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  * Copyright 2009 Henri Verbeet for CodeWeavers
  14  *
  15  * This library is free software; you can redistribute it and/or
  16  * modify it under the terms of the GNU Lesser General Public
  17  * License as published by the Free Software Foundation; either
  18  * version 2.1 of the License, or (at your option) any later version.
  19  *
  20  * This library is distributed in the hope that it will be useful,
  21  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  22  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  23  * Lesser General Public License for more details.
  24  *
  25  * You should have received a copy of the GNU Lesser General Public
  26  * License along with this library; if not, write to the Free Software
  27  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  28  */
  29
  30 #include "config.h"
  31
  32 #include <math.h>
  33 #include <stdio.h>
  34
  35 #include "wined3d_private.h"
  36
  37 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  38 WINE_DECLARE_DEBUG_CHANNEL(d3d_constants);
  39 WINE_DECLARE_DEBUG_CHANNEL(d3d_caps);
  40 WINE_DECLARE_DEBUG_CHANNEL(d3d);
  41
  42 #define GLINFO_LOCATION      (*gl_info)
  43
  44 /* We have to subtract any other PARAMs that we might use in our shader programs.
  45  * ATI seems to count 2 implicit PARAMs when we use fog and NVIDIA counts 1,
  46  * and we reference one row of the PROJECTION matrix which counts as 1 PARAM. */
  47 #define ARB_SHADER_RESERVED_VS_CONSTS 3
  48
  49 /* The arb shader only loads the bump mapping environment matrix into the shader if it finds
  50  * a free constant to do that, so only reduce the number of available constants by 2 for the fog states.
  51  */
  52 #define ARB_SHADER_RESERVED_PS_CONSTS 2
  53
  54 /* Internally used shader constants. Applications can use constants 0 to GL_LIMITS(vshader_constantsF) - 1,
  55  * so upload them above that
  56  */
  57 #define ARB_SHADER_PRIVCONST_BASE (GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS)
  58 #define ARB_SHADER_PRIVCONST_POS ARB_SHADER_PRIVCONST_BASE + 0
  59
  60 /* ARB_program_shader private data */
  61 struct shader_arb_priv {
  62     GLuint                  current_vprogram_id;
  63     GLuint                  current_fprogram_id;
  64     GLuint                  depth_blt_vprogram_id;
  65     GLuint                  depth_blt_fprogram_id[tex_type_count];
  66     BOOL                    use_arbfp_fixed_func;
  67     struct hash_table_t     *fragment_shaders;
  68 };
  69
  70 /********************************************************
  71  * ARB_[vertex/fragment]_program helper functions follow
  72  ********************************************************/
  73
  74 /**
  75  * Loads floating point constants into the currently set ARB_vertex/fragment_program.
  76  * When constant_list == NULL, it will load all the constants.
  77  *
  78  * @target_type should be either GL_VERTEX_PROGRAM_ARB (for vertex shaders)
  79  *  or GL_FRAGMENT_PROGRAM_ARB (for pixel shaders)
  80  */
  81 static unsigned int shader_arb_load_constantsF(IWineD3DBaseShaderImpl* This, const WineD3D_GL_Info *gl_info,
  82         GLuint target_type, unsigned int max_constants, const float *constants, char *dirty_consts)
  83 {
  84     local_constant* lconst;
  85     DWORD i, j;
  86     unsigned int ret;
  87
  88     if (TRACE_ON(d3d_shader)) {
  89         for(i = 0; i < max_constants; i++) {
  90             if(!dirty_consts[i]) continue;
  91             TRACE_(d3d_constants)("Loading constants %i: %f, %f, %f, %f\n", i,
  92                         constants[i * 4 + 0], constants[i * 4 + 1],
  93                         constants[i * 4 + 2], constants[i * 4 + 3]);
  94         }
  95     }
  96     /* In 1.X pixel shaders constants are implicitly clamped in the range [-1;1] */
  97     if (target_type == GL_FRAGMENT_PROGRAM_ARB && This->baseShader.reg_maps.shader_version.major == 1)
  98     {
  99         float lcl_const[4];
 100         for(i = 0; i < max_constants; i++) {
 101             if(!dirty_consts[i]) continue;
 102             dirty_consts[i] = 0;
 103
 104             j = 4 * i;
 105             if(constants[j + 0] > 1.0) lcl_const[0] = 1.0;
 106             else if(constants[j + 0] < -1.0) lcl_const[0] = -1.0;
 107             else lcl_const[0] = constants[j + 0];
 108
 109             if(constants[j + 1] > 1.0) lcl_const[1] = 1.0;
 110             else if(constants[j + 1] < -1.0) lcl_const[1] = -1.0;
 111             else lcl_const[1] = constants[j + 1];
 112
 113             if(constants[j + 2] > 1.0) lcl_const[2] = 1.0;
 114             else if(constants[j + 2] < -1.0) lcl_const[2] = -1.0;
 115             else lcl_const[2] = constants[j + 2];
 116
 117             if(constants[j + 3] > 1.0) lcl_const[3] = 1.0;
 118             else if(constants[j + 3] < -1.0) lcl_const[3] = -1.0;
 119             else lcl_const[3] = constants[j + 3];
 120
 121             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, lcl_const));
 122         }
 123     } else {
 124         if(GL_SUPPORT(EXT_GPU_PROGRAM_PARAMETERS)) {
 125             /* TODO: Benchmark if we're better of with finding the dirty constants ourselves,
 126              * or just reloading *all* constants at once
 127              *
 128             GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, 0, max_constants, constants));
 129              */
 130             for(i = 0; i < max_constants; i++) {
 131                 if(!dirty_consts[i]) continue;
 132
 133                 /* Find the next block of dirty constants */
 134                 dirty_consts[i] = 0;
 135                 j = i;
 136                 for(i++; (i < max_constants) && dirty_consts[i]; i++) {
 137                     dirty_consts[i] = 0;
 138                 }
 139
 140                 GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, j, i - j, constants + (j * 4)));
 141             }
 142         } else {
 143             for(i = 0; i < max_constants; i++) {
 144                 if(dirty_consts[i]) {
 145                     dirty_consts[i] = 0;
 146                     GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, constants + (i * 4)));
 147                 }
 148             }
 149         }
 150     }
 151     checkGLcall("glProgramEnvParameter4fvARB()");
 152
 153     /* Load immediate constants */
 154     if(This->baseShader.load_local_constsF) {
 155         if (TRACE_ON(d3d_shader)) {
 156             LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 157                 GLfloat* values = (GLfloat*)lconst->value;
 158                 TRACE_(d3d_constants)("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
 159                         values[0], values[1], values[2], values[3]);
 160             }
 161         }
 162         /* Immediate constants are clamped for 1.X shaders at loading times */
 163         ret = 0;
 164         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 165             dirty_consts[lconst->idx] = 1; /* Dirtify so the non-immediate constant overwrites it next time */
 166             ret = max(ret, lconst->idx + 1);
 167             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, lconst->idx, (GLfloat*)lconst->value));
 168         }
 169         checkGLcall("glProgramEnvParameter4fvARB()");
 170         return ret; /* The loaded immediate constants need reloading for the next shader */
 171     } else {
 172         return 0; /* No constants are dirty now */
 173     }
 174 }
 175
 176 /**
 177  * Loads the texture dimensions for NP2 fixup into the currently set ARB_[vertex/fragment]_programs.
 178  */
 179 static void shader_arb_load_np2fixup_constants(
 180     IWineD3DDevice* device,
 181     char usePixelShader,
 182     char useVertexShader) {
 183     /* not implemented */
 184 }
 185
 186 /**
 187  * Loads the app-supplied constants into the currently set ARB_[vertex/fragment]_programs.
 188  *
 189  * We only support float constants in ARB at the moment, so don't
 190  * worry about the Integers or Booleans
 191  */
 192 static void shader_arb_load_constants(
 193     IWineD3DDevice* device,
 194     char usePixelShader,
 195     char useVertexShader) {
 196
 197     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
 198     IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
 199     const WineD3D_GL_Info *gl_info = &deviceImpl->adapter->gl_info;
 200     unsigned char i;
 201
 202     if (useVertexShader) {
 203         IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
 204
 205         /* Load DirectX 9 float constants for vertex shader */
 206         deviceImpl->highest_dirty_vs_const = shader_arb_load_constantsF(
 207                 vshader, gl_info, GL_VERTEX_PROGRAM_ARB,
 208                 deviceImpl->highest_dirty_vs_const,
 209                 stateBlock->vertexShaderConstantF,
 210                 deviceImpl->activeContext->vshader_const_dirty);
 211
 212         /* Upload the position fixup */
 213         GL_EXTCALL(glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, ARB_SHADER_PRIVCONST_POS, deviceImpl->posFixup));
 214     }
 215
 216     if (usePixelShader) {
 217
 218         IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
 219         IWineD3DPixelShaderImpl *psi = (IWineD3DPixelShaderImpl *) pshader;
 220
 221         /* Load DirectX 9 float constants for pixel shader */
 222         deviceImpl->highest_dirty_ps_const = shader_arb_load_constantsF(
 223                 pshader, gl_info, GL_FRAGMENT_PROGRAM_ARB,
 224                 deviceImpl->highest_dirty_ps_const,
 225                 stateBlock->pixelShaderConstantF,
 226                 deviceImpl->activeContext->pshader_const_dirty);
 227
 228         for(i = 0; i < psi->numbumpenvmatconsts; i++) {
 229             /* The state manager takes care that this function is always called if the bump env matrix changes
 230              */
 231             const float *data = (const float *)&stateBlock->textureState[(int) psi->bumpenvmatconst[i].texunit][WINED3DTSS_BUMPENVMAT00];
 232             GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->bumpenvmatconst[i].const_num, data));
 233             deviceImpl->activeContext->pshader_const_dirty[psi->bumpenvmatconst[i].const_num] = 1;
 234
 235             if (psi->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED)
 236             {
 237                 /* WINED3DTSS_BUMPENVLSCALE and WINED3DTSS_BUMPENVLOFFSET are next to each other.
 238                  * point gl to the scale, and load 4 floats. x = scale, y = offset, z and w are junk, we
 239                  * don't care about them. The pointers are valid for sure because the stateblock is bigger.
 240                  * (they're WINED3DTSS_TEXTURETRANSFORMFLAGS and WINED3DTSS_ADDRESSW, so most likely 0 or NaN
 241                  */
 242                 const float *scale = (const float *)&stateBlock->textureState[(int) psi->luminanceconst[i].texunit][WINED3DTSS_BUMPENVLSCALE];
 243                 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->luminanceconst[i].const_num, scale));
 244                 deviceImpl->activeContext->pshader_const_dirty[psi->luminanceconst[i].const_num] = 1;
 245             }
 246         }
 247     }
 248 }
 249
 250 static void shader_arb_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count)
 251 {
 252     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 253
 254     /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
 255      * context. On a context switch the old context will be fully dirtified */
 256     memset(This->activeContext->vshader_const_dirty + start, 1,
 257             sizeof(*This->activeContext->vshader_const_dirty) * count);
 258     This->highest_dirty_vs_const = max(This->highest_dirty_vs_const, start + count + 1);
 259 }
 260
 261 static void shader_arb_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count)
 262 {
 263     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 264
 265     /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
 266      * context. On a context switch the old context will be fully dirtified */
 267     memset(This->activeContext->pshader_const_dirty + start, 1,
 268             sizeof(*This->activeContext->pshader_const_dirty) * count);
 269     This->highest_dirty_ps_const = max(This->highest_dirty_ps_const, start + count + 1);
 270 }
 271
 272 static DWORD *local_const_mapping(IWineD3DBaseShaderImpl *This)
 273 {
 274     DWORD *ret;
 275     DWORD idx = 0;
 276     const local_constant *lconst;
 277
 278     if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) return NULL;
 279
 280     ret = HeapAlloc(GetProcessHeap(), 0, sizeof(DWORD) * This->baseShader.limits.temporary);
 281     if(!ret) {
 282         ERR("Out of memory\n");
 283         return NULL;
 284     }
 285
 286     LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 287         ret[lconst->idx] = idx++;
 288     }
 289     return ret;
 290 }
 291
 292 /* Generate the variable & register declarations for the ARB_vertex_program output target */
 293 static void shader_generate_arb_declarations(IWineD3DBaseShader *iface, const shader_reg_maps *reg_maps,
 294         SHADER_BUFFER *buffer, const WineD3D_GL_Info *gl_info, DWORD *lconst_map)
 295 {
 296     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 297     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
 298     DWORD i, cur;
 299     char pshader = shader_is_pshader_version(reg_maps->shader_version.type);
 300     unsigned max_constantsF = min(This->baseShader.limits.constant_float,
 301             (pshader ? GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS :
 302                        GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS));
 303     UINT extra_constants_needed = 0;
 304     const local_constant *lconst;
 305
 306     /* Temporary Output register */
 307     shader_addline(buffer, "TEMP TMP_OUT;\n");
 308
 309     for(i = 0; i < This->baseShader.limits.temporary; i++) {
 310         if (reg_maps->temporary[i])
 311             shader_addline(buffer, "TEMP R%u;\n", i);
 312     }
 313
 314     for (i = 0; i < This->baseShader.limits.address; i++) {
 315         if (reg_maps->address[i])
 316             shader_addline(buffer, "ADDRESS A%d;\n", i);
 317     }
 318
 319     for(i = 0; i < This->baseShader.limits.texcoord; i++) {
 320         if (reg_maps->texcoord[i])
 321             shader_addline(buffer,"TEMP T%u;\n", i);
 322     }
 323
 324     /* Texture coordinate registers must be pre-loaded */
 325     for (i = 0; i < This->baseShader.limits.texcoord; i++) {
 326         if (reg_maps->texcoord[i])
 327             shader_addline(buffer, "MOV T%u, fragment.texcoord[%u];\n", i, i);
 328     }
 329
 330     for(i = 0; i < (sizeof(reg_maps->bumpmat) / sizeof(reg_maps->bumpmat[0])); i++) {
 331         IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *) This;
 332         if(!reg_maps->bumpmat[i]) continue;
 333
 334         cur = ps->numbumpenvmatconsts;
 335         ps->bumpenvmatconst[cur].const_num = -1;
 336         ps->bumpenvmatconst[cur].texunit = i;
 337         ps->luminanceconst[cur].const_num = -1;
 338         ps->luminanceconst[cur].texunit = i;
 339
 340         /* If the shader does not use all available constants, use the next free constant to load the bump mapping environment matrix from
 341          * the stateblock into the shader. If no constant is available don't load, texbem will then just sample the texture without applying
 342          * bump mapping.
