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     DWORD shader_version = This->baseShader.reg_maps.shader_version;
  85     local_constant* lconst;
  86     DWORD i, j;
  87     unsigned int ret;
  88
  89     if (TRACE_ON(d3d_shader)) {
  90         for(i = 0; i < max_constants; i++) {
  91             if(!dirty_consts[i]) continue;
  92             TRACE_(d3d_constants)("Loading constants %i: %f, %f, %f, %f\n", i,
  93                         constants[i * 4 + 0], constants[i * 4 + 1],
  94                         constants[i * 4 + 2], constants[i * 4 + 3]);
  95         }
  96     }
  97     /* In 1.X pixel shaders constants are implicitly clamped in the range [-1;1] */
  98     if (target_type == GL_FRAGMENT_PROGRAM_ARB && WINED3DSHADER_VERSION_MAJOR(shader_version) == 1)
  99     {
 100         float lcl_const[4];
 101         for(i = 0; i < max_constants; i++) {
 102             if(!dirty_consts[i]) continue;
 103             dirty_consts[i] = 0;
 104
 105             j = 4 * i;
 106             if(constants[j + 0] > 1.0) lcl_const[0] = 1.0;
 107             else if(constants[j + 0] < -1.0) lcl_const[0] = -1.0;
 108             else lcl_const[0] = constants[j + 0];
 109
 110             if(constants[j + 1] > 1.0) lcl_const[1] = 1.0;
 111             else if(constants[j + 1] < -1.0) lcl_const[1] = -1.0;
 112             else lcl_const[1] = constants[j + 1];
 113
 114             if(constants[j + 2] > 1.0) lcl_const[2] = 1.0;
 115             else if(constants[j + 2] < -1.0) lcl_const[2] = -1.0;
 116             else lcl_const[2] = constants[j + 2];
 117
 118             if(constants[j + 3] > 1.0) lcl_const[3] = 1.0;
 119             else if(constants[j + 3] < -1.0) lcl_const[3] = -1.0;
 120             else lcl_const[3] = constants[j + 3];
 121
 122             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, lcl_const));
 123         }
 124     } else {
 125         if(GL_SUPPORT(EXT_GPU_PROGRAM_PARAMETERS)) {
 126             /* TODO: Benchmark if we're better of with finding the dirty constants ourselves,
 127              * or just reloading *all* constants at once
 128              *
 129             GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, 0, max_constants, constants));
 130              */
 131             for(i = 0; i < max_constants; i++) {
 132                 if(!dirty_consts[i]) continue;
 133
 134                 /* Find the next block of dirty constants */
 135                 dirty_consts[i] = 0;
 136                 j = i;
 137                 for(i++; (i < max_constants) && dirty_consts[i]; i++) {
 138                     dirty_consts[i] = 0;
 139                 }
 140
 141                 GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, j, i - j, constants + (j * 4)));
 142             }
 143         } else {
 144             for(i = 0; i < max_constants; i++) {
 145                 if(dirty_consts[i]) {
 146                     dirty_consts[i] = 0;
 147                     GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, constants + (i * 4)));
 148                 }
 149             }
 150         }
 151     }
 152     checkGLcall("glProgramEnvParameter4fvARB()");
 153
 154     /* Load immediate constants */
 155     if(This->baseShader.load_local_constsF) {
 156         if (TRACE_ON(d3d_shader)) {
 157             LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 158                 GLfloat* values = (GLfloat*)lconst->value;
 159                 TRACE_(d3d_constants)("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
 160                         values[0], values[1], values[2], values[3]);
 161             }
 162         }
 163         /* Immediate constants are clamped for 1.X shaders at loading times */
 164         ret = 0;
 165         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 166             dirty_consts[lconst->idx] = 1; /* Dirtify so the non-immediate constant overwrites it next time */
 167             ret = max(ret, lconst->idx + 1);
 168             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, lconst->idx, (GLfloat*)lconst->value));
 169         }
 170         checkGLcall("glProgramEnvParameter4fvARB()");
 171         return ret; /* The loaded immediate constants need reloading for the next shader */
 172     } else {
 173         return 0; /* No constants are dirty now */
 174     }
 175 }
 176
 177 /**
 178  * Loads the texture dimensions for NP2 fixup into the currently set ARB_[vertex/fragment]_programs.
 179  */
 180 static void shader_arb_load_np2fixup_constants(
 181     IWineD3DDevice* device,
 182     char usePixelShader,
 183     char useVertexShader) {
 184     /* not implemented */
 185 }
 186
 187 /**
 188  * Loads the app-supplied constants into the currently set ARB_[vertex/fragment]_programs.
 189  *
 190  * We only support float constants in ARB at the moment, so don't
 191  * worry about the Integers or Booleans
 192  */
 193 static void shader_arb_load_constants(
 194     IWineD3DDevice* device,
 195     char usePixelShader,
 196     char useVertexShader) {
 197
 198     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
 199     IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
 200     const WineD3D_GL_Info *gl_info = &deviceImpl->adapter->gl_info;
 201     unsigned char i;
 202
 203     if (useVertexShader) {
 204         IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
 205
 206         /* Load DirectX 9 float constants for vertex shader */
 207         deviceImpl->highest_dirty_vs_const = shader_arb_load_constantsF(
 208                 vshader, gl_info, GL_VERTEX_PROGRAM_ARB,
 209                 deviceImpl->highest_dirty_vs_const,
 210                 stateBlock->vertexShaderConstantF,
 211                 deviceImpl->activeContext->vshader_const_dirty);
 212
 213         /* Upload the position fixup */
 214         GL_EXTCALL(glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, ARB_SHADER_PRIVCONST_POS, deviceImpl->posFixup));
 215     }
 216
 217     if (usePixelShader) {
 218
 219         IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
 220         IWineD3DPixelShaderImpl *psi = (IWineD3DPixelShaderImpl *) pshader;
 221
 222         /* Load DirectX 9 float constants for pixel shader */
 223         deviceImpl->highest_dirty_ps_const = shader_arb_load_constantsF(
 224                 pshader, gl_info, GL_FRAGMENT_PROGRAM_ARB,
 225                 deviceImpl->highest_dirty_ps_const,
 226                 stateBlock->pixelShaderConstantF,
 227                 deviceImpl->activeContext->pshader_const_dirty);
 228
 229         for(i = 0; i < psi->numbumpenvmatconsts; i++) {
 230             /* The state manager takes care that this function is always called if the bump env matrix changes
 231              */
 232             const float *data = (const float *)&stateBlock->textureState[(int) psi->bumpenvmatconst[i].texunit][WINED3DTSS_BUMPENVMAT00];
 233             GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->bumpenvmatconst[i].const_num, data));
 234             deviceImpl->activeContext->pshader_const_dirty[psi->bumpenvmatconst[i].const_num] = 1;
 235
 236             if (psi->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED)
 237             {
 238                 /* WINED3DTSS_BUMPENVLSCALE and WINED3DTSS_BUMPENVLOFFSET are next to each other.
 239                  * point gl to the scale, and load 4 floats. x = scale, y = offset, z and w are junk, we
 240                  * don't care about them. The pointers are valid for sure because the stateblock is bigger.
 241                  * (they're WINED3DTSS_TEXTURETRANSFORMFLAGS and WINED3DTSS_ADDRESSW, so most likely 0 or NaN
 242                  */
 243                 const float *scale = (const float *)&stateBlock->textureState[(int) psi->luminanceconst[i].texunit][WINED3DTSS_BUMPENVLSCALE];
 244                 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->luminanceconst[i].const_num, scale));
 245                 deviceImpl->activeContext->pshader_const_dirty[psi->luminanceconst[i].const_num] = 1;
 246             }
 247         }
 248     }
 249 }
 250
 251 static void shader_arb_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count)
 252 {
 253     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 254
 255     /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
 256      * context. On a context switch the old context will be fully dirtified */
 257     memset(This->activeContext->vshader_const_dirty + start, 1,
 258             sizeof(*This->activeContext->vshader_const_dirty) * count);
 259     This->highest_dirty_vs_const = max(This->highest_dirty_vs_const, start + count + 1);
 260 }
 261
 262 static void shader_arb_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count)
 263 {
 264     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 265
 266     /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
 267      * context. On a context switch the old context will be fully dirtified */
 268     memset(This->activeContext->pshader_const_dirty + start, 1,
 269             sizeof(*This->activeContext->pshader_const_dirty) * count);
 270     This->highest_dirty_ps_const = max(This->highest_dirty_ps_const, start + count + 1);
 271 }
 272
 273 static DWORD *local_const_mapping(IWineD3DBaseShaderImpl *This)
 274 {
 275     DWORD *ret;
 276     DWORD idx = 0;
 277     const local_constant *lconst;
 278
 279     if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) return NULL;
 280
 281     ret = HeapAlloc(GetProcessHeap(), 0, sizeof(DWORD) * This->baseShader.limits.temporary);
 282     if(!ret) {
 283         ERR("Out of memory\n");
 284         return NULL;
 285     }
 286
 287     LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 288         ret[lconst->idx] = idx++;
 289     }
 290     return ret;
 291 }
 292
 293 /* Generate the variable & register declarations for the ARB_vertex_program output target */
 294 static void shader_generate_arb_declarations(IWineD3DBaseShader *iface, const shader_reg_maps *reg_maps,
 295         SHADER_BUFFER *buffer, const WineD3D_GL_Info *gl_info, DWORD *lconst_map)
 296 {
 297     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 298     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
 299     DWORD i, cur;
 300     char pshader = shader_is_pshader_version(reg_maps->shader_version);
 301     unsigned max_constantsF = min(This->baseShader.limits.constant_float,
 302             (pshader ? GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS :
 303                        GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS));
 304     UINT extra_constants_needed = 0;
 305     const local_constant *lconst;
 306
 307     /* Temporary Output register */
 308     shader_addline(buffer, "TEMP TMP_OUT;\n");
 309
 310     for(i = 0; i < This->baseShader.limits.temporary; i++) {
 311         if (reg_maps->temporary[i])
 312             shader_addline(buffer, "TEMP R%u;\n", i);
 313     }
 314
 315     for (i = 0; i < This->baseShader.limits.address; i++) {
 316         if (reg_maps->address[i])
 317             shader_addline(buffer, "ADDRESS A%d;\n", i);
 318     }
 319
 320     for(i = 0; i < This->baseShader.limits.texcoord; i++) {
 321         if (reg_maps->texcoord[i])
 322             shader_addline(buffer,"TEMP T%u;\n", i);
 323     }
 324
 325     /* Texture coordinate registers must be pre-loaded */
 326     for (i = 0; i < This->baseShader.limits.texcoord; i++) {
 327         if (reg_maps->texcoord[i])
 328             shader_addline(buffer, "MOV T%u, fragment.texcoord[%u];\n", i, i);
 329     }
 330
 331     for(i = 0; i < (sizeof(reg_maps->bumpmat) / sizeof(reg_maps->bumpmat[0])); i++) {
 332         IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *) This;
 333         if(!reg_maps->bumpmat[i]) continue;
 334
 335         cur = ps->numbumpenvmatconsts;
 336         ps->bumpenvmatconst[cur].const_num = -1;
 337         ps->bumpenvmatconst[cur].texunit = i;
 338         ps->luminanceconst[cur].const_num = -1;
 339         ps->luminanceconst[cur].texunit = i;
 340
 341         /* If the shader does not use all available constants, use the next free constant to load the bump mapping environment matrix from
 342          * the stateblock into the shader. If no constant is available don't load, texbem will then just sample the texture without applying
 343          * bump mapping.
 344          */
 345         if(max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS) {
 346             ps->bumpenvmatconst[cur].const_num = max_constantsF + extra_constants_needed;
 347             shader_addline(buffer, "PARAM bumpenvmat%d = program.env[%d];\n",
 348                            i, ps->bumpenvmatconst[cur].const_num);
 349             extra_constants_needed++;
 350
 351             if(reg_maps->luminanceparams && max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS) {
 352                 ((IWineD3DPixelShaderImpl *)This)->luminanceconst[cur].const_num = max_constantsF + extra_constants_needed;
 353                 shader_addline(buffer, "PARAM luminance%d = program.env[%d];\n",
 354                                i, ps->luminanceconst[cur].const_num);
 355                 extra_constants_needed++;
 356             } else if(reg_maps->luminanceparams) {
 357                 FIXME("No free constant to load the luminance parameters\n");
 358             }
 359         } else {
 360             FIXME("No free constant found to load environment bump mapping matrix into the shader. texbem instruction will not apply bump mapping\n");
 361         }
 362
 363         ps->numbumpenvmatconsts = cur + 1;
 364     }
 365
 366     if(device->stateBlock->renderState[WINED3DRS_SRGBWRITEENABLE] && pshader) {
 367         shader_addline(buffer, "PARAM srgb_consts1 = {%f, %f, %f, %f};\n",
 368                        srgb_mul_low, srgb_cmp, srgb_pow, srgb_mul_high);
 369         shader_addline(buffer, "PARAM srgb_consts2 = {%f, %f, %f, %f};\n",
 370                        srgb_sub_high, 0.0, 0.0, 0.0);
 371     }
 372
 373     /* Load local constants using the program-local space,
 374      * this avoids reloading them each time the shader is used
 375      */
 376     if(lconst_map) {
 377         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 378             shader_addline(buffer, "PARAM C%u = program.local[%u];\n", lconst->idx,
 379                            lconst_map[lconst->idx]);
 380         }
 381     }
 382
 383     /* we use the array-based constants array if the local constants are marked for loading,
 384      * because then we use indirect addressing, or when the local constant list is empty,
 385      * because then we don't know if we're using indirect addressing or not. If we're hardcoding
 386      * local constants do not declare the loaded constants as an array because ARB compilers usually
 387      * do not optimize unused constants away
 388      */
 389     if(!lconst_map || list_empty(&This->baseShader.constantsF)) {
 390         /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
 391         shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
 392                     max_constantsF, max_constantsF - 1);
 393     } else {
 394         for(i = 0; i < max_constantsF; i++) {
 395             if(!shader_constant_is_local(This, i)) {
 396                 shader_addline(buffer, "PARAM C%d = program.env[%d];\n",i, i);
 397             }
 398         }
 399     }
 400 }
 401
 402 static const char * const shift_tab[] = {
 403     "dummy",     /*  0 (none) */
 404     "coefmul.x", /*  1 (x2)   */
 405     "coefmul.y", /*  2 (x4)   */
 406     "coefmul.z", /*  3 (x8)   */
 407     "coefmul.w", /*  4 (x16)  */
 408     "dummy",     /*  5 (x32)  */
 409     "dummy",     /*  6 (x64)  */
 410     "dummy",     /*  7 (x128) */
 411     "dummy",     /*  8 (d256) */
 412     "dummy",     /*  9 (d128) */
 413     "dummy",     /* 10 (d64)  */
 414     "dummy",     /* 11 (d32)  */
 415     "coefdiv.w", /* 12 (d16)  */
 416     "coefdiv.z", /* 13 (d8)   */
 417     "coefdiv.y", /* 14 (d4)   */
 418     "coefdiv.x"  /* 15 (d2)   */
 419 };
 420
 421 static void shader_arb_get_write_mask(const struct wined3d_shader_instruction *ins,
 422         const struct wined3d_shader_dst_param *dst, char *write_mask)
 423 {
 424     char *ptr = write_mask;
 425     char vshader = shader_is_vshader_version(ins->ctx->reg_maps->shader_version);
 426
 427     if (vshader && dst->register_type == WINED3DSPR_ADDR)
 428     {
 429         *ptr++ = '.';
 430         *ptr++ = 'x';
 431     }
 432     else if (dst->write_mask != WINED3DSP_WRITEMASK_ALL)
 433     {
 434         *ptr++ = '.';
 435         if (dst->write_mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
 436         if (dst->write_mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
 437         if (dst->write_mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
 438         if (dst->write_mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
 439     }
 440
 441     *ptr = '\0';
 442 }
 443
 444 static void shader_arb_get_swizzle(const struct wined3d_shader_src_param *param, BOOL fixup, char *swizzle_str)
 445 {
 446     /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
 447      * but addressed as "rgba". To fix this we need to swap the register's x
 448      * and z components. */
 449     const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
 450     char *ptr = swizzle_str;
 451
 452     /* swizzle bits fields: wwzzyyxx */
 453     DWORD swizzle = param->swizzle;
 454     DWORD swizzle_x = swizzle & 0x03;
 455     DWORD swizzle_y = (swizzle >> 2) & 0x03;
 456     DWORD swizzle_z = (swizzle >> 4) & 0x03;
 457     DWORD swizzle_w = (swizzle >> 6) & 0x03;
 458
 459     /* If the swizzle is the default swizzle (ie, "xyzw"), we don't need to
 460      * generate a swizzle string. Unless we need to our own swizzling. */
 461     if (swizzle != WINED3DSP_NOSWIZZLE || fixup)
 462     {
 463         *ptr++ = '.';
 464         if (swizzle_x == swizzle_y && swizzle_x == swizzle_z && swizzle_x == swizzle_w) {
 465             *ptr++ = swizzle_chars[swizzle_x];
 466         } else {
 467             *ptr++ = swizzle_chars[swizzle_x];
 468             *ptr++ = swizzle_chars[swizzle_y];
 469             *ptr++ = swizzle_chars[swizzle_z];
 470             *ptr++ = swizzle_chars[swizzle_w];
 471         }
 472     }
 473
 474     *ptr = '\0';
 475 }
 476
 477 static void shader_arb_get_register_name(IWineD3DBaseShader *iface, WINED3DSHADER_PARAM_REGISTER_TYPE register_type,
