WINED3DSP_REGNUM_MASK;
1518     char src0[50];
1519
1520     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0);
1521     shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", dst_reg, src0);
1522
1523     /* How to deal with the special case dst_name.g == 0? if r != 0, then
1524      * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1525      * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1526      */
1527     shader_addline(buffer, "RCP TMP.y, TMP.y;\n");
1528     shader_addline(buffer, "MUL TMP.x, TMP.x, TMP.y;\n");
1529     shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1530     shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1531 }
1532
1533 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1534     Vertex/Pixel shaders to ARB_vertex_program codes */
1535 static void shader_hw_mnxn(const SHADER_OPCODE_ARG *arg)
1536 {
1537     int i;
1538     int nComponents = 0;
1539     SHADER_OPCODE_ARG tmpArg;
1540     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)arg->shader;
1541     const SHADER_OPCODE *opcode_table = shader->baseShader.shader_ins;
1542     DWORD shader_version = arg->reg_maps->shader_version;
1543
1544     memset(&tmpArg, 0, sizeof(SHADER_OPCODE_ARG));
1545
1546     /* Set constants for the temporary argument */
1547     tmpArg.shader      = arg->shader;
1548     tmpArg.buffer      = arg->buffer;
1549     tmpArg.src[0]      = arg->src[0];
1550     tmpArg.src_addr[0] = arg->src_addr[0];
1551     tmpArg.src_addr[1] = arg->src_addr[1];
1552     tmpArg.reg_maps = arg->reg_maps;
1553
1554     switch(arg->opcode->opcode) {
1555     case WINED3DSIO_M4x4:
1556         nComponents = 4;
1557         tmpArg.opcode = shader_get_opcode(opcode_table, shader_version, WINED3DSIO_DP4);
1558         break;
1559     case WINED3DSIO_M4x3:
1560         nComponents = 3;
1561         tmpArg.opcode = shader_get_opcode(opcode_table, shader_version, WINED3DSIO_DP4);
1562         break;
1563     case WINED3DSIO_M3x4:
1564         nComponents = 4;
1565         tmpArg.opcode = shader_get_opcode(opcode_table, shader_version, WINED3DSIO_DP3);
1566         break;
1567     case WINED3DSIO_M3x3:
1568         nComponents = 3;
1569         tmpArg.opcode = shader_get_opcode(opcode_table, shader_version, WINED3DSIO_DP3);
1570         break;
1571     case WINED3DSIO_M3x2:
1572         nComponents = 2;
1573         tmpArg.opcode = shader_get_opcode(opcode_table, shader_version, WINED3DSIO_DP3);
1574         break;
1575     default:
1576         break;
1577     }
1578
1579     for (i = 0; i < nComponents; i++) {
1580         tmpArg.dst = ((arg->dst) & ~WINED3DSP_WRITEMASK_ALL)|(WINED3DSP_WRITEMASK_0<<i);
1581         tmpArg.src[1] = arg->src[1]+i;
1582         shader_hw_map2gl(&tmpArg);
1583     }
1584 }
1585
1586 static void vshader_hw_rsq_rcp(const SHADER_OPCODE_ARG *arg)
1587 {
1588     CONST SHADER_OPCODE* curOpcode = arg->opcode;
1589     SHADER_BUFFER* buffer = arg->buffer;
1590     DWORD dst = arg->dst;
1591     DWORD src = arg->src[0];
1592     DWORD swizzle = (src & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
1593     const char *instruction;
1594
1595     char tmpLine[256];
1596
1597     switch(curOpcode->opcode)
1598     {
1599         case WINED3DSIO_RSQ: instruction = "RSQ"; break;
1600         case WINED3DSIO_RCP: instruction = "RCP"; break;
1601         default: instruction = "";
1602             FIXME("Unhandled opcode %s\n", curOpcode->name);
1603             break;
1604     }
1605
1606     strcpy(tmpLine, instruction);
1607     vshader_program_add_param(arg, dst, FALSE, tmpLine); /* Destination */
1608     strcat(tmpLine, ",");
1609     vshader_program_add_param(arg, src, TRUE, tmpLine);
1610     if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) == swizzle) {
1611         /* Dx sdk says .x is used if no swizzle is given, but our test shows that
1612          * .w is used
1613          */
1614         strcat(tmpLine, ".w");
1615     }
1616
1617     shader_addline(buffer, "%s;\n", tmpLine);
1618 }
1619
1620 static void shader_hw_nrm(const SHADER_OPCODE_ARG *arg)
1621 {
1622     SHADER_BUFFER* buffer = arg->buffer;
1623     char dst_name[50];
1624     char src_name[50];
1625     char dst_wmask[20];
1626     DWORD shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
1627     BOOL sat = (arg->dst & WINED3DSP_DSTMOD_MASK) & WINED3DSPDM_SATURATE;
1628
1629     pshader_get_register_name(arg->shader, arg->dst, dst_name);
1630     shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
1631
1632     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_name);
1633     shader_addline(buffer, "DP3 TMP, %s, %s;\n", src_name, src_name);
1634     shader_addline(buffer, "RSQ TMP, TMP.x;\n");
1635     /* dst.w = src[0].w * 1 / (src.x^2 + src.y^2 + src.z^2)^(1/2) according to msdn*/
1636     shader_addline(buffer, "MUL%s %s%s, %s, TMP;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1637                    src_name);
1638
1639     if (shift != 0)
1640         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
1641 }
1642
1643 static void shader_hw_sincos(const SHADER_OPCODE_ARG *arg)
1644 {
1645     /* This instruction exists in ARB, but the d3d instruction takes two extra parameters which
1646      * must contain fixed constants. So we need a separate function to filter those constants and
1647      * can't use map2gl
1648      */
1649     SHADER_BUFFER* buffer = arg->buffer;
1650     char dst_name[50];
1651     char src_name[50];
1652     char dst_wmask[20];
1653     DWORD shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
1654     BOOL sat = (arg->dst & WINED3DSP_DSTMOD_MASK) & WINED3DSPDM_SATURATE;
1655
1656     pshader_get_register_name(arg->shader, arg->dst, dst_name);
1657     shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
1658
1659     pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_name);
1660     shader_addline(buffer, "SCS%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1661                    src_name);
1662
1663     if (shift != 0)
1664         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
1665
1666 }
1667
1668 static GLuint create_arb_blt_vertex_program(const WineD3D_GL_Info *gl_info)
1669 {
1670     GLuint program_id = 0;
1671     const char *blt_vprogram =
1672         "!!ARBvp1.0\n"
1673         "PARAM c[1] = { { 1, 0.5 } };\n"
1674         "MOV result.position, vertex.position;\n"
1675         "MOV result.color, c[0].x;\n"
1676         "MOV result.texcoord[0], vertex.texcoord[0];\n"
1677         "END\n";
1678
1679     GL_EXTCALL(glGenProgramsARB(1, &program_id));
1680     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, program_id));
1681     GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_vprogram), blt_vprogram));
1682
1683     if (glGetError() == GL_INVALID_OPERATION) {
1684         GLint pos;
1685         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1686         FIXME("Vertex program error at position %d: %s\n", pos,
1687             debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1688     }
1689
1690     return program_id;
1691 }
1692
1693 static GLuint create_arb_blt_fragment_program(const WineD3D_GL_Info *gl_info, enum tex_types tex_type)
1694 {
1695     GLuint program_id = 0;
1696     static const char * const blt_fprograms[tex_type_count] =
1697     {
1698         /* tex_1d */
1699         NULL,
1700         /* tex_2d */
1701         "!!ARBfp1.0\n"
1702         "TEMP R0;\n"
1703         "TEX R0.x, fragment.texcoord[0], texture[0], 2D;\n"
1704         "MOV result.depth.z, R0.x;\n"
1705         "END\n",
1706         /* tex_3d */
1707         NULL,
1708         /* tex_cube */
1709         "!!ARBfp1.0\n"
1710         "TEMP R0;\n"
1711         "TEX R0.x, fragment.texcoord[0], texture[0], CUBE;\n"
1712         "MOV result.depth.z, R0.x;\n"
1713         "END\n",
1714         /* tex_rect */
1715         "!!ARBfp1.0\n"
1716         "TEMP R0;\n"
1717         "TEX R0.x, fragment.texcoord[0], texture[0], RECT;\n"
1718         "MOV result.depth.z, R0.x;\n"
1719         "END\n",
1720     };
1721
1722     if (!blt_fprograms[tex_type])
1723     {
1724         FIXME("tex_type %#x not supported\n", tex_type);
1725         tex_type = tex_2d;
1726     }
1727
1728     GL_EXTCALL(glGenProgramsARB(1, &program_id));
1729     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, program_id));
1730     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_fprograms[tex_type]), blt_fprograms[tex_type]));
1731
1732     if (glGetError() == GL_INVALID_OPERATION) {
1733         GLint pos;
1734         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1735         FIXME("Fragment program error at position %d: %s\n", pos,
1736             debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1737     }
1738
1739     return program_id;
1740 }
1741
1742 static void shader_arb_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
1743     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1744     struct shader_arb_priv *priv = (struct shader_arb_priv *) This->shader_priv;
1745     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1746
1747     if (useVS) {
1748         TRACE("Using vertex shader\n");
1749         IWineD3DVertexShaderImpl_CompileShader(This->stateBlock->vertexShader);
1750
1751         priv->current_vprogram_id = ((IWineD3DVertexShaderImpl *)This->stateBlock->vertexShader)->prgId;
1752
1753         /* Bind the vertex program */
1754         GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1755         checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vertexShader->prgId);");
1756
1757         /* Enable OpenGL vertex programs */
1758         glEnable(GL_VERTEX_PROGRAM_ARB);
1759         checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1760         TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1761     } else if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
1762         priv->current_vprogram_id = 0;
1763         glDisable(GL_VERTEX_PROGRAM_ARB);
1764         checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1765     }
1766
1767     if (usePS) {
1768         struct ps_compile_args compile_args;
1769         TRACE("Using pixel shader\n");
1770         find_ps_compile_args((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader, This->stateBlock, &compile_args);
1771         priv->current_fprogram_id = find_gl_pshader((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader,
1772                                                     &compile_args);
1773
1774         /* Bind the fragment program */
1775         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1776         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, pixelShader->prgId);");
1777
1778         if(!priv->use_arbfp_fixed_func) {
1779             /* Enable OpenGL fragment programs */
1780             glEnable(GL_FRAGMENT_PROGRAM_ARB);
1781             checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1782         }
1783         TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1784     } else if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM) && !priv->use_arbfp_fixed_func) {
1785         /* Disable only if we're not using arbfp fixed function fragment processing. If this is used,
1786          * keep GL_FRAGMENT_PROGRAM_ARB enabled, and the fixed function pipeline will bind the fixed function
1787          * replacement shader
1788          */
1789         glDisable(GL_FRAGMENT_PROGRAM_ARB);
1790         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1791         priv->current_fprogram_id = 0;
1792     }
1793 }
1794
1795 static void shader_arb_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {
1796     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1797     struct shader_arb_priv *priv = (struct shader_arb_priv *) This->shader_priv;
1798     GLuint *blt_fprogram = &priv->depth_blt_fprogram_id[tex_type];
1799     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1800
1801     if (!priv->depth_blt_vprogram_id) priv->depth_blt_vprogram_id = create_arb_blt_vertex_program(gl_info);
1802     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->depth_blt_vprogram_id));
1803     glEnable(GL_VERTEX_PROGRAM_ARB);
1804
1805     if (!*blt_fprogram) *blt_fprogram = create_arb_blt_fragment_program(gl_info, tex_type);
1806     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, *blt_fprogram));
1807     glEnable(GL_FRAGMENT_PROGRAM_ARB);
1808 }
1809
1810 static void shader_arb_deselect_depth_blt(IWineD3DDevice *iface) {
1811     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1812     struct shader_arb_priv *priv = (struct shader_arb_priv *) This->shader_priv;
1813     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1814
1815     if (priv->current_vprogram_id) {