 343          */
 344         if(max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS) {
 345             ps->bumpenvmatconst[cur].const_num = max_constantsF + extra_constants_needed;
 346             shader_addline(buffer, "PARAM bumpenvmat%d = program.env[%d];\n",
 347                            i, ps->bumpenvmatconst[cur].const_num);
 348             extra_constants_needed++;
 349
 350             if(reg_maps->luminanceparams && max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS) {
 351                 ((IWineD3DPixelShaderImpl *)This)->luminanceconst[cur].const_num = max_constantsF + extra_constants_needed;
 352                 shader_addline(buffer, "PARAM luminance%d = program.env[%d];\n",
 353                                i, ps->luminanceconst[cur].const_num);
 354                 extra_constants_needed++;
 355             } else if(reg_maps->luminanceparams) {
 356                 FIXME("No free constant to load the luminance parameters\n");
 357             }
 358         } else {
 359             FIXME("No free constant found to load environment bump mapping matrix into the shader. texbem instruction will not apply bump mapping\n");
 360         }
 361
 362         ps->numbumpenvmatconsts = cur + 1;
 363     }
 364
 365     if(device->stateBlock->renderState[WINED3DRS_SRGBWRITEENABLE] && pshader) {
 366         shader_addline(buffer, "PARAM srgb_consts1 = {%f, %f, %f, %f};\n",
 367                        srgb_mul_low, srgb_cmp, srgb_pow, srgb_mul_high);
 368         shader_addline(buffer, "PARAM srgb_consts2 = {%f, %f, %f, %f};\n",
 369                        srgb_sub_high, 0.0, 0.0, 0.0);
 370     }
 371
 372     /* Load local constants using the program-local space,
 373      * this avoids reloading them each time the shader is used
 374      */
 375     if(lconst_map) {
 376         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 377             shader_addline(buffer, "PARAM C%u = program.local[%u];\n", lconst->idx,
 378                            lconst_map[lconst->idx]);
 379         }
 380     }
 381
 382     /* we use the array-based constants array if the local constants are marked for loading,
 383      * because then we use indirect addressing, or when the local constant list is empty,
 384      * because then we don't know if we're using indirect addressing or not. If we're hardcoding
 385      * local constants do not declare the loaded constants as an array because ARB compilers usually
 386      * do not optimize unused constants away
 387      */
 388     if(This->baseShader.reg_maps.usesrelconstF) {
 389         /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
 390         shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
 391                     max_constantsF, max_constantsF - 1);
 392     } else {
 393         for(i = 0; i < max_constantsF; i++) {
 394             DWORD idx, mask;
 395             idx = i >> 5;
 396             mask = 1 << (i & 0x1f);
 397             if(!shader_constant_is_local(This, i) && (This->baseShader.reg_maps.constf[idx] & mask)) {
 398                 shader_addline(buffer, "PARAM C%d = program.env[%d];\n",i, i);
 399             }
 400         }
 401     }
 402 }
 403
 404 static const char * const shift_tab[] = {
 405     "dummy",     /*  0 (none) */
 406     "coefmul.x", /*  1 (x2)   */
 407     "coefmul.y", /*  2 (x4)   */
 408     "coefmul.z", /*  3 (x8)   */
 409     "coefmul.w", /*  4 (x16)  */
 410     "dummy",     /*  5 (x32)  */
 411     "dummy",     /*  6 (x64)  */
 412     "dummy",     /*  7 (x128) */
 413     "dummy",     /*  8 (d256) */
 414     "dummy",     /*  9 (d128) */
 415     "dummy",     /* 10 (d64)  */
 416     "dummy",     /* 11 (d32)  */
 417     "coefdiv.w", /* 12 (d16)  */
 418     "coefdiv.z", /* 13 (d8)   */
 419     "coefdiv.y", /* 14 (d4)   */
 420     "coefdiv.x"  /* 15 (d2)   */
 421 };
 422
 423 static void shader_arb_get_write_mask(const struct wined3d_shader_instruction *ins,
 424         const struct wined3d_shader_dst_param *dst, char *write_mask)
 425 {
 426     char *ptr = write_mask;
 427     char vshader = shader_is_vshader_version(ins->ctx->reg_maps->shader_version.type);
 428
 429     if (vshader && dst->reg.type == WINED3DSPR_ADDR)
 430     {
 431         *ptr++ = '.';
 432         *ptr++ = 'x';
 433     }
 434     else if (dst->write_mask != WINED3DSP_WRITEMASK_ALL)
 435     {
 436         *ptr++ = '.';
 437         if (dst->write_mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
 438         if (dst->write_mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
 439         if (dst->write_mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
 440         if (dst->write_mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
 441     }
 442
 443     *ptr = '\0';
 444 }
 445
 446 static void shader_arb_get_swizzle(const struct wined3d_shader_src_param *param, BOOL fixup, char *swizzle_str)
 447 {
 448     /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
 449      * but addressed as "rgba". To fix this we need to swap the register's x
 450      * and z components. */
 451     const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
 452     char *ptr = swizzle_str;
 453
 454     /* swizzle bits fields: wwzzyyxx */
 455     DWORD swizzle = param->swizzle;
 456     DWORD swizzle_x = swizzle & 0x03;
 457     DWORD swizzle_y = (swizzle >> 2) & 0x03;
 458     DWORD swizzle_z = (swizzle >> 4) & 0x03;
 459     DWORD swizzle_w = (swizzle >> 6) & 0x03;
 460
 461     /* If the swizzle is the default swizzle (ie, "xyzw"), we don't need to
 462      * generate a swizzle string. Unless we need to our own swizzling. */
 463     if (swizzle != WINED3DSP_NOSWIZZLE || fixup)
 464     {
 465         *ptr++ = '.';
 466         if (swizzle_x == swizzle_y && swizzle_x == swizzle_z && swizzle_x == swizzle_w) {
 467             *ptr++ = swizzle_chars[swizzle_x];
 468         } else {
 469             *ptr++ = swizzle_chars[swizzle_x];
 470             *ptr++ = swizzle_chars[swizzle_y];
 471             *ptr++ = swizzle_chars[swizzle_z];
 472             *ptr++ = swizzle_chars[swizzle_w];
 473         }
 474     }
 475
 476     *ptr = '\0';
 477 }
 478
 479 static void shader_arb_get_register_name(IWineD3DBaseShader *iface,
 480         const struct wined3d_shader_register *reg, char *register_name, BOOL *is_color)
 481 {
 482     /* oPos, oFog and oPts in D3D */
 483     static const char * const rastout_reg_names[] = {"TMP_OUT", "result.fogcoord", "result.pointsize"};
 484     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
 485     BOOL pshader = shader_is_pshader_version(This->baseShader.reg_maps.shader_version.type);
 486
 487     *is_color = FALSE;
 488
 489     switch (reg->type)
 490     {
 491         case WINED3DSPR_TEMP:
 492             sprintf(register_name, "R%u", reg->idx);
 493             break;
 494
 495         case WINED3DSPR_INPUT:
 496             if (pshader)
 497             {
 498                 if (reg->idx == 0) strcpy(register_name, "fragment.color.primary");
 499                 else strcpy(register_name, "fragment.color.secondary");
 500             }
 501             else
 502             {
 503                 if (((IWineD3DVertexShaderImpl *)This)->cur_args->swizzle_map & (1 << reg->idx)) *is_color = TRUE;
 504                 sprintf(register_name, "vertex.attrib[%u]", reg->idx);
 505             }
 506             break;
 507
 508         case WINED3DSPR_CONST:
 509             if (!pshader && reg->rel_addr)
 510             {
 511                 UINT rel_offset = ((IWineD3DVertexShaderImpl *)This)->rel_offset;
 512                 if (reg->idx >= rel_offset)
 513                     sprintf(register_name, "C[A0.x + %u]", reg->idx - rel_offset);
 514                 else
 515                     sprintf(register_name, "C[A0.x - %u]", -reg->idx + rel_offset);
 516             }
 517             else
 518             {
 519                 if (This->baseShader.reg_maps.usesrelconstF)
 520                     sprintf(register_name, "C[%u]", reg->idx);
 521                 else
 522                     sprintf(register_name, "C%u", reg->idx);
 523             }
 524             break;
 525
 526         case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
 527             if (pshader) sprintf(register_name, "T%u", reg->idx);
 528             else  sprintf(register_name, "A%u", reg->idx);
 529             break;
 530
 531         case WINED3DSPR_COLOROUT:
 532             if (reg->idx == 0)
 533             {
 534                 strcpy(register_name, "TMP_COLOR");
 535             }
 536             else
 537             {
 538                 /* TODO: See GL_ARB_draw_buffers */
 539                 FIXME("Unsupported write to render target %u\n", reg->idx);
 540                 sprintf(register_name, "unsupported_register");
 541             }
 542             break;
 543
 544         case WINED3DSPR_RASTOUT:
 545             sprintf(register_name, "%s", rastout_reg_names[reg->idx]);
 546             break;
 547
 548         case WINED3DSPR_DEPTHOUT:
 549             strcpy(register_name, "result.depth");
 550             break;
 551
 552         case WINED3DSPR_ATTROUT:
 553             if (pshader) sprintf(register_name, "oD[%u]", reg->idx);
 554             else if (reg->idx == 0) strcpy(register_name, "result.color.primary");
 555             else strcpy(register_name, "result.color.secondary");
 556             break;
 557
 558         case WINED3DSPR_TEXCRDOUT:
 559             if (pshader) sprintf(register_name, "oT[%u]", reg->idx);
 560             else sprintf(register_name, "result.texcoord[%u]", reg->idx);
 561             break;
 562
 563         default:
 564             FIXME("Unhandled register type %#x[%u]\n", reg->type, reg->idx);
 565             sprintf(register_name, "unrecognized_register[%u]", reg->idx);
 566             break;
 567     }
 568 }
 569
 570 static void shader_arb_get_dst_param(const struct wined3d_shader_instruction *ins,
 571         const struct wined3d_shader_dst_param *wined3d_dst, char *str)
 572 {
 573     char register_name[255];
 574     char write_mask[6];
 575     BOOL is_color;
 576
 577     shader_arb_get_register_name(ins->ctx->shader, &wined3d_dst->reg, register_name, &is_color);
 578     strcpy(str, register_name);
 579
 580     shader_arb_get_write_mask(ins, wined3d_dst, write_mask);
 581     strcat(str, write_mask);
 582 }
 583
 584 static const char *shader_arb_get_fixup_swizzle(enum fixup_channel_source channel_source)
 585 {
 586     switch(channel_source)
 587     {
 588         case CHANNEL_SOURCE_ZERO: return "0";
 589         case CHANNEL_SOURCE_ONE: return "1";
 590         case CHANNEL_SOURCE_X: return "x";
 591         case CHANNEL_SOURCE_Y: return "y";
 592         case CHANNEL_SOURCE_Z: return "z";
 593         case CHANNEL_SOURCE_W: return "w";
 594         default:
 595             FIXME("Unhandled channel source %#x\n", channel_source);
 596             return "undefined";
 597     }
 598 }
 599
 600 static void gen_color_correction(SHADER_BUFFER *buffer, const char *reg, DWORD dst_mask,
 601                                  const char *one, const char *two, struct color_fixup_desc fixup)
 602 {
 603     DWORD mask;
 604
 605     if (is_yuv_fixup(fixup))
 606     {
 607         enum yuv_fixup yuv_fixup = get_yuv_fixup(fixup);
 608         FIXME("YUV fixup (%#x) not supported\n", yuv_fixup);
 609         return;
 610     }
 611
 612     mask = 0;
 613     if (fixup.x_source != CHANNEL_SOURCE_X) mask |= WINED3DSP_WRITEMASK_0;
 614     if (fixup.y_source != CHANNEL_SOURCE_Y) mask |= WINED3DSP_WRITEMASK_1;
 615     if (fixup.z_source != CHANNEL_SOURCE_Z) mask |= WINED3DSP_WRITEMASK_2;
 616     if (fixup.w_source != CHANNEL_SOURCE_W) mask |= WINED3DSP_WRITEMASK_3;
 617     mask &= dst_mask;
 618
 619     if (mask)
 620     {
 621         shader_addline(buffer, "SWZ %s, %s, %s, %s, %s, %s;\n", reg, reg,
 622                 shader_arb_get_fixup_swizzle(fixup.x_source), shader_arb_get_fixup_swizzle(fixup.y_source),
 623                 shader_arb_get_fixup_swizzle(fixup.z_source), shader_arb_get_fixup_swizzle(fixup.w_source));
 624     }
 625
 626     mask = 0;
 627     if (fixup.x_sign_fixup) mask |= WINED3DSP_WRITEMASK_0;
 628     if (fixup.y_sign_fixup) mask |= WINED3DSP_WRITEMASK_1;
 629     if (fixup.z_sign_fixup) mask |= WINED3DSP_WRITEMASK_2;
 630     if (fixup.w_sign_fixup) mask |= WINED3DSP_WRITEMASK_3;
 631     mask &= dst_mask;
 632
 633     if (mask)
 634     {
 635         char reg_mask[6];
 636         char *ptr = reg_mask;
 637
 638         if (mask != WINED3DSP_WRITEMASK_ALL)
 639         {
 640             *ptr++ = '.';
 641             if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
 642             if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
 643             if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
 644             if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
 645         }
 646         *ptr = '\0';
 647
 648         shader_addline(buffer, "MAD %s%s, %s, %s, -%s;\n", reg, reg_mask, reg, two, one);
 649     }
 650 }
 651
 652 static void shader_hw_sample(const struct wined3d_shader_instruction *ins, DWORD sampler_idx,
 653         const char *dst_str, const char *coord_reg, BOOL projected, BOOL bias)
 654 {
 655     SHADER_BUFFER *buffer = ins->ctx->buffer;
 656     DWORD sampler_type = ins->ctx->reg_maps->sampler_type[sampler_idx];
 657     const char *tex_type;
 658     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)ins->ctx->shader;
 659     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
 660
 661     switch(sampler_type) {
 662         case WINED3DSTT_1D:
 663             tex_type = "1D";
 664             break;
 665
 666         case WINED3DSTT_2D:
 667             if(device->stateBlock->textures[sampler_idx] &&
 668                IWineD3DBaseTexture_GetTextureDimensions(device->stateBlock->textures[sampler_idx]) == GL_TEXTURE_RECTANGLE_ARB) {
 669                 tex_type = "RECT";
 670             } else {
 671                 tex_type = "2D";
 672             }
 673             if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version.type))
 674             {
 675                const IWineD3DPixelShaderImpl* const ps = (const IWineD3DPixelShaderImpl*)This;
 676                if(ps->cur_args->np2_fixup & (1 << sampler_idx)) {
 677                    FIXME("NP2 texcoord fixup is currently not implemented in ARB mode (use GLSL instead).\n");
 678                }
 679             }
 680             break;
 681
 682         case WINED3DSTT_VOLUME:
 683             tex_type = "3D";
 684             break;
 685
 686         case WINED3DSTT_CUBE:
 687             tex_type = "CUBE";
 688             break;
 689
 690         default:
 691             ERR("Unexpected texture type %d\n", sampler_type);
 692             tex_type = "";
 693     }
 694
 695     if (bias) {
 696         /* Shouldn't be possible, but let's check for it */
 697         if(projected) FIXME("Biased and Projected texture sampling\n");
 698         /* TXB takes the 4th component of the source vector automatically, as d3d. Nothing more to do */
 699         shader_addline(buffer, "TXB %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 700     } else if (projected) {
 701         shader_addline(buffer, "TXP %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 702     } else {
 703         shader_addline(buffer, "TEX %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 704     }
 705
 706     if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version.type))
 707     {
 708         IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
 709         gen_color_correction(buffer, dst_str, ins->dst[0].write_mask,
 710                 "one", "coefmul.x", ps->cur_args->color_fixup[sampler_idx]);
 711     }
 712 }
 713
 714 static void shader_arb_get_src_param(const struct wined3d_shader_instruction *ins,
 715         const struct wined3d_shader_src_param *src, unsigned int tmpreg, char *outregstr)
 716 {
 717     /* Generate a line that does the input modifier computation and return the input register to use */
 718     BOOL is_color = FALSE;
 719     char regstr[256];
 720     char swzstr[20];
 721     int insert_line;
 722     SHADER_BUFFER *buffer = ins->ctx->buffer;
 723
 724     /* Assume a new line will be added */
 725     insert_line = 1;
 726
 727     /* Get register name */
 728     shader_arb_get_register_name(ins->ctx->shader, &src->reg, regstr, &is_color);
 729     shader_arb_get_swizzle(src, is_color, swzstr);
 730
 731     switch (src->modifiers)
 732     {
 733     case WINED3DSPSM_NONE:
 734         sprintf(outregstr, "%s%s", regstr, swzstr);
 735         insert_line = 0;
 736         break;
 737     case WINED3DSPSM_NEG:
 738         sprintf(outregstr, "-%s%s", regstr, swzstr);
 739         insert_line = 0;
 740         break;
 741     case WINED3DSPSM_BIAS:
 742         shader_addline(buffer, "ADD T%c, %s, -coefdiv.x;\n", 'A' + tmpreg, regstr);
 743         break;
 744     case WINED3DSPSM_BIASNEG:
 745         shader_addline(buffer, "ADD T%c, -%s, coefdiv.x;\n", 'A' + tmpreg, regstr);
 746         break;
 747     case WINED3DSPSM_SIGN:
 748         shader_addline(buffer, "MAD T%c, %s, coefmul.x, -one.x;\n", 'A' + tmpreg, regstr);
 749         break;
 750     case WINED3DSPSM_SIGNNEG:
 751         shader_addline(buffer, "MAD T%c, %s, -coefmul.x, one.x;\n", 'A' + tmpreg, regstr);
 752         break;
 753     case WINED3DSPSM_COMP:
 754         shader_addline(buffer, "SUB T%c, one.x, %s;\n", 'A' + tmpreg, regstr);
 755         break;
 756     case WINED3DSPSM_X2:
 757         shader_addline(buffer, "ADD T%c, %s, %s;\n", 'A' + tmpreg, regstr, regstr);
 758         break;
 759     case WINED3DSPSM_X2NEG:
 760         shader_addline(buffer, "ADD T%c, -%s, -%s;\n", 'A' + tmpreg, regstr, regstr);
 761         break;
 762     case WINED3DSPSM_DZ:
 763         shader_addline(buffer, "RCP T%c, %s.z;\n", 'A' + tmpreg, regstr);
 764         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 765         break;
 766     case WINED3DSPSM_DW:
 767         shader_addline(buffer, "RCP T%c, %s.w;\n", 'A' + tmpreg, regstr);
 768         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 769         break;
 770     default:
 771         sprintf(outregstr, "%s%s", regstr, swzstr);
 772         insert_line = 0;
 773     }
 774
 775     /* Return modified or original register, with swizzle */
 776     if (insert_line)
 777         sprintf(outregstr, "T%c%s", 'A' + tmpreg, swzstr);
 778 }
 779
 780 static void pshader_hw_bem(const struct wined3d_shader_instruction *ins)
 781 {
 782     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
 783     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 784     SHADER_BUFFER *buffer = ins->ctx->buffer;
 785     char dst_name[50];
 786     char src_name[2][50];
 787     char dst_wmask[20];
 788     DWORD sampler_code = dst->reg.idx;
 789     BOOL has_bumpmat = FALSE;
 790     BOOL is_color;
 791     int i;
 792
 793     for(i = 0; i < This->numbumpenvmatconsts; i++) {
 794         if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
 795                 && This->bumpenvmatconst[i].texunit == sampler_code)
 796         {
 797             has_bumpmat = TRUE;
 798             break;
 799         }
 800     }
 801
 802     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, dst_name, &is_color);
 803     shader_arb_get_write_mask(ins, dst, dst_wmask);
 804     strcat(dst_name, dst_wmask);
 805
 806     shader_arb_get_src_param(ins, &ins->src[0], 0, src_name[0]);
 807     shader_arb_get_src_param(ins, &ins->src[1], 1, src_name[1]);
 808
 809     if(has_bumpmat) {
 810         /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
 811         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", sampler_code);
 812         shader_addline(buffer, "DP3 TMP.r, TMP2, %s;\n", src_name[1]);
 813         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", sampler_code);
 814         shader_addline(buffer, "DP3 TMP.g, TMP2, %s;\n", src_name[1]);
 815
 816         shader_addline(buffer, "ADD %s, %s, TMP;\n", dst_name, src_name[0]);
 817     } else {
 818         shader_addline(buffer, "MOV %s, %s;\n", dst_name, src_name[0]);
 819     }
 820 }
 821
 822 static void pshader_hw_cnd(const struct wined3d_shader_instruction *ins)
 823 {
 824     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 825     SHADER_BUFFER *buffer = ins->ctx->buffer;
 826     char dst_wmask[20];
 827     char dst_name[50];
 828     char src_name[3][50];
 829     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
 830     BOOL is_color;
 831     DWORD shader_version = WINED3D_SHADER_VERSION(ins->ctx->reg_maps->shader_version.major,
 832             ins->ctx->reg_maps->shader_version.minor);
 833
 834     /* FIXME: support output modifiers */
 835
 836     /* Handle output register */
 837     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, dst_name, &is_color);
 838     shader_arb_get_write_mask(ins, dst, dst_wmask);
 839
 840     /* Generate input register names (with modifiers) */
 841     shader_arb_get_src_param(ins, &ins->src[0], 0, src_name[0]);
 842     shader_arb_get_src_param(ins, &ins->src[1], 1, src_name[1]);
 843     shader_arb_get_src_param(ins, &ins->src[2], 2, src_name[2]);
 844
 845     /* The coissue flag changes the semantic of the cnd instruction in <= 1.3 shaders */
 846     if (shader_version <= WINED3D_SHADER_VERSION(1, 3) && ins->coissue)
 847     {
 848         shader_addline(buffer, "MOV%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1]);
 849     } else {
 850         shader_addline(buffer, "ADD TMP, -%s, coefdiv.x;\n", src_name[0]);
 851         shader_addline(buffer, "CMP%s %s%s, TMP, %s, %s;\n",
 852                                 sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1], src_name[2]);
 853     }
 854 }
 855
 856 static void pshader_hw_cmp(const struct wined3d_shader_instruction *ins)
 857 {
 858     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 859     SHADER_BUFFER *buffer = ins->ctx->buffer;
 860     char dst_wmask[20];
 861     char dst_name[50];
 862     char src_name[3][50];
 863     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
 864     BOOL is_color;
 865
 866     /* FIXME: support output modifiers */
 867
 868     /* Handle output register */
 869     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, dst_name, &is_color);
 870     shader_arb_get_write_mask(ins, dst, dst_wmask);
 871
 872     /* Generate input register names (with modifiers) */
 873     shader_arb_get_src_param(ins, &ins->src[0], 0, src_name[0]);
 874     shader_arb_get_src_param(ins, &ins->src[1], 1, src_name[1]);
 875     shader_arb_get_src_param(ins, &ins->src[2], 2, src_name[2]);
 876
 877     shader_addline(buffer, "CMP%s %s%s, %s, %s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
 878                    src_name[0], src_name[2], src_name[1]);
 879 }
 880
 881 /** Process the WINED3DSIO_DP2ADD instruction in ARB.