 478         UINT register_idx, BOOL rel_addr, char *register_name, BOOL *is_color)
 479 {
 480     /* oPos, oFog and oPts in D3D */
 481     static const char * const rastout_reg_names[] = {"TMP_OUT", "result.fogcoord", "result.pointsize"};
 482     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
 483     DWORD shader_version = This->baseShader.reg_maps.shader_version;
 484     BOOL pshader = shader_is_pshader_version(shader_version);
 485
 486     *is_color = FALSE;
 487
 488     switch (register_type)
 489     {
 490         case WINED3DSPR_TEMP:
 491             sprintf(register_name, "R%u", register_idx);
 492             break;
 493
 494         case WINED3DSPR_INPUT:
 495             if (pshader)
 496             {
 497                 if (register_idx == 0) strcpy(register_name, "fragment.color.primary");
 498                 else strcpy(register_name, "fragment.color.secondary");
 499             }
 500             else
 501             {
 502                 if (((IWineD3DVertexShaderImpl *)This)->cur_args->swizzle_map & (1 << register_idx)) *is_color = TRUE;
 503                 sprintf(register_name, "vertex.attrib[%u]", register_idx);
 504             }
 505             break;
 506
 507         case WINED3DSPR_CONST:
 508             if (!pshader && rel_addr)
 509             {
 510                 UINT rel_offset = ((IWineD3DVertexShaderImpl *)This)->rel_offset;
 511                 if (register_idx >= rel_offset)
 512                     sprintf(register_name, "C[A0.x + %u]", register_idx - rel_offset);
 513                 else
 514                     sprintf(register_name, "C[A0.x - %u]", -register_idx + rel_offset);
 515             }
 516             else
 517             {
 518                 if (This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF))
 519                     sprintf(register_name, "C[%u]", register_idx);
 520                 else
 521                     sprintf(register_name, "C%u", register_idx);
 522             }
 523             break;
 524
 525         case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
 526             if (pshader) sprintf(register_name, "T%u", register_idx);
 527             else  sprintf(register_name, "A%u", register_idx);
 528             break;
 529
 530         case WINED3DSPR_COLOROUT:
 531             if (register_idx == 0)
 532             {
 533                 strcpy(register_name, "TMP_COLOR");
 534             }
 535             else
 536             {
 537                 /* TODO: See GL_ARB_draw_buffers */
 538                 FIXME("Unsupported write to render target %u\n", register_idx);
 539                 sprintf(register_name, "unsupported_register");
 540             }
 541             break;
 542
 543         case WINED3DSPR_RASTOUT:
 544             sprintf(register_name, "%s", rastout_reg_names[register_idx]);
 545             break;
 546
 547         case WINED3DSPR_DEPTHOUT:
 548             strcpy(register_name, "result.depth");
 549             break;
 550
 551         case WINED3DSPR_ATTROUT:
 552             if (pshader) sprintf(register_name, "oD[%u]", register_idx);
 553             else if (register_idx == 0) strcpy(register_name, "result.color.primary");
 554             else strcpy(register_name, "result.color.secondary");
 555             break;
 556
 557         case WINED3DSPR_TEXCRDOUT:
 558             if (pshader) sprintf(register_name, "oT[%u]", register_idx);
 559             else sprintf(register_name, "result.texcoord[%u]", register_idx);
 560             break;
 561
 562         default:
 563             FIXME("Unhandled register type %#x[%u]\n", register_type, register_idx);
 564             sprintf(register_name, "unrecognized_register[%u]", register_idx);
 565             break;
 566     }
 567 }
 568
 569 static void shader_arb_add_src_param(const struct wined3d_shader_instruction *ins,
 570         const struct wined3d_shader_src_param *wined3d_src, char *str)
 571 {
 572     char register_name[255];
 573     char swizzle[6];
 574     BOOL is_color;
 575
 576     if (wined3d_src->modifiers == WINED3DSPSM_NEG) strcat(str, " -");
 577     else strcat(str, " ");
 578
 579     shader_arb_get_register_name(ins->ctx->shader, wined3d_src->register_type,
 580             wined3d_src->register_idx, !!wined3d_src->rel_addr, register_name, &is_color);
 581     strcat(str, register_name);
 582
 583     shader_arb_get_swizzle(wined3d_src, is_color, swizzle);
 584     strcat(str, swizzle);
 585 }
 586
 587 static void shader_arb_add_dst_param(const struct wined3d_shader_instruction *ins,
 588         const struct wined3d_shader_dst_param *wined3d_dst, char *str)
 589 {
 590     char register_name[255];
 591     char write_mask[6];
 592     BOOL is_color;
 593
 594     strcat(str, " ");
 595
 596     shader_arb_get_register_name(ins->ctx->shader, wined3d_dst->register_type,
 597             wined3d_dst->register_idx, !!wined3d_dst->rel_addr, register_name, &is_color);
 598     strcat(str, register_name);
 599
 600     shader_arb_get_write_mask(ins, wined3d_dst, write_mask);
 601     strcat(str, write_mask);
 602 }
 603
 604 static const char *shader_arb_get_fixup_swizzle(enum fixup_channel_source channel_source)
 605 {
 606     switch(channel_source)
 607     {
 608         case CHANNEL_SOURCE_ZERO: return "0";
 609         case CHANNEL_SOURCE_ONE: return "1";
 610         case CHANNEL_SOURCE_X: return "x";
 611         case CHANNEL_SOURCE_Y: return "y";
 612         case CHANNEL_SOURCE_Z: return "z";
 613         case CHANNEL_SOURCE_W: return "w";
 614         default:
 615             FIXME("Unhandled channel source %#x\n", channel_source);
 616             return "undefined";
 617     }
 618 }
 619
 620 static void gen_color_correction(SHADER_BUFFER *buffer, const char *reg, DWORD dst_mask,
 621                                  const char *one, const char *two, struct color_fixup_desc fixup)
 622 {
 623     DWORD mask;
 624
 625     if (is_yuv_fixup(fixup))
 626     {
 627         enum yuv_fixup yuv_fixup = get_yuv_fixup(fixup);
 628         FIXME("YUV fixup (%#x) not supported\n", yuv_fixup);
 629         return;
 630     }
 631
 632     mask = 0;
 633     if (fixup.x_source != CHANNEL_SOURCE_X) mask |= WINED3DSP_WRITEMASK_0;
 634     if (fixup.y_source != CHANNEL_SOURCE_Y) mask |= WINED3DSP_WRITEMASK_1;
 635     if (fixup.z_source != CHANNEL_SOURCE_Z) mask |= WINED3DSP_WRITEMASK_2;
 636     if (fixup.w_source != CHANNEL_SOURCE_W) mask |= WINED3DSP_WRITEMASK_3;
 637     mask &= dst_mask;
 638
 639     if (mask)
 640     {
 641         shader_addline(buffer, "SWZ %s, %s, %s, %s, %s, %s;\n", reg, reg,
 642                 shader_arb_get_fixup_swizzle(fixup.x_source), shader_arb_get_fixup_swizzle(fixup.y_source),
 643                 shader_arb_get_fixup_swizzle(fixup.z_source), shader_arb_get_fixup_swizzle(fixup.w_source));
 644     }
 645
 646     mask = 0;
 647     if (fixup.x_sign_fixup) mask |= WINED3DSP_WRITEMASK_0;
 648     if (fixup.y_sign_fixup) mask |= WINED3DSP_WRITEMASK_1;
 649     if (fixup.z_sign_fixup) mask |= WINED3DSP_WRITEMASK_2;
 650     if (fixup.w_sign_fixup) mask |= WINED3DSP_WRITEMASK_3;
 651     mask &= dst_mask;
 652
 653     if (mask)
 654     {
 655         char reg_mask[6];
 656         char *ptr = reg_mask;
 657
 658         if (mask != WINED3DSP_WRITEMASK_ALL)
 659         {
 660             *ptr++ = '.';
 661             if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
 662             if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
 663             if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
 664             if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
 665         }
 666         *ptr = '\0';
 667
 668         shader_addline(buffer, "MAD %s%s, %s, %s, -%s;\n", reg, reg_mask, reg, two, one);
 669     }
 670 }
 671
 672 static void shader_hw_sample(const struct wined3d_shader_instruction *ins, DWORD sampler_idx,
 673         const char *dst_str, const char *coord_reg, BOOL projected, BOOL bias)
 674 {
 675     SHADER_BUFFER *buffer = ins->ctx->buffer;
 676     DWORD sampler_type = ins->ctx->reg_maps->sampler_type[sampler_idx];
 677     const char *tex_type;
 678     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)ins->ctx->shader;
 679     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
 680
 681     switch(sampler_type) {
 682         case WINED3DSTT_1D:
 683             tex_type = "1D";
 684             break;
 685
 686         case WINED3DSTT_2D:
 687             if(device->stateBlock->textures[sampler_idx] &&
 688                IWineD3DBaseTexture_GetTextureDimensions(device->stateBlock->textures[sampler_idx]) == GL_TEXTURE_RECTANGLE_ARB) {
 689                 tex_type = "RECT";
 690             } else {
 691                 tex_type = "2D";
 692             }
 693             if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version))
 694             {
 695                const IWineD3DPixelShaderImpl* const ps = (const IWineD3DPixelShaderImpl*)This;
 696                if(ps->cur_args->np2_fixup & (1 << sampler_idx)) {
 697                    FIXME("NP2 texcoord fixup is currently not implemented in ARB mode (use GLSL instead).\n");
 698                }
 699             }
 700             break;
 701
 702         case WINED3DSTT_VOLUME:
 703             tex_type = "3D";
 704             break;
 705
 706         case WINED3DSTT_CUBE:
 707             tex_type = "CUBE";
 708             break;
 709
 710         default:
 711             ERR("Unexpected texture type %d\n", sampler_type);
 712             tex_type = "";
 713     }
 714
 715     if (bias) {
 716         /* Shouldn't be possible, but let's check for it */
 717         if(projected) FIXME("Biased and Projected texture sampling\n");
 718         /* TXB takes the 4th component of the source vector automatically, as d3d. Nothing more to do */
 719         shader_addline(buffer, "TXB %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 720     } else if (projected) {
 721         shader_addline(buffer, "TXP %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 722     } else {
 723         shader_addline(buffer, "TEX %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 724     }
 725
 726     if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version))
 727     {
 728         IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
 729         gen_color_correction(buffer, dst_str, ins->dst[0].write_mask,
 730                 "one", "coefmul.x", ps->cur_args->color_fixup[sampler_idx]);
 731     }
 732 }
 733
 734 static void pshader_gen_input_modifier_line(IWineD3DBaseShader *iface, SHADER_BUFFER *buffer,
 735         const struct wined3d_shader_src_param *src, unsigned int tmpreg, char *outregstr)
 736 {
 737     /* Generate a line that does the input modifier computation and return the input register to use */
 738     BOOL is_color = FALSE;
 739     char regstr[256];
 740     char swzstr[20];
 741     int insert_line;
 742
 743     /* Assume a new line will be added */
 744     insert_line = 1;
 745
 746     /* Get register name */
 747     shader_arb_get_register_name(iface, src->register_type,
 748             src->register_idx, !!src->rel_addr, regstr, &is_color);
 749     shader_arb_get_swizzle(src, is_color, swzstr);
 750
 751     switch (src->modifiers)
 752     {
 753     case WINED3DSPSM_NONE:
 754         sprintf(outregstr, "%s%s", regstr, swzstr);
 755         insert_line = 0;
 756         break;
 757     case WINED3DSPSM_NEG:
 758         sprintf(outregstr, "-%s%s", regstr, swzstr);
 759         insert_line = 0;
 760         break;
 761     case WINED3DSPSM_BIAS:
 762         shader_addline(buffer, "ADD T%c, %s, -coefdiv.x;\n", 'A' + tmpreg, regstr);
 763         break;
 764     case WINED3DSPSM_BIASNEG:
 765         shader_addline(buffer, "ADD T%c, -%s, coefdiv.x;\n", 'A' + tmpreg, regstr);
 766         break;
 767     case WINED3DSPSM_SIGN:
 768         shader_addline(buffer, "MAD T%c, %s, coefmul.x, -one.x;\n", 'A' + tmpreg, regstr);
 769         break;
 770     case WINED3DSPSM_SIGNNEG:
 771         shader_addline(buffer, "MAD T%c, %s, -coefmul.x, one.x;\n", 'A' + tmpreg, regstr);
 772         break;
 773     case WINED3DSPSM_COMP:
 774         shader_addline(buffer, "SUB T%c, one.x, %s;\n", 'A' + tmpreg, regstr);
 775         break;
 776     case WINED3DSPSM_X2:
 777         shader_addline(buffer, "ADD T%c, %s, %s;\n", 'A' + tmpreg, regstr, regstr);
 778         break;
 779     case WINED3DSPSM_X2NEG:
 780         shader_addline(buffer, "ADD T%c, -%s, -%s;\n", 'A' + tmpreg, regstr, regstr);
 781         break;
 782     case WINED3DSPSM_DZ:
 783         shader_addline(buffer, "RCP T%c, %s.z;\n", 'A' + tmpreg, regstr);
 784         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 785         break;
 786     case WINED3DSPSM_DW:
 787         shader_addline(buffer, "RCP T%c, %s.w;\n", 'A' + tmpreg, regstr);
 788         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 789         break;
 790     default:
 791         sprintf(outregstr, "%s%s", regstr, swzstr);
 792         insert_line = 0;
 793     }
 794
 795     /* Return modified or original register, with swizzle */
 796     if (insert_line)
 797         sprintf(outregstr, "T%c%s", 'A' + tmpreg, swzstr);
 798 }
 799
 800 static inline void pshader_gen_output_modifier_line(SHADER_BUFFER *buffer, int saturate, const char *write_mask,
 801         int shift, const char *regstr)
 802 {
 803     /* Generate a line that does the output modifier computation */
 804     shader_addline(buffer, "MUL%s %s%s, %s, %s;\n", saturate ? "_SAT" : "",
 805         regstr, write_mask, regstr, shift_tab[shift]);
 806 }
 807
 808 static void pshader_hw_bem(const struct wined3d_shader_instruction *ins)
 809 {
 810     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
 811     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 812     SHADER_BUFFER *buffer = ins->ctx->buffer;
 813     char dst_name[50];
 814     char src_name[2][50];
 815     char dst_wmask[20];
 816     DWORD sampler_code = dst->register_idx;
 817     BOOL has_bumpmat = FALSE;
 818     BOOL is_color;
 819     int i;
 820
 821     for(i = 0; i < This->numbumpenvmatconsts; i++) {
 822         if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
 823                 && This->bumpenvmatconst[i].texunit == sampler_code)
 824         {
 825             has_bumpmat = TRUE;
 826             break;
 827         }
 828     }
 829
 830     shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
 831             dst->register_idx, !!dst->rel_addr, dst_name, &is_color);
 832     shader_arb_get_write_mask(ins, dst, dst_wmask);
 833     strcat(dst_name, dst_wmask);
 834
 835     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_name[0]);
 836     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[1], 1, src_name[1]);
 837
 838     if(has_bumpmat) {
 839         /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
 840         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", sampler_code);
 841         shader_addline(buffer, "DP3 TMP.r, TMP2, %s;\n", src_name[1]);
 842         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", sampler_code);
 843         shader_addline(buffer, "DP3 TMP.g, TMP2, %s;\n", src_name[1]);
 844
 845         shader_addline(buffer, "ADD %s, %s, TMP;\n", dst_name, src_name[0]);
 846     } else {
 847         shader_addline(buffer, "MOV %s, %s;\n", dst_name, src_name[0]);
 848     }
 849 }
 850
 851 static void pshader_hw_cnd(const struct wined3d_shader_instruction *ins)
 852 {
 853     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 854     SHADER_BUFFER *buffer = ins->ctx->buffer;
 855     char dst_wmask[20];
 856     char dst_name[50];
 857     char src_name[3][50];
 858     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
 859     DWORD shift = dst->shift;
 860     BOOL is_color;
 861
 862     /* FIXME: support output modifiers */
 863
 864     /* Handle output register */
 865     shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
 866             dst->register_idx, !!dst->rel_addr, dst_name, &is_color);
 867     shader_arb_get_write_mask(ins, dst, dst_wmask);
 868
 869     /* Generate input register names (with modifiers) */
 870     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_name[0]);
 871     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[1], 1, src_name[1]);
 872     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[2], 2, src_name[2]);
 873
 874     /* The coissue flag changes the semantic of the cnd instruction in <= 1.3 shaders */
 875     if (ins->ctx->reg_maps->shader_version <= WINED3DPS_VERSION(1, 3) && ins->coissue)
 876     {
 877         shader_addline(buffer, "MOV%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1]);
 878     } else {
 879         shader_addline(buffer, "ADD TMP, -%s, coefdiv.x;\n", src_name[0]);
 880         shader_addline(buffer, "CMP%s %s%s, TMP, %s, %s;\n",
 881                                 sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1], src_name[2]);
 882     }
 883     if (shift != 0)
 884         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
 885 }
 886
 887 static void pshader_hw_cmp(const struct wined3d_shader_instruction *ins)
 888 {
 889     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 890     SHADER_BUFFER *buffer = ins->ctx->buffer;
 891     char dst_wmask[20];
 892     char dst_name[50];
 893     char src_name[3][50];
 894     DWORD shift = dst->shift;
 895     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
 896     BOOL is_color;
 897
 898     /* FIXME: support output modifiers */
 899
 900     /* Handle output register */
 901     shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
 902             dst->register_idx, !!dst->rel_addr, dst_name, &is_color);
 903     shader_arb_get_write_mask(ins, dst, dst_wmask);
 904
 905     /* Generate input register names (with modifiers) */
 906     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_name[0]);
 907     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[1], 1, src_name[1]);
 908     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[2], 2, src_name[2]);
 909
 910     shader_addline(buffer, "CMP%s %s%s, %s, %s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
 911                    src_name[0], src_name[2], src_name[1]);
 912
 913     if (shift != 0)
 914         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
 915 }
 916
 917 /** Process the WINED3DSIO_DP2ADD instruction in ARB.