1816         GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1817         checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vertexShader->prgId);");
1818
1819         glEnable(GL_VERTEX_PROGRAM_ARB);
1820         checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1821
1822         TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1823     } else {
1824         glDisable(GL_VERTEX_PROGRAM_ARB);
1825         checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1826     }
1827
1828     if (priv->current_fprogram_id) {
1829         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1830         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, pixelShader->prgId);");
1831
1832         glEnable(GL_FRAGMENT_PROGRAM_ARB);
1833         checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1834
1835         TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1836     } else {
1837         glDisable(GL_FRAGMENT_PROGRAM_ARB);
1838         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1839     }
1840 }
1841
1842 static void shader_arb_destroy(IWineD3DBaseShader *iface) {
1843     IWineD3DBaseShaderImpl *baseShader = (IWineD3DBaseShaderImpl *) iface;
1844     const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)baseShader->baseShader.device)->adapter->gl_info;
1845
1846     if (shader_is_pshader_version(baseShader->baseShader.reg_maps.shader_version))
1847     {
1848         IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *) iface;
1849         UINT i;
1850
1851         ENTER_GL();
1852         for(i = 0; i < This->num_gl_shaders; i++) {
1853             GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1854             checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1855         }
1856         LEAVE_GL();
1857         HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1858         This->gl_shaders = NULL;
1859         This->num_gl_shaders = 0;
1860     } else {
1861         IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *) iface;
1862
1863         ENTER_GL();
1864         GL_EXTCALL(glDeleteProgramsARB(1, &This->prgId));
1865         checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->prgId))");
1866         ((IWineD3DVertexShaderImpl *) This)->prgId = 0;
1867         LEAVE_GL();
1868     }
1869     baseShader->baseShader.is_compiled = FALSE;
1870 }
1871
1872 static HRESULT shader_arb_alloc(IWineD3DDevice *iface) {
1873     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1874     This->shader_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
1875     return WINED3D_OK;
1876 }
1877
1878 static void shader_arb_free(IWineD3DDevice *iface) {
1879     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1880     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1881     struct shader_arb_priv *priv = (struct shader_arb_priv *) This->shader_priv;
1882     int i;
1883
1884     if(priv->depth_blt_vprogram_id) {
1885         GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_vprogram_id));
1886     }
1887     for (i = 0; i < tex_type_count; ++i) {
1888         if (priv->depth_blt_fprogram_id[i]) {
1889             GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_fprogram_id[i]));
1890         }
1891     }
1892
1893     HeapFree(GetProcessHeap(), 0, This->shader_priv);
1894 }
1895
1896 static BOOL shader_arb_dirty_const(IWineD3DDevice *iface) {
1897     return TRUE;
1898 }
1899
1900 static void arbfp_add_sRGB_correction(SHADER_BUFFER *buffer, const char *fragcolor, const char *tmp1,
1901                                       const char *tmp2, const char *tmp3, const char *tmp4) {
1902     /* Perform sRGB write correction. See GLX_EXT_framebuffer_sRGB */
1903
1904     /* Calculate the > 0.0031308 case */
1905     shader_addline(buffer, "POW %s.x, %s.x, srgb_pow.x;\n", tmp1, fragcolor);
1906     shader_addline(buffer, "POW %s.y, %s.y, srgb_pow.y;\n", tmp1, fragcolor);
1907     shader_addline(buffer, "POW %s.z, %s.z, srgb_pow.z;\n", tmp1, fragcolor);
1908     shader_addline(buffer, "MUL %s, %s, srgb_mul_hi;\n", tmp1, tmp1);
1909     shader_addline(buffer, "SUB %s, %s, srgb_sub_hi;\n", tmp1, tmp1);
1910     /* Calculate the < case */
1911     shader_addline(buffer, "MUL %s, srgb_mul_low, %s;\n", tmp2, fragcolor);
1912     /* Get 1.0 / 0.0 masks for > 0.0031308 and < 0.0031308 */
1913     shader_addline(buffer, "SLT %s, srgb_comparison, %s;\n", tmp3, fragcolor);
1914     shader_addline(buffer, "SGE %s, srgb_comparison, %s;\n", tmp4, fragcolor);
1915     /* Store the components > 0.0031308 in the destination */
1916     shader_addline(buffer, "MUL %s, %s, %s;\n", fragcolor, tmp1, tmp3);
1917     /* Add the components that are < 0.0031308 */
1918     shader_addline(buffer, "MAD result.color.xyz, %s, %s, %s;\n", tmp2, tmp4, fragcolor);
1919     /* [0.0;1.0] clamping. Not needed, this is done implicitly */
1920 }
1921
1922 static GLuint shader_arb_generate_pshader(IWineD3DPixelShader *iface, SHADER_BUFFER *buffer, const struct ps_compile_args *args) {
1923     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1924     const shader_reg_maps* reg_maps = &This->baseShader.reg_maps;
1925     CONST DWORD *function = This->baseShader.function;
1926     DWORD shader_version = reg_maps->shader_version;
1927     const char *fragcolor;
1928     const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)This->baseShader.device)->adapter->gl_info;
1929     const local_constant *lconst;
1930     GLuint retval;
1931
1932     /*  Create the hw ARB shader */
1933     shader_addline(buffer, "!!ARBfp1.0\n");
1934
1935     shader_addline(buffer, "TEMP TMP;\n");     /* Used in matrix ops */
1936     shader_addline(buffer, "TEMP TMP2;\n");    /* Used in matrix ops */
1937     shader_addline(buffer, "TEMP TA;\n");      /* Used for modifiers */
1938     shader_addline(buffer, "TEMP TB;\n");      /* Used for modifiers */
1939     shader_addline(buffer, "TEMP TC;\n");      /* Used for modifiers */
1940     shader_addline(buffer, "PARAM coefdiv = { 0.5, 0.25, 0.125, 0.0625 };\n");
1941     shader_addline(buffer, "PARAM coefmul = { 2, 4, 8, 16 };\n");
1942     shader_addline(buffer, "PARAM one = { 1.0, 1.0, 1.0, 1.0 };\n");
1943
1944     /* Base Declarations */
1945     shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION);
1946
1947     /* We need two variables for fog blending */
1948     if(args->fog != FOG_OFF) shader_addline(buffer, "TEMP TMP_FOG;\n");
1949     if (shader_version >= WINED3DPS_VERSION(2,0)) shader_addline(buffer, "TEMP TMP_COLOR;\n");
1950
1951     /* Base Shader Body */
1952     shader_generate_main( (IWineD3DBaseShader*) This, buffer, reg_maps, function);
1953
1954     if (shader_version < WINED3DPS_VERSION(2,0)) {
1955         fragcolor = "R0";
1956     } else {
1957         fragcolor = "TMP_COLOR";
1958     }
1959     if(args->srgb_correction) {
1960         arbfp_add_sRGB_correction(buffer, fragcolor, "TMP", "TMP2", "TA", "TB");
1961     }
1962     if (shader_version < WINED3DPS_VERSION(3,0)) {
1963         /* calculate fog and blend it
1964          * NOTE: state.fog.params.y and state.fog.params.z don't hold fog start s and end e but
1965          * -1/(e-s) and e/(e-s) respectively.
1966          */
1967         switch(args->fog) {
1968             case FOG_OFF:
1969                 shader_addline(buffer, "MOV result.color, %s;\n", fragcolor);
1970                 break;
1971             case FOG_LINEAR:
1972                 shader_addline(buffer, "MAD_SAT TMP_FOG, fragment.fogcoord, state.fog.params.y, state.fog.params.z;\n");
1973                 shader_addline(buffer, "LRP result.color.xyz, TMP_FOG.x, %s, state.fog.color;\n", fragcolor);
1974                 shader_addline(buffer, "MOV result.color.w, %s.w;\n", fragcolor);
1975                 break;
1976             case FOG_EXP:
1977                 FIXME("Implement EXP fog in ARB\n");
1978                 shader_addline(buffer, "MOV result.color, %s;\n", fragcolor);
1979                 break;
1980             case FOG_EXP2:
1981                 FIXME("Implement EXP2 fog in ARB\n");
1982                 shader_addline(buffer, "MOV result.color, %s;\n", fragcolor);
1983                 break;
1984         }
1985     }
1986
1987     shader_addline(buffer, "END\n");
1988
1989     /* TODO: change to resource.glObjectHandle or something like that */
1990     GL_EXTCALL(glGenProgramsARB(1, &retval));
1991
1992     TRACE("Creating a hw pixel shader, prg=%d\n", retval);
1993     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, retval));
1994
1995     TRACE("Created hw pixel shader, prg=%d\n", retval);
1996     /* Create the program and check for errors */
1997     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
1998                buffer->bsize, buffer->buffer));
1999
2000     if (glGetError() == GL_INVALID_OPERATION) {
2001         GLint errPos;
2002         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2003         FIXME("HW PixelShader Error at position %d: %s\n",
2004               errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2005         retval = 0;
2006     }
2007
2008     /* Load immediate constants */
2009     if(!This->baseShader.load_local_constsF) {
2010         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2011             const float *value = (const float *)lconst->value;
2012             GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, lconst->idx, value));
2013             checkGLcall("glProgramLocalParameter4fvARB");
2014         }
2015     }
2016
2017     return retval;
2018 }
2019
2020 static void shader_arb_generate_vshader(IWineD3DVertexShader *iface, SHADER_BUFFER *buffer) {
2021     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
2022     const shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
2023     CONST DWORD *function = This->baseShader.function;
2024     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)This->baseShader.device;
2025     const WineD3D_GL_Info *gl_info = &device->adapter->gl_info;
2026     const local_constant *lconst;
2027
2028     /*  Create the hw ARB shader */
2029     shader_addline(buffer, "!!ARBvp1.0\n");
2030     shader_addline(buffer, "PARAM helper_const = { 2.0, -1.0, %d.0, 0.0 };\n", This->rel_offset);
2031
2032     /* Mesa supports only 95 constants */
2033     if (GL_VEND(MESA) || GL_VEND(WINE))
2034         This->baseShader.limits.constant_float =
2035                 min(95, This->baseShader.limits.constant_float);
2036
2037     shader_addline(buffer, "TEMP TMP;\n");
2038
2039     /* Base Declarations */
2040     shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION);
2041
2042     /* We need a constant to fixup the final position */
2043     shader_addline(buffer, "PARAM posFixup = program.env[%d];\n", ARB_SHADER_PRIVCONST_POS);
2044
2045     /* Initialize output parameters. GL_ARB_vertex_program does not require special initialization values
2046      * for output parameters. D3D in theory does not do that either, but some applications depend on a
2047      * proper initialization of the secondary color, and programs using the fixed function pipeline without
2048      * a replacement shader depend on the texcoord.w being set properly.
2049      *
2050      * GL_NV_vertex_program defines that all output values are initialized to {0.0, 0.0, 0.0, 1.0}. This
2051      * assertion is in effect even when using GL_ARB_vertex_program without any NV specific additions. So
2052      * skip this if NV_vertex_program is supported. Otherwise, initialize the secondary color. For the tex-
2053      * coords, we have a flag in the opengl caps. Many cards do not require the texcoord being set, and
2054      * this can eat a number of instructions, so skip it unless this cap is set as well
2055      */
2056     if(!GL_SUPPORT(NV_VERTEX_PROGRAM)) {
2057         shader_addline(buffer, "MOV result.color.secondary, -helper_const.wwwy;\n");
2058
2059         if((GLINFO_LOCATION).set_texcoord_w && !device->frag_pipe->ffp_proj_control) {
2060             int i;
2061             for(i = 0; i < min(8, MAX_REG_TEXCRD); i++) {
2062                 if(This->baseShader.reg_maps.texcoord_mask[i] != 0 &&
2063                 This->baseShader.reg_maps.texcoord_mask[i] != WINED3DSP_WRITEMASK_ALL) {
2064                     shader_addline(buffer, "MOV result.texcoord[%u].w, -helper_const.y;\n", i);
2065                 }
2066             }
2067         }
2068     }
2069
2070     /* Base Shader Body */
2071     shader_generate_main( (IWineD3DBaseShader*) This, buffer, reg_maps, function);
2072
2073     /* If this shader doesn't use fog copy the z coord to the fog coord so that we can use table fog */
2074     if (!reg_maps->fog)
2075         shader_addline(buffer, "MOV result.fogcoord, TMP_OUT.z;\n");
2076
2077     /* Write the final position.