 882  * dst = dot2(src0, src1) + src2 */
 883 static void pshader_hw_dp2add(const struct wined3d_shader_instruction *ins)
 884 {
 885     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 886     SHADER_BUFFER *buffer = ins->ctx->buffer;
 887     char dst_wmask[20];
 888     char dst_name[50];
 889     char src_name[3][50];
 890     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
 891     BOOL is_color;
 892
 893     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, dst_name, &is_color);
 894     shader_arb_get_write_mask(ins, dst, dst_wmask);
 895
 896     shader_arb_get_src_param(ins, &ins->src[0], 0, src_name[0]);
 897     shader_arb_get_src_param(ins, &ins->src[1], 1, src_name[1]);
 898     shader_arb_get_src_param(ins, &ins->src[2], 2, src_name[2]);
 899
 900     /* Emulate a DP2 with a DP3 and 0.0 */
 901     shader_addline(buffer, "MOV TMP, %s;\n", src_name[0]);
 902     shader_addline(buffer, "MOV TMP.z, 0.0;\n");
 903     shader_addline(buffer, "DP3 TMP2, TMP, %s;\n", src_name[1]);
 904     shader_addline(buffer, "ADD%s %s%s, TMP2, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[2]);
 905 }
 906
 907 /* Map the opcode 1-to-1 to the GL code */
 908 static void shader_hw_map2gl(const struct wined3d_shader_instruction *ins)
 909 {
 910     SHADER_BUFFER *buffer = ins->ctx->buffer;
 911     const char *instruction;
 912     char arguments[256], dst_str[50];
 913     unsigned int i;
 914     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 915     const char *modifier;
 916
 917     switch (ins->handler_idx)
 918     {
 919         case WINED3DSIH_ABS: instruction = "ABS"; break;
 920         case WINED3DSIH_ADD: instruction = "ADD"; break;
 921         case WINED3DSIH_CRS: instruction = "XPD"; break;
 922         case WINED3DSIH_DP3: instruction = "DP3"; break;
 923         case WINED3DSIH_DP4: instruction = "DP4"; break;
 924         case WINED3DSIH_DST: instruction = "DST"; break;
 925         case WINED3DSIH_EXP: instruction = "EX2"; break;
 926         case WINED3DSIH_EXPP: instruction = "EXP"; break;
 927         case WINED3DSIH_FRC: instruction = "FRC"; break;
 928         case WINED3DSIH_LIT: instruction = "LIT"; break;
 929         case WINED3DSIH_LOG: instruction = "LG2"; break;
 930         case WINED3DSIH_LOGP: instruction = "LOG"; break;
 931         case WINED3DSIH_LRP: instruction = "LRP"; break;
 932         case WINED3DSIH_MAD: instruction = "MAD"; break;
 933         case WINED3DSIH_MAX: instruction = "MAX"; break;
 934         case WINED3DSIH_MIN: instruction = "MIN"; break;
 935         case WINED3DSIH_MOV: instruction = "MOV"; break;
 936         case WINED3DSIH_MUL: instruction = "MUL"; break;
 937         case WINED3DSIH_POW: instruction = "POW"; break;
 938         case WINED3DSIH_SGE: instruction = "SGE"; break;
 939         case WINED3DSIH_SLT: instruction = "SLT"; break;
 940         case WINED3DSIH_SUB: instruction = "SUB"; break;
 941         default: instruction = "";
 942             FIXME("Unhandled opcode %#x\n", ins->handler_idx);
 943             break;
 944     }
 945
 946     /* All instructions handled by this function have a destination parameter */
 947     if(dst->modifiers & WINED3DSPDM_SATURATE) modifier = "_SAT";
 948     else modifier = "";
 949
 950     /* Note that shader_arb_add_dst_param() adds spaces. */
 951     arguments[0] = '\0';
 952     shader_arb_get_dst_param(ins, dst, dst_str);
 953     for (i = 0; i < ins->src_count; ++i)
 954     {
 955         char operand[100];
 956         strcat(arguments, ", ");
 957         shader_arb_get_src_param(ins, &ins->src[i], i, operand);
 958         strcat(arguments, operand);
 959     }
 960     shader_addline(buffer, "%s%s %s%s;\n", instruction, modifier, dst_str, arguments);
 961 }
 962
 963 static void shader_hw_nop(const struct wined3d_shader_instruction *ins)
 964 {
 965     SHADER_BUFFER *buffer = ins->ctx->buffer;
 966     shader_addline(buffer, "NOP;\n");
 967 }
 968
 969 static void shader_hw_mov(const struct wined3d_shader_instruction *ins)
 970 {
 971     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)ins->ctx->shader;
 972
 973     if ((ins->ctx->reg_maps->shader_version.major == 1
 974             && !shader_is_pshader_version(ins->ctx->reg_maps->shader_version.type)
 975             && ins->dst[0].reg.type == WINED3DSPR_ADDR)
 976             || ins->handler_idx == WINED3DSIH_MOVA)
 977     {
 978         SHADER_BUFFER *buffer = ins->ctx->buffer;
 979         char src0_param[256];
 980
 981         if (ins->handler_idx == WINED3DSIH_MOVA)
 982             FIXME("mova should round\n");
 983
 984         src0_param[0] = '\0';
 985         if (((IWineD3DVertexShaderImpl *)shader)->rel_offset)
 986         {
 987             shader_arb_get_src_param(ins, &ins->src[0], 0, src0_param);
 988             shader_addline(buffer, "ADD TMP.x, %s, helper_const.z;\n", src0_param);
 989             shader_addline(buffer, "ARL A0.x, TMP.x;\n");
 990         }
 991         else
 992         {
 993             /* Apple's ARB_vertex_program implementation does not accept an ARL source argument
 994              * with more than one component. Thus replicate the first source argument over all
 995              * 4 components. For example, .xyzw -> .x (or better: .xxxx), .zwxy -> .z, etc) */
 996             struct wined3d_shader_src_param tmp_src = ins->src[0];
 997             tmp_src.swizzle = (tmp_src.swizzle & 0x3) * 0x55;
 998             shader_arb_get_src_param(ins, &tmp_src, 0, src0_param);
 999             shader_addline(buffer, "ARL A0.x, %s;\n", src0_param);
1000         }
1001     }
1002     else
1003     {
1004         shader_hw_map2gl(ins);
1005     }
1006 }
1007
1008 static void pshader_hw_texkill(const struct wined3d_shader_instruction *ins)
1009 {
1010     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1011     SHADER_BUFFER *buffer = ins->ctx->buffer;
1012     char reg_dest[40];
1013
1014     /* No swizzles are allowed in d3d's texkill. PS 1.x ignores the 4th component as documented,
1015      * but >= 2.0 honors it(undocumented, but tested by the d3d9 testsuit)
1016      */
1017     shader_arb_get_dst_param(ins, dst, reg_dest);
1018
1019     if (ins->ctx->reg_maps->shader_version.major >= 2)
1020     {
1021         /* The arb backend doesn't claim ps 2.0 support, but try to eat what the app feeds to us */
1022         shader_addline(buffer, "KIL %s;\n", reg_dest);
1023     } else {
1024         /* ARB fp doesn't like swizzles on the parameter of the KIL instruction. To mask the 4th component,
1025          * copy the register into our general purpose TMP variable, overwrite .w and pass TMP to KIL
1026          */
1027         shader_addline(buffer, "SWZ TMP, %s, x, y, z, 1;\n", reg_dest);
1028         shader_addline(buffer, "KIL TMP;\n");
1029     }
1030 }
1031
1032 static void pshader_hw_tex(const struct wined3d_shader_instruction *ins)
1033 {
1034     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1035     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1036     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1037     BOOL is_color;
1038     DWORD shader_version = WINED3D_SHADER_VERSION(ins->ctx->reg_maps->shader_version.major,
1039             ins->ctx->reg_maps->shader_version.minor);
1040     BOOL projected = FALSE, bias = FALSE;
1041
1042     char reg_dest[40];
1043     char reg_coord[40];
1044     DWORD reg_sampler_code;
1045
1046     /* All versions have a destination register */
1047     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, reg_dest, &is_color);
1048
1049     /* 1.0-1.3: Use destination register as coordinate source.
1050        1.4+: Use provided coordinate source register. */
1051    if (shader_version < WINED3D_SHADER_VERSION(1,4))
1052       strcpy(reg_coord, reg_dest);
1053    else
1054       shader_arb_get_src_param(ins, &ins->src[0], 0, reg_coord);
1055
1056   /* 1.0-1.4: Use destination register number as texture code.
1057      2.0+: Use provided sampler number as texure code. */
1058   if (shader_version < WINED3D_SHADER_VERSION(2,0))
1059      reg_sampler_code = dst->reg.idx;
1060   else
1061      reg_sampler_code = ins->src[1].reg.idx;
1062
1063   /* projection flag:
1064    * 1.1, 1.2, 1.3: Use WINED3DTSS_TEXTURETRANSFORMFLAGS
1065    * 1.4: Use WINED3DSPSM_DZ or WINED3DSPSM_DW on src[0]
1066    * 2.0+: Use WINED3DSI_TEXLD_PROJECT on the opcode
1067    */
1068   if (shader_version < WINED3D_SHADER_VERSION(1,4))
1069   {
1070       DWORD flags = 0;
1071       if(reg_sampler_code < MAX_TEXTURES) {
1072         flags = deviceImpl->stateBlock->textureState[reg_sampler_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1073       }
1074       if (flags & WINED3DTTFF_PROJECTED) {
1075           projected = TRUE;
1076       }
1077   }
1078   else if (shader_version < WINED3D_SHADER_VERSION(2,0))
1079   {
1080       DWORD src_mod = ins->src[0].modifiers;
1081       if (src_mod == WINED3DSPSM_DZ) {
1082           projected = TRUE;
1083       } else if(src_mod == WINED3DSPSM_DW) {
1084           projected = TRUE;
1085       }
1086   } else {
1087       if (ins->flags & WINED3DSI_TEXLD_PROJECT) projected = TRUE;
1088       if (ins->flags & WINED3DSI_TEXLD_BIAS) bias = TRUE;
1089   }
1090   shader_hw_sample(ins, reg_sampler_code, reg_dest, reg_coord, projected, bias);
1091 }
1092
1093 static void pshader_hw_texcoord(const struct wined3d_shader_instruction *ins)
1094 {
1095     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1096     SHADER_BUFFER *buffer = ins->ctx->buffer;
1097     DWORD shader_version = WINED3D_SHADER_VERSION(ins->ctx->reg_maps->shader_version.major,
1098             ins->ctx->reg_maps->shader_version.minor);
1099
1100     char tmp[20];
1101     shader_arb_get_write_mask(ins, dst, tmp);
1102     if (shader_version != WINED3D_SHADER_VERSION(1,4))
1103     {
1104         DWORD reg = dst->reg.idx;
1105         shader_addline(buffer, "MOV_SAT T%u%s, fragment.texcoord[%u];\n", reg, tmp, reg);
1106     } else {
1107         char reg_src[40];
1108
1109         shader_arb_get_src_param(ins, &ins->src[0], 0, reg_src);
1110         shader_addline(buffer, "MOV R%u%s, %s;\n", dst->reg.idx, tmp, reg_src);
1111    }
1112 }
1113
1114 static void pshader_hw_texreg2ar(const struct wined3d_shader_instruction *ins)
1115 {
1116      SHADER_BUFFER *buffer = ins->ctx->buffer;
1117      IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1118      IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1119      DWORD flags;
1120
1121      DWORD reg1 = ins->dst[0].reg.idx;
1122      char dst_str[8];
1123      char src_str[50];
1124
1125      sprintf(dst_str, "T%u", reg1);
1126      shader_arb_get_src_param(ins, &ins->src[0], 0, src_str);
1127      shader_addline(buffer, "MOV TMP.x, %s.w;\n", src_str);
1128      shader_addline(buffer, "MOV TMP.y, %s.x;\n", src_str);
1129      flags = reg1 < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg1][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1130      shader_hw_sample(ins, reg1, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1131 }
1132
1133 static void pshader_hw_texreg2gb(const struct wined3d_shader_instruction *ins)
1134 {
1135      SHADER_BUFFER *buffer = ins->ctx->buffer;
1136
1137      DWORD reg1 = ins->dst[0].reg.idx;
1138      char dst_str[8];
1139      char src_str[50];
1140
1141      sprintf(dst_str, "T%u", reg1);
1142      shader_arb_get_src_param(ins, &ins->src[0], 0, src_str);
1143      shader_addline(buffer, "MOV TMP.x, %s.y;\n", src_str);
1144      shader_addline(buffer, "MOV TMP.y, %s.z;\n", src_str);
1145      shader_hw_sample(ins, reg1, dst_str, "TMP", FALSE, FALSE);
1146 }
1147
1148 static void pshader_hw_texreg2rgb(const struct wined3d_shader_instruction *ins)
1149 {
1150     DWORD reg1 = ins->dst[0].reg.idx;
1151     char dst_str[8];
1152     char src_str[50];
1153
1154     sprintf(dst_str, "T%u", reg1);
1155     shader_arb_get_src_param(ins, &ins->src[0], 0, src_str);
1156     shader_hw_sample(ins, reg1, dst_str, src_str, FALSE, FALSE);
1157 }
1158
1159 static void pshader_hw_texbem(const struct wined3d_shader_instruction *ins)
1160 {
1161     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1162     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1163     BOOL has_bumpmat = FALSE;
1164     BOOL has_luminance = FALSE;
1165     BOOL is_color;
1166     int i;
1167
1168     SHADER_BUFFER *buffer = ins->ctx->buffer;
1169
1170     char reg_coord[40];
1171     DWORD reg_dest_code;
1172
1173     /* All versions have a destination register */
1174     reg_dest_code = dst->reg.idx;
1175     /* Can directly use the name because texbem is only valid for <= 1.3 shaders */
1176     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, reg_coord, &is_color);
1177
1178     for(i = 0; i < This->numbumpenvmatconsts; i++) {
1179         if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1180                 && reg_dest_code == This->bumpenvmatconst[i].texunit)
1181         {
1182             has_bumpmat = TRUE;
1183             break;
1184         }
1185     }
1186     for(i = 0; i < This->numbumpenvmatconsts; i++) {
1187         if (This->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1188                 && reg_dest_code == This->luminanceconst[i].texunit)
1189         {
1190             has_luminance = TRUE;
1191             break;
1192         }
1193     }
1194
1195     if(has_bumpmat) {
1196         DWORD src = ins->src[0].reg.idx;
1197
1198         /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
1199
1200         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", reg_dest_code);
1201         shader_addline(buffer, "DP3 TMP.x, TMP2, T%u;\n", src);
1202         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", reg_dest_code);
1203         shader_addline(buffer, "DP3 TMP.y, TMP2, T%u;\n", src);
1204
1205         /* with projective textures, texbem only divides the static texture coord, not the displacement,
1206          * so we can't let the GL handle this.
1207          */
1208         if (((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS]
1209               & WINED3DTTFF_PROJECTED) {
1210             shader_addline(buffer, "RCP TMP2.w, %s.w;\n", reg_coord);
1211             shader_addline(buffer, "MUL TMP2.xy, %s, TMP2.w;\n", reg_coord);
1212             shader_addline(buffer, "ADD TMP.xy, TMP, TMP2;\n");
1213         } else {
1214             shader_addline(buffer, "ADD TMP.xy, TMP, %s;\n", reg_coord);
1215         }
1216
1217         shader_hw_sample(ins, reg_dest_code, reg_coord, "TMP", FALSE, FALSE);
1218
1219         if (ins->handler_idx == WINED3DSIH_TEXBEML && has_luminance)
1220         {
1221             shader_addline(buffer, "MAD TMP, T%u.z, luminance%d.x, luminance%d.y;\n",
1222                            src, reg_dest_code, reg_dest_code);
1223             shader_addline(buffer, "MUL %s, %s, TMP;\n", reg_coord, reg_coord);
1224         }
1225
1226     } else {
1227         DWORD tf;
1228         if(reg_dest_code < MAX_TEXTURES) {
1229             tf = ((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1230         } else {
1231             tf = 0;
1232         }
1233         /* Without a bump matrix loaded, just sample with the unmodified coordinates */
1234         shader_hw_sample(ins, reg_dest_code, reg_coord, reg_coord, tf & WINED3DTTFF_PROJECTED, FALSE);
1235     }
1236 }
1237
1238 static void pshader_hw_texm3x2pad(const struct wined3d_shader_instruction *ins)
1239 {
1240     DWORD reg = ins->dst[0].reg.idx;
1241     SHADER_BUFFER *buffer = ins->ctx->buffer;
1242     char src0_name[50];
1243
1244     shader_arb_get_src_param(ins, &ins->src[0], 0, src0_name);
1245     shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", reg, src0_name);
1246 }
1247
1248 static void pshader_hw_texm3x2tex(const struct wined3d_shader_instruction *ins)
1249 {
1250     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1251     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1252     DWORD flags;
1253     DWORD reg = ins->dst[0].reg.idx;
1254     SHADER_BUFFER *buffer = ins->ctx->buffer;
1255     char dst_str[8];
1256     char src0_name[50];
1257
1258     sprintf(dst_str, "T%u", reg);
1259     shader_arb_get_src_param(ins, &ins->src[0], 0, src0_name);
1260     shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", reg, src0_name);
1261     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1262     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1263 }
1264
1265 static void pshader_hw_texm3x3pad(const struct wined3d_shader_instruction *ins)
1266 {
1267     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1268     DWORD reg = ins->dst[0].reg.idx;
1269     SHADER_BUFFER *buffer = ins->ctx->buffer;
1270     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1271     char src0_name[50];
1272
1273     shader_arb_get_src_param(ins, &ins->src[0], 0, src0_name);
1274     shader_addline(buffer, "DP3 TMP.%c, T%u, %s;\n", 'x' + current_state->current_row, reg, src0_name);
1275     current_state->texcoord_w[current_state->current_row++] = reg;
1276 }
1277
1278 static void pshader_hw_texm3x3tex(const struct wined3d_shader_instruction *ins)
1279 {
1280     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1281     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1282     DWORD flags;
1283     DWORD reg = ins->dst[0].reg.idx;
1284     SHADER_BUFFER *buffer = ins->ctx->buffer;
1285     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1286     char dst_str[8];
1287     char src0_name[50];
1288
1289     shader_arb_get_src_param(ins, &ins->src[0], 0, src0_name);
1290     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1291
1292     /* Sample the texture using the calculated coordinates */
1293     sprintf(dst_str, "T%u", reg);
1294     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1295     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1296     current_state->current_row = 0;
1297 }
1298
1299 static void pshader_hw_texm3x3vspec(const struct wined3d_shader_instruction *ins)
1300 {
1301     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1302     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1303     DWORD flags;
1304     DWORD reg = ins->dst[0].reg.idx;
1305     SHADER_BUFFER *buffer = ins->ctx->buffer;
1306     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1307     char dst_str[8];
1308     char src0_name[50];
1309
1310     shader_arb_get_src_param(ins, &ins->src[0], 0, src0_name);
1311     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1312
1313     /* Construct the eye-ray vector from w coordinates */
1314     shader_addline(buffer, "MOV TMP2.x, fragment.texcoord[%u].w;\n", current_state->texcoord_w[0]);
1315     shader_addline(buffer, "MOV TMP2.y, fragment.texcoord[%u].w;\n", current_state->texcoord_w[1]);
1316     shader_addline(buffer, "MOV TMP2.z, fragment.texcoord[%u].w;\n", reg);
1317
1318     /* Calculate reflection vector
1319      */
1320     shader_addline(buffer, "DP3 TMP.w, TMP, TMP2;\n");
1321     /* The .w is ignored when sampling, so I can use TMP2.w to calculate dot(N, N) */
1322     shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1323     shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1324     shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1325     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1326     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -TMP2;\n");
1327
1328     /* Sample the texture using the calculated coordinates */
1329     sprintf(dst_str, "T%u", reg);
1330     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1331     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1332     current_state->current_row = 0;
1333 }
1334
1335 static void pshader_hw_texm3x3spec(const struct wined3d_shader_instruction *ins)
1336 {
1337     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1338     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1339     DWORD flags;
1340     DWORD reg = ins->dst[0].reg.idx;
1341     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1342     SHADER_BUFFER *buffer = ins->ctx->buffer;
1343     char dst_str[8];
1344     char src0_name[50];
1345     char src1_name[50];
1346
1347     shader_arb_get_src_param(ins, &ins->src[0], 0, src0_name);
1348     shader_arb_get_src_param(ins, &ins->src[0], 1, src1_name);
1349     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1350
1351     /* Calculate reflection vector.