 918  * dst = dot2(src0, src1) + src2 */
 919 static void pshader_hw_dp2add(const struct wined3d_shader_instruction *ins)
 920 {
 921     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 922     SHADER_BUFFER *buffer = ins->ctx->buffer;
 923     char dst_wmask[20];
 924     char dst_name[50];
 925     char src_name[3][50];
 926     DWORD shift = dst->shift;
 927     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
 928     BOOL is_color;
 929
 930     shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
 931             dst->register_idx, !!dst->rel_addr, dst_name, &is_color);
 932     shader_arb_get_write_mask(ins, dst, dst_wmask);
 933
 934     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_name[0]);
 935     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[1], 1, src_name[1]);
 936     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[2], 2, src_name[2]);
 937
 938     /* Emulate a DP2 with a DP3 and 0.0 */
 939     shader_addline(buffer, "MOV TMP, %s;\n", src_name[0]);
 940     shader_addline(buffer, "MOV TMP.z, 0.0;\n");
 941     shader_addline(buffer, "DP3 TMP2, TMP, %s;\n", src_name[1]);
 942     shader_addline(buffer, "ADD%s %s%s, TMP2, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[2]);
 943
 944     if (shift != 0)
 945         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
 946 }
 947
 948 /* Map the opcode 1-to-1 to the GL code */
 949 static void shader_hw_map2gl(const struct wined3d_shader_instruction *ins)
 950 {
 951     SHADER_BUFFER *buffer = ins->ctx->buffer;
 952     const char *instruction;
 953     char arguments[256];
 954     unsigned int i;
 955
 956     switch (ins->handler_idx)
 957     {
 958         case WINED3DSIH_ABS: instruction = "ABS"; break;
 959         case WINED3DSIH_ADD: instruction = "ADD"; break;
 960         case WINED3DSIH_CRS: instruction = "XPD"; break;
 961         case WINED3DSIH_DP3: instruction = "DP3"; break;
 962         case WINED3DSIH_DP4: instruction = "DP4"; break;
 963         case WINED3DSIH_DST: instruction = "DST"; break;
 964         case WINED3DSIH_EXP: instruction = "EX2"; break;
 965         case WINED3DSIH_EXPP: instruction = "EXP"; break;
 966         case WINED3DSIH_FRC: instruction = "FRC"; break;
 967         case WINED3DSIH_LIT: instruction = "LIT"; break;
 968         case WINED3DSIH_LOG: instruction = "LG2"; break;
 969         case WINED3DSIH_LOGP: instruction = "LOG"; break;
 970         case WINED3DSIH_LRP: instruction = "LRP"; break;
 971         case WINED3DSIH_MAD: instruction = "MAD"; break;
 972         case WINED3DSIH_MAX: instruction = "MAX"; break;
 973         case WINED3DSIH_MIN: instruction = "MIN"; break;
 974         case WINED3DSIH_MOV: instruction = "MOV"; break;
 975         case WINED3DSIH_MUL: instruction = "MUL"; break;
 976         case WINED3DSIH_NOP: instruction = "NOP"; break;
 977         case WINED3DSIH_POW: instruction = "POW"; break;
 978         case WINED3DSIH_SGE: instruction = "SGE"; break;
 979         case WINED3DSIH_SLT: instruction = "SLT"; break;
 980         case WINED3DSIH_SUB: instruction = "SUB"; break;
 981         default: instruction = "";
 982             FIXME("Unhandled opcode %#x\n", ins->handler_idx);
 983             break;
 984     }
 985
 986     if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version))
 987     {
 988         /* Output token related */
 989         const struct wined3d_shader_dst_param *dst;
 990         char output_rname[256];
 991         char output_wmask[20];
 992         char operands[4][100];
 993         BOOL saturate = FALSE;
 994         BOOL centroid = FALSE;
 995         BOOL partialprecision = FALSE;
 996         const char *modifier;
 997         BOOL is_color;
 998         DWORD shift;
 999
1000         if (!(ins->dst_count + ins->src_count))
1001         {
1002             ERR("Opcode \"%#x\" has no parameters\n", ins->handler_idx);
1003             return;
1004         }
1005         dst = &ins->dst[0];
1006
1007         /* Process modifiers */
1008         if (dst->modifiers)
1009         {
1010             DWORD mask = dst->modifiers;
1011
1012             saturate = mask & WINED3DSPDM_SATURATE;
1013             centroid = mask & WINED3DSPDM_MSAMPCENTROID;
1014             partialprecision = mask & WINED3DSPDM_PARTIALPRECISION;
1015             mask &= ~(WINED3DSPDM_MSAMPCENTROID | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_SATURATE);
1016             if (mask)
1017                 FIXME("Unrecognized modifier(%#x)\n", mask);
1018
1019             if (centroid)
1020                 FIXME("Unhandled modifier(%#x)\n", mask);
1021         }
1022         shift = dst->shift;
1023         modifier = (saturate && !shift) ? "_SAT" : "";
1024
1025         /* Generate input register names (with modifiers) */
1026         for (i = 0; i < ins->src_count; ++i)
1027         {
1028             pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[i], i, operands[i + 1]);
1029         }
1030
1031         /* Handle output register */
1032         shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1033                 dst->register_idx, !!dst->rel_addr, output_rname, &is_color);
1034         strcpy(operands[0], output_rname);
1035         shader_arb_get_write_mask(ins, dst, output_wmask);
1036         strcat(operands[0], output_wmask);
1037
1038         arguments[0] = '\0';
1039         strcat(arguments, operands[0]);
1040         for (i = 0; i < ins->src_count; ++i)
1041         {
1042             strcat(arguments, ", ");
1043             strcat(arguments, operands[i + 1]);
1044         }
1045         shader_addline(buffer, "%s%s %s;\n", instruction, modifier, arguments);
1046
1047         /* A shift requires another line. */
1048         if (shift) pshader_gen_output_modifier_line(buffer, saturate, output_wmask, shift, output_rname);
1049     } else {
1050         /* Note that shader_arb_add_*_param() adds spaces. */
1051
1052         arguments[0] = '\0';
1053         if (ins->dst_count)
1054         {
1055             shader_arb_add_dst_param(ins, &ins->dst[0], arguments);
1056             for (i = 0; i < ins->src_count; ++i)
1057             {
1058                 strcat(arguments, ",");
1059                 shader_arb_add_src_param(ins, &ins->src[i], arguments);
1060             }
1061         }
1062         shader_addline(buffer, "%s%s;\n", instruction, arguments);
1063     }
1064 }
1065
1066 static void shader_hw_mov(const struct wined3d_shader_instruction *ins)
1067 {
1068     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)ins->ctx->shader;
1069
1070     if ((WINED3DSHADER_VERSION_MAJOR(ins->ctx->reg_maps->shader_version) == 1
1071             && !shader_is_pshader_version(ins->ctx->reg_maps->shader_version)
1072             && ins->dst[0].register_type == WINED3DSPR_ADDR)
1073             || ins->handler_idx == WINED3DSIH_MOVA)
1074     {
1075         SHADER_BUFFER *buffer = ins->ctx->buffer;
1076         char src0_param[256];
1077
1078         if (ins->handler_idx == WINED3DSIH_MOVA)
1079             FIXME("mova should round\n");
1080
1081         src0_param[0] = '\0';
1082         if (((IWineD3DVertexShaderImpl *)shader)->rel_offset)
1083         {
1084             shader_arb_add_src_param(ins, &ins->src[0], src0_param);
1085             shader_addline(buffer, "ADD TMP.x, %s, helper_const.z;\n", src0_param);
1086             shader_addline(buffer, "ARL A0.x, TMP.x;\n");
1087         }
1088         else
1089         {
1090             /* Apple's ARB_vertex_program implementation does not accept an ARL source argument
1091              * with more than one component. Thus replicate the first source argument over all
1092              * 4 components. For example, .xyzw -> .x (or better: .xxxx), .zwxy -> .z, etc) */
1093             struct wined3d_shader_src_param tmp_src = ins->src[0];
1094             tmp_src.swizzle = (tmp_src.swizzle & 0x3) * 0x55;
1095             shader_arb_add_src_param(ins, &tmp_src, src0_param);
1096             shader_addline(buffer, "ARL A0.x, %s;\n", src0_param);
1097         }
1098     }
1099     else
1100     {
1101         shader_hw_map2gl(ins);
1102     }
1103 }
1104
1105 static void pshader_hw_texkill(const struct wined3d_shader_instruction *ins)
1106 {
1107     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1108     DWORD shader_version = ins->ctx->reg_maps->shader_version;
1109     SHADER_BUFFER *buffer = ins->ctx->buffer;
1110     char reg_dest[40];
1111     BOOL is_color;
1112
1113     /* No swizzles are allowed in d3d's texkill. PS 1.x ignores the 4th component as documented,
1114      * but >= 2.0 honors it(undocumented, but tested by the d3d9 testsuit)
1115      */
1116     shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1117             dst->register_idx, !!dst->rel_addr, reg_dest, &is_color);
1118
1119     if (shader_version >= WINED3DPS_VERSION(2,0))
1120     {
1121         /* The arb backend doesn't claim ps 2.0 support, but try to eat what the app feeds to us */
1122         shader_addline(buffer, "KIL %s;\n", reg_dest);
1123     } else {
1124         /* ARB fp doesn't like swizzles on the parameter of the KIL instruction. To mask the 4th component,
1125          * copy the register into our general purpose TMP variable, overwrite .w and pass TMP to KIL
1126          */
1127         shader_addline(buffer, "MOV TMP, %s;\n", reg_dest);
1128         shader_addline(buffer, "MOV TMP.w, one.w;\n");
1129         shader_addline(buffer, "KIL TMP;\n");
1130     }
1131 }
1132
1133 static void pshader_hw_tex(const struct wined3d_shader_instruction *ins)
1134 {
1135     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1136     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1137     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1138     BOOL is_color;
1139     SHADER_BUFFER *buffer = ins->ctx->buffer;
1140     DWORD shader_version = ins->ctx->reg_maps->shader_version;
1141     BOOL projected = FALSE, bias = FALSE;
1142
1143     char reg_dest[40];
1144     char reg_coord[40];
1145     DWORD reg_sampler_code;
1146
1147     /* All versions have a destination register */
1148     shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1149             dst->register_idx, !!dst->rel_addr, reg_dest, &is_color);
1150
1151     /* 1.0-1.3: Use destination register as coordinate source.
1152        1.4+: Use provided coordinate source register. */
1153    if (shader_version < WINED3DPS_VERSION(1,4))
1154       strcpy(reg_coord, reg_dest);
1155    else
1156       pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, reg_coord);
1157
1158   /* 1.0-1.4: Use destination register number as texture code.
1159      2.0+: Use provided sampler number as texure code. */
1160   if (shader_version < WINED3DPS_VERSION(2,0))
1161      reg_sampler_code = dst->register_idx;
1162   else
1163      reg_sampler_code = ins->src[1].register_idx;
1164
1165   /* projection flag:
1166    * 1.1, 1.2, 1.3: Use WINED3DTSS_TEXTURETRANSFORMFLAGS
1167    * 1.4: Use WINED3DSPSM_DZ or WINED3DSPSM_DW on src[0]
1168    * 2.0+: Use WINED3DSI_TEXLD_PROJECT on the opcode
1169    */
1170   if (shader_version < WINED3DPS_VERSION(1,4))
1171   {
1172       DWORD flags = 0;
1173       if(reg_sampler_code < MAX_TEXTURES) {
1174         flags = deviceImpl->stateBlock->textureState[reg_sampler_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1175       }
1176       if (flags & WINED3DTTFF_PROJECTED) {
1177           projected = TRUE;
1178       }
1179   }
1180   else if (shader_version < WINED3DPS_VERSION(2,0))
1181   {
1182       DWORD src_mod = ins->src[0].modifiers;
1183       if (src_mod == WINED3DSPSM_DZ) {
1184           projected = TRUE;
1185       } else if(src_mod == WINED3DSPSM_DW) {
1186           projected = TRUE;
1187       }
1188   } else {
1189       if (ins->flags & WINED3DSI_TEXLD_PROJECT) projected = TRUE;
1190       if (ins->flags & WINED3DSI_TEXLD_BIAS) bias = TRUE;
1191   }
1192   shader_hw_sample(ins, reg_sampler_code, reg_dest, reg_coord, projected, bias);
1193 }
1194
1195 static void pshader_hw_texcoord(const struct wined3d_shader_instruction *ins)
1196 {
1197     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1198     SHADER_BUFFER *buffer = ins->ctx->buffer;
1199
1200     char tmp[20];
1201     shader_arb_get_write_mask(ins, dst, tmp);
1202     if (ins->ctx->reg_maps->shader_version != WINED3DPS_VERSION(1,4))
1203     {
1204         DWORD reg = dst->register_idx;
1205         shader_addline(buffer, "MOV_SAT T%u%s, fragment.texcoord[%u];\n", reg, tmp, reg);
1206     } else {
1207         char reg_src[40];
1208
1209         pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, reg_src);
1210         shader_addline(buffer, "MOV R%u%s, %s;\n", dst->register_idx, tmp, reg_src);
1211    }
1212 }
1213
1214 static void pshader_hw_texreg2ar(const struct wined3d_shader_instruction *ins)
1215 {
1216      SHADER_BUFFER *buffer = ins->ctx->buffer;
1217      IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1218      IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1219      DWORD flags;
1220
1221      DWORD reg1 = ins->dst[0].register_idx;
1222      char dst_str[8];
1223      char src_str[50];
1224
1225      sprintf(dst_str, "T%u", reg1);
1226      pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_str);
1227      shader_addline(buffer, "MOV TMP.x, %s.w;\n", src_str);
1228      shader_addline(buffer, "MOV TMP.y, %s.x;\n", src_str);
1229      flags = reg1 < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg1][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1230      shader_hw_sample(ins, reg1, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1231 }
1232
1233 static void pshader_hw_texreg2gb(const struct wined3d_shader_instruction *ins)
1234 {
1235      SHADER_BUFFER *buffer = ins->ctx->buffer;
1236
1237      DWORD reg1 = ins->dst[0].register_idx;
1238      char dst_str[8];
1239      char src_str[50];
1240
1241      sprintf(dst_str, "T%u", reg1);
1242      pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_str);
1243      shader_addline(buffer, "MOV TMP.x, %s.y;\n", src_str);
1244      shader_addline(buffer, "MOV TMP.y, %s.z;\n", src_str);
1245      shader_hw_sample(ins, reg1, dst_str, "TMP", FALSE, FALSE);
1246 }
1247
1248 static void pshader_hw_texreg2rgb(const struct wined3d_shader_instruction *ins)
1249 {
1250     SHADER_BUFFER *buffer = ins->ctx->buffer;
1251     DWORD reg1 = ins->dst[0].register_idx;
1252     char dst_str[8];
1253     char src_str[50];
1254
1255     sprintf(dst_str, "T%u", reg1);
1256     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_str);
1257     shader_hw_sample(ins, reg1, dst_str, src_str, FALSE, FALSE);
1258 }
1259
1260 static void pshader_hw_texbem(const struct wined3d_shader_instruction *ins)
1261 {
1262     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1263     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1264     BOOL has_bumpmat = FALSE;
1265     BOOL has_luminance = FALSE;
1266     BOOL is_color;
1267     int i;
1268
1269     SHADER_BUFFER *buffer = ins->ctx->buffer;
1270
1271     char reg_coord[40];
1272     DWORD reg_dest_code;
1273
1274     /* All versions have a destination register */
1275     reg_dest_code = dst->register_idx;
1276     /* Can directly use the name because texbem is only valid for <= 1.3 shaders */
1277     shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1278             dst->register_idx, !!dst->rel_addr, reg_coord, &is_color);
1279
1280     for(i = 0; i < This->numbumpenvmatconsts; i++) {
1281         if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1282                 && reg_dest_code == This->bumpenvmatconst[i].texunit)
1283         {
1284             has_bumpmat = TRUE;
1285             break;
1286         }
1287     }
1288     for(i = 0; i < This->numbumpenvmatconsts; i++) {
1289         if (This->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1290                 && reg_dest_code == This->luminanceconst[i].texunit)
1291         {
1292             has_luminance = TRUE;
1293             break;
1294         }
1295     }
1296
1297     if(has_bumpmat) {
1298         DWORD src = ins->src[0].register_idx;
1299
1300         /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
1301
1302         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", reg_dest_code);
1303         shader_addline(buffer, "DP3 TMP.x, TMP2, T%u;\n", src);
1304         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", reg_dest_code);
1305         shader_addline(buffer, "DP3 TMP.y, TMP2, T%u;\n", src);
1306
1307         /* with projective textures, texbem only divides the static texture coord, not the displacement,
1308          * so we can't let the GL handle this.