2078      *
2079      * OpenGL coordinates specify the center of the pixel while d3d coords specify
2080      * the corner. The offsets are stored in z and w in posFixup. posFixup.y contains
2081      * 1.0 or -1.0 to turn the rendering upside down for offscreen rendering. PosFixup.x
2082      * contains 1.0 to allow a mad, but arb vs swizzles are too restricted for that.
2083      */
2084     shader_addline(buffer, "MUL TMP, posFixup, TMP_OUT.w;\n");
2085     shader_addline(buffer, "ADD TMP_OUT.x, TMP_OUT.x, TMP.z;\n");
2086     shader_addline(buffer, "MAD TMP_OUT.y, TMP_OUT.y, posFixup.y, TMP.w;\n");
2087
2088     /* Z coord [0;1]->[-1;1] mapping, see comment in transform_projection in state.c
2089      * and the glsl equivalent
2090      */
2091     shader_addline(buffer, "MAD TMP_OUT.z, TMP_OUT.z, helper_const.x, -TMP_OUT.w;\n");
2092
2093     shader_addline(buffer, "MOV result.position, TMP_OUT;\n");
2094
2095     shader_addline(buffer, "END\n");
2096
2097     /* TODO: change to resource.glObjectHandle or something like that */
2098     GL_EXTCALL(glGenProgramsARB(1, &This->prgId));
2099
2100     TRACE("Creating a hw vertex shader, prg=%d\n", This->prgId);
2101     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, This->prgId));
2102
2103     TRACE("Created hw vertex shader, prg=%d\n", This->prgId);
2104     /* Create the program and check for errors */
2105     GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2106                buffer->bsize, buffer->buffer));
2107
2108     if (glGetError() == GL_INVALID_OPERATION) {
2109         GLint errPos;
2110         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2111         FIXME("HW VertexShader Error at position %d: %s\n",
2112               errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2113         This->prgId = -1;
2114     }
2115
2116     /* Load immediate constants */
2117     if(!This->baseShader.load_local_constsF) {
2118         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2119             const float *value = (const float *)lconst->value;
2120             GL_EXTCALL(glProgramLocalParameter4fvARB(GL_VERTEX_PROGRAM_ARB, lconst->idx, value));
2121         }
2122     }
2123 }
2124
2125 static void shader_arb_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct shader_caps *pCaps)
2126 {
2127     /* We don't have an ARB fixed function pipeline yet, so let the none backend set its caps,
2128      * then overwrite the shader specific ones
2129      */
2130     none_shader_backend.shader_get_caps(devtype, gl_info, pCaps);
2131
2132     if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
2133         pCaps->VertexShaderVersion = WINED3DVS_VERSION(1,1);
2134         TRACE_(d3d_caps)("Hardware vertex shader version 1.1 enabled (ARB_PROGRAM)\n");
2135         pCaps->MaxVertexShaderConst = GL_LIMITS(vshader_constantsF);
2136     }
2137
2138     if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM)) {
2139         pCaps->PixelShaderVersion    = WINED3DPS_VERSION(1,4);
2140         pCaps->PixelShader1xMaxValue = 8.0;
2141         TRACE_(d3d_caps)("Hardware pixel shader version 1.4 enabled (ARB_PROGRAM)\n");
2142     }
2143 }
2144
2145 static BOOL shader_arb_color_fixup_supported(struct color_fixup_desc fixup)
2146 {
2147     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
2148     {
2149         TRACE("Checking support for color_fixup:\n");
2150         dump_color_fixup_desc(fixup);
2151     }
2152
2153     /* We support everything except YUV conversions. */
2154     if (!is_yuv_fixup(fixup))
2155     {
2156         TRACE("[OK]\n");
2157         return TRUE;
2158     }
2159
2160     TRACE("[FAILED]\n");
2161     return FALSE;
2162 }
2163
2164 static const SHADER_HANDLER shader_arb_instruction_handler_table[WINED3DSIH_TABLE_SIZE] =
2165 {
2166     /* WINED3DSIH_ABS           */ shader_hw_map2gl,
2167     /* WINED3DSIH_ADD           */ shader_hw_map2gl,
2168     /* WINED3DSIH_BEM           */ pshader_hw_bem,
2169     /* WINED3DSIH_BREAK         */ NULL,
2170     /* WINED3DSIH_BREAKC        */ NULL,
2171     /* WINED3DSIH_BREAKP        */ NULL,
2172     /* WINED3DSIH_CALL          */ NULL,
2173     /* WINED3DSIH_CALLNZ        */ NULL,
2174     /* WINED3DSIH_CMP           */ pshader_hw_cmp,
2175     /* WINED3DSIH_CND           */ pshader_hw_cnd,
2176     /* WINED3DSIH_CRS           */ shader_hw_map2gl,
2177     /* WINED3DSIH_DCL           */ NULL,
2178     /* WINED3DSIH_DEF           */ NULL,
2179     /* WINED3DSIH_DEFB          */ NULL,
2180     /* WINED3DSIH_DEFI          */ NULL,
2181     /* WINED3DSIH_DP2ADD        */ pshader_hw_dp2add,
2182     /* WINED3DSIH_DP3           */ shader_hw_map2gl,
2183     /* WINED3DSIH_DP4           */ shader_hw_map2gl,
2184     /* WINED3DSIH_DST           */ shader_hw_map2gl,
2185     /* WINED3DSIH_DSX           */ NULL,
2186     /* WINED3DSIH_DSY           */ NULL,
2187     /* WINED3DSIH_ELSE          */ NULL,
2188     /* WINED3DSIH_ENDIF         */ NULL,
2189     /* WINED3DSIH_ENDLOOP       */ NULL,
2190     /* WINED3DSIH_ENDREP        */ NULL,
2191     /* WINED3DSIH_EXP           */ shader_hw_map2gl,
2192     /* WINED3DSIH_EXPP          */ shader_hw_map2gl,
2193     /* WINED3DSIH_FRC           */ shader_hw_map2gl,
2194     /* WINED3DSIH_IF            */ NULL,
2195     /* WINED3DSIH_IFC           */ NULL,
2196     /* WINED3DSIH_LABEL         */ NULL,
2197     /* WINED3DSIH_LIT           */ shader_hw_map2gl,
2198     /* WINED3DSIH_LOG           */ shader_hw_map2gl,
2199     /* WINED3DSIH_LOGP          */ shader_hw_map2gl,
2200     /* WINED3DSIH_LOOP          */ NULL,
2201     /* WINED3DSIH_LRP           */ shader_hw_map2gl,
2202     /* WINED3DSIH_M3x2          */ shader_hw_mnxn,
2203     /* WINED3DSIH_M3x3          */ shader_hw_mnxn,
2204     /* WINED3DSIH_M3x4          */ shader_hw_mnxn,
2205     /* WINED3DSIH_M4x3          */ shader_hw_mnxn,
2206     /* WINED3DSIH_M4x4          */ shader_hw_mnxn,
2207     /* WINED3DSIH_MAD           */ shader_hw_map2gl,
2208     /* WINED3DSIH_MAX           */ shader_hw_map2gl,
2209     /* WINED3DSIH_MIN           */ shader_hw_map2gl,
2210     /* WINED3DSIH_MOV           */ shader_hw_mov,
2211     /* WINED3DSIH_MOVA          */ shader_hw_mov,
2212     /* WINED3DSIH_MUL           */ shader_hw_map2gl,
2213     /* WINED3DSIH_NOP           */ shader_hw_map2gl,
2214     /* WINED3DSIH_NRM           */ shader_hw_nrm,
2215     /* WINED3DSIH_PHASE         */ NULL,
2216     /* WINED3DSIH_POW           */ shader_hw_map2gl,
2217     /* WINED3DSIH_RCP           */ vshader_hw_rsq_rcp,
2218     /* WINED3DSIH_REP           */ NULL,
2219     /* WINED3DSIH_RET           */ NULL,
2220     /* WINED3DSIH_RSQ           */ vshader_hw_rsq_rcp,
2221     /* WINED3DSIH_SETP          */ NULL,
2222     /* WINED3DSIH_SGE           */ shader_hw_map2gl,
2223     /* WINED3DSIH_SGN           */ NULL,
2224     /* WINED3DSIH_SINCOS        */ shader_hw_sincos,
2225     /* WINED3DSIH_SLT           */ shader_hw_map2gl,
2226     /* WINED3DSIH_SUB           */ shader_hw_map2gl,
2227     /* WINED3DSIH_TEX           */ pshader_hw_tex,
2228     /* WINED3DSIH_TEXBEM        */ pshader_hw_texbem,
2229     /* WINED3DSIH_TEXBEML       */ pshader_hw_texbem,
2230     /* WINED3DSIH_TEXCOORD      */ pshader_hw_texcoord,
2231     /* WINED3DSIH_TEXDEPTH      */ pshader_hw_texdepth,
2232     /* WINED3DSIH_TEXDP3        */ pshader_hw_texdp3,
2233     /* WINED3DSIH_TEXDP3TEX     */ pshader_hw_texdp3tex,
2234     /* WINED3DSIH_TEXKILL       */ pshader_hw_texkill,
2235     /* WINED3DSIH_TEXLDD        */ NULL,
2236     /* WINED3DSIH_TEXLDL        */ NULL,
2237     /* WINED3DSIH_TEXM3x2DEPTH  */ pshader_hw_texm3x2depth,
2238     /* WINED3DSIH_TEXM3x2PAD    */ pshader_hw_texm3x2pad,
2239     /* WINED3DSIH_TEXM3x2TEX    */ pshader_hw_texm3x2tex,
2240     /* WINED3DSIH_TEXM3x3       */ pshader_hw_texm3x3,
2241     /* WINED3DSIH_TEXM3x3DIFF   */ NULL,
2242     /* WINED3DSIH_TEXM3x3PAD    */ pshader_hw_texm3x3pad,
2243     /* WINED3DSIH_TEXM3x3SPEC   */ pshader_hw_texm3x3spec,
2244     /* WINED3DSIH_TEXM3x3TEX    */ pshader_hw_texm3x3tex,
2245     /* WINED3DSIH_TEXM3x3VSPEC  */ pshader_hw_texm3x3vspec,
2246     /* WINED3DSIH_TEXREG2AR     */ pshader_hw_texreg2ar,
2247     /* WINED3DSIH_TEXREG2GB     */ pshader_hw_texreg2gb,
2248     /* WINED3DSIH_TEXREG2RGB    */ pshader_hw_texreg2rgb,
2249 };
2250
2251 const shader_backend_t arb_program_shader_backend = {
2252     shader_arb_instruction_handler_table,
2253     shader_arb_select,
2254     shader_arb_select_depth_blt,
2255     shader_arb_deselect_depth_blt,
2256     shader_arb_update_float_vertex_constants,
2257     shader_arb_update_float_pixel_constants,
2258     shader_arb_load_constants,
2259     shader_arb_color_correction,
2260     shader_arb_destroy,
2261     shader_arb_alloc,
2262     shader_arb_free,
2263     shader_arb_dirty_const,
2264     shader_arb_generate_pshader,
2265     shader_arb_generate_vshader,
2266     shader_arb_get_caps,
2267     shader_arb_color_fixup_supported,
2268 };
2269
2270 /* ARB_fragment_program fixed function pipeline replacement definitions */
2271 #define ARB_FFP_CONST_TFACTOR           0
2272 #define ARB_FFP_CONST_SPECULAR_ENABLE   ((ARB_FFP_CONST_TFACTOR) + 1)
2273 #define ARB_FFP_CONST_CONSTANT(i)       ((ARB_FFP_CONST_SPECULAR_ENABLE) + 1 + i)
2274 #define ARB_FFP_CONST_BUMPMAT(i)        ((ARB_FFP_CONST_CONSTANT(7)) + 1 + i)
2275 #define ARB_FFP_CONST_LUMINANCE(i)      ((ARB_FFP_CONST_BUMPMAT(7)) + 1 + i)
2276
2277 struct arbfp_ffp_desc
2278 {
2279     struct ffp_frag_desc parent;
2280     GLuint shader;
2281     unsigned int num_textures_used;
2282 };
2283
2284 static void arbfp_enable(IWineD3DDevice *iface, BOOL enable) {
2285     if(enable) {
2286         glEnable(GL_FRAGMENT_PROGRAM_ARB);
2287         checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
2288     } else {
2289         glDisable(GL_FRAGMENT_PROGRAM_ARB);
2290         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
2291     }
2292 }
2293
2294 static HRESULT arbfp_alloc(IWineD3DDevice *iface) {
2295     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2296     struct shader_arb_priv *priv;
2297     /* Share private data between the shader backend and the pipeline replacement, if both
2298      * are the arb implementation. This is needed to figure out whether ARBfp should be disabled
2299      * if no pixel shader is bound or not
2300      */
2301     if(This->shader_backend == &arb_program_shader_backend) {
2302         This->fragment_priv = This->shader_priv;
2303     } else {
2304         This->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
2305         if(!This->fragment_priv) return E_OUTOFMEMORY;
2306     }
2307     priv = (struct shader_arb_priv *) This->fragment_priv;
2308     priv->fragment_shaders = hash_table_create(ffp_frag_program_key_hash, ffp_frag_program_key_compare);
2309     priv->use_arbfp_fixed_func = TRUE;
2310     return WINED3D_OK;
2311 }
2312
2313 static void arbfp_free_ffpshader(void *value, void *gli) {
2314     const WineD3D_GL_Info *gl_info = gli;
2315     struct arbfp_ffp_desc *entry_arb = value;
2316
2317     ENTER_GL();
2318     GL_EXTCALL(glDeleteProgramsARB(1, &entry_arb->shader));
2319     checkGLcall("glDeleteProgramsARB(1, &entry_arb->shader)");
2320     HeapFree(GetProcessHeap(), 0, entry_arb);
2321     LEAVE_GL();
2322 }
2323
2324 static void arbfp_free(IWineD3DDevice *iface) {
2325     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2326     struct shader_arb_priv *priv = (struct shader_arb_priv *) This->fragment_priv;
2327
2328     hash_table_destroy(priv->fragment_shaders, arbfp_free_ffpshader, &This->adapter->gl_info);
2329     priv->use_arbfp_fixed_func = FALSE;
2330
2331     if(This->shader_backend != &arb_program_shader_backend) {
2332         HeapFree(GetProcessHeap(), 0, This->fragment_priv);
2333     }
2334 }
2335
2336 static void arbfp_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct fragment_caps *caps)
2337 {
2338     caps->TextureOpCaps =  WINED3DTEXOPCAPS_DISABLE                     |
2339                            WINED3DTEXOPCAPS_SELECTARG1                  |
2340                            WINED3DTEXOPCAPS_SELECTARG2                  |
2341                            WINED3DTEXOPCAPS_MODULATE4X                  |
2342                            WINED3DTEXOPCAPS_MODULATE2X                  |
2343                            WINED3DTEXOPCAPS_MODULATE                    |
2344                            WINED3DTEXOPCAPS_ADDSIGNED2X                 |
2345                            WINED3DTEXOPCAPS_ADDSIGNED                   |
2346                            WINED3DTEXOPCAPS_ADD                         |
2347                            WINED3DTEXOPCAPS_SUBTRACT                    |
2348                            WINED3DTEXOPCAPS_ADDSMOOTH                   |
2349                            WINED3DTEXOPCAPS_BLENDCURRENTALPHA           |
2350                            WINED3DTEXOPCAPS_BLENDFACTORALPHA            |
2351                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHA           |
2352                            WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA           |
2353                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM         |
2354                            WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR      |
2355                            WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA      |
2356                            WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA   |
2357                            WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR   |
2358                            WINED3DTEXOPCAPS_DOTPRODUCT3                 |
2359                            WINED3DTEXOPCAPS_MULTIPLYADD                 |
2360                            WINED3DTEXOPCAPS_LERP                        |
2361                            WINED3DTEXOPCAPS_BUMPENVMAP                  |
2362                            WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE;
2363
2364     /* TODO: Implement WINED3DTEXOPCAPS_PREMODULATE */
2365
2366     caps->MaxTextureBlendStages   = 8;
2367     caps->MaxSimultaneousTextures = min(GL_LIMITS(fragment_samplers), 8);
2368
2369     caps->PrimitiveMiscCaps |= WINED3DPMISCCAPS_TSSARGTEMP;
2370 }
2371 #undef GLINFO_LOCATION
2372
2373 #define GLINFO_LOCATION stateblock->wineD3DDevice->adapter->gl_info