1352      *
1353      *               dot(N, E)
1354      * TMP.xyz = 2 * --------- * N - E
1355      *               dot(N, N)
1356      *
1357      * Which normalizes the normal vector
1358      */
1359     shader_addline(buffer, "DP3 TMP.w, TMP, %s;\n", src1_name);
1360     shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1361     shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1362     shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1363     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1364     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -%s;\n", src1_name);
1365
1366     /* Sample the texture using the calculated coordinates */
1367     sprintf(dst_str, "T%u", reg);
1368     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1369     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1370     current_state->current_row = 0;
1371 }
1372
1373 static void pshader_hw_texdepth(const struct wined3d_shader_instruction *ins)
1374 {
1375     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1376     SHADER_BUFFER *buffer = ins->ctx->buffer;
1377     char dst_name[50];
1378     BOOL is_color;
1379
1380     /* texdepth has an implicit destination, the fragment depth value. It's only parameter,
1381      * which is essentially an input, is the destination register because it is the first
1382      * parameter. According to the msdn, this must be register r5, but let's keep it more flexible
1383      * here
1384      */
1385     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, dst_name, &is_color);
1386
1387     /* According to the msdn, the source register(must be r5) is unusable after
1388      * the texdepth instruction, so we're free to modify it
1389      */
1390     shader_addline(buffer, "MIN %s.y, %s.y, one.y;\n", dst_name, dst_name);
1391
1392     /* How to deal with the special case dst_name.g == 0? if r != 0, then
1393      * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1394      * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1395      */
1396     shader_addline(buffer, "RCP %s.y, %s.y;\n", dst_name, dst_name);
1397     shader_addline(buffer, "MUL TMP.x, %s.x, %s.y;\n", dst_name, dst_name);
1398     shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1399     shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1400 }
1401
1402 /** Process the WINED3DSIO_TEXDP3TEX instruction in ARB:
1403  * Take a 3-component dot product of the TexCoord[dstreg] and src,
1404  * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1405 static void pshader_hw_texdp3tex(const struct wined3d_shader_instruction *ins)
1406 {
1407     SHADER_BUFFER *buffer = ins->ctx->buffer;
1408     DWORD sampler_idx = ins->dst[0].reg.idx;
1409     char src0[50];
1410     char dst_str[8];
1411
1412     shader_arb_get_src_param(ins, &ins->src[0], 0, src0);
1413     shader_addline(buffer, "MOV TMP, 0.0;\n");
1414     shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", sampler_idx, src0);
1415
1416     sprintf(dst_str, "T%u", sampler_idx);
1417     shader_hw_sample(ins, sampler_idx, dst_str, "TMP", FALSE /* Only one coord, can't be projected */, FALSE);
1418 }
1419
1420 /** Process the WINED3DSIO_TEXDP3 instruction in ARB:
1421  * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1422 static void pshader_hw_texdp3(const struct wined3d_shader_instruction *ins)
1423 {
1424     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1425     char src0[50];
1426     char dst_str[50];
1427     char dst_mask[6];
1428     SHADER_BUFFER *buffer = ins->ctx->buffer;
1429     BOOL is_color;
1430
1431     /* Handle output register */
1432     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, dst_str, &is_color);
1433     shader_arb_get_write_mask(ins, dst, dst_mask);
1434
1435     shader_arb_get_src_param(ins, &ins->src[0], 0, src0);
1436     shader_addline(buffer, "DP3 %s%s, T%u, %s;\n", dst_str, dst_mask, dst->reg.idx, src0);
1437
1438     /* TODO: Handle output modifiers */
1439 }
1440
1441 /** Process the WINED3DSIO_TEXM3X3 instruction in ARB
1442  * Perform the 3rd row of a 3x3 matrix multiply */
1443 static void pshader_hw_texm3x3(const struct wined3d_shader_instruction *ins)
1444 {
1445     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1446     SHADER_BUFFER *buffer = ins->ctx->buffer;
1447     char dst_str[50];
1448     char dst_mask[6];
1449     char src0[50];
1450     BOOL is_color;
1451
1452     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, dst_str, &is_color);
1453     shader_arb_get_write_mask(ins, dst, dst_mask);
1454
1455     shader_arb_get_src_param(ins, &ins->src[0], 0, src0);
1456     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", dst->reg.idx, src0);
1457     shader_addline(buffer, "MOV %s%s, TMP;\n", dst_str, dst_mask);
1458
1459     /* TODO: Handle output modifiers */
1460 }
1461
1462 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in ARB:
1463  * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1464  * Calculate tmp0.y = TexCoord[dstreg] . src.xyz;  (tmp0.x has already been calculated)
1465  * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1466  */
1467 static void pshader_hw_texm3x2depth(const struct wined3d_shader_instruction *ins)
1468 {
1469     SHADER_BUFFER *buffer = ins->ctx->buffer;
1470     DWORD dst_reg = ins->dst[0].reg.idx;
1471     char src0[50];
1472
1473     shader_arb_get_src_param(ins, &ins->src[0], 0, src0);
1474     shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", dst_reg, src0);
1475
1476     /* How to deal with the special case dst_name.g == 0? if r != 0, then
1477      * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1478      * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1479      */
1480     shader_addline(buffer, "RCP TMP.y, TMP.y;\n");
1481     shader_addline(buffer, "MUL TMP.x, TMP.x, TMP.y;\n");
1482     shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1483     shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1484 }
1485
1486 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1487     Vertex/Pixel shaders to ARB_vertex_program codes */
1488 static void shader_hw_mnxn(const struct wined3d_shader_instruction *ins)
1489 {
1490     int i;
1491     int nComponents = 0;
1492     struct wined3d_shader_dst_param tmp_dst = {{0}};
1493     struct wined3d_shader_src_param tmp_src[2] = {{{0}}};
1494     struct wined3d_shader_instruction tmp_ins;
1495
1496     memset(&tmp_ins, 0, sizeof(tmp_ins));
1497
1498     /* Set constants for the temporary argument */
1499     tmp_ins.ctx = ins->ctx;
1500     tmp_ins.dst_count = 1;
1501     tmp_ins.dst = &tmp_dst;
1502     tmp_ins.src_count = 2;
1503     tmp_ins.src = tmp_src;
1504
1505     switch(ins->handler_idx)
1506     {
1507         case WINED3DSIH_M4x4:
1508             nComponents = 4;
1509             tmp_ins.handler_idx = WINED3DSIH_DP4;
1510             break;
1511         case WINED3DSIH_M4x3:
1512             nComponents = 3;
1513             tmp_ins.handler_idx = WINED3DSIH_DP4;
1514             break;
1515         case WINED3DSIH_M3x4:
1516             nComponents = 4;
1517             tmp_ins.handler_idx = WINED3DSIH_DP3;
1518             break;
1519         case WINED3DSIH_M3x3:
1520             nComponents = 3;
1521             tmp_ins.handler_idx = WINED3DSIH_DP3;
1522             break;
1523         case WINED3DSIH_M3x2:
1524             nComponents = 2;
1525             tmp_ins.handler_idx = WINED3DSIH_DP3;
1526             break;
1527         default:
1528             FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1529             break;
1530     }
1531
1532     tmp_dst = ins->dst[0];
1533     tmp_src[0] = ins->src[0];
1534     tmp_src[1] = ins->src[1];
1535     for (i = 0; i < nComponents; i++) {
1536         tmp_dst.write_mask = WINED3DSP_WRITEMASK_0 << i;
1537         shader_hw_map2gl(&tmp_ins);
1538         ++tmp_src[1].reg.idx;
1539     }
1540 }
1541
1542 static void shader_hw_rsq_rcp(const struct wined3d_shader_instruction *ins)
1543 {
1544     SHADER_BUFFER *buffer = ins->ctx->buffer;
1545     const char *instruction, *sat;
1546
1547     char dst[50];
1548     char src[50];
1549
1550     switch(ins->handler_idx)
1551     {
1552         case WINED3DSIH_RSQ: instruction = "RSQ"; break;
1553         case WINED3DSIH_RCP: instruction = "RCP"; break;
1554         default: instruction = "";
1555             FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1556             break;
1557     }
1558
1559     if(ins->dst[0].modifiers & WINED3DSPDM_SATURATE) sat = "_SAT";
1560     else sat = "";
1561
1562     shader_arb_get_dst_param(ins, &ins->dst[0], dst); /* Destination */
1563     shader_arb_get_src_param(ins, &ins->src[0], 0, src);
1564     if (ins->src[0].swizzle == WINED3DSP_NOSWIZZLE)
1565     {
1566         /* Dx sdk says .x is used if no swizzle is given, but our test shows that
1567          * .w is used
1568          */
1569         strcat(src, ".w");
1570     }
1571
1572     shader_addline(buffer, "%s%s %s, %s;\n", instruction, sat, dst, src);
1573 }
1574
1575 static void shader_hw_nrm(const struct wined3d_shader_instruction *ins)
1576 {
1577     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1578     SHADER_BUFFER *buffer = ins->ctx->buffer;
1579     char dst_name[50];
1580     char src_name[50];
1581     char dst_wmask[20];
1582     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
1583     BOOL is_color;
1584
1585     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, dst_name, &is_color);
1586     shader_arb_get_write_mask(ins, dst, dst_wmask);
1587
1588     shader_arb_get_src_param(ins, &ins->src[0], 0, src_name);
1589     shader_addline(buffer, "DP3 TMP, %s, %s;\n", src_name, src_name);
1590     shader_addline(buffer, "RSQ TMP, TMP.x;\n");
1591     /* dst.w = src[0].w * 1 / (src.x^2 + src.y^2 + src.z^2)^(1/2) according to msdn*/
1592     shader_addline(buffer, "MUL%s %s%s, %s, TMP;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1593                    src_name);
1594 }
1595
1596 static void shader_hw_sincos(const struct wined3d_shader_instruction *ins)
1597 {
1598     /* This instruction exists in ARB, but the d3d instruction takes two extra parameters which
1599      * must contain fixed constants. So we need a separate function to filter those constants and
1600      * can't use map2gl
1601      */
1602     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1603     SHADER_BUFFER *buffer = ins->ctx->buffer;
1604     char dst_name[50];
1605     char src_name[50];
1606     char dst_wmask[20];
1607     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
1608     BOOL is_color;
1609
1610     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, dst_name, &is_color);
1611     shader_arb_get_write_mask(ins, dst, dst_wmask);
1612
1613     shader_arb_get_src_param(ins, &ins->src[0], 0, src_name);
1614     shader_addline(buffer, "SCS%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1615                    src_name);
1616 }
1617
1618 static GLuint create_arb_blt_vertex_program(const WineD3D_GL_Info *gl_info)
1619 {
1620     GLuint program_id = 0;
1621     const char *blt_vprogram =
1622         "!!ARBvp1.0\n"
1623         "PARAM c[1] = { { 1, 0.5 } };\n"
1624         "MOV result.position, vertex.position;\n"
1625         "MOV result.color, c[0].x;\n"
1626         "MOV result.texcoord[0], vertex.texcoord[0];\n"
1627         "END\n";
1628
1629     GL_EXTCALL(glGenProgramsARB(1, &program_id));
1630     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, program_id));
1631     GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_vprogram), blt_vprogram));
1632
1633     if (glGetError() == GL_INVALID_OPERATION) {
1634         GLint pos;
1635         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1636         FIXME("Vertex program error at position %d: %s\n", pos,
1637             debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1638     }
1639
1640     return program_id;
1641 }
1642
1643 static GLuint create_arb_blt_fragment_program(const WineD3D_GL_Info *gl_info, enum tex_types tex_type)
1644 {
1645     GLuint program_id = 0;
1646     static const char * const blt_fprograms[tex_type_count] =
1647     {
1648         /* tex_1d */
1649         NULL,
1650         /* tex_2d */
1651         "!!ARBfp1.0\n"
1652         "TEMP R0;\n"
1653         "TEX R0.x, fragment.texcoord[0], texture[0], 2D;\n"
1654         "MOV result.depth.z, R0.x;\n"
1655         "END\n",
1656         /* tex_3d */
1657         NULL,
1658         /* tex_cube */
1659         "!!ARBfp1.0\n"
1660         "TEMP R0;\n"
1661         "TEX R0.x, fragment.texcoord[0], texture[0], CUBE;\n"
1662         "MOV result.depth.z, R0.x;\n"
1663         "END\n",
1664         /* tex_rect */
1665         "!!ARBfp1.0\n"
1666         "TEMP R0;\n"
1667         "TEX R0.x, fragment.texcoord[0], texture[0], RECT;\n"
1668         "MOV result.depth.z, R0.x;\n"
1669         "END\n",
1670     };
1671
1672     if (!blt_fprograms[tex_type])
1673     {
1674         FIXME("tex_type %#x not supported\n", tex_type);
1675         tex_type = tex_2d;
1676     }
1677
1678     GL_EXTCALL(glGenProgramsARB(1, &program_id));
1679     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, program_id));
1680     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_fprograms[tex_type]), blt_fprograms[tex_type]));
1681
1682     if (glGetError() == GL_INVALID_OPERATION) {
1683         GLint pos;
1684         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1685         FIXME("Fragment program error at position %d: %s\n", pos,
1686             debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1687     }
1688
1689     return program_id;
1690 }
1691
1692 static void shader_arb_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
1693     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1694     struct shader_arb_priv *priv = This->shader_priv;
1695     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1696
1697     if (useVS) {
1698         struct vs_compile_args compile_args;
1699
1700         TRACE("Using vertex shader\n");
1701         find_vs_compile_args((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, This->stateBlock, &compile_args);
1702         priv->current_vprogram_id = find_gl_vshader((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, &compile_args);
1703
1704         /* Bind the vertex program */
1705         GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1706         checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id);");
1707
1708         /* Enable OpenGL vertex programs */
1709         glEnable(GL_VERTEX_PROGRAM_ARB);
1710         checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1711         TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1712     } else if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
1713         priv->current_vprogram_id = 0;
1714         glDisable(GL_VERTEX_PROGRAM_ARB);
1715         checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1716     }
1717
1718     if (usePS) {
1719         struct ps_compile_args compile_args;
1720         TRACE("Using pixel shader\n");
1721         find_ps_compile_args((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader, This->stateBlock, &compile_args);
1722         priv->current_fprogram_id = find_gl_pshader((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader,
1723                                                     &compile_args);
1724
1725         /* Bind the fragment program */
1726         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1727         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id);");
1728
1729         if(!priv->use_arbfp_fixed_func) {
1730             /* Enable OpenGL fragment programs */
1731             glEnable(GL_FRAGMENT_PROGRAM_ARB);
1732             checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1733         }
1734         TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1735     } else if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM) && !priv->use_arbfp_fixed_func) {
1736         /* Disable only if we're not using arbfp fixed function fragment processing. If this is used,
1737          * keep GL_FRAGMENT_PROGRAM_ARB enabled, and the fixed function pipeline will bind the fixed function
1738          * replacement shader
1739          */
1740         glDisable(GL_FRAGMENT_PROGRAM_ARB);
1741         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1742         priv->current_fprogram_id = 0;
1743     }
1744 }
1745
1746 static void shader_arb_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {
1747     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1748     struct shader_arb_priv *priv = This->shader_priv;
1749     GLuint *blt_fprogram = &priv->depth_blt_fprogram_id[tex_type];
1750     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1751
1752     if (!priv->depth_blt_vprogram_id) priv->depth_blt_vprogram_id = create_arb_blt_vertex_program(gl_info);
1753     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->depth_blt_vprogram_id));
1754     glEnable(GL_VERTEX_PROGRAM_ARB);
1755
1756     if (!*blt_fprogram) *blt_fprogram = create_arb_blt_fragment_program(gl_info, tex_type);
1757     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, *blt_fprogram));
1758     glEnable(GL_FRAGMENT_PROGRAM_ARB);
1759 }
1760
1761 static void shader_arb_deselect_depth_blt(IWineD3DDevice *iface) {
1762     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1763     struct shader_arb_priv *priv = This->shader_priv;
1764     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1765
1766     if (priv->current_vprogram_id) {
1767         GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1768         checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vertexShader->prgId);");
1769
1770         glEnable(GL_VERTEX_PROGRAM_ARB);
1771         checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1772
1773         TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1774     } else {
1775         glDisable(GL_VERTEX_PROGRAM_ARB);
1776         checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1777     }
1778
1779     if (priv->current_fprogram_id) {
1780         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1781         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, pixelShader->prgId);");
1782
1783         glEnable(GL_FRAGMENT_PROGRAM_ARB);
1784         checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1785
1786         TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1787     } else {
1788         glDisable(GL_FRAGMENT_PROGRAM_ARB);
1789         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1790     }
1791 }
1792
1793 static void shader_arb_destroy(IWineD3DBaseShader *iface) {
1794     IWineD3DBaseShaderImpl *baseShader = (IWineD3DBaseShaderImpl *) iface;
1795     const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)baseShader->baseShader.device)->adapter->gl_info;
1796
1797     if (shader_is_pshader_version(baseShader->baseShader.reg_maps.shader_version.type))
1798     {
1799         IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *) iface;
1800         UINT i;
1801
1802         ENTER_GL();
1803         for(i = 0; i < This->num_gl_shaders; i++) {
1804             GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1805             checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1806         }
1807         LEAVE_GL();
1808         HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1809         This->gl_shaders = NULL;
1810         This->num_gl_shaders = 0;
1811         This->shader_array_size = 0;
1812     } else {
1813         IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *) iface;
1814         UINT i;
1815
1816         ENTER_GL();
1817         for(i = 0; i < This->num_gl_shaders; i++) {
1818             GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1819             checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1820         }
1821         LEAVE_GL();
1822         HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1823         This->gl_shaders = NULL;
1824         This->num_gl_shaders = 0;
1825         This->shader_array_size = 0;
1826     }
1827 }
1828
1829 static HRESULT shader_arb_alloc(IWineD3DDevice *iface) {
1830     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1831     This->shader_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
1832     return WINED3D_OK;
1833 }
1834
1835 static void shader_arb_free(IWineD3DDevice *iface) {
1836     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1837     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1838     struct shader_arb_priv *priv = This->shader_priv;
1839     int i;
1840
1841     if(priv->depth_blt_vprogram_id) {
1842         GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_vprogram_id));
1843     }
1844     for (i = 0; i < tex_type_count; ++i) {
1845         if (priv->depth_blt_fprogram_id[i]) {
1846             GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_fprogram_id[i]));
1847         }
1848     }
1849
1850     HeapFree(GetProcessHeap(), 0, This->shader_priv);
1851 }
1852
1853 static BOOL shader_arb_dirty_const(IWineD3DDevice *iface) {
1854     return TRUE;
1855 }
1856
1857 static void arbfp_add_sRGB_correction(SHADER_BUFFER *buffer, const char *fragcolor, const char *tmp1,
1858                                       const char *tmp2, const char *tmp3, const char *tmp4) {
1859     /* Perform sRGB write correction. See GLX_EXT_framebuffer_sRGB */
1860
1861     /* Calculate the > 0.0031308 case */
1862     shader_addline(buffer, "POW %s.x, %s.x, srgb_consts1.z;\n", tmp1, fragcolor);
1863     shader_addline(buffer, "POW %s.y, %s.y, srgb_consts1.z;\n", tmp1, fragcolor);
1864     shader_addline(buffer, "POW %s.z, %s.z, srgb_consts1.z;\n", tmp1, fragcolor);
1865     shader_addline(buffer, "MUL %s, %s, srgb_consts1.w;\n", tmp1, tmp1);
1866     shader_addline(buffer, "SUB %s, %s, srgb_consts2.x;\n", tmp1, tmp1);
1867     /* Calculate the < case */
1868     shader_addline(buffer, "MUL %s, srgb_consts1.x, %s;\n", tmp2, fragcolor);
1869     /* Get 1.0 / 0.0 masks for > 0.0031308 and < 0.0031308 */
1870     shader_addline(buffer, "SLT %s, srgb_consts1.y, %s;\n", tmp3, fragcolor);
1871     shader_addline(buffer, "SGE %s, srgb_consts1.y, %s;\n", tmp4, fragcolor);
1872     /* Store the components > 0.0031308 in the destination */
1873     shader_addline(buffer, "MUL %s, %s, %s;\n", fragcolor, tmp1, tmp3);
1874     /* Add the components that are < 0.0031308 */
1875     shader_addline(buffer, "MAD result.color.xyz, %s, %s, %s;\n", tmp2, tmp4, fragcolor);
1876     /* [0.0;1.0] clamping. Not needed, this is done implicitly */
1877 }
1878
1879 static GLuint shader_arb_generate_pshader(IWineD3DPixelShader *iface,
1880         SHADER_BUFFER *buffer, const struct ps_compile_args *args)
1881 {
1882     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1883     const shader_reg_maps* reg_maps = &This->baseShader.reg_maps;
1884     CONST DWORD *function = This->baseShader.function;
1885     const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)This->baseShader.device)->adapter->gl_info;
1886     const local_constant *lconst;
1887     GLuint retval;
1888     const char *fragcolor;
1889     DWORD *lconst_map = local_const_mapping((IWineD3DBaseShaderImpl *) This);
1890
1891     /*  Create the hw ARB shader */
1892     shader_addline(buffer, "!!ARBfp1.0\n");
1893
1894     if (reg_maps->shader_version.major < 3)
1895     {
1896         switch(args->fog) {
1897             case FOG_OFF:
1898                 break;
1899             case FOG_LINEAR:
1900                 shader_addline(buffer, "OPTION ARB_fog_linear;\n");
1901                 break;
1902             case FOG_EXP:
1903                 shader_addline(buffer, "OPTION ARB_fog_exp;\n");
1904                 break;
1905             case FOG_EXP2:
1906                 shader_addline(buffer, "OPTION ARB_fog_exp2;\n");
1907                 break;
1908         }
1909     }
1910
1911     shader_addline(buffer, "TEMP TMP;\n");     /* Used in matrix ops */
1912     shader_addline(buffer, "TEMP TMP2;\n");    /* Used in matrix ops */
1913     shader_addline(buffer, "TEMP TA;\n");      /* Used for modifiers */
1914     shader_addline(buffer, "TEMP TB;\n");      /* Used for modifiers */
1915     shader_addline(buffer, "TEMP TC;\n");      /* Used for modifiers */
1916     shader_addline(buffer, "PARAM coefdiv = { 0.5, 0.25, 0.125, 0.0625 };\n");
1917     shader_addline(buffer, "PARAM coefmul = { 2, 4, 8, 16 };\n");
1918     shader_addline(buffer, "PARAM one = { 1.0, 1.0, 1.0, 1.0 };\n");
1919
1920     if (reg_maps->shader_version.major < 2)
1921     {
1922         fragcolor = "R0";
1923     } else {
1924         shader_addline(buffer, "TEMP TMP_COLOR;\n");
1925         fragcolor = "TMP_COLOR";
1926     }
1927
1928     /* Base Declarations */
1929     shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION, lconst_map);
1930
1931     /* Base Shader Body */
1932     shader_generate_main((IWineD3DBaseShader *)This, buffer, reg_maps, function);
1933
1934     if(args->srgb_correction) {
1935         arbfp_add_sRGB_correction(buffer, fragcolor, "TMP", "TMP2", "TA", "TB");
1936     }
1937     shader_addline(buffer, "MOV result.color, %s;\n", fragcolor);
1938     shader_addline(buffer, "END\n");
1939
1940     /* TODO: change to resource.glObjectHandle or something like that */
1941     GL_EXTCALL(glGenProgramsARB(1, &retval));
1942
1943     TRACE("Creating a hw pixel shader, prg=%d\n", retval);
1944     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, retval));
1945
1946     TRACE("Created hw pixel shader, prg=%d\n", retval);
1947     /* Create the program and check for errors */
1948     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
1949                buffer->bsize, buffer->buffer));
1950
1951     if (glGetError() == GL_INVALID_OPERATION) {
1952         GLint errPos;
1953         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
1954         FIXME("HW PixelShader Error at position %d: %s\n",
1955               errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1956         retval = 0;
1957     }
1958
1959     /* Load immediate constants */
1960     if(lconst_map) {
1961         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
1962             const float *value = (const float *)lconst->value;
1963             GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, lconst_map[lconst->idx], value));
1964             checkGLcall("glProgramLocalParameter4fvARB");
1965         }
1966         HeapFree(GetProcessHeap(), 0, lconst_map);
1967     }
1968
1969     return retval;
1970 }
1971
1972 static GLuint shader_arb_generate_vshader(IWineD3DVertexShader *iface,
1973         SHADER_BUFFER *buffer, const struct vs_compile_args *args)
1974 {
1975     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
1976     const shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
1977     CONST DWORD *function = This->baseShader.function;
1978     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)This->baseShader.device;
1979     const WineD3D_GL_Info *gl_info = &device->adapter->gl_info;
1980     const local_constant *lconst;
1981     GLuint ret;
1982     DWORD *lconst_map = local_const_mapping((IWineD3DBaseShaderImpl *) This);
1983
1984     /*  Create the hw ARB shader */
1985     shader_addline(buffer, "!!ARBvp1.0\n");
1986     shader_addline(buffer, "PARAM helper_const = { 2.0, -1.0, %d.0, 0.0 };\n", This->rel_offset);
1987
1988     /* Mesa supports only 95 constants */
1989     if (GL_VEND(MESA) || GL_VEND(WINE))
1990         This->baseShader.limits.constant_float =
1991                 min(95, This->baseShader.limits.constant_float);
1992
1993     shader_addline(buffer, "TEMP TMP;\n");
1994
1995     /* Base Declarations */
1996     shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION, lconst_map);
1997
1998     /* We need a constant to fixup the final position */
1999     shader_addline(buffer, "PARAM posFixup = program.env[%d];\n", ARB_SHADER_PRIVCONST_POS);
2000
2001     /* Initialize output parameters. GL_ARB_vertex_program does not require special initialization values
2002      * for output parameters. D3D in theory does not do that either, but some applications depend on a
2003      * proper initialization of the secondary color, and programs using the fixed function pipeline without
2004      * a replacement shader depend on the texcoord.w being set properly.