1309          */
1310         if (((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS]
1311               & WINED3DTTFF_PROJECTED) {
1312             shader_addline(buffer, "RCP TMP2.w, %s.w;\n", reg_coord);
1313             shader_addline(buffer, "MUL TMP2.xy, %s, TMP2.w;\n", reg_coord);
1314             shader_addline(buffer, "ADD TMP.xy, TMP, TMP2;\n");
1315         } else {
1316             shader_addline(buffer, "ADD TMP.xy, TMP, %s;\n", reg_coord);
1317         }
1318
1319         shader_hw_sample(ins, reg_dest_code, reg_coord, "TMP", FALSE, FALSE);
1320
1321         if (ins->handler_idx == WINED3DSIH_TEXBEML && has_luminance)
1322         {
1323             shader_addline(buffer, "MAD TMP, T%u.z, luminance%d.x, luminance%d.y;\n",
1324                            src, reg_dest_code, reg_dest_code);
1325             shader_addline(buffer, "MUL %s, %s, TMP;\n", reg_coord, reg_coord);
1326         }
1327
1328     } else {
1329         DWORD tf;
1330         if(reg_dest_code < MAX_TEXTURES) {
1331             tf = ((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1332         } else {
1333             tf = 0;
1334         }
1335         /* Without a bump matrix loaded, just sample with the unmodified coordinates */
1336         shader_hw_sample(ins, reg_dest_code, reg_coord, reg_coord, tf & WINED3DTTFF_PROJECTED, FALSE);
1337     }
1338 }
1339
1340 static void pshader_hw_texm3x2pad(const struct wined3d_shader_instruction *ins)
1341 {
1342     DWORD reg = ins->dst[0].register_idx;
1343     SHADER_BUFFER *buffer = ins->ctx->buffer;
1344     char src0_name[50];
1345
1346     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0_name);
1347     shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", reg, src0_name);
1348 }
1349
1350 static void pshader_hw_texm3x2tex(const struct wined3d_shader_instruction *ins)
1351 {
1352     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1353     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1354     DWORD flags;
1355     DWORD reg = ins->dst[0].register_idx;
1356     SHADER_BUFFER *buffer = ins->ctx->buffer;
1357     char dst_str[8];
1358     char src0_name[50];
1359
1360     sprintf(dst_str, "T%u", reg);
1361     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0_name);
1362     shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", reg, src0_name);
1363     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1364     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1365 }
1366
1367 static void pshader_hw_texm3x3pad(const struct wined3d_shader_instruction *ins)
1368 {
1369     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1370     DWORD reg = ins->dst[0].register_idx;
1371     SHADER_BUFFER *buffer = ins->ctx->buffer;
1372     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1373     char src0_name[50];
1374
1375     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0_name);
1376     shader_addline(buffer, "DP3 TMP.%c, T%u, %s;\n", 'x' + current_state->current_row, reg, src0_name);
1377     current_state->texcoord_w[current_state->current_row++] = reg;
1378 }
1379
1380 static void pshader_hw_texm3x3tex(const struct wined3d_shader_instruction *ins)
1381 {
1382     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1383     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1384     DWORD flags;
1385     DWORD reg = ins->dst[0].register_idx;
1386     SHADER_BUFFER *buffer = ins->ctx->buffer;
1387     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1388     char dst_str[8];
1389     char src0_name[50];
1390
1391     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0_name);
1392     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1393
1394     /* Sample the texture using the calculated coordinates */
1395     sprintf(dst_str, "T%u", reg);
1396     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1397     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1398     current_state->current_row = 0;
1399 }
1400
1401 static void pshader_hw_texm3x3vspec(const struct wined3d_shader_instruction *ins)
1402 {
1403     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1404     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1405     DWORD flags;
1406     DWORD reg = ins->dst[0].register_idx;
1407     SHADER_BUFFER *buffer = ins->ctx->buffer;
1408     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1409     char dst_str[8];
1410     char src0_name[50];
1411
1412     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0_name);
1413     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1414
1415     /* Construct the eye-ray vector from w coordinates */
1416     shader_addline(buffer, "MOV TMP2.x, fragment.texcoord[%u].w;\n", current_state->texcoord_w[0]);
1417     shader_addline(buffer, "MOV TMP2.y, fragment.texcoord[%u].w;\n", current_state->texcoord_w[1]);
1418     shader_addline(buffer, "MOV TMP2.z, fragment.texcoord[%u].w;\n", reg);
1419
1420     /* Calculate reflection vector
1421      */
1422     shader_addline(buffer, "DP3 TMP.w, TMP, TMP2;\n");
1423     /* The .w is ignored when sampling, so I can use TMP2.w to calculate dot(N, N) */
1424     shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1425     shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1426     shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1427     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1428     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -TMP2;\n");
1429
1430     /* Sample the texture using the calculated coordinates */
1431     sprintf(dst_str, "T%u", reg);
1432     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1433     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1434     current_state->current_row = 0;
1435 }
1436
1437 static void pshader_hw_texm3x3spec(const struct wined3d_shader_instruction *ins)
1438 {
1439     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1440     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1441     DWORD flags;
1442     DWORD reg = ins->dst[0].register_idx;
1443     DWORD reg3 = ins->src[1].register_idx;
1444     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1445     SHADER_BUFFER *buffer = ins->ctx->buffer;
1446     char dst_str[8];
1447     char src0_name[50];
1448
1449     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0_name);
1450     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1451
1452     /* Calculate reflection vector.
1453      *
1454      *               dot(N, E)
1455      * TMP.xyz = 2 * --------- * N - E
1456      *               dot(N, N)
1457      *
1458      * Which normalizes the normal vector
1459      */
1460     shader_addline(buffer, "DP3 TMP.w, TMP, C[%u];\n", reg3);
1461     shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1462     shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1463     shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1464     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1465     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -C[%u];\n", reg3);
1466
1467     /* Sample the texture using the calculated coordinates */
1468     sprintf(dst_str, "T%u", reg);
1469     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1470     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1471     current_state->current_row = 0;
1472 }
1473
1474 static void pshader_hw_texdepth(const struct wined3d_shader_instruction *ins)
1475 {
1476     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1477     SHADER_BUFFER *buffer = ins->ctx->buffer;
1478     char dst_name[50];
1479     BOOL is_color;
1480
1481     /* texdepth has an implicit destination, the fragment depth value. It's only parameter,
1482      * which is essentially an input, is the destination register because it is the first
1483      * parameter. According to the msdn, this must be register r5, but let's keep it more flexible
1484      * here
1485      */
1486     shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1487             dst->register_idx, !!dst->rel_addr, dst_name, &is_color);
1488
1489     /* According to the msdn, the source register(must be r5) is unusable after
1490      * the texdepth instruction, so we're free to modify it
1491      */
1492     shader_addline(buffer, "MIN %s.y, %s.y, one.y;\n", dst_name, dst_name);
1493
1494     /* How to deal with the special case dst_name.g == 0? if r != 0, then
1495      * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1496      * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1497      */
1498     shader_addline(buffer, "RCP %s.y, %s.y;\n", dst_name, dst_name);
1499     shader_addline(buffer, "MUL TMP.x, %s.x, %s.y;\n", dst_name, dst_name);
1500     shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1501     shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1502 }
1503
1504 /** Process the WINED3DSIO_TEXDP3TEX instruction in ARB:
1505  * Take a 3-component dot product of the TexCoord[dstreg] and src,
1506  * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1507 static void pshader_hw_texdp3tex(const struct wined3d_shader_instruction *ins)
1508 {
1509     SHADER_BUFFER *buffer = ins->ctx->buffer;
1510     DWORD sampler_idx = ins->dst[0].register_idx;
1511     char src0[50];
1512     char dst_str[8];
1513
1514     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0);
1515     shader_addline(buffer, "MOV TMP, 0.0;\n");
1516     shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", sampler_idx, src0);
1517
1518     sprintf(dst_str, "T%u", sampler_idx);
1519     shader_hw_sample(ins, sampler_idx, dst_str, "TMP", FALSE /* Only one coord, can't be projected */, FALSE);
1520 }
1521
1522 /** Process the WINED3DSIO_TEXDP3 instruction in ARB:
1523  * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1524 static void pshader_hw_texdp3(const struct wined3d_shader_instruction *ins)
1525 {
1526     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1527     char src0[50];
1528     char dst_str[50];
1529     char dst_mask[6];
1530     SHADER_BUFFER *buffer = ins->ctx->buffer;
1531     BOOL is_color;
1532
1533     /* Handle output register */
1534     shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1535             dst->register_idx, !!dst->rel_addr, dst_str, &is_color);
1536     shader_arb_get_write_mask(ins, dst, dst_mask);
1537
1538     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0);
1539     shader_addline(buffer, "DP3 %s%s, T%u, %s;\n", dst_str, dst_mask, dst->register_idx, src0);
1540
1541     /* TODO: Handle output modifiers */
1542 }
1543
1544 /** Process the WINED3DSIO_TEXM3X3 instruction in ARB
1545  * Perform the 3rd row of a 3x3 matrix multiply */
1546 static void pshader_hw_texm3x3(const struct wined3d_shader_instruction *ins)
1547 {
1548     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1549     SHADER_BUFFER *buffer = ins->ctx->buffer;
1550     char dst_str[50];
1551     char dst_mask[6];
1552     char src0[50];
1553     BOOL is_color;
1554
1555     shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1556             dst->register_idx, !!dst->rel_addr, dst_str, &is_color);
1557     shader_arb_get_write_mask(ins, dst, dst_mask);
1558
1559     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0);
1560     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", dst->register_idx, src0);
1561     shader_addline(buffer, "MOV %s%s, TMP;\n", dst_str, dst_mask);
1562
1563     /* TODO: Handle output modifiers */
1564 }
1565
1566 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in ARB:
1567  * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1568  * Calculate tmp0.y = TexCoord[dstreg] . src.xyz;  (tmp0.x has already been calculated)
1569  * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1570  */
1571 static void pshader_hw_texm3x2depth(const struct wined3d_shader_instruction *ins)
1572 {
1573     SHADER_BUFFER *buffer = ins->ctx->buffer;
1574     DWORD dst_reg = ins->dst[0].register_idx;
1575     char src0[50];
1576
1577     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0);
1578     shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", dst_reg, src0);
1579
1580     /* How to deal with the special case dst_name.g == 0? if r != 0, then
1581      * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1582      * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1583      */
1584     shader_addline(buffer, "RCP TMP.y, TMP.y;\n");
1585     shader_addline(buffer, "MUL TMP.x, TMP.x, TMP.y;\n");
1586     shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1587     shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1588 }
1589
1590 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1591     Vertex/Pixel shaders to ARB_vertex_program codes */
1592 static void shader_hw_mnxn(const struct wined3d_shader_instruction *ins)
1593 {
1594     int i;
1595     int nComponents = 0;
1596     struct wined3d_shader_dst_param tmp_dst = {0};
1597     struct wined3d_shader_src_param tmp_src[2] = {{0}};
1598     struct wined3d_shader_instruction tmp_ins;
1599
1600     memset(&tmp_ins, 0, sizeof(tmp_ins));
1601
1602     /* Set constants for the temporary argument */
1603     tmp_ins.ctx = ins->ctx;
1604     tmp_ins.dst_count = 1;
1605     tmp_ins.dst = &tmp_dst;
1606     tmp_ins.src_count = 2;
1607     tmp_ins.src = tmp_src;
1608
1609     switch(ins->handler_idx)
1610     {
1611         case WINED3DSIH_M4x4:
1612             nComponents = 4;
1613             tmp_ins.handler_idx = WINED3DSIH_DP4;
1614             break;
1615         case WINED3DSIH_M4x3:
1616             nComponents = 3;
1617             tmp_ins.handler_idx = WINED3DSIH_DP4;
1618             break;
1619         case WINED3DSIH_M3x4:
1620             nComponents = 4;
1621             tmp_ins.handler_idx = WINED3DSIH_DP3;
1622             break;
1623         case WINED3DSIH_M3x3:
1624             nComponents = 3;
1625             tmp_ins.handler_idx = WINED3DSIH_DP3;
1626             break;
1627         case WINED3DSIH_M3x2:
1628             nComponents = 2;
1629             tmp_ins.handler_idx = WINED3DSIH_DP3;
1630             break;
1631         default:
1632             FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1633             break;
1634     }
1635
1636     tmp_dst = ins->dst[0];
1637     tmp_src[0] = ins->src[0];
1638     tmp_src[1] = ins->src[1];
1639     for (i = 0; i < nComponents; i++) {
1640         tmp_dst.write_mask = WINED3DSP_WRITEMASK_0 << i;
1641         shader_hw_map2gl(&tmp_ins);
1642         ++tmp_src[1].register_idx;
1643     }
1644 }
1645
1646 static void vshader_hw_rsq_rcp(const struct wined3d_shader_instruction *ins)
1647 {
1648     SHADER_BUFFER *buffer = ins->ctx->buffer;
1649     const char *instruction;
1650
1651     char tmpLine[256];
1652
1653     switch(ins->handler_idx)
1654     {
1655         case WINED3DSIH_RSQ: instruction = "RSQ"; break;
1656         case WINED3DSIH_RCP: instruction = "RCP"; break;
1657         default: instruction = "";
1658             FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1659             break;
1660     }
1661
1662     strcpy(tmpLine, instruction);
1663     shader_arb_add_dst_param(ins, &ins->dst[0], tmpLine); /* Destination */
1664     strcat(tmpLine, ",");
1665     shader_arb_add_src_param(ins, &ins->src[0], tmpLine);
1666     if (ins->src[0].swizzle == WINED3DSP_NOSWIZZLE)
1667     {
1668         /* Dx sdk says .x is used if no swizzle is given, but our test shows that
1669          * .w is used
1670          */
1671         strcat(tmpLine, ".w");
1672     }
1673
1674     shader_addline(buffer, "%s;\n", tmpLine);
1675 }
1676
1677 static void shader_hw_nrm(const struct wined3d_shader_instruction *ins)
1678 {
1679     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1680     SHADER_BUFFER *buffer = ins->ctx->buffer;
1681     char dst_name[50];
1682     char src_name[50];
1683     char dst_wmask[20];
1684     DWORD shift = dst->shift;
1685     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
1686     BOOL is_color;
1687
1688     shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1689             dst->register_idx, !!dst->rel_addr, dst_name, &is_color);
1690     shader_arb_get_write_mask(ins, dst, dst_wmask);
1691
1692     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_name);
1693     shader_addline(buffer, "DP3 TMP, %s, %s;\n", src_name, src_name);
1694     shader_addline(buffer, "RSQ TMP, TMP.x;\n");
1695     /* dst.w = src[0].w * 1 / (src.x^2 + src.y^2 + src.z^2)^(1/2) according to msdn*/
1696     shader_addline(buffer, "MUL%s %s%s, %s, TMP;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1697                    src_name);
1698
1699     if (shift != 0)
1700         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
1701 }
1702
1703 static void shader_hw_sincos(const struct wined3d_shader_instruction *ins)
1704 {
1705     /* This instruction exists in ARB, but the d3d instruction takes two extra parameters which
1706      * must contain fixed constants. So we need a separate function to filter those constants and
1707      * can't use map2gl
1708      */
1709     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1710     SHADER_BUFFER *buffer = ins->ctx->buffer;
1711     char dst_name[50];
1712     char src_name[50];
1713     char dst_wmask[20];
1714     DWORD shift = dst->shift;
1715     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
1716     BOOL is_color;
1717
1718     shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1719             dst->register_idx, !!dst->rel_addr, dst_name, &is_color);
1720     shader_arb_get_write_mask(ins, dst, dst_wmask);
1721
1722     pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_name);
1723     shader_addline(buffer, "SCS%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1724                    src_name);