2374 static void state_texfactor_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2375     float col[4];
2376     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2377
2378     /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2379      * application provided constants
2380      */
2381     if(device->shader_backend == &arb_program_shader_backend) {
2382         if (use_ps(stateblock)) return;
2383
2384         device = stateblock->wineD3DDevice;
2385         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_TFACTOR] = 1;
2386         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_TFACTOR + 1);
2387     }
2388
2389     D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
2390     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col));
2391     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col)");
2392
2393 }
2394
2395 static void state_arb_specularenable(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2396     float col[4];
2397     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2398
2399     /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2400      * application provided constants
2401      */
2402     if(device->shader_backend == &arb_program_shader_backend) {
2403         if (use_ps(stateblock)) return;
2404
2405         device = stateblock->wineD3DDevice;
2406         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_SPECULAR_ENABLE] = 1;
2407         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_SPECULAR_ENABLE + 1);
2408     }
2409
2410     if(stateblock->renderState[WINED3DRS_SPECULARENABLE]) {
2411         /* The specular color has no alpha */
2412         col[0] = 1.0; col[1] = 1.0;
2413         col[2] = 1.0; col[3] = 0.0;
2414     } else {
2415         col[0] = 0.0; col[1] = 0.0;
2416         col[2] = 0.0; col[3] = 0.0;
2417     }
2418     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col));
2419     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col)");
2420 }
2421
2422 static void set_bumpmat_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2423     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / WINED3D_HIGHEST_TEXTURE_STATE;
2424     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2425     float mat[2][2];
2426
2427     if (use_ps(stateblock))
2428     {
2429         if(stage != 0 &&
2430            ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.bumpmat[stage]) {
2431             /* The pixel shader has to know the bump env matrix. Do a constants update if it isn't scheduled
2432              * anyway
2433              */
2434             if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2435                 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2436             }
2437         }
2438
2439         if(device->shader_backend == &arb_program_shader_backend) {
2440             /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2441             return;
2442         }
2443     } else if(device->shader_backend == &arb_program_shader_backend) {
2444         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_BUMPMAT(stage)] = 1;
2445         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_BUMPMAT(stage) + 1);
2446     }
2447
2448     mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
2449     mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
2450     mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
2451     mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
2452
2453     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0]));
2454     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0])");
2455 }
2456
2457 static void tex_bumpenvlum_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2458     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / WINED3D_HIGHEST_TEXTURE_STATE;
2459     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2460     float param[4];
2461
2462     if (use_ps(stateblock))
2463     {
2464         if(stage != 0 &&
2465            ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.luminanceparams[stage]) {
2466             /* The pixel shader has to know the luminance offset. Do a constants update if it
2467              * isn't scheduled anyway
2468              */
2469             if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2470                 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2471             }
2472         }
2473
2474         if(device->shader_backend == &arb_program_shader_backend) {
2475             /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2476             return;
2477         }
2478     } else if(device->shader_backend == &arb_program_shader_backend) {
2479         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_LUMINANCE(stage)] = 1;
2480         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_LUMINANCE(stage) + 1);
2481     }
2482
2483     param[0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLSCALE]);
2484     param[1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLOFFSET]);
2485     param[2] = 0.0;
2486     param[3] = 0.0;
2487
2488     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param));
2489     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param)");
2490 }
2491
2492 static const char *get_argreg(SHADER_BUFFER *buffer, DWORD argnum, unsigned int stage, DWORD arg) {
2493     const char *ret;
2494
2495     if(arg == ARG_UNUSED) return "unused"; /* This is the marker for unused registers */
2496
2497     switch(arg & WINED3DTA_SELECTMASK) {
2498         case WINED3DTA_DIFFUSE:
2499             ret = "fragment.color.primary"; break;
2500
2501         case WINED3DTA_CURRENT:
2502             if(stage == 0) ret = "fragment.color.primary";
2503             else ret = "ret";
2504             break;
2505
2506         case WINED3DTA_TEXTURE:
2507             switch(stage) {
2508                 case 0: ret = "tex0"; break;
2509                 case 1: ret = "tex1"; break;
2510                 case 2: ret = "tex2"; break;
2511                 case 3: ret = "tex3"; break;
2512                 case 4: ret = "tex4"; break;
2513                 case 5: ret = "tex5"; break;
2514                 case 6: ret = "tex6"; break;
2515                 case 7: ret = "tex7"; break;
2516                 default: ret = "unknown texture";
2517             }
2518             break;
2519
2520         case WINED3DTA_TFACTOR:
2521             ret = "tfactor"; break;
2522
2523         case WINED3DTA_SPECULAR:
2524             ret = "fragment.color.secondary"; break;
2525
2526         case WINED3DTA_TEMP:
2527             ret = "tempreg"; break;
2528
2529         case WINED3DTA_CONSTANT:
2530             FIXME("Implement perstage constants\n");
2531             switch(stage) {
2532                 case 0: ret = "const0"; break;
2533                 case 1: ret = "const1"; break;
2534                 case 2: ret = "const2"; break;
2535                 case 3: ret = "const3"; break;
2536                 case 4: ret = "const4"; break;
2537                 case 5: ret = "const5"; break;
2538                 case 6: ret = "const6"; break;
2539                 case 7: ret = "const7"; break;
2540                 default: ret = "unknown constant";
2541             }
2542             break;
2543
2544         default:
2545             return "unknown";
2546     }
2547
2548     if(arg & WINED3DTA_COMPLEMENT) {
2549         shader_addline(buffer, "SUB arg%u, const.x, %s;\n", argnum, ret);
2550         if(argnum == 0) ret = "arg0";
2551         if(argnum == 1) ret = "arg1";
2552         if(argnum == 2) ret = "arg2";
2553     }
2554     if(arg & WINED3DTA_ALPHAREPLICATE) {
2555         shader_addline(buffer, "MOV arg%u, %s.w;\n", argnum, ret);
2556         if(argnum == 0) ret = "arg0";
2557         if(argnum == 1) ret = "arg1";
2558         if(argnum == 2) ret = "arg2";
2559     }
2560     return ret;
2561 }
2562
2563 static void gen_ffp_instr(SHADER_BUFFER *buffer, unsigned int stage, BOOL color, BOOL alpha,
2564                           DWORD dst, DWORD op, DWORD dw_arg0, DWORD dw_arg1, DWORD dw_arg2) {
2565     const char *dstmask, *dstreg, *arg0, *arg1, *arg2;
2566     unsigned int mul = 1;
2567     BOOL mul_final_dest = FALSE;
2568
2569     if(color && alpha) dstmask = "";
2570     else if(color) dstmask = ".xyz";
2571     else dstmask = ".w";
2572
2573     if(dst == tempreg) dstreg = "tempreg";
2574     else dstreg = "ret";
2575
2576     arg0 = get_argreg(buffer, 0, stage, dw_arg0);
2577     arg1 = get_argreg(buffer, 1, stage, dw_arg1);
2578     arg2 = get_argreg(buffer, 2, stage, dw_arg2);
2579
2580     switch(op) {
2581         case WINED3DTOP_DISABLE:
2582             if(stage == 0) shader_addline(buffer, "MOV %s%s, fragment.color.primary;\n", dstreg, dstmask);
2583             break;
2584
2585         case WINED3DTOP_SELECTARG2:
2586             arg1 = arg2;
2587         case WINED3DTOP_SELECTARG1:
2588             shader_addline(buffer, "MOV %s%s, %s;\n", dstreg, dstmask, arg1);
2589             break;
2590
2591         case WINED3DTOP_MODULATE4X:
2592             mul = 2;
2593         case WINED3DTOP_MODULATE2X:
2594             mul *= 2;
2595             if(strcmp(dstreg, "result.color") == 0) {
2596                 dstreg = "ret";
2597                 mul_final_dest = TRUE;
2598             }
2599         case WINED3DTOP_MODULATE:
2600             shader_addline(buffer, "MUL %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2601             break;
2602
2603         case WINED3DTOP_ADDSIGNED2X:
2604             mul = 2;
2605             if(strcmp(dstreg, "result.color") == 0) {
2606                 dstreg = "ret";
2607                 mul_final_dest = TRUE;
2608             }
2609         case WINED3DTOP_ADDSIGNED:
2610             shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2611             arg2 = "arg2";
2612         case WINED3DTOP_ADD:
2613             shader_addline(buffer, "ADD_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2614             break;
2615
2616         case WINED3DTOP_SUBTRACT:
2617             shader_addline(buffer, "SUB_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2618             break;
2619
2620         case WINED3DTOP_ADDSMOOTH:
2621             shader_addline(buffer, "SUB arg1, const.x, %s;\n", arg1);
2622             shader_addline(buffer, "MAD_SAT %s%s, arg1, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2623             break;
2624
2625         case WINED3DTOP_BLENDCURRENTALPHA:
2626             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_CURRENT);
2627             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2628             break;
2629         case WINED3DTOP_BLENDFACTORALPHA:
2630             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TFACTOR);
2631             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2632             break;
2633         case WINED3DTOP_BLENDTEXTUREALPHA:
2634             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2635             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2636             break;
2637         case WINED3DTOP_BLENDDIFFUSEALPHA:
2638             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_DIFFUSE);
2639             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2640             break;
2641
2642         case WINED3DTOP_BLENDTEXTUREALPHAPM:
2643             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2644             shader_addline(buffer, "SUB arg0.w, const.x, %s.w;\n", arg0);
2645             shader_addline(buffer, "MAD_SAT %s%s, %s, arg0.w, %s;\n", dstreg, dstmask, arg2, arg1);
2646             break;
2647
2648         /* D3DTOP_PREMODULATE ???? */
2649
2650         case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
2651             shader_addline(buffer, "SUB arg0.w, const.x, %s;\n", arg1);
2652             shader_addline(buffer, "MAD_SAT %s%s, arg0.w, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2653             break;
2654         case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
2655             shader_addline(buffer, "MAD_SAT %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg1);
2656             break;
2657         case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
2658             shader_addline(buffer, "SUB arg0, const.x, %s;\n", arg1);
2659             shader_addline(buffer, "MAD_SAT %s%s, arg0, %s, %s.w;\n", dstreg, dstmask, arg2, arg1);
2660             break;
2661         case WINED3DTOP_MODULATECOLOR_ADDALPHA:
2662             shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s.w;\n", dstreg, dstmask, arg1, arg2, arg1);
2663             break;
2664
2665         case WINED3DTOP_DOTPRODUCT3:
2666             mul = 4;
2667             if(strcmp(dstreg, "result.color") == 0) {
2668                 dstreg = "ret";
2669                 mul_final_dest = TRUE;
2670             }
2671             shader_addline(buffer, "SUB arg1, %s, const.w;\n", arg1);
2672             shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2673             shader_addline(buffer, "DP3_SAT %s%s, arg1, arg2;\n", dstreg, dstmask);
2674             break;
2675
2676         case WINED3DTOP_MULTIPLYADD:
2677             shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg0);
2678             break;
2679
2680         case WINED3DTOP_LERP:
2681             /* The msdn is not quite right here */
2682             shader_addline(buffer, "LRP %s%s, %s, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2683             break;
2684
2685         case WINED3DTOP_BUMPENVMAP:
2686         case WINED3DTOP_BUMPENVMAPLUMINANCE:
2687             /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
2688             break;
2689
2690         default:
2691             FIXME("Unhandled texture op %08x\n", op);
2692     }
2693
2694     if(mul == 2) {
2695         shader_addline(buffer, "MUL_SAT %s%s, %s, const.y;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2696     } else if(mul == 4) {
2697         shader_addline(buffer, "MUL_SAT %s%s, %s, const.z;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2698     }
2699 }
2700
2701 /* The stateblock is passed for GLINFO_LOCATION */
2702 static GLuint gen_arbfp_ffp_shader(const struct ffp_frag_settings *settings, IWineD3DStateBlockImpl *stateblock)
2703 {
2704     unsigned int stage;
2705     SHADER_BUFFER buffer;