2005      *
2006      * GL_NV_vertex_program defines that all output values are initialized to {0.0, 0.0, 0.0, 1.0}. This
2007      * assertion is in effect even when using GL_ARB_vertex_program without any NV specific additions. So
2008      * skip this if NV_vertex_program is supported. Otherwise, initialize the secondary color. For the tex-
2009      * coords, we have a flag in the opengl caps. Many cards do not require the texcoord being set, and
2010      * this can eat a number of instructions, so skip it unless this cap is set as well
2011      */
2012     if(!GL_SUPPORT(NV_VERTEX_PROGRAM)) {
2013         shader_addline(buffer, "MOV result.color.secondary, -helper_const.wwwy;\n");
2014
2015         if((GLINFO_LOCATION).set_texcoord_w && !device->frag_pipe->ffp_proj_control) {
2016             int i;
2017             for(i = 0; i < min(8, MAX_REG_TEXCRD); i++) {
2018                 if(This->baseShader.reg_maps.texcoord_mask[i] != 0 &&
2019                 This->baseShader.reg_maps.texcoord_mask[i] != WINED3DSP_WRITEMASK_ALL) {
2020                     shader_addline(buffer, "MOV result.texcoord[%u].w, -helper_const.y;\n", i);
2021                 }
2022             }
2023         }
2024     }
2025
2026     /* Base Shader Body */
2027     shader_generate_main((IWineD3DBaseShader *)This, buffer, reg_maps, function);
2028
2029     /* The D3DRS_FOGTABLEMODE render state defines if the shader-generated fog coord is used
2030      * or if the fragment depth is used. If the fragment depth is used(FOGTABLEMODE != NONE),
2031      * the fog frag coord is thrown away. If the fog frag coord is used, but not written by
2032      * the shader, it is set to 0.0(fully fogged, since start = 1.0, end = 0.0)
2033      */
2034     if(args->fog_src == VS_FOG_Z) {
2035         shader_addline(buffer, "MOV result.fogcoord, TMP_OUT.z;\n");
2036     } else if (!reg_maps->fog) {
2037         shader_addline(buffer, "MOV result.fogcoord, helper_const.w;\n");
2038     }
2039
2040     /* Write the final position.
2041      *
2042      * OpenGL coordinates specify the center of the pixel while d3d coords specify
2043      * the corner. The offsets are stored in z and w in posFixup. posFixup.y contains
2044      * 1.0 or -1.0 to turn the rendering upside down for offscreen rendering. PosFixup.x
2045      * contains 1.0 to allow a mad, but arb vs swizzles are too restricted for that.
2046      */
2047     shader_addline(buffer, "MUL TMP, posFixup, TMP_OUT.w;\n");
2048     shader_addline(buffer, "ADD TMP_OUT.x, TMP_OUT.x, TMP.z;\n");
2049     shader_addline(buffer, "MAD TMP_OUT.y, TMP_OUT.y, posFixup.y, TMP.w;\n");
2050
2051     /* Z coord [0;1]->[-1;1] mapping, see comment in transform_projection in state.c
2052      * and the glsl equivalent
2053      */
2054     shader_addline(buffer, "MAD TMP_OUT.z, TMP_OUT.z, helper_const.x, -TMP_OUT.w;\n");
2055
2056     shader_addline(buffer, "MOV result.position, TMP_OUT;\n");
2057
2058     shader_addline(buffer, "END\n");
2059
2060     /* TODO: change to resource.glObjectHandle or something like that */
2061     GL_EXTCALL(glGenProgramsARB(1, &ret));
2062
2063     TRACE("Creating a hw vertex shader, prg=%d\n", ret);
2064     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, ret));
2065
2066     TRACE("Created hw vertex shader, prg=%d\n", ret);
2067     /* Create the program and check for errors */
2068     GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2069                buffer->bsize, buffer->buffer));
2070
2071     if (glGetError() == GL_INVALID_OPERATION) {
2072         GLint errPos;
2073         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2074         FIXME("HW VertexShader Error at position %d: %s\n",
2075               errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2076         ret = -1;
2077     } else {
2078         /* Load immediate constants */
2079         if(lconst_map) {
2080             LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2081                 const float *value = (const float *)lconst->value;
2082                 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_VERTEX_PROGRAM_ARB, lconst_map[lconst->idx], value));
2083             }
2084         }
2085     }
2086     HeapFree(GetProcessHeap(), 0, lconst_map);
2087
2088     return ret;
2089 }
2090
2091 static void shader_arb_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct shader_caps *pCaps)
2092 {
2093     /* We don't have an ARB fixed function pipeline yet, so let the none backend set its caps,
2094      * then overwrite the shader specific ones
2095      */
2096     none_shader_backend.shader_get_caps(devtype, gl_info, pCaps);
2097
2098     if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
2099         pCaps->VertexShaderVersion = WINED3DVS_VERSION(1,1);
2100         TRACE_(d3d_caps)("Hardware vertex shader version 1.1 enabled (ARB_PROGRAM)\n");
2101         pCaps->MaxVertexShaderConst = GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS;
2102     }
2103
2104     if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM)) {
2105         pCaps->PixelShaderVersion    = WINED3DPS_VERSION(1,4);
2106         pCaps->PixelShader1xMaxValue = 8.0;
2107         TRACE_(d3d_caps)("Hardware pixel shader version 1.4 enabled (ARB_PROGRAM)\n");
2108         pCaps->MaxPixelShaderConst = GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS;
2109     }
2110 }
2111
2112 static BOOL shader_arb_color_fixup_supported(struct color_fixup_desc fixup)
2113 {
2114     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
2115     {
2116         TRACE("Checking support for color_fixup:\n");
2117         dump_color_fixup_desc(fixup);
2118     }
2119
2120     /* We support everything except YUV conversions. */
2121     if (!is_yuv_fixup(fixup))
2122     {
2123         TRACE("[OK]\n");
2124         return TRUE;
2125     }
2126
2127     TRACE("[FAILED]\n");
2128     return FALSE;
2129 }
2130
2131 static void shader_arb_add_instruction_modifiers(const struct wined3d_shader_instruction *ins) {
2132     BOOL saturate;
2133     DWORD shift;
2134     char write_mask[20], regstr[50];
2135     SHADER_BUFFER *buffer = ins->ctx->buffer;
2136     BOOL is_color = FALSE;
2137     const struct wined3d_shader_dst_param *dst;
2138
2139     if (!ins->dst_count) return;
2140
2141     dst = &ins->dst[0];
2142     shift = dst->shift;
2143     if(shift == 0) return; /* Saturate alone is handled by the instructions */
2144     saturate = dst->modifiers & WINED3DSPDM_SATURATE;
2145
2146     shader_arb_get_write_mask(ins, dst, write_mask);
2147     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, regstr, &is_color);
2148
2149     /* Generate a line that does the output modifier computation */
2150     shader_addline(buffer, "MUL%s %s%s, %s, %s;\n", saturate ? "_SAT" : "",
2151                    regstr, write_mask, regstr, shift_tab[shift]);
2152 }
2153
2154 static const SHADER_HANDLER shader_arb_instruction_handler_table[WINED3DSIH_TABLE_SIZE] =
2155 {
2156     /* WINED3DSIH_ABS           */ shader_hw_map2gl,
2157     /* WINED3DSIH_ADD           */ shader_hw_map2gl,
2158     /* WINED3DSIH_BEM           */ pshader_hw_bem,
2159     /* WINED3DSIH_BREAK         */ NULL,
2160     /* WINED3DSIH_BREAKC        */ NULL,
2161     /* WINED3DSIH_BREAKP        */ NULL,
2162     /* WINED3DSIH_CALL          */ NULL,
2163     /* WINED3DSIH_CALLNZ        */ NULL,
2164     /* WINED3DSIH_CMP           */ pshader_hw_cmp,
2165     /* WINED3DSIH_CND           */ pshader_hw_cnd,
2166     /* WINED3DSIH_CRS           */ shader_hw_map2gl,
2167     /* WINED3DSIH_DCL           */ NULL,
2168     /* WINED3DSIH_DEF           */ NULL,
2169     /* WINED3DSIH_DEFB          */ NULL,
2170     /* WINED3DSIH_DEFI          */ NULL,
2171     /* WINED3DSIH_DP2ADD        */ pshader_hw_dp2add,
2172     /* WINED3DSIH_DP3           */ shader_hw_map2gl,
2173     /* WINED3DSIH_DP4           */ shader_hw_map2gl,
2174     /* WINED3DSIH_DST           */ shader_hw_map2gl,
2175     /* WINED3DSIH_DSX           */ NULL,
2176     /* WINED3DSIH_DSY           */ NULL,
2177     /* WINED3DSIH_ELSE          */ NULL,
2178     /* WINED3DSIH_ENDIF         */ NULL,
2179     /* WINED3DSIH_ENDLOOP       */ NULL,
2180     /* WINED3DSIH_ENDREP        */ NULL,
2181     /* WINED3DSIH_EXP           */ shader_hw_map2gl,
2182     /* WINED3DSIH_EXPP          */ shader_hw_map2gl,
2183     /* WINED3DSIH_FRC           */ shader_hw_map2gl,
2184     /* WINED3DSIH_IF            */ NULL,
2185     /* WINED3DSIH_IFC           */ NULL,
2186     /* WINED3DSIH_LABEL         */ NULL,
2187     /* WINED3DSIH_LIT           */ shader_hw_map2gl,
2188     /* WINED3DSIH_LOG           */ shader_hw_map2gl,
2189     /* WINED3DSIH_LOGP          */ shader_hw_map2gl,
2190     /* WINED3DSIH_LOOP          */ NULL,
2191     /* WINED3DSIH_LRP           */ shader_hw_map2gl,
2192     /* WINED3DSIH_M3x2          */ shader_hw_mnxn,
2193     /* WINED3DSIH_M3x3          */ shader_hw_mnxn,
2194     /* WINED3DSIH_M3x4          */ shader_hw_mnxn,
2195     /* WINED3DSIH_M4x3          */ shader_hw_mnxn,
2196     /* WINED3DSIH_M4x4          */ shader_hw_mnxn,
2197     /* WINED3DSIH_MAD           */ shader_hw_map2gl,
2198     /* WINED3DSIH_MAX           */ shader_hw_map2gl,
2199     /* WINED3DSIH_MIN           */ shader_hw_map2gl,
2200     /* WINED3DSIH_MOV           */ shader_hw_mov,
2201     /* WINED3DSIH_MOVA          */ shader_hw_mov,
2202     /* WINED3DSIH_MUL           */ shader_hw_map2gl,
2203     /* WINED3DSIH_NOP           */ shader_hw_nop,
2204     /* WINED3DSIH_NRM           */ shader_hw_nrm,
2205     /* WINED3DSIH_PHASE         */ NULL,
2206     /* WINED3DSIH_POW           */ shader_hw_map2gl,
2207     /* WINED3DSIH_RCP           */ shader_hw_rsq_rcp,
2208     /* WINED3DSIH_REP           */ NULL,
2209     /* WINED3DSIH_RET           */ NULL,
2210     /* WINED3DSIH_RSQ           */ shader_hw_rsq_rcp,
2211     /* WINED3DSIH_SETP          */ NULL,
2212     /* WINED3DSIH_SGE           */ shader_hw_map2gl,
2213     /* WINED3DSIH_SGN           */ NULL,
2214     /* WINED3DSIH_SINCOS        */ shader_hw_sincos,
2215     /* WINED3DSIH_SLT           */ shader_hw_map2gl,
2216     /* WINED3DSIH_SUB           */ shader_hw_map2gl,
2217     /* WINED3DSIH_TEX           */ pshader_hw_tex,
2218     /* WINED3DSIH_TEXBEM        */ pshader_hw_texbem,
2219     /* WINED3DSIH_TEXBEML       */ pshader_hw_texbem,
2220     /* WINED3DSIH_TEXCOORD      */ pshader_hw_texcoord,
2221     /* WINED3DSIH_TEXDEPTH      */ pshader_hw_texdepth,
2222     /* WINED3DSIH_TEXDP3        */ pshader_hw_texdp3,
2223     /* WINED3DSIH_TEXDP3TEX     */ pshader_hw_texdp3tex,
2224     /* WINED3DSIH_TEXKILL       */ pshader_hw_texkill,
2225     /* WINED3DSIH_TEXLDD        */ NULL,
2226     /* WINED3DSIH_TEXLDL        */ NULL,
2227     /* WINED3DSIH_TEXM3x2DEPTH  */ pshader_hw_texm3x2depth,
2228     /* WINED3DSIH_TEXM3x2PAD    */ pshader_hw_texm3x2pad,
2229     /* WINED3DSIH_TEXM3x2TEX    */ pshader_hw_texm3x2tex,
2230     /* WINED3DSIH_TEXM3x3       */ pshader_hw_texm3x3,
2231     /* WINED3DSIH_TEXM3x3DIFF   */ NULL,
2232     /* WINED3DSIH_TEXM3x3PAD    */ pshader_hw_texm3x3pad,
2233     /* WINED3DSIH_TEXM3x3SPEC   */ pshader_hw_texm3x3spec,
2234     /* WINED3DSIH_TEXM3x3TEX    */ pshader_hw_texm3x3tex,
2235     /* WINED3DSIH_TEXM3x3VSPEC  */ pshader_hw_texm3x3vspec,
2236     /* WINED3DSIH_TEXREG2AR     */ pshader_hw_texreg2ar,
2237     /* WINED3DSIH_TEXREG2GB     */ pshader_hw_texreg2gb,
2238     /* WINED3DSIH_TEXREG2RGB    */ pshader_hw_texreg2rgb,
2239 };
2240
2241 const shader_backend_t arb_program_shader_backend = {
2242     shader_arb_instruction_handler_table,
2243     shader_arb_select,
2244     shader_arb_select_depth_blt,
2245     shader_arb_deselect_depth_blt,
2246     shader_arb_update_float_vertex_constants,
2247     shader_arb_update_float_pixel_constants,
2248     shader_arb_load_constants,
2249     shader_arb_load_np2fixup_constants,
2250     shader_arb_destroy,
2251     shader_arb_alloc,
2252     shader_arb_free,
2253     shader_arb_dirty_const,
2254     shader_arb_generate_pshader,
2255     shader_arb_generate_vshader,
2256     shader_arb_get_caps,
2257     shader_arb_color_fixup_supported,
2258     shader_arb_add_instruction_modifiers,
2259 };
2260
2261 /* ARB_fragment_program fixed function pipeline replacement definitions */
2262 #define ARB_FFP_CONST_TFACTOR           0
2263 #define ARB_FFP_CONST_SPECULAR_ENABLE   ((ARB_FFP_CONST_TFACTOR) + 1)
2264 #define ARB_FFP_CONST_CONSTANT(i)       ((ARB_FFP_CONST_SPECULAR_ENABLE) + 1 + i)
2265 #define ARB_FFP_CONST_BUMPMAT(i)        ((ARB_FFP_CONST_CONSTANT(7)) + 1 + i)
2266 #define ARB_FFP_CONST_LUMINANCE(i)      ((ARB_FFP_CONST_BUMPMAT(7)) + 1 + i)
2267
2268 struct arbfp_ffp_desc
2269 {
2270     struct ffp_frag_desc parent;
2271     GLuint shader;
2272     unsigned int num_textures_used;
2273 };
2274
2275 static void arbfp_enable(IWineD3DDevice *iface, BOOL enable) {
2276     if(enable) {
2277         glEnable(GL_FRAGMENT_PROGRAM_ARB);
2278         checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
2279     } else {
2280         glDisable(GL_FRAGMENT_PROGRAM_ARB);
2281         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
2282     }
2283 }
2284
2285 static HRESULT arbfp_alloc(IWineD3DDevice *iface) {
2286     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2287     struct shader_arb_priv *priv;
2288     /* Share private data between the shader backend and the pipeline replacement, if both
2289      * are the arb implementation. This is needed to figure out whether ARBfp should be disabled
2290      * if no pixel shader is bound or not
2291      */
2292     if(This->shader_backend == &arb_program_shader_backend) {
2293         This->fragment_priv = This->shader_priv;
2294     } else {
2295         This->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
2296         if(!This->fragment_priv) return E_OUTOFMEMORY;
2297     }
2298     priv = This->fragment_priv;
2299     priv->fragment_shaders = hash_table_create(ffp_frag_program_key_hash, ffp_frag_program_key_compare);
2300     priv->use_arbfp_fixed_func = TRUE;
2301     return WINED3D_OK;
2302 }
2303
2304 static void arbfp_free_ffpshader(void *value, void *gli) {
2305     const WineD3D_GL_Info *gl_info = gli;
2306     struct arbfp_ffp_desc *entry_arb = value;
2307
2308     ENTER_GL();
2309     GL_EXTCALL(glDeleteProgramsARB(1, &entry_arb->shader));
2310     checkGLcall("glDeleteProgramsARB(1, &entry_arb->shader)");
2311     HeapFree(GetProcessHeap(), 0, entry_arb);
2312     LEAVE_GL();
2313 }
2314
2315 static void arbfp_free(IWineD3DDevice *iface) {
2316     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2317     struct shader_arb_priv *priv = This->fragment_priv;
2318
2319     hash_table_destroy(priv->fragment_shaders, arbfp_free_ffpshader, &This->adapter->gl_info);
2320     priv->use_arbfp_fixed_func = FALSE;
2321
2322     if(This->shader_backend != &arb_program_shader_backend) {
2323         HeapFree(GetProcessHeap(), 0, This->fragment_priv);
2324     }
2325 }
2326
2327 static void arbfp_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct fragment_caps *caps)
2328 {
2329     caps->TextureOpCaps =  WINED3DTEXOPCAPS_DISABLE                     |
2330                            WINED3DTEXOPCAPS_SELECTARG1                  |
2331                            WINED3DTEXOPCAPS_SELECTARG2                  |
2332                            WINED3DTEXOPCAPS_MODULATE4X                  |
2333                            WINED3DTEXOPCAPS_MODULATE2X                  |
2334                            WINED3DTEXOPCAPS_MODULATE                    |
2335                            WINED3DTEXOPCAPS_ADDSIGNED2X                 |
2336                            WINED3DTEXOPCAPS_ADDSIGNED                   |
2337                            WINED3DTEXOPCAPS_ADD                         |
2338                            WINED3DTEXOPCAPS_SUBTRACT                    |
2339                            WINED3DTEXOPCAPS_ADDSMOOTH                   |
2340                            WINED3DTEXOPCAPS_BLENDCURRENTALPHA           |
2341                            WINED3DTEXOPCAPS_BLENDFACTORALPHA            |
2342                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHA           |
2343                            WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA           |
2344                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM         |
2345                            WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR      |
2346                            WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA      |
2347                            WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA   |
2348                            WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR   |
2349                            WINED3DTEXOPCAPS_DOTPRODUCT3                 |
2350                            WINED3DTEXOPCAPS_MULTIPLYADD                 |
2351                            WINED3DTEXOPCAPS_LERP                        |
2352                            WINED3DTEXOPCAPS_BUMPENVMAP                  |
2353                            WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE;
2354
2355     /* TODO: Implement WINED3DTEXOPCAPS_PREMODULATE */
2356
2357     caps->MaxTextureBlendStages   = 8;
2358     caps->MaxSimultaneousTextures = min(GL_LIMITS(fragment_samplers), 8);
2359
2360     caps->PrimitiveMiscCaps |= WINED3DPMISCCAPS_TSSARGTEMP;
2361 }
2362 #undef GLINFO_LOCATION
2363
2364 #define GLINFO_LOCATION stateblock->wineD3DDevice->adapter->gl_info
2365 static void state_texfactor_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2366     float col[4];
2367     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2368
2369     /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2370      * application provided constants
2371      */
2372     if(device->shader_backend == &arb_program_shader_backend) {
2373         if (use_ps(stateblock)) return;
2374
2375         device = stateblock->wineD3DDevice;
2376         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_TFACTOR] = 1;
2377         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_TFACTOR + 1);
2378     }
2379
2380     D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
2381     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col));
2382     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col)");
2383
2384 }
2385
2386 static void state_arb_specularenable(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2387     float col[4];
2388     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2389
2390     /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2391      * application provided constants
2392      */
2393     if(device->shader_backend == &arb_program_shader_backend) {
2394         if (use_ps(stateblock)) return;
2395
2396         device = stateblock->wineD3DDevice;
2397         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_SPECULAR_ENABLE] = 1;
2398         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_SPECULAR_ENABLE + 1);
2399     }
2400
2401     if(stateblock->renderState[WINED3DRS_SPECULARENABLE]) {
2402         /* The specular color has no alpha */
2403         col[0] = 1.0; col[1] = 1.0;
2404         col[2] = 1.0; col[3] = 0.0;
2405     } else {
2406         col[0] = 0.0; col[1] = 0.0;
2407         col[2] = 0.0; col[3] = 0.0;
2408     }
2409     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col));
2410     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col)");
2411 }
2412
2413 static void set_bumpmat_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2414     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2415     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2416     float mat[2][2];
2417
2418     if (use_ps(stateblock))
2419     {
2420         if(stage != 0 &&
2421            ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.bumpmat[stage]) {
2422             /* The pixel shader has to know the bump env matrix. Do a constants update if it isn't scheduled
2423              * anyway
2424              */
2425             if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2426                 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2427             }
2428         }
2429
2430         if(device->shader_backend == &arb_program_shader_backend) {
2431             /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2432             return;
2433         }
2434     } else if(device->shader_backend == &arb_program_shader_backend) {
2435         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_BUMPMAT(stage)] = 1;
2436         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_BUMPMAT(stage) + 1);
2437     }
2438
2439     mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
2440     mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
2441     mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
2442     mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
2443
2444     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0]));
2445     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0])");
2446 }
2447
2448 static void tex_bumpenvlum_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2449     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2450     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2451     float param[4];
2452
2453     if (use_ps(stateblock))
2454     {
2455         if(stage != 0 &&
2456            ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.luminanceparams[stage]) {
2457             /* The pixel shader has to know the luminance offset. Do a constants update if it
2458              * isn't scheduled anyway
2459              */
2460             if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2461                 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2462             }
2463         }
2464
2465         if(device->shader_backend == &arb_program_shader_backend) {
2466             /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2467             return;
2468         }
2469     } else if(device->shader_backend == &arb_program_shader_backend) {
2470         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_LUMINANCE(stage)] = 1;