1725
1726     if (shift != 0)
1727         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
1728
1729 }
1730
1731 static GLuint create_arb_blt_vertex_program(const WineD3D_GL_Info *gl_info)
1732 {
1733     GLuint program_id = 0;
1734     const char *blt_vprogram =
1735         "!!ARBvp1.0\n"
1736         "PARAM c[1] = { { 1, 0.5 } };\n"
1737         "MOV result.position, vertex.position;\n"
1738         "MOV result.color, c[0].x;\n"
1739         "MOV result.texcoord[0], vertex.texcoord[0];\n"
1740         "END\n";
1741
1742     GL_EXTCALL(glGenProgramsARB(1, &program_id));
1743     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, program_id));
1744     GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_vprogram), blt_vprogram));
1745
1746     if (glGetError() == GL_INVALID_OPERATION) {
1747         GLint pos;
1748         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1749         FIXME("Vertex program error at position %d: %s\n", pos,
1750             debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1751     }
1752
1753     return program_id;
1754 }
1755
1756 static GLuint create_arb_blt_fragment_program(const WineD3D_GL_Info *gl_info, enum tex_types tex_type)
1757 {
1758     GLuint program_id = 0;
1759     static const char * const blt_fprograms[tex_type_count] =
1760     {
1761         /* tex_1d */
1762         NULL,
1763         /* tex_2d */
1764         "!!ARBfp1.0\n"
1765         "TEMP R0;\n"
1766         "TEX R0.x, fragment.texcoord[0], texture[0], 2D;\n"
1767         "MOV result.depth.z, R0.x;\n"
1768         "END\n",
1769         /* tex_3d */
1770         NULL,
1771         /* tex_cube */
1772         "!!ARBfp1.0\n"
1773         "TEMP R0;\n"
1774         "TEX R0.x, fragment.texcoord[0], texture[0], CUBE;\n"
1775         "MOV result.depth.z, R0.x;\n"
1776         "END\n",
1777         /* tex_rect */
1778         "!!ARBfp1.0\n"
1779         "TEMP R0;\n"
1780         "TEX R0.x, fragment.texcoord[0], texture[0], RECT;\n"
1781         "MOV result.depth.z, R0.x;\n"
1782         "END\n",
1783     };
1784
1785     if (!blt_fprograms[tex_type])
1786     {
1787         FIXME("tex_type %#x not supported\n", tex_type);
1788         tex_type = tex_2d;
1789     }
1790
1791     GL_EXTCALL(glGenProgramsARB(1, &program_id));
1792     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, program_id));
1793     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_fprograms[tex_type]), blt_fprograms[tex_type]));
1794
1795     if (glGetError() == GL_INVALID_OPERATION) {
1796         GLint pos;
1797         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1798         FIXME("Fragment program error at position %d: %s\n", pos,
1799             debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1800     }
1801
1802     return program_id;
1803 }
1804
1805 static void shader_arb_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
1806     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1807     struct shader_arb_priv *priv = This->shader_priv;
1808     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1809
1810     if (useVS) {
1811         struct vs_compile_args compile_args;
1812
1813         TRACE("Using vertex shader\n");
1814         find_vs_compile_args((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, This->stateBlock, &compile_args);
1815         priv->current_vprogram_id = find_gl_vshader((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, &compile_args);
1816
1817         /* Bind the vertex program */
1818         GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1819         checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id);");
1820
1821         /* Enable OpenGL vertex programs */
1822         glEnable(GL_VERTEX_PROGRAM_ARB);
1823         checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1824         TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1825     } else if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
1826         priv->current_vprogram_id = 0;
1827         glDisable(GL_VERTEX_PROGRAM_ARB);
1828         checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1829     }
1830
1831     if (usePS) {
1832         struct ps_compile_args compile_args;
1833         TRACE("Using pixel shader\n");
1834         find_ps_compile_args((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader, This->stateBlock, &compile_args);
1835         priv->current_fprogram_id = find_gl_pshader((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader,
1836                                                     &compile_args);
1837
1838         /* Bind the fragment program */
1839         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1840         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id);");
1841
1842         if(!priv->use_arbfp_fixed_func) {
1843             /* Enable OpenGL fragment programs */
1844             glEnable(GL_FRAGMENT_PROGRAM_ARB);
1845             checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1846         }
1847         TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1848     } else if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM) && !priv->use_arbfp_fixed_func) {
1849         /* Disable only if we're not using arbfp fixed function fragment processing. If this is used,
1850          * keep GL_FRAGMENT_PROGRAM_ARB enabled, and the fixed function pipeline will bind the fixed function
1851          * replacement shader
1852          */
1853         glDisable(GL_FRAGMENT_PROGRAM_ARB);
1854         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1855         priv->current_fprogram_id = 0;
1856     }
1857 }
1858
1859 static void shader_arb_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {
1860     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1861     struct shader_arb_priv *priv = This->shader_priv;
1862     GLuint *blt_fprogram = &priv->depth_blt_fprogram_id[tex_type];
1863     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1864
1865     if (!priv->depth_blt_vprogram_id) priv->depth_blt_vprogram_id = create_arb_blt_vertex_program(gl_info);
1866     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->depth_blt_vprogram_id));
1867     glEnable(GL_VERTEX_PROGRAM_ARB);
1868
1869     if (!*blt_fprogram) *blt_fprogram = create_arb_blt_fragment_program(gl_info, tex_type);
1870     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, *blt_fprogram));
1871     glEnable(GL_FRAGMENT_PROGRAM_ARB);
1872 }
1873
1874 static void shader_arb_deselect_depth_blt(IWineD3DDevice *iface) {
1875     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1876     struct shader_arb_priv *priv = This->shader_priv;
1877     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1878
1879     if (priv->current_vprogram_id) {
1880         GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1881         checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vertexShader->prgId);");
1882
1883         glEnable(GL_VERTEX_PROGRAM_ARB);
1884         checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1885
1886         TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1887     } else {
1888         glDisable(GL_VERTEX_PROGRAM_ARB);
1889         checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1890     }
1891
1892     if (priv->current_fprogram_id) {
1893         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1894         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, pixelShader->prgId);");
1895
1896         glEnable(GL_FRAGMENT_PROGRAM_ARB);
1897         checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1898
1899         TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1900     } else {
1901         glDisable(GL_FRAGMENT_PROGRAM_ARB);
1902         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1903     }
1904 }
1905
1906 static void shader_arb_destroy(IWineD3DBaseShader *iface) {
1907     IWineD3DBaseShaderImpl *baseShader = (IWineD3DBaseShaderImpl *) iface;
1908     const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)baseShader->baseShader.device)->adapter->gl_info;
1909
1910     if (shader_is_pshader_version(baseShader->baseShader.reg_maps.shader_version))
1911     {
1912         IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *) iface;
1913         UINT i;
1914
1915         ENTER_GL();
1916         for(i = 0; i < This->num_gl_shaders; i++) {
1917             GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1918             checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1919         }
1920         LEAVE_GL();
1921         HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1922         This->gl_shaders = NULL;
1923         This->num_gl_shaders = 0;
1924         This->shader_array_size = 0;
1925     } else {
1926         IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *) iface;
1927         UINT i;
1928
1929         ENTER_GL();
1930         for(i = 0; i < This->num_gl_shaders; i++) {
1931             GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1932             checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1933         }
1934         LEAVE_GL();
1935         HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1936         This->gl_shaders = NULL;
1937         This->num_gl_shaders = 0;
1938         This->shader_array_size = 0;
1939     }
1940 }
1941
1942 static HRESULT shader_arb_alloc(IWineD3DDevice *iface) {
1943     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1944     This->shader_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
1945     return WINED3D_OK;
1946 }
1947
1948 static void shader_arb_free(IWineD3DDevice *iface) {
1949     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1950     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1951     struct shader_arb_priv *priv = This->shader_priv;
1952     int i;
1953
1954     if(priv->depth_blt_vprogram_id) {
1955         GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_vprogram_id));
1956     }
1957     for (i = 0; i < tex_type_count; ++i) {
1958         if (priv->depth_blt_fprogram_id[i]) {
1959             GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_fprogram_id[i]));
1960         }
1961     }
1962
1963     HeapFree(GetProcessHeap(), 0, This->shader_priv);
1964 }
1965
1966 static BOOL shader_arb_dirty_const(IWineD3DDevice *iface) {
1967     return TRUE;
1968 }
1969
1970 static void arbfp_add_sRGB_correction(SHADER_BUFFER *buffer, const char *fragcolor, const char *tmp1,
1971                                       const char *tmp2, const char *tmp3, const char *tmp4) {
1972     /* Perform sRGB write correction. See GLX_EXT_framebuffer_sRGB */
1973
1974     /* Calculate the > 0.0031308 case */
1975     shader_addline(buffer, "POW %s.x, %s.x, srgb_consts1.z;\n", tmp1, fragcolor);
1976     shader_addline(buffer, "POW %s.y, %s.y, srgb_consts1.z;\n", tmp1, fragcolor);
1977     shader_addline(buffer, "POW %s.z, %s.z, srgb_consts1.z;\n", tmp1, fragcolor);
1978     shader_addline(buffer, "MUL %s, %s, srgb_consts1.w;\n", tmp1, tmp1);
1979     shader_addline(buffer, "SUB %s, %s, srgb_consts2.x;\n", tmp1, tmp1);
1980     /* Calculate the < case */
1981     shader_addline(buffer, "MUL %s, srgb_consts1.x, %s;\n", tmp2, fragcolor);
1982     /* Get 1.0 / 0.0 masks for > 0.0031308 and < 0.0031308 */
1983     shader_addline(buffer, "SLT %s, srgb_consts1.y, %s;\n", tmp3, fragcolor);
1984     shader_addline(buffer, "SGE %s, srgb_consts1.y, %s;\n", tmp4, fragcolor);
1985     /* Store the components > 0.0031308 in the destination */
1986     shader_addline(buffer, "MUL %s, %s, %s;\n", fragcolor, tmp1, tmp3);
1987     /* Add the components that are < 0.0031308 */
1988     shader_addline(buffer, "MAD result.color.xyz, %s, %s, %s;\n", tmp2, tmp4, fragcolor);
1989     /* [0.0;1.0] clamping. Not needed, this is done implicitly */
1990 }
1991
1992 static GLuint shader_arb_generate_pshader(IWineD3DPixelShader *iface,
1993         SHADER_BUFFER *buffer, const struct ps_compile_args *args)
1994 {
1995     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1996     const shader_reg_maps* reg_maps = &This->baseShader.reg_maps;
1997     CONST DWORD *function = This->baseShader.function;
1998     DWORD shader_version = reg_maps->shader_version;
1999     const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)This->baseShader.device)->adapter->gl_info;
2000     const local_constant *lconst;
2001     GLuint retval;
2002     const char *fragcolor;
2003     DWORD *lconst_map = local_const_mapping((IWineD3DBaseShaderImpl *) This);
2004
2005     /*  Create the hw ARB shader */
2006     shader_addline(buffer, "!!ARBfp1.0\n");
2007
2008     if (shader_version < WINED3DPS_VERSION(3,0)) {
2009         switch(args->fog) {
2010             case FOG_OFF:
2011                 break;
2012             case FOG_LINEAR:
2013                 shader_addline(buffer, "OPTION ARB_fog_linear;\n");
2014                 break;
2015             case FOG_EXP:
2016                 shader_addline(buffer, "OPTION ARB_fog_exp;\n");
2017                 break;
2018             case FOG_EXP2:
2019                 shader_addline(buffer, "OPTION ARB_fog_exp2;\n");
2020                 break;
2021         }
2022     }
2023
2024     shader_addline(buffer, "TEMP TMP;\n");     /* Used in matrix ops */
2025     shader_addline(buffer, "TEMP TMP2;\n");    /* Used in matrix ops */
2026     shader_addline(buffer, "TEMP TA;\n");      /* Used for modifiers */
2027     shader_addline(buffer, "TEMP TB;\n");      /* Used for modifiers */
2028     shader_addline(buffer, "TEMP TC;\n");      /* Used for modifiers */
2029     shader_addline(buffer, "PARAM coefdiv = { 0.5, 0.25, 0.125, 0.0625 };\n");
2030     shader_addline(buffer, "PARAM coefmul = { 2, 4, 8, 16 };\n");
2031     shader_addline(buffer, "PARAM one = { 1.0, 1.0, 1.0, 1.0 };\n");
2032
2033     if (shader_version < WINED3DPS_VERSION(2,0)) {
2034         fragcolor = "R0";
2035     } else {
2036         shader_addline(buffer, "TEMP TMP_COLOR;\n");
2037         fragcolor = "TMP_COLOR";
2038     }
2039
2040     /* Base Declarations */
2041     shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION, lconst_map);
2042
2043     /* Base Shader Body */
2044     shader_generate_main((IWineD3DBaseShader *)This, buffer, reg_maps, function);
2045
2046     if(args->srgb_correction) {
2047         arbfp_add_sRGB_correction(buffer, fragcolor, "TMP", "TMP2", "TA", "TB");
2048     }
2049     shader_addline(buffer, "MOV result.color, %s;\n", fragcolor);
2050     shader_addline(buffer, "END\n");
2051
2052     /* TODO: change to resource.glObjectHandle or something like that */
2053     GL_EXTCALL(glGenProgramsARB(1, &retval));
2054
2055     TRACE("Creating a hw pixel shader, prg=%d\n", retval);
2056     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, retval));
2057
2058     TRACE("Created hw pixel shader, prg=%d\n", retval);
2059     /* Create the program and check for errors */
2060     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2061                buffer->bsize, buffer->buffer));
2062
2063     if (glGetError() == GL_INVALID_OPERATION) {
2064         GLint errPos;
2065         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2066         FIXME("HW PixelShader Error at position %d: %s\n",
2067               errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2068         retval = 0;
2069     }
2070
2071     /* Load immediate constants */
2072     if(lconst_map) {
2073         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2074             const float *value = (const float *)lconst->value;
2075             GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, lconst_map[lconst->idx], value));
2076             checkGLcall("glProgramLocalParameter4fvARB");
2077         }
2078         HeapFree(GetProcessHeap(), 0, lconst_map);
2079     }
2080
2081     return retval;
2082 }
2083
2084 static GLuint shader_arb_generate_vshader(IWineD3DVertexShader *iface,
2085         SHADER_BUFFER *buffer, const struct vs_compile_args *args)
2086 {
2087     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
2088     const shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
2089     CONST DWORD *function = This->baseShader.function;
2090     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)This->baseShader.device;
2091     const WineD3D_GL_Info *gl_info = &device->adapter->gl_info;
2092     const local_constant *lconst;
2093     GLuint ret;
2094     DWORD *lconst_map = local_const_mapping((IWineD3DBaseShaderImpl *) This);
2095
2096     /*  Create the hw ARB shader */
2097     shader_addline(buffer, "!!ARBvp1.0\n");
2098     shader_addline(buffer, "PARAM helper_const = { 2.0, -1.0, %d.0, 0.0 };\n", This->rel_offset);
2099
2100     /* Mesa supports only 95 constants */
2101     if (GL_VEND(MESA) || GL_VEND(WINE))
2102         This->baseShader.limits.constant_float =
2103                 min(95, This->baseShader.limits.constant_float);
2104
2105     shader_addline(buffer, "TEMP TMP;\n");
2106
2107     /* Base Declarations */
2108     shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION, lconst_map);
2109
2110     /* We need a constant to fixup the final position */
2111     shader_addline(buffer, "PARAM posFixup = program.env[%d];\n", ARB_SHADER_PRIVCONST_POS);
2112
2113     /* Initialize output parameters. GL_ARB_vertex_program does not require special initialization values
2114      * for output parameters. D3D in theory does not do that either, but some applications depend on a
2115      * proper initialization of the secondary color, and programs using the fixed function pipeline without
2116      * a replacement shader depend on the texcoord.w being set properly.