2706     BOOL tex_read[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2707     BOOL bump_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2708     BOOL luminance_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2709     const char *textype;
2710     const char *instr, *sat;
2711     char colorcor_dst[8];
2712     GLuint ret;
2713     DWORD arg0, arg1, arg2;
2714     BOOL tempreg_used = FALSE, tfactor_used = FALSE;
2715     BOOL op_equal;
2716     const char *final_combiner_src = "ret";
2717
2718     /* Find out which textures are read */
2719     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2720         if(settings->op[stage].cop == WINED3DTOP_DISABLE) break;
2721         arg0 = settings->op[stage].carg0 & WINED3DTA_SELECTMASK;
2722         arg1 = settings->op[stage].carg1 & WINED3DTA_SELECTMASK;
2723         arg2 = settings->op[stage].carg2 & WINED3DTA_SELECTMASK;
2724         if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2725         if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2726         if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2727
2728         if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) tex_read[stage] = TRUE;
2729         if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHAPM) tex_read[stage] = TRUE;
2730         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP) {
2731             bump_used[stage] = TRUE;
2732             tex_read[stage] = TRUE;
2733         }
2734         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2735             bump_used[stage] = TRUE;
2736             tex_read[stage] = TRUE;
2737             luminance_used[stage] = TRUE;
2738         } else if(settings->op[stage].cop == WINED3DTOP_BLENDFACTORALPHA) {
2739             tfactor_used = TRUE;
2740         }
2741
2742         if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2743             tfactor_used = TRUE;
2744         }
2745
2746         if(settings->op[stage].dst == tempreg) tempreg_used = TRUE;
2747         if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2748             tempreg_used = TRUE;
2749         }
2750
2751         if(settings->op[stage].aop == WINED3DTOP_DISABLE) continue;
2752         arg0 = settings->op[stage].aarg0 & WINED3DTA_SELECTMASK;
2753         arg1 = settings->op[stage].aarg1 & WINED3DTA_SELECTMASK;
2754         arg2 = settings->op[stage].aarg2 & WINED3DTA_SELECTMASK;
2755         if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2756         if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2757         if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2758
2759         if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2760             tempreg_used = TRUE;
2761         }
2762         if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2763             tfactor_used = TRUE;
2764         }
2765     }
2766
2767     /* Shader header */
2768     shader_buffer_init(&buffer);
2769
2770     shader_addline(&buffer, "!!ARBfp1.0\n");
2771
2772     switch(settings->fog) {
2773         case FOG_OFF:                                                         break;
2774         case FOG_LINEAR: shader_addline(&buffer, "OPTION ARB_fog_linear;\n"); break;
2775         case FOG_EXP:    shader_addline(&buffer, "OPTION ARB_fog_exp;\n");    break;
2776         case FOG_EXP2:   shader_addline(&buffer, "OPTION ARB_fog_exp2;\n");   break;
2777         default: FIXME("Unexpected fog setting %d\n", settings->fog);
2778     }
2779
2780     shader_addline(&buffer, "PARAM const = {1, 2, 4, 0.5};\n");
2781     shader_addline(&buffer, "TEMP TMP;\n");
2782     shader_addline(&buffer, "TEMP ret;\n");
2783     if(tempreg_used || settings->sRGB_write) shader_addline(&buffer, "TEMP tempreg;\n");
2784     shader_addline(&buffer, "TEMP arg0;\n");
2785     shader_addline(&buffer, "TEMP arg1;\n");
2786     shader_addline(&buffer, "TEMP arg2;\n");
2787     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2788         if(!tex_read[stage]) continue;
2789         shader_addline(&buffer, "TEMP tex%u;\n", stage);
2790         if(!bump_used[stage]) continue;
2791         shader_addline(&buffer, "PARAM bumpmat%u = program.env[%u];\n", stage, ARB_FFP_CONST_BUMPMAT(stage));
2792         if(!luminance_used[stage]) continue;
2793         shader_addline(&buffer, "PARAM luminance%u = program.env[%u];\n", stage, ARB_FFP_CONST_LUMINANCE(stage));
2794     }
2795     if(tfactor_used) {
2796         shader_addline(&buffer, "PARAM tfactor = program.env[%u];\n", ARB_FFP_CONST_TFACTOR);
2797     }
2798         shader_addline(&buffer, "PARAM specular_enable = program.env[%u];\n", ARB_FFP_CONST_SPECULAR_ENABLE);
2799
2800     if(settings->sRGB_write) {
2801         shader_addline(&buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n",
2802                        srgb_mul_low, srgb_mul_low, srgb_mul_low);
2803         shader_addline(&buffer, "PARAM srgb_comparison =  {%f, %f, %f, %f};\n",
2804                        srgb_cmp, srgb_cmp, srgb_cmp, srgb_cmp);
2805         shader_addline(&buffer, "PARAM srgb_pow =  {%f, %f, %f, 1.0};\n",
2806                        srgb_pow, srgb_pow, srgb_pow);
2807         shader_addline(&buffer, "PARAM srgb_mul_hi =  {%f, %f, %f, 1.0};\n",
2808                        srgb_mul_high, srgb_mul_high, srgb_mul_high);
2809         shader_addline(&buffer, "PARAM srgb_sub_hi =  {%f, %f, %f, 0.0};\n",
2810                        srgb_sub_high, srgb_sub_high, srgb_sub_high);
2811     }
2812
2813     /* Generate texture sampling instructions) */
2814     for(stage = 0; stage < MAX_TEXTURES && settings->op[stage].cop != WINED3DTOP_DISABLE; stage++) {
2815         if(!tex_read[stage]) continue;
2816
2817         switch(settings->op[stage].tex_type) {
2818             case tex_1d:                    textype = "1D";     break;
2819             case tex_2d:                    textype = "2D";     break;
2820             case tex_3d:                    textype = "3D";     break;
2821             case tex_cube:                  textype = "CUBE";   break;
2822             case tex_rect:                  textype = "RECT";   break;
2823             default: textype = "unexpected_textype";   break;
2824         }
2825
2826         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP ||
2827            settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2828             sat = "";
2829         } else {
2830             sat = "_SAT";
2831         }
2832
2833         if(settings->op[stage].projected == proj_none) {
2834             instr = "TEX";
2835         } else if(settings->op[stage].projected == proj_count4 ||
2836                   settings->op[stage].projected == proj_count3) {
2837             instr = "TXP";
2838         } else {
2839             FIXME("Unexpected projection mode %d\n", settings->op[stage].projected);
2840             instr = "TXP";
2841         }
2842
2843         if(stage > 0 &&
2844            (settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
2845             settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
2846             shader_addline(&buffer, "SWZ arg1, bumpmat%u, x, z, 0, 0;\n", stage - 1);
2847             shader_addline(&buffer, "DP3 ret.x, arg1, tex%u;\n", stage - 1);
2848             shader_addline(&buffer, "SWZ arg1, bumpmat%u, y, w, 0, 0;\n", stage - 1);
2849             shader_addline(&buffer, "DP3 ret.y, arg1, tex%u;\n", stage - 1);
2850
2851             /* with projective textures, texbem only divides the static texture coord, not the displacement,
2852              * so multiply the displacement with the dividing parameter before passing it to TXP
2853              */
2854             if (settings->op[stage].projected != proj_none) {
2855                 if(settings->op[stage].projected == proj_count4) {
2856                     shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].w;\n", stage);
2857                     shader_addline(&buffer, "MUL ret.xyz, ret, fragment.texcoord[%u].w, fragment.texcoord[%u];\n", stage, stage);
2858                 } else {
2859                     shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].z;\n", stage);
2860                     shader_addline(&buffer, "MAD ret.xyz, ret, fragment.texcoord[%u].z, fragment.texcoord[%u];\n", stage, stage);
2861                 }
2862             } else {
2863                 shader_addline(&buffer, "ADD ret, ret, fragment.texcoord[%u];\n", stage);
2864             }
2865
2866             shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2867                            instr, sat, stage, stage, textype);
2868             if(settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2869                 shader_addline(&buffer, "MAD_SAT ret.x, tex%u.z, luminance%u.x, luminance%u.y;\n",
2870                                stage - 1, stage - 1, stage - 1);
2871                 shader_addline(&buffer, "MUL tex%u, tex%u, ret.x;\n", stage, stage);
2872             }
2873         } else if(settings->op[stage].projected == proj_count3) {
2874             shader_addline(&buffer, "MOV ret, fragment.texcoord[%u];\n", stage);
2875             shader_addline(&buffer, "MOV ret.w, ret.z;\n");
2876             shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2877                             instr, sat, stage, stage, textype);
2878         } else {
2879             shader_addline(&buffer, "%s%s tex%u, fragment.texcoord[%u], texture[%u], %s;\n",
2880                             instr, sat, stage, stage, stage, textype);
2881         }
2882
2883         sprintf(colorcor_dst, "tex%u", stage);
2884         gen_color_correction(&buffer, colorcor_dst, WINED3DSP_WRITEMASK_ALL, "const.x", "const.y",
2885                 settings->op[stage].color_fixup);
2886     }
2887
2888     /* Generate the main shader */
2889     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2890         if(settings->op[stage].cop == WINED3DTOP_DISABLE) {
2891             if(stage == 0) {
2892                 final_combiner_src = "fragment.color.primary";
2893             }
2894             break;
2895         }
2896
2897         if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2898            settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2899             op_equal = settings->op[stage].carg1 == settings->op[stage].aarg1;
2900         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2901                   settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2902             op_equal = settings->op[stage].carg1 == settings->op[stage].aarg2;
2903         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2904                   settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2905             op_equal = settings->op[stage].carg2 == settings->op[stage].aarg1;
2906         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2907                   settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2908             op_equal = settings->op[stage].carg2 == settings->op[stage].aarg2;
2909         } else {
2910             op_equal = settings->op[stage].aop   == settings->op[stage].cop &&
2911                        settings->op[stage].carg0 == settings->op[stage].aarg0 &&
2912                        settings->op[stage].carg1 == settings->op[stage].aarg1 &&
2913                        settings->op[stage].carg2 == settings->op[stage].aarg2;
2914         }
2915
2916         if(settings->op[stage].aop == WINED3DTOP_DISABLE) {
2917             gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2918                           settings->op[stage].cop, settings->op[stage].carg0,
2919                           settings->op[stage].carg1, settings->op[stage].carg2);
2920             if(stage == 0) {
2921                 shader_addline(&buffer, "MOV ret.w, fragment.color.primary.w;\n");
2922             }
2923         } else if(op_equal) {
2924             gen_ffp_instr(&buffer, stage, TRUE, TRUE, settings->op[stage].dst,
2925                           settings->op[stage].cop, settings->op[stage].carg0,
2926                           settings->op[stage].carg1, settings->op[stage].carg2);
2927         } else {
2928             gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2929                           settings->op[stage].cop, settings->op[stage].carg0,
2930                           settings->op[stage].carg1, settings->op[stage].carg2);
2931             gen_ffp_instr(&buffer, stage, FALSE, TRUE, settings->op[stage].dst,
2932                           settings->op[stage].aop, settings->op[stage].aarg0,
2933                           settings->op[stage].aarg1, settings->op[stage].aarg2);
2934         }
2935     }
2936
2937     if(settings->sRGB_write) {
2938         shader_addline(&buffer, "MAD ret, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
2939         arbfp_add_sRGB_correction(&buffer, "ret", "arg0", "arg1", "arg2", "tempreg");
2940         shader_addline(&buffer, "MOV result.color.w, ret.w;\n");
2941     } else {
2942         shader_addline(&buffer, "MAD result.color, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
2943     }
2944
2945     /* Footer */
2946     shader_addline(&buffer, "END\n");
2947
2948     /* Generate the shader */
2949     GL_EXTCALL(glGenProgramsARB(1, &ret));
2950     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, ret));
2951     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
2952
2953     if (glGetError() == GL_INVALID_OPERATION) {
2954         GLint pos;
2955         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
2956         FIXME("Fragment program error at position %d: %s\n", pos,
2957               debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2958     }
2959     shader_buffer_free(&buffer);
2960     return ret;
2961 }
2962
2963 static void fragment_prog_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2964     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2965     struct shader_arb_priv *priv = (struct shader_arb_priv *) device->fragment_priv;
2966     BOOL use_pshader = use_ps(stateblock);
2967     BOOL use_vshader = use_vs(stateblock);
2968     struct ffp_frag_settings settings;
2969     const struct arbfp_ffp_desc *desc;
2970     unsigned int i;
2971
2972     if(isStateDirty(context, STATE_RENDER(WINED3DRS_FOGENABLE))) {
2973         if(!use_pshader && device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
2974             /* Reload fixed function constants since they collide with the pixel shader constants */