2471         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_LUMINANCE(stage) + 1);
2472     }
2473
2474     param[0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLSCALE]);
2475     param[1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLOFFSET]);
2476     param[2] = 0.0;
2477     param[3] = 0.0;
2478
2479     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param));
2480     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param)");
2481 }
2482
2483 static const char *get_argreg(SHADER_BUFFER *buffer, DWORD argnum, unsigned int stage, DWORD arg) {
2484     const char *ret;
2485
2486     if(arg == ARG_UNUSED) return "unused"; /* This is the marker for unused registers */
2487
2488     switch(arg & WINED3DTA_SELECTMASK) {
2489         case WINED3DTA_DIFFUSE:
2490             ret = "fragment.color.primary"; break;
2491
2492         case WINED3DTA_CURRENT:
2493             if(stage == 0) ret = "fragment.color.primary";
2494             else ret = "ret";
2495             break;
2496
2497         case WINED3DTA_TEXTURE:
2498             switch(stage) {
2499                 case 0: ret = "tex0"; break;
2500                 case 1: ret = "tex1"; break;
2501                 case 2: ret = "tex2"; break;
2502                 case 3: ret = "tex3"; break;
2503                 case 4: ret = "tex4"; break;
2504                 case 5: ret = "tex5"; break;
2505                 case 6: ret = "tex6"; break;
2506                 case 7: ret = "tex7"; break;
2507                 default: ret = "unknown texture";
2508             }
2509             break;
2510
2511         case WINED3DTA_TFACTOR:
2512             ret = "tfactor"; break;
2513
2514         case WINED3DTA_SPECULAR:
2515             ret = "fragment.color.secondary"; break;
2516
2517         case WINED3DTA_TEMP:
2518             ret = "tempreg"; break;
2519
2520         case WINED3DTA_CONSTANT:
2521             FIXME("Implement perstage constants\n");
2522             switch(stage) {
2523                 case 0: ret = "const0"; break;
2524                 case 1: ret = "const1"; break;
2525                 case 2: ret = "const2"; break;
2526                 case 3: ret = "const3"; break;
2527                 case 4: ret = "const4"; break;
2528                 case 5: ret = "const5"; break;
2529                 case 6: ret = "const6"; break;
2530                 case 7: ret = "const7"; break;
2531                 default: ret = "unknown constant";
2532             }
2533             break;
2534
2535         default:
2536             return "unknown";
2537     }
2538
2539     if(arg & WINED3DTA_COMPLEMENT) {
2540         shader_addline(buffer, "SUB arg%u, const.x, %s;\n", argnum, ret);
2541         if(argnum == 0) ret = "arg0";
2542         if(argnum == 1) ret = "arg1";
2543         if(argnum == 2) ret = "arg2";
2544     }
2545     if(arg & WINED3DTA_ALPHAREPLICATE) {
2546         shader_addline(buffer, "MOV arg%u, %s.w;\n", argnum, ret);
2547         if(argnum == 0) ret = "arg0";
2548         if(argnum == 1) ret = "arg1";
2549         if(argnum == 2) ret = "arg2";
2550     }
2551     return ret;
2552 }
2553
2554 static void gen_ffp_instr(SHADER_BUFFER *buffer, unsigned int stage, BOOL color, BOOL alpha,
2555                           DWORD dst, DWORD op, DWORD dw_arg0, DWORD dw_arg1, DWORD dw_arg2) {
2556     const char *dstmask, *dstreg, *arg0, *arg1, *arg2;
2557     unsigned int mul = 1;
2558     BOOL mul_final_dest = FALSE;
2559
2560     if(color && alpha) dstmask = "";
2561     else if(color) dstmask = ".xyz";
2562     else dstmask = ".w";
2563
2564     if(dst == tempreg) dstreg = "tempreg";
2565     else dstreg = "ret";
2566
2567     arg0 = get_argreg(buffer, 0, stage, dw_arg0);
2568     arg1 = get_argreg(buffer, 1, stage, dw_arg1);
2569     arg2 = get_argreg(buffer, 2, stage, dw_arg2);
2570
2571     switch(op) {
2572         case WINED3DTOP_DISABLE:
2573             if(stage == 0) shader_addline(buffer, "MOV %s%s, fragment.color.primary;\n", dstreg, dstmask);
2574             break;
2575
2576         case WINED3DTOP_SELECTARG2:
2577             arg1 = arg2;
2578         case WINED3DTOP_SELECTARG1:
2579             shader_addline(buffer, "MOV %s%s, %s;\n", dstreg, dstmask, arg1);
2580             break;
2581
2582         case WINED3DTOP_MODULATE4X:
2583             mul = 2;
2584         case WINED3DTOP_MODULATE2X:
2585             mul *= 2;
2586             if(strcmp(dstreg, "result.color") == 0) {
2587                 dstreg = "ret";
2588                 mul_final_dest = TRUE;
2589             }
2590         case WINED3DTOP_MODULATE:
2591             shader_addline(buffer, "MUL %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2592             break;
2593
2594         case WINED3DTOP_ADDSIGNED2X:
2595             mul = 2;
2596             if(strcmp(dstreg, "result.color") == 0) {
2597                 dstreg = "ret";
2598                 mul_final_dest = TRUE;
2599             }
2600         case WINED3DTOP_ADDSIGNED:
2601             shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2602             arg2 = "arg2";
2603         case WINED3DTOP_ADD:
2604             shader_addline(buffer, "ADD_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2605             break;
2606
2607         case WINED3DTOP_SUBTRACT:
2608             shader_addline(buffer, "SUB_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2609             break;
2610
2611         case WINED3DTOP_ADDSMOOTH:
2612             shader_addline(buffer, "SUB arg1, const.x, %s;\n", arg1);
2613             shader_addline(buffer, "MAD_SAT %s%s, arg1, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2614             break;
2615
2616         case WINED3DTOP_BLENDCURRENTALPHA:
2617             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_CURRENT);
2618             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2619             break;
2620         case WINED3DTOP_BLENDFACTORALPHA:
2621             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TFACTOR);
2622             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2623             break;
2624         case WINED3DTOP_BLENDTEXTUREALPHA:
2625             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2626             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2627             break;
2628         case WINED3DTOP_BLENDDIFFUSEALPHA:
2629             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_DIFFUSE);
2630             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2631             break;
2632
2633         case WINED3DTOP_BLENDTEXTUREALPHAPM:
2634             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2635             shader_addline(buffer, "SUB arg0.w, const.x, %s.w;\n", arg0);
2636             shader_addline(buffer, "MAD_SAT %s%s, %s, arg0.w, %s;\n", dstreg, dstmask, arg2, arg1);
2637             break;
2638
2639         /* D3DTOP_PREMODULATE ???? */
2640
2641         case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
2642             shader_addline(buffer, "SUB arg0.w, const.x, %s;\n", arg1);
2643             shader_addline(buffer, "MAD_SAT %s%s, arg0.w, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2644             break;
2645         case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
2646             shader_addline(buffer, "MAD_SAT %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg1);
2647             break;
2648         case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
2649             shader_addline(buffer, "SUB arg0, const.x, %s;\n", arg1);
2650             shader_addline(buffer, "MAD_SAT %s%s, arg0, %s, %s.w;\n", dstreg, dstmask, arg2, arg1);
2651             break;
2652         case WINED3DTOP_MODULATECOLOR_ADDALPHA:
2653             shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s.w;\n", dstreg, dstmask, arg1, arg2, arg1);
2654             break;
2655
2656         case WINED3DTOP_DOTPRODUCT3:
2657             mul = 4;
2658             if(strcmp(dstreg, "result.color") == 0) {
2659                 dstreg = "ret";
2660                 mul_final_dest = TRUE;
2661             }
2662             shader_addline(buffer, "SUB arg1, %s, const.w;\n", arg1);
2663             shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2664             shader_addline(buffer, "DP3_SAT %s%s, arg1, arg2;\n", dstreg, dstmask);
2665             break;
2666
2667         case WINED3DTOP_MULTIPLYADD:
2668             shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg0);
2669             break;
2670
2671         case WINED3DTOP_LERP:
2672             /* The msdn is not quite right here */
2673             shader_addline(buffer, "LRP %s%s, %s, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2674             break;
2675
2676         case WINED3DTOP_BUMPENVMAP:
2677         case WINED3DTOP_BUMPENVMAPLUMINANCE:
2678             /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
2679             break;
2680
2681         default:
2682             FIXME("Unhandled texture op %08x\n", op);
2683     }
2684
2685     if(mul == 2) {
2686         shader_addline(buffer, "MUL_SAT %s%s, %s, const.y;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2687     } else if(mul == 4) {
2688         shader_addline(buffer, "MUL_SAT %s%s, %s, const.z;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2689     }
2690 }
2691
2692 /* The stateblock is passed for GLINFO_LOCATION */
2693 static GLuint gen_arbfp_ffp_shader(const struct ffp_frag_settings *settings, IWineD3DStateBlockImpl *stateblock)
2694 {
2695     unsigned int stage;
2696     SHADER_BUFFER buffer;
2697     BOOL tex_read[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2698     BOOL bump_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2699     BOOL luminance_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2700     const char *textype;
2701     const char *instr, *sat;
2702     char colorcor_dst[8];
2703     GLuint ret;
2704     DWORD arg0, arg1, arg2;
2705     BOOL tempreg_used = FALSE, tfactor_used = FALSE;
2706     BOOL op_equal;
2707     const char *final_combiner_src = "ret";
2708
2709     /* Find out which textures are read */
2710     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2711         if(settings->op[stage].cop == WINED3DTOP_DISABLE) break;
2712         arg0 = settings->op[stage].carg0 & WINED3DTA_SELECTMASK;
2713         arg1 = settings->op[stage].carg1 & WINED3DTA_SELECTMASK;
2714         arg2 = settings->op[stage].carg2 & WINED3DTA_SELECTMASK;
2715         if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2716         if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2717         if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2718
2719         if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) tex_read[stage] = TRUE;
2720         if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHAPM) tex_read[stage] = TRUE;
2721         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP) {
2722             bump_used[stage] = TRUE;
2723             tex_read[stage] = TRUE;
2724         }
2725         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2726             bump_used[stage] = TRUE;
2727             tex_read[stage] = TRUE;
2728             luminance_used[stage] = TRUE;
2729         } else if(settings->op[stage].cop == WINED3DTOP_BLENDFACTORALPHA) {
2730             tfactor_used = TRUE;
2731         }
2732
2733         if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2734             tfactor_used = TRUE;
2735         }
2736
2737         if(settings->op[stage].dst == tempreg) tempreg_used = TRUE;
2738         if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2739             tempreg_used = TRUE;
2740         }
2741
2742         if(settings->op[stage].aop == WINED3DTOP_DISABLE) continue;
2743         arg0 = settings->op[stage].aarg0 & WINED3DTA_SELECTMASK;
2744         arg1 = settings->op[stage].aarg1 & WINED3DTA_SELECTMASK;
2745         arg2 = settings->op[stage].aarg2 & WINED3DTA_SELECTMASK;
2746         if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2747         if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2748         if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2749
2750         if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2751             tempreg_used = TRUE;
2752         }
2753         if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2754             tfactor_used = TRUE;
2755         }
2756     }
2757
2758     /* Shader header */
2759     shader_buffer_init(&buffer);
2760
2761     shader_addline(&buffer, "!!ARBfp1.0\n");
2762
2763     switch(settings->fog) {
2764         case FOG_OFF:                                                         break;
2765         case FOG_LINEAR: shader_addline(&buffer, "OPTION ARB_fog_linear;\n"); break;
2766         case FOG_EXP:    shader_addline(&buffer, "OPTION ARB_fog_exp;\n");    break;
2767         case FOG_EXP2:   shader_addline(&buffer, "OPTION ARB_fog_exp2;\n");   break;
2768         default: FIXME("Unexpected fog setting %d\n", settings->fog);
2769     }
2770
2771     shader_addline(&buffer, "PARAM const = {1, 2, 4, 0.5};\n");
2772     shader_addline(&buffer, "TEMP TMP;\n");
2773     shader_addline(&buffer, "TEMP ret;\n");
2774     if(tempreg_used || settings->sRGB_write) shader_addline(&buffer, "TEMP tempreg;\n");
2775     shader_addline(&buffer, "TEMP arg0;\n");
2776     shader_addline(&buffer, "TEMP arg1;\n");
2777     shader_addline(&buffer, "TEMP arg2;\n");
2778     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2779         if(!tex_read[stage]) continue;
2780         shader_addline(&buffer, "TEMP tex%u;\n", stage);
2781         if(!bump_used[stage]) continue;
2782         shader_addline(&buffer, "PARAM bumpmat%u = program.env[%u];\n", stage, ARB_FFP_CONST_BUMPMAT(stage));
2783         if(!luminance_used[stage]) continue;
2784         shader_addline(&buffer, "PARAM luminance%u = program.env[%u];\n", stage, ARB_FFP_CONST_LUMINANCE(stage));
2785     }
2786     if(tfactor_used) {
2787         shader_addline(&buffer, "PARAM tfactor = program.env[%u];\n", ARB_FFP_CONST_TFACTOR);
2788     }
2789         shader_addline(&buffer, "PARAM specular_enable = program.env[%u];\n", ARB_FFP_CONST_SPECULAR_ENABLE);
2790
2791     if(settings->sRGB_write) {
2792         shader_addline(&buffer, "PARAM srgb_consts1 = {%f, %f, %f, %f};\n",
2793                        srgb_mul_low, srgb_cmp, srgb_pow, srgb_mul_high);
2794         shader_addline(&buffer, "PARAM srgb_consts2 = {%f, %f, %f, %f};\n",
2795                        srgb_sub_high, 0.0, 0.0, 0.0);
2796     }
2797
2798     /* Generate texture sampling instructions) */
2799     for(stage = 0; stage < MAX_TEXTURES && settings->op[stage].cop != WINED3DTOP_DISABLE; stage++) {
2800         if(!tex_read[stage]) continue;
2801
2802         switch(settings->op[stage].tex_type) {
2803             case tex_1d:                    textype = "1D";     break;
2804             case tex_2d:                    textype = "2D";     break;
2805             case tex_3d:                    textype = "3D";     break;
2806             case tex_cube:                  textype = "CUBE";   break;
2807             case tex_rect:                  textype = "RECT";   break;
2808             default: textype = "unexpected_textype";   break;
2809         }
2810
2811         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP ||
2812            settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2813             sat = "";
2814         } else {
2815             sat = "_SAT";
2816         }
2817
2818         if(settings->op[stage].projected == proj_none) {
2819             instr = "TEX";
2820         } else if(settings->op[stage].projected == proj_count4 ||
2821                   settings->op[stage].projected == proj_count3) {
2822             instr = "TXP";
2823         } else {
2824             FIXME("Unexpected projection mode %d\n", settings->op[stage].projected);
2825             instr = "TXP";
2826         }
2827
2828         if(stage > 0 &&
2829            (settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
2830             settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
2831             shader_addline(&buffer, "SWZ arg1, bumpmat%u, x, z, 0, 0;\n", stage - 1);
2832             shader_addline(&buffer, "DP3 ret.x, arg1, tex%u;\n", stage - 1);
2833             shader_addline(&buffer, "SWZ arg1, bumpmat%u, y, w, 0, 0;\n", stage - 1);
2834             shader_addline(&buffer, "DP3 ret.y, arg1, tex%u;\n", stage - 1);
2835
2836             /* with projective textures, texbem only divides the static texture coord, not the displacement,
2837              * so multiply the displacement with the dividing parameter before passing it to TXP
2838              */
2839             if (settings->op[stage].projected != proj_none) {
2840                 if(settings->op[stage].projected == proj_count4) {
2841                     shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].w;\n", stage);
2842                     shader_addline(&buffer, "MUL ret.xyz, ret, fragment.texcoord[%u].w, fragment.texcoord[%u];\n", stage, stage);
2843                 } else {
2844                     shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].z;\n", stage);
2845                     shader_addline(&buffer, "MAD ret.xyz, ret, fragment.texcoord[%u].z, fragment.texcoord[%u];\n", stage, stage);
2846                 }
2847             } else {
2848                 shader_addline(&buffer, "ADD ret, ret, fragment.texcoord[%u];\n", stage);
2849             }
2850
2851             shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2852                            instr, sat, stage, stage, textype);
2853             if(settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2854                 shader_addline(&buffer, "MAD_SAT ret.x, tex%u.z, luminance%u.x, luminance%u.y;\n",
2855                                stage - 1, stage - 1, stage - 1);
2856                 shader_addline(&buffer, "MUL tex%u, tex%u, ret.x;\n", stage, stage);
2857             }
2858         } else if(settings->op[stage].projected == proj_count3) {
2859             shader_addline(&buffer, "MOV ret, fragment.texcoord[%u];\n", stage);
2860             shader_addline(&buffer, "MOV ret.w, ret.z;\n");
2861             shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2862                             instr, sat, stage, stage, textype);
2863         } else {
2864             shader_addline(&buffer, "%s%s tex%u, fragment.texcoord[%u], texture[%u], %s;\n",
2865                             instr, sat, stage, stage, stage, textype);
2866         }
2867
2868         sprintf(colorcor_dst, "tex%u", stage);
2869         gen_color_correction(&buffer, colorcor_dst, WINED3DSP_WRITEMASK_ALL, "const.x", "const.y",
2870                 settings->op[stage].color_fixup);
2871     }
2872
2873     /* Generate the main shader */
2874     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2875         if(settings->op[stage].cop == WINED3DTOP_DISABLE) {
2876             if(stage == 0) {
2877                 final_combiner_src = "fragment.color.primary";
2878             }
2879             break;
2880         }
2881
2882         if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2883            settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2884             op_equal = settings->op[stage].carg1 == settings->op[stage].aarg1;
2885         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2886                   settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2887             op_equal = settings->op[stage].carg1 == settings->op[stage].aarg2;
2888         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2889                   settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2890             op_equal = settings->op[stage].carg2 == settings->op[stage].aarg1;
2891         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2892                   settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2893             op_equal = settings->op[stage].carg2 == settings->op[stage].aarg2;
2894         } else {
2895             op_equal = settings->op[stage].aop   == settings->op[stage].cop &&
2896                        settings->op[stage].carg0 == settings->op[stage].aarg0 &&
2897                        settings->op[stage].carg1 == settings->op[stage].aarg1 &&
2898                        settings->op[stage].carg2 == settings->op[stage].aarg2;
2899         }
2900
2901         if(settings->op[stage].aop == WINED3DTOP_DISABLE) {
2902             gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2903                           settings->op[stage].cop, settings->op[stage].carg0,
2904                           settings->op[stage].carg1, settings->op[stage].carg2);
2905             if(stage == 0) {
2906                 shader_addline(&buffer, "MOV ret.w, fragment.color.primary.w;\n");
2907             }
2908         } else if(op_equal) {
2909             gen_ffp_instr(&buffer, stage, TRUE, TRUE, settings->op[stage].dst,
2910                           settings->op[stage].cop, settings->op[stage].carg0,
2911                           settings->op[stage].carg1, settings->op[stage].carg2);
2912         } else {
2913             gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2914                           settings->op[stage].cop, settings->op[stage].carg0,
2915                           settings->op[stage].carg1, settings->op[stage].carg2);
2916             gen_ffp_instr(&buffer, stage, FALSE, TRUE, settings->op[stage].dst,
2917                           settings->op[stage].aop, settings->op[stage].aarg0,
2918                           settings->op[stage].aarg1, settings->op[stage].aarg2);
2919         }
2920     }
2921
2922     if(settings->sRGB_write) {
2923         shader_addline(&buffer, "MAD ret, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
2924         arbfp_add_sRGB_correction(&buffer, "ret", "arg0", "arg1", "arg2", "tempreg");
2925         shader_addline(&buffer, "MOV result.color.w, ret.w;\n");
2926     } else {
2927         shader_addline(&buffer, "MAD result.color, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
2928     }
2929
2930     /* Footer */
2931     shader_addline(&buffer, "END\n");
2932