2117      *
2118      * GL_NV_vertex_program defines that all output values are initialized to {0.0, 0.0, 0.0, 1.0}. This
2119      * assertion is in effect even when using GL_ARB_vertex_program without any NV specific additions. So
2120      * skip this if NV_vertex_program is supported. Otherwise, initialize the secondary color. For the tex-
2121      * coords, we have a flag in the opengl caps. Many cards do not require the texcoord being set, and
2122      * this can eat a number of instructions, so skip it unless this cap is set as well
2123      */
2124     if(!GL_SUPPORT(NV_VERTEX_PROGRAM)) {
2125         shader_addline(buffer, "MOV result.color.secondary, -helper_const.wwwy;\n");
2126
2127         if((GLINFO_LOCATION).set_texcoord_w && !device->frag_pipe->ffp_proj_control) {
2128             int i;
2129             for(i = 0; i < min(8, MAX_REG_TEXCRD); i++) {
2130                 if(This->baseShader.reg_maps.texcoord_mask[i] != 0 &&
2131                 This->baseShader.reg_maps.texcoord_mask[i] != WINED3DSP_WRITEMASK_ALL) {
2132                     shader_addline(buffer, "MOV result.texcoord[%u].w, -helper_const.y;\n", i);
2133                 }
2134             }
2135         }
2136     }
2137
2138     /* Base Shader Body */
2139     shader_generate_main((IWineD3DBaseShader *)This, buffer, reg_maps, function);
2140
2141     /* The D3DRS_FOGTABLEMODE render state defines if the shader-generated fog coord is used
2142      * or if the fragment depth is used. If the fragment depth is used(FOGTABLEMODE != NONE),
2143      * the fog frag coord is thrown away. If the fog frag coord is used, but not written by
2144      * the shader, it is set to 0.0(fully fogged, since start = 1.0, end = 0.0)
2145      */
2146     if(args->fog_src == VS_FOG_Z) {
2147         shader_addline(buffer, "MOV result.fogcoord, TMP_OUT.z;\n");
2148     } else if (!reg_maps->fog) {
2149         shader_addline(buffer, "MOV result.fogcoord, helper_const.w;\n");
2150     }
2151
2152     /* Write the final position.
2153      *
2154      * OpenGL coordinates specify the center of the pixel while d3d coords specify
2155      * the corner. The offsets are stored in z and w in posFixup. posFixup.y contains
2156      * 1.0 or -1.0 to turn the rendering upside down for offscreen rendering. PosFixup.x
2157      * contains 1.0 to allow a mad, but arb vs swizzles are too restricted for that.
2158      */
2159     shader_addline(buffer, "MUL TMP, posFixup, TMP_OUT.w;\n");
2160     shader_addline(buffer, "ADD TMP_OUT.x, TMP_OUT.x, TMP.z;\n");
2161     shader_addline(buffer, "MAD TMP_OUT.y, TMP_OUT.y, posFixup.y, TMP.w;\n");
2162
2163     /* Z coord [0;1]->[-1;1] mapping, see comment in transform_projection in state.c
2164      * and the glsl equivalent
2165      */
2166     shader_addline(buffer, "MAD TMP_OUT.z, TMP_OUT.z, helper_const.x, -TMP_OUT.w;\n");
2167
2168     shader_addline(buffer, "MOV result.position, TMP_OUT;\n");
2169
2170     shader_addline(buffer, "END\n");
2171
2172     /* TODO: change to resource.glObjectHandle or something like that */
2173     GL_EXTCALL(glGenProgramsARB(1, &ret));
2174
2175     TRACE("Creating a hw vertex shader, prg=%d\n", ret);
2176     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, ret));
2177
2178     TRACE("Created hw vertex shader, prg=%d\n", ret);
2179     /* Create the program and check for errors */
2180     GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2181                buffer->bsize, buffer->buffer));
2182
2183     if (glGetError() == GL_INVALID_OPERATION) {
2184         GLint errPos;
2185         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2186         FIXME("HW VertexShader Error at position %d: %s\n",
2187               errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2188         ret = -1;
2189     } else {
2190         /* Load immediate constants */
2191         if(lconst_map) {
2192             LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2193                 const float *value = (const float *)lconst->value;
2194                 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_VERTEX_PROGRAM_ARB, lconst_map[lconst->idx], value));
2195             }
2196         }
2197     }
2198     HeapFree(GetProcessHeap(), 0, lconst_map);
2199
2200     return ret;
2201 }
2202
2203 static void shader_arb_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct shader_caps *pCaps)
2204 {
2205     /* We don't have an ARB fixed function pipeline yet, so let the none backend set its caps,
2206      * then overwrite the shader specific ones
2207      */
2208     none_shader_backend.shader_get_caps(devtype, gl_info, pCaps);
2209
2210     if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
2211         pCaps->VertexShaderVersion = WINED3DVS_VERSION(1,1);
2212         TRACE_(d3d_caps)("Hardware vertex shader version 1.1 enabled (ARB_PROGRAM)\n");
2213         pCaps->MaxVertexShaderConst = GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS;
2214     }
2215
2216     if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM)) {
2217         pCaps->PixelShaderVersion    = WINED3DPS_VERSION(1,4);
2218         pCaps->PixelShader1xMaxValue = 8.0;
2219         TRACE_(d3d_caps)("Hardware pixel shader version 1.4 enabled (ARB_PROGRAM)\n");
2220         pCaps->MaxPixelShaderConst = GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS;
2221     }
2222 }
2223
2224 static BOOL shader_arb_color_fixup_supported(struct color_fixup_desc fixup)
2225 {
2226     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
2227     {
2228         TRACE("Checking support for color_fixup:\n");
2229         dump_color_fixup_desc(fixup);
2230     }
2231
2232     /* We support everything except YUV conversions. */
2233     if (!is_yuv_fixup(fixup))
2234     {
2235         TRACE("[OK]\n");
2236         return TRUE;
2237     }
2238
2239     TRACE("[FAILED]\n");
2240     return FALSE;
2241 }
2242
2243 static const SHADER_HANDLER shader_arb_instruction_handler_table[WINED3DSIH_TABLE_SIZE] =
2244 {
2245     /* WINED3DSIH_ABS           */ shader_hw_map2gl,
2246     /* WINED3DSIH_ADD           */ shader_hw_map2gl,
2247     /* WINED3DSIH_BEM           */ pshader_hw_bem,
2248     /* WINED3DSIH_BREAK         */ NULL,
2249     /* WINED3DSIH_BREAKC        */ NULL,
2250     /* WINED3DSIH_BREAKP        */ NULL,
2251     /* WINED3DSIH_CALL          */ NULL,
2252     /* WINED3DSIH_CALLNZ        */ NULL,
2253     /* WINED3DSIH_CMP           */ pshader_hw_cmp,
2254     /* WINED3DSIH_CND           */ pshader_hw_cnd,
2255     /* WINED3DSIH_CRS           */ shader_hw_map2gl,
2256     /* WINED3DSIH_DCL           */ NULL,
2257     /* WINED3DSIH_DEF           */ NULL,
2258     /* WINED3DSIH_DEFB          */ NULL,
2259     /* WINED3DSIH_DEFI          */ NULL,
2260     /* WINED3DSIH_DP2ADD        */ pshader_hw_dp2add,
2261     /* WINED3DSIH_DP3           */ shader_hw_map2gl,
2262     /* WINED3DSIH_DP4           */ shader_hw_map2gl,
2263     /* WINED3DSIH_DST           */ shader_hw_map2gl,
2264     /* WINED3DSIH_DSX           */ NULL,
2265     /* WINED3DSIH_DSY           */ NULL,
2266     /* WINED3DSIH_ELSE          */ NULL,
2267     /* WINED3DSIH_ENDIF         */ NULL,
2268     /* WINED3DSIH_ENDLOOP       */ NULL,
2269     /* WINED3DSIH_ENDREP        */ NULL,
2270     /* WINED3DSIH_EXP           */ shader_hw_map2gl,
2271     /* WINED3DSIH_EXPP          */ shader_hw_map2gl,
2272     /* WINED3DSIH_FRC           */ shader_hw_map2gl,
2273     /* WINED3DSIH_IF            */ NULL,
2274     /* WINED3DSIH_IFC           */ NULL,
2275     /* WINED3DSIH_LABEL         */ NULL,
2276     /* WINED3DSIH_LIT           */ shader_hw_map2gl,
2277     /* WINED3DSIH_LOG           */ shader_hw_map2gl,
2278     /* WINED3DSIH_LOGP          */ shader_hw_map2gl,
2279     /* WINED3DSIH_LOOP          */ NULL,
2280     /* WINED3DSIH_LRP           */ shader_hw_map2gl,
2281     /* WINED3DSIH_M3x2          */ shader_hw_mnxn,
2282     /* WINED3DSIH_M3x3          */ shader_hw_mnxn,
2283     /* WINED3DSIH_M3x4          */ shader_hw_mnxn,
2284     /* WINED3DSIH_M4x3          */ shader_hw_mnxn,
2285     /* WINED3DSIH_M4x4          */ shader_hw_mnxn,
2286     /* WINED3DSIH_MAD           */ shader_hw_map2gl,
2287     /* WINED3DSIH_MAX           */ shader_hw_map2gl,
2288     /* WINED3DSIH_MIN           */ shader_hw_map2gl,
2289     /* WINED3DSIH_MOV           */ shader_hw_mov,
2290     /* WINED3DSIH_MOVA          */ shader_hw_mov,
2291     /* WINED3DSIH_MUL           */ shader_hw_map2gl,
2292     /* WINED3DSIH_NOP           */ shader_hw_map2gl,
2293     /* WINED3DSIH_NRM           */ shader_hw_nrm,
2294     /* WINED3DSIH_PHASE         */ NULL,
2295     /* WINED3DSIH_POW           */ shader_hw_map2gl,
2296     /* WINED3DSIH_RCP           */ vshader_hw_rsq_rcp,
2297     /* WINED3DSIH_REP           */ NULL,
2298     /* WINED3DSIH_RET           */ NULL,
2299     /* WINED3DSIH_RSQ           */ vshader_hw_rsq_rcp,
2300     /* WINED3DSIH_SETP          */ NULL,
2301     /* WINED3DSIH_SGE           */ shader_hw_map2gl,
2302     /* WINED3DSIH_SGN           */ NULL,
2303     /* WINED3DSIH_SINCOS        */ shader_hw_sincos,
2304     /* WINED3DSIH_SLT           */ shader_hw_map2gl,
2305     /* WINED3DSIH_SUB           */ shader_hw_map2gl,
2306     /* WINED3DSIH_TEX           */ pshader_hw_tex,
2307     /* WINED3DSIH_TEXBEM        */ pshader_hw_texbem,
2308     /* WINED3DSIH_TEXBEML       */ pshader_hw_texbem,
2309     /* WINED3DSIH_TEXCOORD      */ pshader_hw_texcoord,
2310     /* WINED3DSIH_TEXDEPTH      */ pshader_hw_texdepth,
2311     /* WINED3DSIH_TEXDP3        */ pshader_hw_texdp3,
2312     /* WINED3DSIH_TEXDP3TEX     */ pshader_hw_texdp3tex,
2313     /* WINED3DSIH_TEXKILL       */ pshader_hw_texkill,
2314     /* WINED3DSIH_TEXLDD        */ NULL,
2315     /* WINED3DSIH_TEXLDL        */ NULL,
2316     /* WINED3DSIH_TEXM3x2DEPTH  */ pshader_hw_texm3x2depth,
2317     /* WINED3DSIH_TEXM3x2PAD    */ pshader_hw_texm3x2pad,
2318     /* WINED3DSIH_TEXM3x2TEX    */ pshader_hw_texm3x2tex,
2319     /* WINED3DSIH_TEXM3x3       */ pshader_hw_texm3x3,
2320     /* WINED3DSIH_TEXM3x3DIFF   */ NULL,
2321     /* WINED3DSIH_TEXM3x3PAD    */ pshader_hw_texm3x3pad,
2322     /* WINED3DSIH_TEXM3x3SPEC   */ pshader_hw_texm3x3spec,
2323     /* WINED3DSIH_TEXM3x3TEX    */ pshader_hw_texm3x3tex,
2324     /* WINED3DSIH_TEXM3x3VSPEC  */ pshader_hw_texm3x3vspec,
2325     /* WINED3DSIH_TEXREG2AR     */ pshader_hw_texreg2ar,
2326     /* WINED3DSIH_TEXREG2GB     */ pshader_hw_texreg2gb,
2327     /* WINED3DSIH_TEXREG2RGB    */ pshader_hw_texreg2rgb,
2328 };
2329
2330 const shader_backend_t arb_program_shader_backend = {
2331     shader_arb_instruction_handler_table,
2332     shader_arb_select,
2333     shader_arb_select_depth_blt,
2334     shader_arb_deselect_depth_blt,
2335     shader_arb_update_float_vertex_constants,
2336     shader_arb_update_float_pixel_constants,
2337     shader_arb_load_constants,
2338     shader_arb_load_np2fixup_constants,
2339     shader_arb_destroy,
2340     shader_arb_alloc,
2341     shader_arb_free,
2342     shader_arb_dirty_const,
2343     shader_arb_generate_pshader,
2344     shader_arb_generate_vshader,
2345     shader_arb_get_caps,
2346     shader_arb_color_fixup_supported,
2347 };
2348
2349 /* ARB_fragment_program fixed function pipeline replacement definitions */
2350 #define ARB_FFP_CONST_TFACTOR           0
2351 #define ARB_FFP_CONST_SPECULAR_ENABLE   ((ARB_FFP_CONST_TFACTOR) + 1)
2352 #define ARB_FFP_CONST_CONSTANT(i)       ((ARB_FFP_CONST_SPECULAR_ENABLE) + 1 + i)
2353 #define ARB_FFP_CONST_BUMPMAT(i)        ((ARB_FFP_CONST_CONSTANT(7)) + 1 + i)
2354 #define ARB_FFP_CONST_LUMINANCE(i)      ((ARB_FFP_CONST_BUMPMAT(7)) + 1 + i)
2355
2356 struct arbfp_ffp_desc
2357 {
2358     struct ffp_frag_desc parent;
2359     GLuint shader;
2360     unsigned int num_textures_used;
2361 };
2362
2363 static void arbfp_enable(IWineD3DDevice *iface, BOOL enable) {
2364     if(enable) {
2365         glEnable(GL_FRAGMENT_PROGRAM_ARB);
2366         checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
2367     } else {
2368         glDisable(GL_FRAGMENT_PROGRAM_ARB);
2369         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
2370     }
2371 }
2372
2373 static HRESULT arbfp_alloc(IWineD3DDevice *iface) {
2374     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2375     struct shader_arb_priv *priv;
2376     /* Share private data between the shader backend and the pipeline replacement, if both
2377      * are the arb implementation. This is needed to figure out whether ARBfp should be disabled
2378      * if no pixel shader is bound or not
2379      */
2380     if(This->shader_backend == &arb_program_shader_backend) {
2381         This->fragment_priv = This->shader_priv;
2382     } else {
2383         This->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
2384         if(!This->fragment_priv) return E_OUTOFMEMORY;
2385     }
2386     priv = This->fragment_priv;
2387     priv->fragment_shaders = hash_table_create(ffp_frag_program_key_hash, ffp_frag_program_key_compare);
2388     priv->use_arbfp_fixed_func = TRUE;
2389     return WINED3D_OK;
2390 }
2391
2392 static void arbfp_free_ffpshader(void *value, void *gli) {
2393     const WineD3D_GL_Info *gl_info = gli;
2394     struct arbfp_ffp_desc *entry_arb = value;
2395
2396     ENTER_GL();
2397     GL_EXTCALL(glDeleteProgramsARB(1, &entry_arb->shader));
2398     checkGLcall("glDeleteProgramsARB(1, &entry_arb->shader)");
2399     HeapFree(GetProcessHeap(), 0, entry_arb);
2400     LEAVE_GL();
2401 }
2402
2403 static void arbfp_free(IWineD3DDevice *iface) {
2404     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2405     struct shader_arb_priv *priv = This->fragment_priv;
2406
2407     hash_table_destroy(priv->fragment_shaders, arbfp_free_ffpshader, &This->adapter->gl_info);
2408     priv->use_arbfp_fixed_func = FALSE;
2409
2410     if(This->shader_backend != &arb_program_shader_backend) {
2411         HeapFree(GetProcessHeap(), 0, This->fragment_priv);
2412     }
2413 }
2414
2415 static void arbfp_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct fragment_caps *caps)
2416 {
2417     caps->TextureOpCaps =  WINED3DTEXOPCAPS_DISABLE                     |
2418                            WINED3DTEXOPCAPS_SELECTARG1                  |
2419                            WINED3DTEXOPCAPS_SELECTARG2                  |
2420                            WINED3DTEXOPCAPS_MODULATE4X                  |
2421                            WINED3DTEXOPCAPS_MODULATE2X                  |
2422                            WINED3DTEXOPCAPS_MODULATE                    |
2423                            WINED3DTEXOPCAPS_ADDSIGNED2X                 |
2424                            WINED3DTEXOPCAPS_ADDSIGNED                   |
2425                            WINED3DTEXOPCAPS_ADD                         |
2426                            WINED3DTEXOPCAPS_SUBTRACT                    |
2427                            WINED3DTEXOPCAPS_ADDSMOOTH                   |
2428                            WINED3DTEXOPCAPS_BLENDCURRENTALPHA           |
2429                            WINED3DTEXOPCAPS_BLENDFACTORALPHA            |
2430                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHA           |
2431                            WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA           |
2432                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM         |
2433                            WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR      |
2434                            WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA      |
2435                            WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA   |
2436                            WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR   |
2437                            WINED3DTEXOPCAPS_DOTPRODUCT3                 |
2438                            WINED3DTEXOPCAPS_MULTIPLYADD                 |
2439                            WINED3DTEXOPCAPS_LERP                        |
2440                            WINED3DTEXOPCAPS_BUMPENVMAP                  |
2441                            WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE;
2442
2443     /* TODO: Implement WINED3DTEXOPCAPS_PREMODULATE */
2444
2445     caps->MaxTextureBlendStages   = 8;
2446     caps->MaxSimultaneousTextures = min(GL_LIMITS(fragment_samplers), 8);
2447
2448     caps->PrimitiveMiscCaps |= WINED3DPMISCCAPS_TSSARGTEMP;
2449 }
2450 #undef GLINFO_LOCATION
2451
2452 #define GLINFO_LOCATION stateblock->wineD3DDevice->adapter->gl_info
2453 static void state_texfactor_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2454     float col[4];
2455     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2456
2457     /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2458      * application provided constants
2459      */
2460     if(device->shader_backend == &arb_program_shader_backend) {
2461         if (use_ps(stateblock)) return;
2462
2463         device = stateblock->wineD3DDevice;
2464         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_TFACTOR] = 1;
2465         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_TFACTOR + 1);
2466     }
2467
2468     D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
2469     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col));
2470     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col)");
2471
2472 }
2473
2474 static void state_arb_specularenable(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2475     float col[4];
2476     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2477
2478     /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2479      * application provided constants
2480      */
2481     if(device->shader_backend == &arb_program_shader_backend) {
2482         if (use_ps(stateblock)) return;
2483
2484         device = stateblock->wineD3DDevice;
2485         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_SPECULAR_ENABLE] = 1;
2486         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_SPECULAR_ENABLE + 1);
2487     }
2488
2489     if(stateblock->renderState[WINED3DRS_SPECULARENABLE]) {
2490         /* The specular color has no alpha */
2491         col[0] = 1.0; col[1] = 1.0;
2492         col[2] = 1.0; col[3] = 0.0;
2493     } else {
2494         col[0] = 0.0; col[1] = 0.0;
2495         col[2] = 0.0; col[3] = 0.0;
2496     }
2497     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col));
2498     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col)");
2499 }
2500
2501 static void set_bumpmat_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2502     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2503     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2504     float mat[2][2];
2505
2506     if (use_ps(stateblock))
2507     {
2508         if(stage != 0 &&
2509            ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.bumpmat[stage]) {
2510             /* The pixel shader has to know the bump env matrix. Do a constants update if it isn't scheduled
2511              * anyway
2512              */