2975             for(i = 0; i < MAX_TEXTURES; i++) {
2976                 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
2977             }
2978             state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
2979             state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
2980         } else if(use_pshader && !isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
2981             device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
2982         }
2983         return;
2984     }
2985
2986     if(!use_pshader) {
2987         /* Find or create a shader implementing the fixed function pipeline settings, then activate it */
2988         gen_ffp_frag_op(stateblock, &settings, FALSE);
2989         desc = (const struct arbfp_ffp_desc *)find_ffp_frag_shader(priv->fragment_shaders, &settings);
2990         if(!desc) {
2991             struct arbfp_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
2992             if (!new_desc)
2993             {
2994                 ERR("Out of memory\n");
2995                 return;
2996             }
2997             new_desc->num_textures_used = 0;
2998             for(i = 0; i < GL_LIMITS(texture_stages); i++) {
2999                 if(settings.op[i].cop == WINED3DTOP_DISABLE) break;
3000                 new_desc->num_textures_used = i;
3001             }
3002
3003             memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
3004             new_desc->shader = gen_arbfp_ffp_shader(&settings, stateblock);
3005             add_ffp_frag_shader(priv->fragment_shaders, &new_desc->parent);
3006             TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
3007             desc = new_desc;
3008         }
3009
3010         /* Now activate the replacement program. GL_FRAGMENT_PROGRAM_ARB is already active(however, note the
3011          * comment above the shader_select call below). If e.g. GLSL is active, the shader_select call will
3012          * deactivate it.
3013          */
3014         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader));
3015         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader)");
3016         priv->current_fprogram_id = desc->shader;
3017
3018         if(device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3019             /* Reload fixed function constants since they collide with the pixel shader constants */
3020             for(i = 0; i < MAX_TEXTURES; i++) {
3021                 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3022             }
3023             state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3024             state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3025         }
3026     }
3027
3028     /* Finally, select the shader. If a pixel shader is used, it will be set and enabled by the shader backend.
3029      * If this shader backend is arbfp(most likely), then it will simply overwrite the last fixed function replace-
3030      * ment shader. If the shader backend is not ARB, it currently is important that the opengl implementation
3031      * type overwrites GL_ARB_fragment_program. This is currently the case with GLSL. If we really want to use
3032      * atifs or nvrc pixel shaders with arb fragment programs we'd have to disable GL_FRAGMENT_PROGRAM_ARB here
3033      *
3034      * Don't call shader_select if the vertex shader is dirty, because some shader backends(GLSL) need both shaders
3035      * to be compiled before activating them(needs some cleanups in the shader backend interface)
3036      */
3037     if(!isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3038         device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3039
3040         if (!isStateDirty(context, STATE_VERTEXSHADERCONSTANT) && (use_vshader || use_pshader)) {
3041             device->StateTable[STATE_VERTEXSHADERCONSTANT].apply(STATE_VERTEXSHADERCONSTANT, stateblock, context);
3042         }
3043     }
3044     if(use_pshader) {
3045         device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
3046     }
3047 }
3048
3049 /* We can't link the fog states to the fragment state directly since the vertex pipeline links them
3050  * to FOGENABLE. A different linking in different pipeline parts can't be expressed in the combined
3051  * state table, so we need to handle that with a forwarding function. The other invisible side effect
3052  * is that changing the fog start and fog end(which links to FOGENABLE in vertex) results in the
3053  * fragment_prog_arbfp function being called because FOGENABLE is dirty, which calls this function here
3054  */
3055 static void state_arbfp_fog(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3056     if(!isStateDirty(context, STATE_PIXELSHADER)) {
3057         fragment_prog_arbfp(state, stateblock, context);
3058     }
3059 }
3060
3061 static void textransform(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3062     if(!isStateDirty(context, STATE_PIXELSHADER)) {
3063         fragment_prog_arbfp(state, stateblock, context);
3064     }
3065 }
3066
3067 #undef GLINFO_LOCATION
3068
3069 static const struct StateEntryTemplate arbfp_fragmentstate_template[] = {
3070     {STATE_RENDER(WINED3DRS_TEXTUREFACTOR),               { STATE_RENDER(WINED3DRS_TEXTUREFACTOR),              state_texfactor_arbfp   }, 0                               },
3071     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3072     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3073     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3074     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3075     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3076     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3077     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3078     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3079     {STATE_TEXTURESTAGE(0, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3080     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3081     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3082     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3083     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3084     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3085     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3086     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3087     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3088     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3089     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3090     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3091     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3092     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3093     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3094     {STATE_TEXTURESTAGE(1, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3095     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3096     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3097     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3098     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3099     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3100     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3101     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3102     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3103     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3104     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3105     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3106     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3107     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3108     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3109     {STATE_TEXTURESTAGE(2, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3110     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3111     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3112     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3113     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3114     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3115     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3116     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3117     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3118     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3119     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3120     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3121     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3122     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3123     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3124     {STATE_TEXTURESTAGE(3, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3125     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3126     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3127     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3128     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3129     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3130     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3131     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3132     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3133     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3134     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3135     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3136     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3137     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3138     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3139     {STATE_TEXTURESTAGE(4, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3140     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3141     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3142     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3143     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3144     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3145     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3146     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3147     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3148     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3149     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3150     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3151     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3152     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3153     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3154     {STATE_TEXTURESTAGE(5, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3155     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3156     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3157     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3158     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3159     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3160     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3161     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3162     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3163     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3164     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3165     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3166     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3167     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3168     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3169     {STATE_TEXTURESTAGE(6, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3170     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3171     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3172     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3173     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3174     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3175     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3176     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3177     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3178     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3179     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3180     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3181     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3182     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3183     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3184     {STATE_TEXTURESTAGE(7, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3185     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3186     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3187     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3188     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3189     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3190     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3191     { STATE_SAMPLER(0),                                   { STATE_SAMPLER(0),                                   sampler_texdim          }, 0                               },
3192     { STATE_SAMPLER(1),                                   { STATE_SAMPLER(1),                                   sampler_texdim          }, 0                               },
3193     { STATE_SAMPLER(2),                                   { STATE_SAMPLER(2),                                   sampler_texdim          }, 0                               },
3194     { STATE_SAMPLER(3),                                   { STATE_SAMPLER(3),                                   sampler_texdim          }, 0                               },
3195     { STATE_SAMPLER(4),                                   { STATE_SAMPLER(4),                                   sampler_texdim          }, 0                               },
3196     { STATE_SAMPLER(5),                                   { STATE_SAMPLER(5),                                   sampler_texdim          }, 0                               },
3197     { STATE_SAMPLER(6),                                   { STATE_SAMPLER(6),                                   sampler_texdim          }, 0                               },
3198     { STATE_SAMPLER(7),                                   { STATE_SAMPLER(7),                                   sampler_texdim          }, 0                               },