2933     /* Generate the shader */
2934     GL_EXTCALL(glGenProgramsARB(1, &ret));
2935     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, ret));
2936     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
2937
2938     if (glGetError() == GL_INVALID_OPERATION) {
2939         GLint pos;
2940         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
2941         FIXME("Fragment program error at position %d: %s\n", pos,
2942               debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2943     }
2944     shader_buffer_free(&buffer);
2945     return ret;
2946 }
2947
2948 static void fragment_prog_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2949     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2950     struct shader_arb_priv *priv = device->fragment_priv;
2951     BOOL use_pshader = use_ps(stateblock);
2952     BOOL use_vshader = use_vs(stateblock);
2953     struct ffp_frag_settings settings;
2954     const struct arbfp_ffp_desc *desc;
2955     unsigned int i;
2956
2957     TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
2958
2959     if(isStateDirty(context, STATE_RENDER(WINED3DRS_FOGENABLE))) {
2960         if(!use_pshader && device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
2961             /* Reload fixed function constants since they collide with the pixel shader constants */
2962             for(i = 0; i < MAX_TEXTURES; i++) {
2963                 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
2964             }
2965             state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
2966             state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
2967         } else if(use_pshader && !isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
2968             device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
2969         }
2970         return;
2971     }
2972
2973     if(!use_pshader) {
2974         /* Find or create a shader implementing the fixed function pipeline settings, then activate it */
2975         gen_ffp_frag_op(stateblock, &settings, FALSE);
2976         desc = (const struct arbfp_ffp_desc *)find_ffp_frag_shader(priv->fragment_shaders, &settings);
2977         if(!desc) {
2978             struct arbfp_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
2979             if (!new_desc)
2980             {
2981                 ERR("Out of memory\n");
2982                 return;
2983             }
2984             new_desc->num_textures_used = 0;
2985             for(i = 0; i < GL_LIMITS(texture_stages); i++) {
2986                 if(settings.op[i].cop == WINED3DTOP_DISABLE) break;
2987                 new_desc->num_textures_used = i;
2988             }
2989
2990             memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
2991             new_desc->shader = gen_arbfp_ffp_shader(&settings, stateblock);
2992             add_ffp_frag_shader(priv->fragment_shaders, &new_desc->parent);
2993             TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
2994             desc = new_desc;
2995         }
2996
2997         /* Now activate the replacement program. GL_FRAGMENT_PROGRAM_ARB is already active(however, note the
2998          * comment above the shader_select call below). If e.g. GLSL is active, the shader_select call will
2999          * deactivate it.
3000          */
3001         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader));
3002         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader)");
3003         priv->current_fprogram_id = desc->shader;
3004
3005         if(device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3006             /* Reload fixed function constants since they collide with the pixel shader constants */
3007             for(i = 0; i < MAX_TEXTURES; i++) {
3008                 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3009             }
3010             state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3011             state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3012         }
3013         context->last_was_pshader = FALSE;
3014     } else {
3015         context->last_was_pshader = TRUE;
3016     }
3017
3018     /* Finally, select the shader. If a pixel shader is used, it will be set and enabled by the shader backend.
3019      * If this shader backend is arbfp(most likely), then it will simply overwrite the last fixed function replace-
3020      * ment shader. If the shader backend is not ARB, it currently is important that the opengl implementation
3021      * type overwrites GL_ARB_fragment_program. This is currently the case with GLSL. If we really want to use
3022      * atifs or nvrc pixel shaders with arb fragment programs we'd have to disable GL_FRAGMENT_PROGRAM_ARB here
3023      *
3024      * Don't call shader_select if the vertex shader is dirty, because it will be called later on by the vertex
3025      * shader handler
3026      */
3027     if(!isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3028         device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3029
3030         if (!isStateDirty(context, STATE_VERTEXSHADERCONSTANT) && (use_vshader || use_pshader)) {
3031             device->StateTable[STATE_VERTEXSHADERCONSTANT].apply(STATE_VERTEXSHADERCONSTANT, stateblock, context);
3032         }
3033     }
3034     if(use_pshader) {
3035         device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
3036     }
3037 }
3038
3039 /* We can't link the fog states to the fragment state directly since the vertex pipeline links them
3040  * to FOGENABLE. A different linking in different pipeline parts can't be expressed in the combined
3041  * state table, so we need to handle that with a forwarding function. The other invisible side effect
3042  * is that changing the fog start and fog end(which links to FOGENABLE in vertex) results in the
3043  * fragment_prog_arbfp function being called because FOGENABLE is dirty, which calls this function here
3044  */
3045 static void state_arbfp_fog(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3046     enum fogsource new_source;
3047
3048     TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
3049
3050     if(!isStateDirty(context, STATE_PIXELSHADER)) {
3051         fragment_prog_arbfp(state, stateblock, context);
3052     }
3053
3054     if(!stateblock->renderState[WINED3DRS_FOGENABLE]) return;
3055
3056     if(stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
3057         if(use_vs(stateblock)) {
3058             new_source = FOGSOURCE_VS;
3059         } else {
3060             if(stateblock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || context->last_was_rhw) {
3061                 new_source = FOGSOURCE_COORD;
3062             } else {
3063                 new_source = FOGSOURCE_FFP;
3064             }
3065         }
3066     } else {
3067         new_source = FOGSOURCE_FFP;
3068     }
3069     if(new_source != context->fog_source) {
3070         context->fog_source = new_source;
3071         state_fogstartend(STATE_RENDER(WINED3DRS_FOGSTART), stateblock, context);
3072     }
3073 }
3074
3075 static void textransform(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3076     if(!isStateDirty(context, STATE_PIXELSHADER)) {
3077         fragment_prog_arbfp(state, stateblock, context);
3078     }
3079 }
3080
3081 #undef GLINFO_LOCATION
3082
3083 static const struct StateEntryTemplate arbfp_fragmentstate_template[] = {
3084     {STATE_RENDER(WINED3DRS_TEXTUREFACTOR),               { STATE_RENDER(WINED3DRS_TEXTUREFACTOR),              state_texfactor_arbfp   }, 0                               },
3085     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3086     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3087     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3088     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3089     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3090     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3091     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3092     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3093     {STATE_TEXTURESTAGE(0, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3094     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3095     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3096     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3097     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3098     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3099     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3100     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3101     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3102     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3103     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3104     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3105     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3106     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3107     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3108     {STATE_TEXTURESTAGE(1, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3109     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3110     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3111     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3112     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3113     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3114     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3115     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3116     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3117     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3118     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3119     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3120     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3121     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3122     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3123     {STATE_TEXTURESTAGE(2, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3124     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3125     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3126     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3127     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3128     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3129     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3130     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3131     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3132     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3133     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3134     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3135     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3136     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3137     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3138     {STATE_TEXTURESTAGE(3, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3139     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3140     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3141     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3142     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3143     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3144     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3145     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3146     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3147     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3148     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3149     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3150     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3151     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3152     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3153     {STATE_TEXTURESTAGE(4, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3154     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3155     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3156     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3157     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3158     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3159     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3160     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3161     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3162     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3163     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3164     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3165     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3166     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3167     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3168     {STATE_TEXTURESTAGE(5, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3169     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3170     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3171     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3172     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3173     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3174     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3175     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3176     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3177     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3178     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3179     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3180     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3181     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3182     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3183     {STATE_TEXTURESTAGE(6, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3184     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3185     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3186     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3187     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3188     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3189     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3190     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3191     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3192     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3193     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3194     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3195     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3196     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3197     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3198     {STATE_TEXTURESTAGE(7, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3199     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3200     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3201     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3202     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3203     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3204     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3205     {STATE_SAMPLER(0),                                    { STATE_SAMPLER(0),                                   sampler_texdim          }, 0                               },
3206     {STATE_SAMPLER(1),                                    { STATE_SAMPLER(1),                                   sampler_texdim          }, 0                               },
3207     {STATE_SAMPLER(2),                                    { STATE_SAMPLER(2),                                   sampler_texdim          }, 0                               },
3208     {STATE_SAMPLER(3),                                    { STATE_SAMPLER(3),                                   sampler_texdim          }, 0                               },
3209     {STATE_SAMPLER(4),                                    { STATE_SAMPLER(4),                                   sampler_texdim          }, 0                               },
3210     {STATE_SAMPLER(5),                                    { STATE_SAMPLER(5),                                   sampler_texdim          }, 0                               },
3211     {STATE_SAMPLER(6),                                    { STATE_SAMPLER(6),                                   sampler_texdim          }, 0                               },
3212     {STATE_SAMPLER(7),                                    { STATE_SAMPLER(7),                                   sampler_texdim          }, 0                               },
3213     {STATE_PIXELSHADER,                                   { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3214     {STATE_RENDER(WINED3DRS_FOGENABLE),                   { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3215     {STATE_RENDER(WINED3DRS_FOGTABLEMODE),                { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3216     {STATE_RENDER(WINED3DRS_FOGVERTEXMODE),               { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3217     {STATE_RENDER(WINED3DRS_FOGSTART),                    { STATE_RENDER(WINED3DRS_FOGSTART),                   state_fogstartend       }, 0                               },
3218     {STATE_RENDER(WINED3DRS_FOGEND),                      { STATE_RENDER(WINED3DRS_FOGSTART),                   state_fogstartend       }, 0                               },
3219     {STATE_RENDER(WINED3DRS_SRGBWRITEENABLE),             { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3220     {STATE_RENDER(WINED3DRS_FOGCOLOR),                    { STATE_RENDER(WINED3DRS_FOGCOLOR),                   state_fogcolor          }, 0                               },
3221     {STATE_RENDER(WINED3DRS_FOGDENSITY),                  { STATE_RENDER(WINED3DRS_FOGDENSITY),                 state_fogdensity        }, 0                               },
3222     {STATE_TEXTURESTAGE(0,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(0, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3223     {STATE_TEXTURESTAGE(1,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(1, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3224     {STATE_TEXTURESTAGE(2,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(2, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3225     {STATE_TEXTURESTAGE(3,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(3, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3226     {STATE_TEXTURESTAGE(4,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(4, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3227     {STATE_TEXTURESTAGE(5,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(5, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3228     {STATE_TEXTURESTAGE(6,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(6, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3229     {STATE_TEXTURESTAGE(7,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(7, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3230     {STATE_RENDER(WINED3DRS_SPECULARENABLE),              { STATE_RENDER(WINED3DRS_SPECULARENABLE),             state_arb_specularenable}, 0                               },
3231     {0 /* Terminate */,                                   { 0,                                                  0                       }, 0                               },
3232 };
3233
3234 const struct fragment_pipeline arbfp_fragment_pipeline = {
3235     arbfp_enable,
3236     arbfp_get_caps,
3237     arbfp_alloc,
3238     arbfp_free,
3239     shader_arb_color_fixup_supported,
3240     arbfp_fragmentstate_template,
3241     TRUE /* We can disable projected textures */
3242 };
3243
3244 #define GLINFO_LOCATION device->adapter->gl_info
3245
3246 struct arbfp_blit_priv {
3247     GLenum yuy2_rect_shader, yuy2_2d_shader;
3248     GLenum uyvy_rect_shader, uyvy_2d_shader;
3249     GLenum yv12_rect_shader, yv12_2d_shader;
3250 };
3251
3252 static HRESULT arbfp_blit_alloc(IWineD3DDevice *iface) {
3253     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3254     device->blit_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct arbfp_blit_priv));
3255     if(!device->blit_priv) {
3256         ERR("Out of memory\n");
3257         return E_OUTOFMEMORY;
3258     }
3259     return WINED3D_OK;
3260 }
3261 static void arbfp_blit_free(IWineD3DDevice *iface) {
3262     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3263     struct arbfp_blit_priv *priv = device->blit_priv;
3264
3265     ENTER_GL();
3266     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_rect_shader));
3267     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_2d_shader));
3268     GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_rect_shader));
3269     GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_2d_shader));
3270     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_rect_shader));
3271     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_2d_shader));
3272     checkGLcall("Delete yuv programs\n");
3273     LEAVE_GL();
3274 }
3275
3276 static BOOL gen_planar_yuv_read(SHADER_BUFFER *buffer, enum yuv_fixup yuv_fixup, GLenum textype, char *luminance)
3277 {
3278     char chroma;
3279     const char *tex, *texinstr;
3280
3281     if (yuv_fixup == YUV_FIXUP_UYVY) {
3282         chroma = 'x';
3283         *luminance = 'w';
3284     } else {
3285         chroma = 'w';
3286         *luminance = 'x';
3287     }
3288     switch(textype) {
3289         case GL_TEXTURE_2D:             tex = "2D";     texinstr = "TXP"; break;
3290         case GL_TEXTURE_RECTANGLE_ARB:  tex = "RECT";   texinstr = "TEX"; break;
3291         default:
3292             /* This is more tricky than just replacing the texture type - we have to navigate
3293              * properly in the texture to find the correct chroma values
3294              */
3295             FIXME("Implement yuv correction for non-2d, non-rect textures\n");
3296             return FALSE;
3297     }
3298
3299     /* First we have to read the chroma values. This means we need at least two pixels(no filtering),
3300      * or 4 pixels(with filtering). To get the unmodified chromas, we have to rid ourselves of the
3301      * filtering when we sample the texture.