2513             if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2514                 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2515             }
2516         }
2517
2518         if(device->shader_backend == &arb_program_shader_backend) {
2519             /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2520             return;
2521         }
2522     } else if(device->shader_backend == &arb_program_shader_backend) {
2523         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_BUMPMAT(stage)] = 1;
2524         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_BUMPMAT(stage) + 1);
2525     }
2526
2527     mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
2528     mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
2529     mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
2530     mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
2531
2532     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0]));
2533     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0])");
2534 }
2535
2536 static void tex_bumpenvlum_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2537     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2538     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2539     float param[4];
2540
2541     if (use_ps(stateblock))
2542     {
2543         if(stage != 0 &&
2544            ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.luminanceparams[stage]) {
2545             /* The pixel shader has to know the luminance offset. Do a constants update if it
2546              * isn't scheduled anyway
2547              */
2548             if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2549                 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2550             }
2551         }
2552
2553         if(device->shader_backend == &arb_program_shader_backend) {
2554             /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2555             return;
2556         }
2557     } else if(device->shader_backend == &arb_program_shader_backend) {
2558         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_LUMINANCE(stage)] = 1;
2559         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_LUMINANCE(stage) + 1);
2560     }
2561
2562     param[0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLSCALE]);
2563     param[1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLOFFSET]);
2564     param[2] = 0.0;
2565     param[3] = 0.0;
2566
2567     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param));
2568     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param)");
2569 }
2570
2571 static const char *get_argreg(SHADER_BUFFER *buffer, DWORD argnum, unsigned int stage, DWORD arg) {
2572     const char *ret;
2573
2574     if(arg == ARG_UNUSED) return "unused"; /* This is the marker for unused registers */
2575
2576     switch(arg & WINED3DTA_SELECTMASK) {
2577         case WINED3DTA_DIFFUSE:
2578             ret = "fragment.color.primary"; break;
2579
2580         case WINED3DTA_CURRENT:
2581             if(stage == 0) ret = "fragment.color.primary";
2582             else ret = "ret";
2583             break;
2584
2585         case WINED3DTA_TEXTURE:
2586             switch(stage) {
2587                 case 0: ret = "tex0"; break;
2588                 case 1: ret = "tex1"; break;
2589                 case 2: ret = "tex2"; break;
2590                 case 3: ret = "tex3"; break;
2591                 case 4: ret = "tex4"; break;
2592                 case 5: ret = "tex5"; break;
2593                 case 6: ret = "tex6"; break;
2594                 case 7: ret = "tex7"; break;
2595                 default: ret = "unknown texture";
2596             }
2597             break;
2598
2599         case WINED3DTA_TFACTOR:
2600             ret = "tfactor"; break;
2601
2602         case WINED3DTA_SPECULAR:
2603             ret = "fragment.color.secondary"; break;
2604
2605         case WINED3DTA_TEMP:
2606             ret = "tempreg"; break;
2607
2608         case WINED3DTA_CONSTANT:
2609             FIXME("Implement perstage constants\n");
2610             switch(stage) {
2611                 case 0: ret = "const0"; break;
2612                 case 1: ret = "const1"; break;
2613                 case 2: ret = "const2"; break;
2614                 case 3: ret = "const3"; break;
2615                 case 4: ret = "const4"; break;
2616                 case 5: ret = "const5"; break;
2617                 case 6: ret = "const6"; break;
2618                 case 7: ret = "const7"; break;
2619                 default: ret = "unknown constant";
2620             }
2621             break;
2622
2623         default:
2624             return "unknown";
2625     }
2626
2627     if(arg & WINED3DTA_COMPLEMENT) {
2628         shader_addline(buffer, "SUB arg%u, const.x, %s;\n", argnum, ret);
2629         if(argnum == 0) ret = "arg0";
2630         if(argnum == 1) ret = "arg1";
2631         if(argnum == 2) ret = "arg2";
2632     }
2633     if(arg & WINED3DTA_ALPHAREPLICATE) {
2634         shader_addline(buffer, "MOV arg%u, %s.w;\n", argnum, ret);
2635         if(argnum == 0) ret = "arg0";
2636         if(argnum == 1) ret = "arg1";
2637         if(argnum == 2) ret = "arg2";
2638     }
2639     return ret;
2640 }
2641
2642 static void gen_ffp_instr(SHADER_BUFFER *buffer, unsigned int stage, BOOL color, BOOL alpha,
2643                           DWORD dst, DWORD op, DWORD dw_arg0, DWORD dw_arg1, DWORD dw_arg2) {
2644     const char *dstmask, *dstreg, *arg0, *arg1, *arg2;
2645     unsigned int mul = 1;
2646     BOOL mul_final_dest = FALSE;
2647
2648     if(color && alpha) dstmask = "";
2649     else if(color) dstmask = ".xyz";
2650     else dstmask = ".w";
2651
2652     if(dst == tempreg) dstreg = "tempreg";
2653     else dstreg = "ret";
2654
2655     arg0 = get_argreg(buffer, 0, stage, dw_arg0);
2656     arg1 = get_argreg(buffer, 1, stage, dw_arg1);
2657     arg2 = get_argreg(buffer, 2, stage, dw_arg2);
2658
2659     switch(op) {
2660         case WINED3DTOP_DISABLE:
2661             if(stage == 0) shader_addline(buffer, "MOV %s%s, fragment.color.primary;\n", dstreg, dstmask);
2662             break;
2663
2664         case WINED3DTOP_SELECTARG2:
2665             arg1 = arg2;
2666         case WINED3DTOP_SELECTARG1:
2667             shader_addline(buffer, "MOV %s%s, %s;\n", dstreg, dstmask, arg1);
2668             break;
2669
2670         case WINED3DTOP_MODULATE4X:
2671             mul = 2;
2672         case WINED3DTOP_MODULATE2X:
2673             mul *= 2;
2674             if(strcmp(dstreg, "result.color") == 0) {
2675                 dstreg = "ret";
2676                 mul_final_dest = TRUE;
2677             }
2678         case WINED3DTOP_MODULATE:
2679             shader_addline(buffer, "MUL %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2680             break;
2681
2682         case WINED3DTOP_ADDSIGNED2X:
2683             mul = 2;
2684             if(strcmp(dstreg, "result.color") == 0) {
2685                 dstreg = "ret";
2686                 mul_final_dest = TRUE;
2687             }
2688         case WINED3DTOP_ADDSIGNED:
2689             shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2690             arg2 = "arg2";
2691         case WINED3DTOP_ADD:
2692             shader_addline(buffer, "ADD_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2693             break;
2694
2695         case WINED3DTOP_SUBTRACT:
2696             shader_addline(buffer, "SUB_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2697             break;
2698
2699         case WINED3DTOP_ADDSMOOTH:
2700             shader_addline(buffer, "SUB arg1, const.x, %s;\n", arg1);
2701             shader_addline(buffer, "MAD_SAT %s%s, arg1, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2702             break;
2703
2704         case WINED3DTOP_BLENDCURRENTALPHA:
2705             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_CURRENT);
2706             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2707             break;
2708         case WINED3DTOP_BLENDFACTORALPHA:
2709             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TFACTOR);
2710             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2711             break;
2712         case WINED3DTOP_BLENDTEXTUREALPHA:
2713             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2714             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2715             break;
2716         case WINED3DTOP_BLENDDIFFUSEALPHA:
2717             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_DIFFUSE);
2718             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2719             break;
2720
2721         case WINED3DTOP_BLENDTEXTUREALPHAPM:
2722             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2723             shader_addline(buffer, "SUB arg0.w, const.x, %s.w;\n", arg0);
2724             shader_addline(buffer, "MAD_SAT %s%s, %s, arg0.w, %s;\n", dstreg, dstmask, arg2, arg1);
2725             break;
2726
2727         /* D3DTOP_PREMODULATE ???? */
2728
2729         case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
2730             shader_addline(buffer, "SUB arg0.w, const.x, %s;\n", arg1);
2731             shader_addline(buffer, "MAD_SAT %s%s, arg0.w, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2732             break;
2733         case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
2734             shader_addline(buffer, "MAD_SAT %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg1);
2735             break;
2736         case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
2737             shader_addline(buffer, "SUB arg0, const.x, %s;\n", arg1);
2738             shader_addline(buffer, "MAD_SAT %s%s, arg0, %s, %s.w;\n", dstreg, dstmask, arg2, arg1);
2739             break;
2740         case WINED3DTOP_MODULATECOLOR_ADDALPHA:
2741             shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s.w;\n", dstreg, dstmask, arg1, arg2, arg1);
2742             break;
2743
2744         case WINED3DTOP_DOTPRODUCT3:
2745             mul = 4;
2746             if(strcmp(dstreg, "result.color") == 0) {
2747                 dstreg = "ret";
2748                 mul_final_dest = TRUE;
2749             }
2750             shader_addline(buffer, "SUB arg1, %s, const.w;\n", arg1);
2751             shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2752             shader_addline(buffer, "DP3_SAT %s%s, arg1, arg2;\n", dstreg, dstmask);
2753             break;
2754
2755         case WINED3DTOP_MULTIPLYADD:
2756             shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg0);
2757             break;
2758
2759         case WINED3DTOP_LERP:
2760             /* The msdn is not quite right here */
2761             shader_addline(buffer, "LRP %s%s, %s, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2762             break;
2763
2764         case WINED3DTOP_BUMPENVMAP:
2765         case WINED3DTOP_BUMPENVMAPLUMINANCE:
2766             /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
2767             break;
2768
2769         default:
2770             FIXME("Unhandled texture op %08x\n", op);
2771     }
2772
2773     if(mul == 2) {
2774         shader_addline(buffer, "MUL_SAT %s%s, %s, const.y;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2775     } else if(mul == 4) {
2776         shader_addline(buffer, "MUL_SAT %s%s, %s, const.z;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2777     }
2778 }
2779
2780 /* The stateblock is passed for GLINFO_LOCATION */
2781 static GLuint gen_arbfp_ffp_shader(const struct ffp_frag_settings *settings, IWineD3DStateBlockImpl *stateblock)
2782 {
2783     unsigned int stage;
2784     SHADER_BUFFER buffer;
2785     BOOL tex_read[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2786     BOOL bump_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2787     BOOL luminance_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2788     const char *textype;
2789     const char *instr, *sat;
2790     char colorcor_dst[8];
2791     GLuint ret;
2792     DWORD arg0, arg1, arg2;
2793     BOOL tempreg_used = FALSE, tfactor_used = FALSE;
2794     BOOL op_equal;
2795     const char *final_combiner_src = "ret";
2796
2797     /* Find out which textures are read */
2798     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2799         if(settings->op[stage].cop == WINED3DTOP_DISABLE) break;
2800         arg0 = settings->op[stage].carg0 & WINED3DTA_SELECTMASK;
2801         arg1 = settings->op[stage].carg1 & WINED3DTA_SELECTMASK;
2802         arg2 = settings->op[stage].carg2 & WINED3DTA_SELECTMASK;
2803         if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2804         if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2805         if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2806
2807         if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) tex_read[stage] = TRUE;
2808         if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHAPM) tex_read[stage] = TRUE;
2809         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP) {
2810             bump_used[stage] = TRUE;
2811             tex_read[stage] = TRUE;
2812         }
2813         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2814             bump_used[stage] = TRUE;
2815             tex_read[stage] = TRUE;
2816             luminance_used[stage] = TRUE;
2817         } else if(settings->op[stage].cop == WINED3DTOP_BLENDFACTORALPHA) {
2818             tfactor_used = TRUE;
2819         }
2820
2821         if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2822             tfactor_used = TRUE;
2823         }
2824
2825         if(settings->op[stage].dst == tempreg) tempreg_used = TRUE;
2826         if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2827             tempreg_used = TRUE;
2828         }
2829
2830         if(settings->op[stage].aop == WINED3DTOP_DISABLE) continue;
2831         arg0 = settings->op[stage].aarg0 & WINED3DTA_SELECTMASK;
2832         arg1 = settings->op[stage].aarg1 & WINED3DTA_SELECTMASK;
2833         arg2 = settings->op[stage].aarg2 & WINED3DTA_SELECTMASK;
2834         if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2835         if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2836         if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2837
2838         if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2839             tempreg_used = TRUE;
2840         }
2841         if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2842             tfactor_used = TRUE;
2843         }
2844     }
2845
2846     /* Shader header */
2847     shader_buffer_init(&buffer);
2848
2849     shader_addline(&buffer, "!!ARBfp1.0\n");
2850
2851     switch(settings->fog) {
2852         case FOG_OFF:                                                         break;
2853         case FOG_LINEAR: shader_addline(&buffer, "OPTION ARB_fog_linear;\n"); break;
2854         case FOG_EXP:    shader_addline(&buffer, "OPTION ARB_fog_exp;\n");    break;
2855         case FOG_EXP2:   shader_addline(&buffer, "OPTION ARB_fog_exp2;\n");   break;
2856         default: FIXME("Unexpected fog setting %d\n", settings->fog);
2857     }
2858
2859     shader_addline(&buffer, "PARAM const = {1, 2, 4, 0.5};\n");
2860     shader_addline(&buffer, "TEMP TMP;\n");
2861     shader_addline(&buffer, "TEMP ret;\n");
2862     if(tempreg_used || settings->sRGB_write) shader_addline(&buffer, "TEMP tempreg;\n");
2863     shader_addline(&buffer, "TEMP arg0;\n");
2864     shader_addline(&buffer, "TEMP arg1;\n");
2865     shader_addline(&buffer, "TEMP arg2;\n");
2866     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2867         if(!tex_read[stage]) continue;
2868         shader_addline(&buffer, "TEMP tex%u;\n", stage);
2869         if(!bump_used[stage]) continue;
2870         shader_addline(&buffer, "PARAM bumpmat%u = program.env[%u];\n", stage, ARB_FFP_CONST_BUMPMAT(stage));
2871         if(!luminance_used[stage]) continue;
2872         shader_addline(&buffer, "PARAM luminance%u = program.env[%u];\n", stage, ARB_FFP_CONST_LUMINANCE(stage));
2873     }
2874     if(tfactor_used) {
2875         shader_addline(&buffer, "PARAM tfactor = program.env[%u];\n", ARB_FFP_CONST_TFACTOR);
2876     }
2877         shader_addline(&buffer, "PARAM specular_enable = program.env[%u];\n", ARB_FFP_CONST_SPECULAR_ENABLE);
2878
2879     if(settings->sRGB_write) {
2880         shader_addline(&buffer, "PARAM srgb_consts1 = {%f, %f, %f, %f};\n",
2881                        srgb_mul_low, srgb_cmp, srgb_pow, srgb_mul_high);
2882         shader_addline(&buffer, "PARAM srgb_consts2 = {%f, %f, %f, %f};\n",
2883                        srgb_sub_high, 0.0, 0.0, 0.0);
2884     }
2885
2886     /* Generate texture sampling instructions) */
2887     for(stage = 0; stage < MAX_TEXTURES && settings->op[stage].cop != WINED3DTOP_DISABLE; stage++) {
2888         if(!tex_read[stage]) continue;
2889
2890         switch(settings->op[stage].tex_type) {
2891             case tex_1d:                    textype = "1D";     break;
2892             case tex_2d:                    textype = "2D";     break;
2893             case tex_3d:                    textype = "3D";     break;
2894             case tex_cube:                  textype = "CUBE";   break;
2895             case tex_rect:                  textype = "RECT";   break;
2896             default: textype = "unexpected_textype";   break;
2897         }
2898
2899         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP ||
2900            settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2901             sat = "";
2902         } else {
2903             sat = "_SAT";
2904         }
2905
2906         if(settings->op[stage].projected == proj_none) {
2907             instr = "TEX";
2908         } else if(settings->op[stage].projected == proj_count4 ||
2909                   settings->op[stage].projected == proj_count3) {
2910             instr = "TXP";
2911         } else {
2912             FIXME("Unexpected projection mode %d\n", settings->op[stage].projected);
2913             instr = "TXP";
2914         }
2915
2916         if(stage > 0 &&
2917            (settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
2918             settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
2919             shader_addline(&buffer, "SWZ arg1, bumpmat%u, x, z, 0, 0;\n", stage - 1);
2920             shader_addline(&buffer, "DP3 ret.x, arg1, tex%u;\n", stage - 1);
2921             shader_addline(&buffer, "SWZ arg1, bumpmat%u, y, w, 0, 0;\n", stage - 1);
2922             shader_addline(&buffer, "DP3 ret.y, arg1, tex%u;\n", stage - 1);
2923
2924             /* with projective textures, texbem only divides the static texture coord, not the displacement,
2925              * so multiply the displacement with the dividing parameter before passing it to TXP
2926              */
2927             if (settings->op[stage].projected != proj_none) {
2928                 if(settings->op[stage].projected == proj_count4) {
2929                     shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].w;\n", stage);