3199     { STATE_PIXELSHADER,                                  { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3200     { STATE_RENDER(WINED3DRS_FOGENABLE),                  { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3201     { STATE_RENDER(WINED3DRS_FOGTABLEMODE),               { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3202     { STATE_RENDER(WINED3DRS_FOGVERTEXMODE),              { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3203     { STATE_RENDER(WINED3DRS_SRGBWRITEENABLE),            { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3204     {STATE_TEXTURESTAGE(0,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(0, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3205     {STATE_TEXTURESTAGE(1,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(1, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3206     {STATE_TEXTURESTAGE(2,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(2, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3207     {STATE_TEXTURESTAGE(3,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(3, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3208     {STATE_TEXTURESTAGE(4,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(4, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3209     {STATE_TEXTURESTAGE(5,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(5, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3210     {STATE_TEXTURESTAGE(6,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(6, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3211     {STATE_TEXTURESTAGE(7,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(7, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3212     { STATE_RENDER(WINED3DRS_SPECULARENABLE),             { STATE_RENDER(WINED3DRS_SPECULARENABLE),             state_arb_specularenable}, 0                               },
3213     {0 /* Terminate */,                                   { 0,                                                  0                       }, 0                               },
3214 };
3215
3216 const struct fragment_pipeline arbfp_fragment_pipeline = {
3217     arbfp_enable,
3218     arbfp_get_caps,
3219     arbfp_alloc,
3220     arbfp_free,
3221     shader_arb_color_fixup_supported,
3222     arbfp_fragmentstate_template,
3223     TRUE /* We can disable projected textures */
3224 };
3225
3226 #define GLINFO_LOCATION device->adapter->gl_info
3227
3228 struct arbfp_blit_priv {
3229     GLenum yuy2_rect_shader, yuy2_2d_shader;
3230     GLenum uyvy_rect_shader, uyvy_2d_shader;
3231     GLenum yv12_rect_shader, yv12_2d_shader;
3232 };
3233
3234 static HRESULT arbfp_blit_alloc(IWineD3DDevice *iface) {
3235     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3236     device->blit_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct arbfp_blit_priv));
3237     if(!device->blit_priv) {
3238         ERR("Out of memory\n");
3239         return E_OUTOFMEMORY;
3240     }
3241     return WINED3D_OK;
3242 }
3243 static void arbfp_blit_free(IWineD3DDevice *iface) {
3244     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3245     struct arbfp_blit_priv *priv = (struct arbfp_blit_priv *) device->blit_priv;
3246
3247     ENTER_GL();
3248     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_rect_shader));
3249     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_2d_shader));
3250     GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_rect_shader));
3251     GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_2d_shader));
3252     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_rect_shader));
3253     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_2d_shader));
3254     checkGLcall("Delete yuv programs\n");
3255     LEAVE_GL();
3256 }
3257
3258 static BOOL gen_planar_yuv_read(SHADER_BUFFER *buffer, enum yuv_fixup yuv_fixup, GLenum textype, char *luminance)
3259 {
3260     char chroma;
3261     const char *tex, *texinstr;
3262
3263     if (yuv_fixup == YUV_FIXUP_UYVY) {
3264         chroma = 'x';
3265         *luminance = 'w';
3266     } else {
3267         chroma = 'w';
3268         *luminance = 'x';
3269     }
3270     switch(textype) {
3271         case GL_TEXTURE_2D:             tex = "2D";     texinstr = "TXP"; break;
3272         case GL_TEXTURE_RECTANGLE_ARB:  tex = "RECT";   texinstr = "TEX"; break;
3273         default:
3274             /* This is more tricky than just replacing the texture type - we have to navigate
3275              * properly in the texture to find the correct chroma values
3276              */
3277             FIXME("Implement yuv correction for non-2d, non-rect textures\n");
3278             return FALSE;
3279     }
3280
3281     /* First we have to read the chroma values. This means we need at least two pixels(no filtering),
3282      * or 4 pixels(with filtering). To get the unmodified chromas, we have to rid ourselves of the
3283      * filtering when we sample the texture.
3284      *
3285      * These are the rules for reading the chroma:
3286      *
3287      * Even pixel: Cr
3288      * Even pixel: U
3289      * Odd pixel: V
3290      *
3291      * So we have to get the sampling x position in non-normalized coordinates in integers
3292      */
3293     if(textype != GL_TEXTURE_RECTANGLE_ARB) {
3294         shader_addline(buffer, "MUL texcrd.xy, fragment.texcoord[0], size.x;\n");
3295         shader_addline(buffer, "MOV texcrd.w, size.x;\n");
3296     } else {
3297         shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3298     }
3299     /* We must not allow filtering between pixel x and x+1, this would mix U and V
3300      * Vertical filtering is ok. However, bear in mind that the pixel center is at
3301      * 0.5, so add 0.5.
3302      */
3303     shader_addline(buffer, "FLR texcrd.x, texcrd.x;\n");
3304     shader_addline(buffer, "ADD texcrd.x, texcrd.x, coef.y;\n");
3305
3306     /* Divide the x coordinate by 0.5 and get the fraction. This gives 0.25 and 0.75 for the
3307      * even and odd pixels respectively
3308      */
3309     shader_addline(buffer, "MUL texcrd2, texcrd, coef.y;\n");
3310     shader_addline(buffer, "FRC texcrd2, texcrd2;\n");
3311
3312     /* Sample Pixel 1 */
3313     shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3314
3315     /* Put the value into either of the chroma values */
3316     shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3317     shader_addline(buffer, "MUL chroma.x, luminance.%c, temp.x;\n", chroma);
3318     shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3319     shader_addline(buffer, "MUL chroma.y, luminance.%c, temp.x;\n", chroma);
3320
3321     /* Sample pixel 2. If we read an even pixel(SLT above returned 1), sample
3322      * the pixel right to the current one. Otherwise, sample the left pixel.
3323      * Bias and scale the SLT result to -1;1 and add it to the texcrd.x.
3324      */
3325     shader_addline(buffer, "MAD temp.x, temp.x, coef.z, -coef.x;\n");
3326     shader_addline(buffer, "ADD texcrd.x, texcrd, temp.x;\n");
3327     shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3328
3329     /* Put the value into the other chroma */
3330     shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3331     shader_addline(buffer, "MAD chroma.y, luminance.%c, temp.x, chroma.y;\n", chroma);
3332     shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3333     shader_addline(buffer, "MAD chroma.x, luminance.%c, temp.x, chroma.x;\n", chroma);
3334
3335     /* TODO: If filtering is enabled, sample a 2nd pair of pixels left or right of
3336      * the current one and lerp the two U and V values
3337      */
3338
3339     /* This gives the correctly filtered luminance value */
3340     shader_addline(buffer, "TEX luminance, fragment.texcoord[0], texture[0], %s;\n", tex);
3341
3342     return TRUE;
3343 }
3344
3345 static BOOL gen_yv12_read(SHADER_BUFFER *buffer, GLenum textype, char *luminance)
3346 {
3347     const char *tex;
3348
3349     switch(textype) {
3350         case GL_TEXTURE_2D:             tex = "2D";     break;
3351         case GL_TEXTURE_RECTANGLE_ARB:  tex = "RECT";   break;
3352         default:
3353             FIXME("Implement yv12 correction for non-2d, non-rect textures\n");
3354             return FALSE;
3355     }
3356
3357     /* YV12 surfaces contain a WxH sized luminance plane, followed by a (W/2)x(H/2)
3358      * V and a (W/2)x(H/2) U plane, each with 8 bit per pixel. So the effective
3359      * bitdepth is 12 bits per pixel. Since the U and V planes have only half the
3360      * pitch of the luminance plane, the packing into the gl texture is a bit
3361      * unfortunate. If the whole texture is interpreted as luminance data it looks
3362      * approximately like this:
3363      *
3364      *        +----------------------------------+----
3365      *        |                                  |
3366      *        |                                  |
3367      *        |                                  |
3368      *        |                                  |
3369      *        |                                  |   2
3370      *        |            LUMINANCE             |   -
3371      *        |                                  |   3
3372      *        |                                  |
3373      *        |                                  |
3374      *        |                                  |
3375      *        |                                  |
3376      *        +----------------+-----------------+----
3377      *        |                |                 |
3378      *        |  U even rows   |  U odd rows     |
3379      *        |                |                 |   1
3380      *        +----------------+------------------   -
3381      *        |                |                 |   3
3382      *        |  V even rows   |  V odd rows     |
3383      *        |                |                 |
3384      *        +----------------+-----------------+----
3385      *        |                |                 |
3386      *        |     0.5        |       0.5       |
3387      *
3388      * So it appears as if there are 4 chroma images, but in fact the odd rows
3389      * in the chroma images are in the same row as the even ones. So its is
3390      * kinda tricky to read
3391      *
3392      * When reading from rectangle textures, keep in mind that the input y coordinates
3393      * go from 0 to d3d_height, whereas the opengl texture height is 1.5 * d3d_height
3394      */
3395     shader_addline(buffer, "PARAM yv12_coef = {%f, %f, %f, %f};\n",
3396                    2.0 / 3.0, 1.0 / 6.0, (2.0 / 3.0) + (1.0 / 6.0), 1.0 / 3.0);
3397
3398     shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3399     /* the chroma planes have only half the width */
3400     shader_addline(buffer, "MUL texcrd.x, texcrd.x, coef.y;\n");
3401
3402     /* The first value is between 2/3 and 5/6th of the texture's height, so scale+bias
3403      * the coordinate. Also read the right side of the image when reading odd lines
3404      *
3405      * Don't forget to clamp the y values in into the range, otherwise we'll get filtering
3406      * bleeding
3407      */
3408     if(textype == GL_TEXTURE_2D) {
3409
3410         shader_addline(buffer, "RCP chroma.w, size.y;\n");
3411
3412         shader_addline(buffer, "MUL texcrd2.y, texcrd.y, size.y;\n");
3413
3414         shader_addline(buffer, "FLR texcrd2.y, texcrd2.y;\n");
3415         shader_addline(buffer, "MAD texcrd.y, texcrd.y, yv12_coef.y, yv12_coef.x;\n");
3416
3417         /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3418         shader_addline(buffer, "ADD texcrd2.x, texcrd2.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3419         shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3420         shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3421         shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3422
3423         /* clamp, keep the half pixel origin in mind */
3424         shader_addline(buffer, "MAD temp.y, coef.y, chroma.w, yv12_coef.x;\n");
3425         shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3426         shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.z;\n");
3427         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3428     } else {
3429         /* Read from [size - size+size/4] */
3430         shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3431         shader_addline(buffer, "MAD texcrd.y, texcrd.y, coef.w, size.y;\n");
3432
3433         /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3434         shader_addline(buffer, "ADD texcrd2.x, texcrd.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3435         shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3436         shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3437         shader_addline(buffer, "MUL texcrd2.x, texcrd2.x, size.x;\n");
3438         shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3439
3440         /* Make sure to read exactly from the pixel center */
3441         shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3442         shader_addline(buffer, "ADD texcrd.y, texcrd.y, coef.y;\n");
3443
3444         /* Clamp */
3445         shader_addline(buffer, "MAD temp.y, size.y, coef.w, size.y;\n");
3446         shader_addline(buffer, "ADD temp.y, temp.y, -coef.y;\n");
3447         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3448         shader_addline(buffer, "ADD temp.y, size.y, -coef.y;\n");
3449         shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3450     }
3451     /* Read the texture, put the result into the output register */
3452     shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3453     shader_addline(buffer, "MOV chroma.x, temp.w;\n");
3454
3455     /* The other chroma value is 1/6th of the texture lower, from 5/6th to 6/6th
3456      * No need to clamp because we're just reusing the already clamped value from above
3457      */
3458     if(textype == GL_TEXTURE_2D) {
3459         shader_addline(buffer, "ADD texcrd.y, texcrd.y, yv12_coef.y;\n");
3460     } else {
3461         shader_addline(buffer, "MAD texcrd.y, size.y, coef.w, texcrd.y;\n");
3462     }
3463     shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3464     shader_addline(buffer, "MOV chroma.y, temp.w;\n");
3465
3466     /* Sample the luminance value. It is in the top 2/3rd of the texture, so scale the y coordinate.