3302      *
3303      * These are the rules for reading the chroma:
3304      *
3305      * Even pixel: Cr
3306      * Even pixel: U
3307      * Odd pixel: V
3308      *
3309      * So we have to get the sampling x position in non-normalized coordinates in integers
3310      */
3311     if(textype != GL_TEXTURE_RECTANGLE_ARB) {
3312         shader_addline(buffer, "MUL texcrd.xy, fragment.texcoord[0], size.x;\n");
3313         shader_addline(buffer, "MOV texcrd.w, size.x;\n");
3314     } else {
3315         shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3316     }
3317     /* We must not allow filtering between pixel x and x+1, this would mix U and V
3318      * Vertical filtering is ok. However, bear in mind that the pixel center is at
3319      * 0.5, so add 0.5.
3320      */
3321     shader_addline(buffer, "FLR texcrd.x, texcrd.x;\n");
3322     shader_addline(buffer, "ADD texcrd.x, texcrd.x, coef.y;\n");
3323
3324     /* Divide the x coordinate by 0.5 and get the fraction. This gives 0.25 and 0.75 for the
3325      * even and odd pixels respectively
3326      */
3327     shader_addline(buffer, "MUL texcrd2, texcrd, coef.y;\n");
3328     shader_addline(buffer, "FRC texcrd2, texcrd2;\n");
3329
3330     /* Sample Pixel 1 */
3331     shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3332
3333     /* Put the value into either of the chroma values */
3334     shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3335     shader_addline(buffer, "MUL chroma.x, luminance.%c, temp.x;\n", chroma);
3336     shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3337     shader_addline(buffer, "MUL chroma.y, luminance.%c, temp.x;\n", chroma);
3338
3339     /* Sample pixel 2. If we read an even pixel(SLT above returned 1), sample
3340      * the pixel right to the current one. Otherwise, sample the left pixel.
3341      * Bias and scale the SLT result to -1;1 and add it to the texcrd.x.
3342      */
3343     shader_addline(buffer, "MAD temp.x, temp.x, coef.z, -coef.x;\n");
3344     shader_addline(buffer, "ADD texcrd.x, texcrd, temp.x;\n");
3345     shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3346
3347     /* Put the value into the other chroma */
3348     shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3349     shader_addline(buffer, "MAD chroma.y, luminance.%c, temp.x, chroma.y;\n", chroma);
3350     shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3351     shader_addline(buffer, "MAD chroma.x, luminance.%c, temp.x, chroma.x;\n", chroma);
3352
3353     /* TODO: If filtering is enabled, sample a 2nd pair of pixels left or right of
3354      * the current one and lerp the two U and V values
3355      */
3356
3357     /* This gives the correctly filtered luminance value */
3358     shader_addline(buffer, "TEX luminance, fragment.texcoord[0], texture[0], %s;\n", tex);
3359
3360     return TRUE;
3361 }
3362
3363 static BOOL gen_yv12_read(SHADER_BUFFER *buffer, GLenum textype, char *luminance)
3364 {
3365     const char *tex;
3366
3367     switch(textype) {
3368         case GL_TEXTURE_2D:             tex = "2D";     break;
3369         case GL_TEXTURE_RECTANGLE_ARB:  tex = "RECT";   break;
3370         default:
3371             FIXME("Implement yv12 correction for non-2d, non-rect textures\n");
3372             return FALSE;
3373     }
3374
3375     /* YV12 surfaces contain a WxH sized luminance plane, followed by a (W/2)x(H/2)
3376      * V and a (W/2)x(H/2) U plane, each with 8 bit per pixel. So the effective
3377      * bitdepth is 12 bits per pixel. Since the U and V planes have only half the
3378      * pitch of the luminance plane, the packing into the gl texture is a bit
3379      * unfortunate. If the whole texture is interpreted as luminance data it looks
3380      * approximately like this:
3381      *
3382      *        +----------------------------------+----
3383      *        |                                  |
3384      *        |                                  |
3385      *        |                                  |
3386      *        |                                  |
3387      *        |                                  |   2
3388      *        |            LUMINANCE             |   -
3389      *        |                                  |   3
3390      *        |                                  |
3391      *        |                                  |
3392      *        |                                  |
3393      *        |                                  |
3394      *        +----------------+-----------------+----
3395      *        |                |                 |
3396      *        |  U even rows   |  U odd rows     |
3397      *        |                |                 |   1
3398      *        +----------------+------------------   -
3399      *        |                |                 |   3
3400      *        |  V even rows   |  V odd rows     |
3401      *        |                |                 |
3402      *        +----------------+-----------------+----
3403      *        |                |                 |
3404      *        |     0.5        |       0.5       |
3405      *
3406      * So it appears as if there are 4 chroma images, but in fact the odd rows
3407      * in the chroma images are in the same row as the even ones. So its is
3408      * kinda tricky to read
3409      *
3410      * When reading from rectangle textures, keep in mind that the input y coordinates
3411      * go from 0 to d3d_height, whereas the opengl texture height is 1.5 * d3d_height
3412      */
3413     shader_addline(buffer, "PARAM yv12_coef = {%f, %f, %f, %f};\n",
3414                    2.0 / 3.0, 1.0 / 6.0, (2.0 / 3.0) + (1.0 / 6.0), 1.0 / 3.0);
3415
3416     shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3417     /* the chroma planes have only half the width */
3418     shader_addline(buffer, "MUL texcrd.x, texcrd.x, coef.y;\n");
3419
3420     /* The first value is between 2/3 and 5/6th of the texture's height, so scale+bias
3421      * the coordinate. Also read the right side of the image when reading odd lines
3422      *
3423      * Don't forget to clamp the y values in into the range, otherwise we'll get filtering
3424      * bleeding
3425      */
3426     if(textype == GL_TEXTURE_2D) {
3427
3428         shader_addline(buffer, "RCP chroma.w, size.y;\n");
3429
3430         shader_addline(buffer, "MUL texcrd2.y, texcrd.y, size.y;\n");
3431
3432         shader_addline(buffer, "FLR texcrd2.y, texcrd2.y;\n");
3433         shader_addline(buffer, "MAD texcrd.y, texcrd.y, yv12_coef.y, yv12_coef.x;\n");
3434
3435         /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3436         shader_addline(buffer, "ADD texcrd2.x, texcrd2.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3437         shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3438         shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3439         shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3440
3441         /* clamp, keep the half pixel origin in mind */
3442         shader_addline(buffer, "MAD temp.y, coef.y, chroma.w, yv12_coef.x;\n");
3443         shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3444         shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.z;\n");
3445         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3446     } else {
3447         /* Read from [size - size+size/4] */
3448         shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3449         shader_addline(buffer, "MAD texcrd.y, texcrd.y, coef.w, size.y;\n");
3450
3451         /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3452         shader_addline(buffer, "ADD texcrd2.x, texcrd.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3453         shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3454         shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3455         shader_addline(buffer, "MUL texcrd2.x, texcrd2.x, size.x;\n");
3456         shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3457
3458         /* Make sure to read exactly from the pixel center */
3459         shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3460         shader_addline(buffer, "ADD texcrd.y, texcrd.y, coef.y;\n");
3461
3462         /* Clamp */
3463         shader_addline(buffer, "MAD temp.y, size.y, coef.w, size.y;\n");
3464         shader_addline(buffer, "ADD temp.y, temp.y, -coef.y;\n");
3465         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3466         shader_addline(buffer, "ADD temp.y, size.y, -coef.y;\n");
3467         shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3468     }
3469     /* Read the texture, put the result into the output register */
3470     shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3471     shader_addline(buffer, "MOV chroma.x, temp.w;\n");
3472
3473     /* The other chroma value is 1/6th of the texture lower, from 5/6th to 6/6th
3474      * No need to clamp because we're just reusing the already clamped value from above
3475      */
3476     if(textype == GL_TEXTURE_2D) {
3477         shader_addline(buffer, "ADD texcrd.y, texcrd.y, yv12_coef.y;\n");
3478     } else {
3479         shader_addline(buffer, "MAD texcrd.y, size.y, coef.w, texcrd.y;\n");
3480     }
3481     shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3482     shader_addline(buffer, "MOV chroma.y, temp.w;\n");
3483
3484     /* Sample the luminance value. It is in the top 2/3rd of the texture, so scale the y coordinate.
3485      * Clamp the y coordinate to prevent the chroma values from bleeding into the sampled luminance
3486      * values due to filtering
3487      */
3488     shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3489     if(textype == GL_TEXTURE_2D) {
3490         /* Multiply the y coordinate by 2/3 and clamp it */
3491         shader_addline(buffer, "MUL texcrd.y, texcrd.y, yv12_coef.x;\n");
3492         shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.x;\n");
3493         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3494         shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3495     } else {
3496         /* Reading from texture_rectangles is pretty straightforward, just use the unmodified
3497          * texture coordinate. It is still a good idea to clamp it though, since the opengl texture
3498          * is bigger
3499          */
3500         shader_addline(buffer, "ADD temp.x, size.y, -coef.y;\n");
3501         shader_addline(buffer, "MIN texcrd.y, texcrd.y, size.x;\n");
3502         shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3503     }
3504     *luminance = 'a';
3505
3506     return TRUE;
3507 }
3508
3509 static GLuint gen_yuv_shader(IWineD3DDeviceImpl *device, enum yuv_fixup yuv_fixup, GLenum textype)
3510 {
3511     GLenum shader;
3512     SHADER_BUFFER buffer;
3513     char luminance_component;
3514     struct arbfp_blit_priv *priv = device->blit_priv;
3515
3516     /* Shader header */
3517     shader_buffer_init(&buffer);
3518
3519     ENTER_GL();
3520     GL_EXTCALL(glGenProgramsARB(1, &shader));
3521     checkGLcall("GL_EXTCALL(glGenProgramsARB(1, &shader))");
3522     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3523     checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3524     LEAVE_GL();
3525     if(!shader) {
3526         shader_buffer_free(&buffer);
3527         return 0;
3528     }
3529
3530     /* The YUY2 and UYVY formats contain two pixels packed into a 32 bit macropixel,
3531      * giving effectively 16 bit per pixel. The color consists of a luminance(Y) and
3532      * two chroma(U and V) values. Each macropixel has two luminance values, one for
3533      * each single pixel it contains, and one U and one V value shared between both
3534      * pixels.
3535      *
3536      * The data is loaded into an A8L8 texture. With YUY2, the luminance component
3537      * contains the luminance and alpha the chroma. With UYVY it is vice versa. Thus
3538      * take the format into account when generating the read swizzles
3539      *
3540      * Reading the Y value is straightforward - just sample the texture. The hardware
3541      * takes care of filtering in the horizontal and vertical direction.
3542      *
3543      * Reading the U and V values is harder. We have to avoid filtering horizontally,
3544      * because that would mix the U and V values of one pixel or two adjacent pixels.
3545      * Thus floor the texture coordinate and add 0.5 to get an unfiltered read,
3546      * regardless of the filtering setting. Vertical filtering works automatically
3547      * though - the U and V values of two rows are mixed nicely.
3548      *
3549      * Appart of avoiding filtering issues, the code has to know which value it just
3550      * read, and where it can find the other one. To determine this, it checks if
3551      * it sampled an even or odd pixel, and shifts the 2nd read accordingly.
3552      *
3553      * Handling horizontal filtering of U and V values requires reading a 2nd pair
3554      * of pixels, extracting U and V and mixing them. This is not implemented yet.
3555      *
3556      * An alternative implementation idea is to load the texture as A8R8G8B8 texture,
3557      * with width / 2. This way one read gives all 3 values, finding U and V is easy
3558      * in an unfiltered situation. Finding the luminance on the other hand requires
3559      * finding out if it is an odd or even pixel. The real drawback of this approach
3560      * is filtering. This would have to be emulated completely in the shader, reading
3561      * up two 2 packed pixels in up to 2 rows and interpolating both horizontally and
3562      * vertically. Beyond that it would require adjustments to the texture handling
3563      * code to deal with the width scaling
3564      */
3565     shader_addline(&buffer, "!!ARBfp1.0\n");
3566     shader_addline(&buffer, "TEMP luminance;\n");
3567     shader_addline(&buffer, "TEMP temp;\n");
3568     shader_addline(&buffer, "TEMP chroma;\n");
3569     shader_addline(&buffer, "TEMP texcrd;\n");
3570     shader_addline(&buffer, "TEMP texcrd2;\n");
3571     shader_addline(&buffer, "PARAM coef = {1.0, 0.5, 2.0, 0.25};\n");
3572     shader_addline(&buffer, "PARAM yuv_coef = {1.403, 0.344, 0.714, 1.770};\n");
3573     shader_addline(&buffer, "PARAM size = program.local[0];\n");
3574
3575     switch (yuv_fixup)
3576     {
3577         case YUV_FIXUP_UYVY:
3578         case YUV_FIXUP_YUY2:
3579             if (!gen_planar_yuv_read(&buffer, yuv_fixup, textype, &luminance_component))
3580             {
3581                 shader_buffer_free(&buffer);
3582                 return 0;
3583             }
3584             break;
3585
3586         case YUV_FIXUP_YV12:
3587             if (!gen_yv12_read(&buffer, textype, &luminance_component))
3588             {
3589                 shader_buffer_free(&buffer);
3590                 return 0;
3591             }
3592             break;
3593
3594         default:
3595             FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3596             shader_buffer_free(&buffer);
3597             return 0;
3598     }
3599
3600     /* Calculate the final result. Formula is taken from
3601      * http://www.fourcc.org/fccyvrgb.php. Note that the chroma
3602      * ranges from -0.5 to 0.5
3603      */
3604     shader_addline(&buffer, "SUB chroma.xy, chroma, coef.y;\n");
3605
3606     shader_addline(&buffer, "MAD result.color.x, chroma.x, yuv_coef.x, luminance.%c;\n", luminance_component);
3607     shader_addline(&buffer, "MAD temp.x, -chroma.y, yuv_coef.y, luminance.%c;\n", luminance_component);
3608     shader_addline(&buffer, "MAD result.color.y, -chroma.x, yuv_coef.z, temp.x;\n");
3609     shader_addline(&buffer, "MAD result.color.z, chroma.y, yuv_coef.w, luminance.%c;\n", luminance_component);
3610     shader_addline(&buffer, "END\n");
3611
3612     ENTER_GL();
3613     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3614
3615     if (glGetError() == GL_INVALID_OPERATION) {
3616         GLint pos;
3617         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3618         FIXME("Fragment program error at position %d: %s\n", pos,
3619               debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3620     }
3621     shader_buffer_free(&buffer);
3622     LEAVE_GL();
3623
3624     switch (yuv_fixup)
3625     {
3626         case YUV_FIXUP_YUY2:
3627             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yuy2_rect_shader = shader;
3628             else priv->yuy2_2d_shader = shader;
3629             break;
3630
3631         case YUV_FIXUP_UYVY:
3632             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->uyvy_rect_shader = shader;
3633             else priv->uyvy_2d_shader = shader;
3634             break;
3635
3636         case YUV_FIXUP_YV12:
3637             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yv12_rect_shader = shader;
3638             else priv->yv12_2d_shader = shader;
3639             break;
3640     }
3641
3642     return shader;
3643 }
3644
3645 static HRESULT arbfp_blit_set(IWineD3DDevice *iface, const struct GlPixelFormatDesc *format_desc,
3646         GLenum textype, UINT width, UINT height)
3647 {
3648     GLenum shader;
3649     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3650     float size[4] = {width, height, 1, 1};
3651     struct arbfp_blit_priv *priv = device->blit_priv;
3652     enum yuv_fixup yuv_fixup;
3653
3654     if (!is_yuv_fixup(format_desc->color_fixup))
3655     {
3656         TRACE("Fixup:\n");
3657         dump_color_fixup_desc(format_desc->color_fixup);
3658         /* Don't bother setting up a shader for unconverted formats */
3659         ENTER_GL();
3660         glEnable(textype);
3661         checkGLcall("glEnable(textype)");
3662         LEAVE_GL();
3663         return WINED3D_OK;
3664     }
3665
3666     yuv_fixup = get_yuv_fixup(format_desc->color_fixup);
3667
3668     switch(yuv_fixup)
3669     {
3670         case YUV_FIXUP_YUY2:
3671             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yuy2_rect_shader : priv->yuy2_2d_shader;
3672             break;
3673
3674         case YUV_FIXUP_UYVY:
3675             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->uyvy_rect_shader : priv->uyvy_2d_shader;
3676             break;
3677
3678         case YUV_FIXUP_YV12:
3679             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yv12_rect_shader : priv->yv12_2d_shader;
3680             break;
3681
3682         default:
3683             FIXME("Unsupported YUV fixup %#x, not setting a shader\n", yuv_fixup);
3684             ENTER_GL();
3685             glEnable(textype);
3686             checkGLcall("glEnable(textype)");
3687             LEAVE_GL();
3688             return E_NOTIMPL;
3689     }
3690
3691     if (!shader) shader = gen_yuv_shader(device, yuv_fixup, textype);
3692
3693     ENTER_GL();
3694     glEnable(GL_FRAGMENT_PROGRAM_ARB);
3695     checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
3696     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3697     checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3698     GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, 0, size));
3699     checkGLcall("glProgramLocalParameter4fvARB");
3700     LEAVE_GL();
3701
3702     return WINED3D_OK;
3703 }
3704
3705 static void arbfp_blit_unset(IWineD3DDevice *iface) {
3706     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3707
3708     ENTER_GL();
3709     glDisable(GL_FRAGMENT_PROGRAM_ARB);
3710     checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
3711     glDisable(GL_TEXTURE_2D);
3712     checkGLcall("glDisable(GL_TEXTURE_2D)");
3713     if(GL_SUPPORT(ARB_TEXTURE_CUBE_MAP)) {
3714         glDisable(GL_TEXTURE_CUBE_MAP_ARB);
3715         checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
3716     }
3717     if(GL_SUPPORT(ARB_TEXTURE_RECTANGLE)) {
3718         glDisable(GL_TEXTURE_RECTANGLE_ARB);
3719         checkGLcall("glDisable(GL_TEXTURE_RECTANGLE_ARB)");
3720     }
3721     LEAVE_GL();
3722 }
3723
3724 static BOOL arbfp_blit_color_fixup_supported(struct color_fixup_desc fixup)
3725 {
3726     enum yuv_fixup yuv_fixup;
3727
3728     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
3729     {
3730         TRACE("Checking support for fixup:\n");
3731         dump_color_fixup_desc(fixup);
3732     }
3733
3734     if (is_identity_fixup(fixup))
3735     {
3736         TRACE("[OK]\n");
3737         return TRUE;
3738     }
3739
3740     /* We only support YUV conversions. */
3741     if (!is_yuv_fixup(fixup))
3742     {
3743         TRACE("[FAILED]\n");
3744         return FALSE;
3745     }
3746
3747     yuv_fixup = get_yuv_fixup(fixup);
3748     switch(yuv_fixup)
3749     {
3750         case YUV_FIXUP_YUY2:
3751         case YUV_FIXUP_UYVY:
3752         case YUV_FIXUP_YV12:
3753             TRACE("[OK]\n");
3754             return TRUE;
3755
3756         default:
3757             FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3758             TRACE("[FAILED]\n");
3759             return FALSE;
3760     }
3761 }
3762
3763 const struct blit_shader arbfp_blit = {
3764     arbfp_blit_alloc,
3765     arbfp_blit_free,
3766     arbfp_blit_set,
3767     arbfp_blit_unset,
3768     arbfp_blit_color_fixup_supported,
3769 };
3770
3771 #undef GLINFO_LOCATION