3467      * Clamp the y coordinate to prevent the chroma values from bleeding into the sampled luminance
3468      * values due to filtering
3469      */
3470     shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3471     if(textype == GL_TEXTURE_2D) {
3472         /* Multiply the y coordinate by 2/3 and clamp it */
3473         shader_addline(buffer, "MUL texcrd.y, texcrd.y, yv12_coef.x;\n");
3474         shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.x;\n");
3475         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3476         shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3477     } else {
3478         /* Reading from texture_rectangles is pretty straightforward, just use the unmodified
3479          * texture coordinate. It is still a good idea to clamp it though, since the opengl texture
3480          * is bigger
3481          */
3482         shader_addline(buffer, "ADD temp.x, size.y, -coef.y;\n");
3483         shader_addline(buffer, "MIN texcrd.y, texcrd.y, size.x;\n");
3484         shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3485     }
3486     *luminance = 'a';
3487
3488     return TRUE;
3489 }
3490
3491 static GLuint gen_yuv_shader(IWineD3DDeviceImpl *device, enum yuv_fixup yuv_fixup, GLenum textype)
3492 {
3493     GLenum shader;
3494     SHADER_BUFFER buffer;
3495     char luminance_component;
3496     struct arbfp_blit_priv *priv = (struct arbfp_blit_priv *) device->blit_priv;
3497
3498     /* Shader header */
3499     shader_buffer_init(&buffer);
3500
3501     ENTER_GL();
3502     GL_EXTCALL(glGenProgramsARB(1, &shader));
3503     checkGLcall("GL_EXTCALL(glGenProgramsARB(1, &shader))");
3504     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3505     checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3506     LEAVE_GL();
3507     if(!shader) {
3508         shader_buffer_free(&buffer);
3509         return 0;
3510     }
3511
3512     /* The YUY2 and UYVY formats contain two pixels packed into a 32 bit macropixel,
3513      * giving effectively 16 bit per pixel. The color consists of a luminance(Y) and
3514      * two chroma(U and V) values. Each macropixel has two luminance values, one for
3515      * each single pixel it contains, and one U and one V value shared between both
3516      * pixels.
3517      *
3518      * The data is loaded into an A8L8 texture. With YUY2, the luminance component
3519      * contains the luminance and alpha the chroma. With UYVY it is vice versa. Thus
3520      * take the format into account when generating the read swizzles
3521      *
3522      * Reading the Y value is straightforward - just sample the texture. The hardware
3523      * takes care of filtering in the horizontal and vertical direction.
3524      *
3525      * Reading the U and V values is harder. We have to avoid filtering horizontally,
3526      * because that would mix the U and V values of one pixel or two adjacent pixels.
3527      * Thus floor the texture coordinate and add 0.5 to get an unfiltered read,
3528      * regardless of the filtering setting. Vertical filtering works automatically
3529      * though - the U and V values of two rows are mixed nicely.
3530      *
3531      * Appart of avoiding filtering issues, the code has to know which value it just
3532      * read, and where it can find the other one. To determine this, it checks if
3533      * it sampled an even or odd pixel, and shifts the 2nd read accordingly.
3534      *
3535      * Handling horizontal filtering of U and V values requires reading a 2nd pair
3536      * of pixels, extracting U and V and mixing them. This is not implemented yet.
3537      *
3538      * An alternative implementation idea is to load the texture as A8R8G8B8 texture,
3539      * with width / 2. This way one read gives all 3 values, finding U and V is easy
3540      * in an unfiltered situation. Finding the luminance on the other hand requires
3541      * finding out if it is an odd or even pixel. The real drawback of this approach
3542      * is filtering. This would have to be emulated completely in the shader, reading
3543      * up two 2 packed pixels in up to 2 rows and interpolating both horizontally and
3544      * vertically. Beyond that it would require adjustments to the texture handling
3545      * code to deal with the width scaling
3546      */
3547     shader_addline(&buffer, "!!ARBfp1.0\n");
3548     shader_addline(&buffer, "TEMP luminance;\n");
3549     shader_addline(&buffer, "TEMP temp;\n");
3550     shader_addline(&buffer, "TEMP chroma;\n");
3551     shader_addline(&buffer, "TEMP texcrd;\n");
3552     shader_addline(&buffer, "TEMP texcrd2;\n");
3553     shader_addline(&buffer, "PARAM coef = {1.0, 0.5, 2.0, 0.25};\n");
3554     shader_addline(&buffer, "PARAM yuv_coef = {1.403, 0.344, 0.714, 1.770};\n");
3555     shader_addline(&buffer, "PARAM size = program.local[0];\n");
3556
3557     switch (yuv_fixup)
3558     {
3559         case YUV_FIXUP_UYVY:
3560         case YUV_FIXUP_YUY2:
3561             if (!gen_planar_yuv_read(&buffer, yuv_fixup, textype, &luminance_component))
3562             {
3563                 shader_buffer_free(&buffer);
3564                 return 0;
3565             }
3566             break;
3567
3568         case YUV_FIXUP_YV12:
3569             if (!gen_yv12_read(&buffer, textype, &luminance_component))
3570             {
3571                 shader_buffer_free(&buffer);
3572                 return 0;
3573             }
3574             break;
3575
3576         default:
3577             FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3578             shader_buffer_free(&buffer);
3579             return 0;
3580     }
3581
3582     /* Calculate the final result. Formula is taken from
3583      * http://www.fourcc.org/fccyvrgb.php. Note that the chroma
3584      * ranges from -0.5 to 0.5
3585      */
3586     shader_addline(&buffer, "SUB chroma.xy, chroma, coef.y;\n");
3587
3588     shader_addline(&buffer, "MAD result.color.x, chroma.x, yuv_coef.x, luminance.%c;\n", luminance_component);
3589     shader_addline(&buffer, "MAD temp.x, -chroma.y, yuv_coef.y, luminance.%c;\n", luminance_component);
3590     shader_addline(&buffer, "MAD result.color.y, -chroma.x, yuv_coef.z, temp.x;\n");
3591     shader_addline(&buffer, "MAD result.color.z, chroma.y, yuv_coef.w, luminance.%c;\n", luminance_component);
3592     shader_addline(&buffer, "END\n");
3593
3594     ENTER_GL();
3595     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3596
3597     if (glGetError() == GL_INVALID_OPERATION) {
3598         GLint pos;
3599         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3600         FIXME("Fragment program error at position %d: %s\n", pos,
3601               debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3602     }
3603     shader_buffer_free(&buffer);
3604     LEAVE_GL();
3605
3606     switch (yuv_fixup)
3607     {
3608         case YUV_FIXUP_YUY2:
3609             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yuy2_rect_shader = shader;
3610             else priv->yuy2_2d_shader = shader;
3611             break;
3612
3613         case YUV_FIXUP_UYVY:
3614             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->uyvy_rect_shader = shader;
3615             else priv->uyvy_2d_shader = shader;
3616             break;
3617
3618         case YUV_FIXUP_YV12:
3619             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yv12_rect_shader = shader;
3620             else priv->yv12_2d_shader = shader;
3621             break;
3622     }
3623
3624     return shader;
3625 }
3626
3627 static HRESULT arbfp_blit_set(IWineD3DDevice *iface, WINED3DFORMAT fmt, GLenum textype, UINT width, UINT height) {
3628     GLenum shader;
3629     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3630     float size[4] = {width, height, 1, 1};
3631     struct arbfp_blit_priv *priv = (struct arbfp_blit_priv *) device->blit_priv;
3632     const struct GlPixelFormatDesc *glDesc;
3633     enum yuv_fixup yuv_fixup;
3634
3635     getFormatDescEntry(fmt, &GLINFO_LOCATION, &glDesc);
3636
3637     if (!is_yuv_fixup(glDesc->color_fixup))
3638     {
3639         TRACE("Fixup:\n");
3640         dump_color_fixup_desc(glDesc->color_fixup);
3641         /* Don't bother setting up a shader for unconverted formats */
3642         ENTER_GL();
3643         glEnable(textype);
3644         checkGLcall("glEnable(textype)");
3645         LEAVE_GL();
3646         return WINED3D_OK;
3647     }
3648
3649     yuv_fixup = get_yuv_fixup(glDesc->color_fixup);
3650
3651     switch(yuv_fixup)
3652     {
3653         case YUV_FIXUP_YUY2:
3654             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yuy2_rect_shader : priv->yuy2_2d_shader;
3655             break;
3656
3657         case YUV_FIXUP_UYVY:
3658             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->uyvy_rect_shader : priv->uyvy_2d_shader;
3659             break;
3660
3661         case YUV_FIXUP_YV12:
3662             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yv12_rect_shader : priv->yv12_2d_shader;
3663             break;
3664
3665         default:
3666             FIXME("Unsupported YUV fixup %#x, not setting a shader\n", yuv_fixup);
3667             ENTER_GL();
3668             glEnable(textype);
3669             checkGLcall("glEnable(textype)");
3670             LEAVE_GL();
3671             return E_NOTIMPL;
3672     }
3673
3674     if (!shader) shader = gen_yuv_shader(device, yuv_fixup, textype);
3675
3676     ENTER_GL();
3677     glEnable(GL_FRAGMENT_PROGRAM_ARB);
3678     checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
3679     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3680     checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3681     GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, 0, size));
3682     checkGLcall("glProgramLocalParameter4fvARB");
3683     LEAVE_GL();
3684
3685     return WINED3D_OK;
3686 }
3687
3688 static void arbfp_blit_unset(IWineD3DDevice *iface) {
3689     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3690
3691     ENTER_GL();
3692     glDisable(GL_FRAGMENT_PROGRAM_ARB);
3693     checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
3694     glDisable(GL_TEXTURE_2D);
3695     checkGLcall("glDisable(GL_TEXTURE_2D)");
3696     if(GL_SUPPORT(ARB_TEXTURE_CUBE_MAP)) {
3697         glDisable(GL_TEXTURE_CUBE_MAP_ARB);
3698         checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
3699     }
3700     if(GL_SUPPORT(ARB_TEXTURE_RECTANGLE)) {
3701         glDisable(GL_TEXTURE_RECTANGLE_ARB);
3702         checkGLcall("glDisable(GL_TEXTURE_RECTANGLE_ARB)");
3703     }
3704     LEAVE_GL();
3705 }
3706
3707 static BOOL arbfp_blit_color_fixup_supported(struct color_fixup_desc fixup)
3708 {
3709     enum yuv_fixup yuv_fixup;
3710
3711     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
3712     {
3713         TRACE("Checking support for fixup:\n");
3714         dump_color_fixup_desc(fixup);
3715     }
3716
3717     if (is_identity_fixup(fixup))
3718     {
3719         TRACE("[OK]\n");
3720         return TRUE;
3721     }
3722
3723     /* We only support YUV conversions. */
3724     if (!is_yuv_fixup(fixup))
3725     {
3726         TRACE("[FAILED]\n");
3727         return FALSE;
3728     }
3729
3730     yuv_fixup = get_yuv_fixup(fixup);
3731     switch(yuv_fixup)
3732     {
3733         case YUV_FIXUP_YUY2:
3734         case YUV_FIXUP_UYVY:
3735         case YUV_FIXUP_YV12:
3736             TRACE("[OK]\n");
3737             return TRUE;
3738
3739         default:
3740             FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3741             TRACE("[FAILED]\n");
3742             return FALSE;
3743     }
3744 }
3745
3746 const struct blit_shader arbfp_blit = {
3747     arbfp_blit_alloc,
3748     arbfp_blit_free,
3749     arbfp_blit_set,
3750     arbfp_blit_unset,
3751     arbfp_blit_color_fixup_supported,
3752 };
3753
3754 #undef GLINFO_LOCATION