2930                     shader_addline(&buffer, "MUL ret.xyz, ret, fragment.texcoord[%u].w, fragment.texcoord[%u];\n", stage, stage);
2931                 } else {
2932                     shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].z;\n", stage);
2933                     shader_addline(&buffer, "MAD ret.xyz, ret, fragment.texcoord[%u].z, fragment.texcoord[%u];\n", stage, stage);
2934                 }
2935             } else {
2936                 shader_addline(&buffer, "ADD ret, ret, fragment.texcoord[%u];\n", stage);
2937             }
2938
2939             shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2940                            instr, sat, stage, stage, textype);
2941             if(settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2942                 shader_addline(&buffer, "MAD_SAT ret.x, tex%u.z, luminance%u.x, luminance%u.y;\n",
2943                                stage - 1, stage - 1, stage - 1);
2944                 shader_addline(&buffer, "MUL tex%u, tex%u, ret.x;\n", stage, stage);
2945             }
2946         } else if(settings->op[stage].projected == proj_count3) {
2947             shader_addline(&buffer, "MOV ret, fragment.texcoord[%u];\n", stage);
2948             shader_addline(&buffer, "MOV ret.w, ret.z;\n");
2949             shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2950                             instr, sat, stage, stage, textype);
2951         } else {
2952             shader_addline(&buffer, "%s%s tex%u, fragment.texcoord[%u], texture[%u], %s;\n",
2953                             instr, sat, stage, stage, stage, textype);
2954         }
2955
2956         sprintf(colorcor_dst, "tex%u", stage);
2957         gen_color_correction(&buffer, colorcor_dst, WINED3DSP_WRITEMASK_ALL, "const.x", "const.y",
2958                 settings->op[stage].color_fixup);
2959     }
2960
2961     /* Generate the main shader */
2962     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2963         if(settings->op[stage].cop == WINED3DTOP_DISABLE) {
2964             if(stage == 0) {
2965                 final_combiner_src = "fragment.color.primary";
2966             }
2967             break;
2968         }
2969
2970         if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2971            settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2972             op_equal = settings->op[stage].carg1 == settings->op[stage].aarg1;
2973         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2974                   settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2975             op_equal = settings->op[stage].carg1 == settings->op[stage].aarg2;
2976         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2977                   settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2978             op_equal = settings->op[stage].carg2 == settings->op[stage].aarg1;
2979         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2980                   settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2981             op_equal = settings->op[stage].carg2 == settings->op[stage].aarg2;
2982         } else {
2983             op_equal = settings->op[stage].aop   == settings->op[stage].cop &&
2984                        settings->op[stage].carg0 == settings->op[stage].aarg0 &&
2985                        settings->op[stage].carg1 == settings->op[stage].aarg1 &&
2986                        settings->op[stage].carg2 == settings->op[stage].aarg2;
2987         }
2988
2989         if(settings->op[stage].aop == WINED3DTOP_DISABLE) {
2990             gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2991                           settings->op[stage].cop, settings->op[stage].carg0,
2992                           settings->op[stage].carg1, settings->op[stage].carg2);
2993             if(stage == 0) {
2994                 shader_addline(&buffer, "MOV ret.w, fragment.color.primary.w;\n");
2995             }
2996         } else if(op_equal) {
2997             gen_ffp_instr(&buffer, stage, TRUE, TRUE, settings->op[stage].dst,
2998                           settings->op[stage].cop, settings->op[stage].carg0,
2999                           settings->op[stage].carg1, settings->op[stage].carg2);
3000         } else {
3001             gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
3002                           settings->op[stage].cop, settings->op[stage].carg0,
3003                           settings->op[stage].carg1, settings->op[stage].carg2);
3004             gen_ffp_instr(&buffer, stage, FALSE, TRUE, settings->op[stage].dst,
3005                           settings->op[stage].aop, settings->op[stage].aarg0,
3006                           settings->op[stage].aarg1, settings->op[stage].aarg2);
3007         }
3008     }
3009
3010     if(settings->sRGB_write) {
3011         shader_addline(&buffer, "MAD ret, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
3012         arbfp_add_sRGB_correction(&buffer, "ret", "arg0", "arg1", "arg2", "tempreg");
3013         shader_addline(&buffer, "MOV result.color.w, ret.w;\n");
3014     } else {
3015         shader_addline(&buffer, "MAD result.color, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
3016     }
3017
3018     /* Footer */
3019     shader_addline(&buffer, "END\n");
3020
3021     /* Generate the shader */
3022     GL_EXTCALL(glGenProgramsARB(1, &ret));
3023     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, ret));
3024     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3025
3026     if (glGetError() == GL_INVALID_OPERATION) {
3027         GLint pos;
3028         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3029         FIXME("Fragment program error at position %d: %s\n", pos,
3030               debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3031     }
3032     shader_buffer_free(&buffer);
3033     return ret;
3034 }
3035
3036 static void fragment_prog_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3037     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
3038     struct shader_arb_priv *priv = device->fragment_priv;
3039     BOOL use_pshader = use_ps(stateblock);
3040     BOOL use_vshader = use_vs(stateblock);
3041     struct ffp_frag_settings settings;
3042     const struct arbfp_ffp_desc *desc;
3043     unsigned int i;
3044
3045     TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
3046
3047     if(isStateDirty(context, STATE_RENDER(WINED3DRS_FOGENABLE))) {
3048         if(!use_pshader && device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3049             /* Reload fixed function constants since they collide with the pixel shader constants */
3050             for(i = 0; i < MAX_TEXTURES; i++) {
3051                 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3052             }
3053             state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3054             state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3055         } else if(use_pshader && !isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3056             device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3057         }
3058         return;
3059     }
3060
3061     if(!use_pshader) {
3062         /* Find or create a shader implementing the fixed function pipeline settings, then activate it */
3063         gen_ffp_frag_op(stateblock, &settings, FALSE);
3064         desc = (const struct arbfp_ffp_desc *)find_ffp_frag_shader(priv->fragment_shaders, &settings);
3065         if(!desc) {
3066             struct arbfp_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
3067             if (!new_desc)
3068             {
3069                 ERR("Out of memory\n");
3070                 return;
3071             }
3072             new_desc->num_textures_used = 0;
3073             for(i = 0; i < GL_LIMITS(texture_stages); i++) {
3074                 if(settings.op[i].cop == WINED3DTOP_DISABLE) break;
3075                 new_desc->num_textures_used = i;
3076             }
3077
3078             memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
3079             new_desc->shader = gen_arbfp_ffp_shader(&settings, stateblock);
3080             add_ffp_frag_shader(priv->fragment_shaders, &new_desc->parent);
3081             TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
3082             desc = new_desc;
3083         }
3084
3085         /* Now activate the replacement program. GL_FRAGMENT_PROGRAM_ARB is already active(however, note the
3086          * comment above the shader_select call below). If e.g. GLSL is active, the shader_select call will
3087          * deactivate it.
3088          */
3089         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader));
3090         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader)");
3091         priv->current_fprogram_id = desc->shader;
3092
3093         if(device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3094             /* Reload fixed function constants since they collide with the pixel shader constants */
3095             for(i = 0; i < MAX_TEXTURES; i++) {
3096                 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3097             }
3098             state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3099             state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3100         }
3101         context->last_was_pshader = FALSE;
3102     } else {
3103         context->last_was_pshader = TRUE;
3104     }
3105
3106     /* Finally, select the shader. If a pixel shader is used, it will be set and enabled by the shader backend.
3107      * If this shader backend is arbfp(most likely), then it will simply overwrite the last fixed function replace-
3108      * ment shader. If the shader backend is not ARB, it currently is important that the opengl implementation
3109      * type overwrites GL_ARB_fragment_program. This is currently the case with GLSL. If we really want to use
3110      * atifs or nvrc pixel shaders with arb fragment programs we'd have to disable GL_FRAGMENT_PROGRAM_ARB here
3111      *
3112      * Don't call shader_select if the vertex shader is dirty, because it will be called later on by the vertex
3113      * shader handler
3114      */
3115     if(!isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3116         device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3117
3118         if (!isStateDirty(context, STATE_VERTEXSHADERCONSTANT) && (use_vshader || use_pshader)) {
3119             device->StateTable[STATE_VERTEXSHADERCONSTANT].apply(STATE_VERTEXSHADERCONSTANT, stateblock, context);
3120         }
3121     }
3122     if(use_pshader) {
3123         device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
3124     }
3125 }
3126
3127 /* We can't link the fog states to the fragment state directly since the vertex pipeline links them
3128  * to FOGENABLE. A different linking in different pipeline parts can't be expressed in the combined
3129  * state table, so we need to handle that with a forwarding function. The other invisible side effect
3130  * is that changing the fog start and fog end(which links to FOGENABLE in vertex) results in the
3131  * fragment_prog_arbfp function being called because FOGENABLE is dirty, which calls this function here
3132  */
3133 static void state_arbfp_fog(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3134     enum fogsource new_source;
3135
3136     TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
3137
3138     if(!isStateDirty(context, STATE_PIXELSHADER)) {
3139         fragment_prog_arbfp(state, stateblock, context);
3140     }
3141
3142     if(!stateblock->renderState[WINED3DRS_FOGENABLE]) return;
3143
3144     if(stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
3145         if(use_vs(stateblock)) {
3146             new_source = FOGSOURCE_VS;
3147         } else {
3148             if(stateblock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || context->last_was_rhw) {
3149                 new_source = FOGSOURCE_COORD;
3150             } else {
3151                 new_source = FOGSOURCE_FFP;
3152             }
3153         }
3154     } else {
3155         new_source = FOGSOURCE_FFP;
3156     }
3157     if(new_source != context->fog_source) {
3158         context->fog_source = new_source;
3159         state_fogstartend(STATE_RENDER(WINED3DRS_FOGSTART), stateblock, context);
3160     }
3161 }
3162
3163 static void textransform(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3164     if(!isStateDirty(context, STATE_PIXELSHADER)) {
3165         fragment_prog_arbfp(state, stateblock, context);
3166     }
3167 }
3168
3169 #undef GLINFO_LOCATION
3170
3171 static const struct StateEntryTemplate arbfp_fragmentstate_template[] = {
3172     {STATE_RENDER(WINED3DRS_TEXTUREFACTOR),               { STATE_RENDER(WINED3DRS_TEXTUREFACTOR),              state_texfactor_arbfp   }, 0                               },
3173     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3174     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3175     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3176     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3177     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3178     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3179     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3180     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3181     {STATE_TEXTURESTAGE(0, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3182     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3183     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3184     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3185     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3186     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3187     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3188     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3189     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3190     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3191     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3192     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3193     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3194     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3195     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3196     {STATE_TEXTURESTAGE(1, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3197     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3198     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3199     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3200     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3201     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3202     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3203     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3204     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3205     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3206     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3207     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3208     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3209     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3210     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3211     {STATE_TEXTURESTAGE(2, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3212     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3213     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3214     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3215     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3216     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3217     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3218     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3219     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3220     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3221     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3222     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3223     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3224     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3225     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3226     {STATE_TEXTURESTAGE(3, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3227     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3228     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3229     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3230     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3231     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3232     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3233     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3234     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3235     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3236     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3237     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3238     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3239     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3240     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3241     {STATE_TEXTURESTAGE(4, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3242     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3243     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3244     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3245     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3246     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3247     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3248     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3249     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3250     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3251     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3252     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3253     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3254     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3255     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3256     {STATE_TEXTURESTAGE(5, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3257     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3258     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3259     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3260     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3261     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3262     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3263     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3264     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3265     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3266     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3267     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3268     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3269     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3270     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3271     {STATE_TEXTURESTAGE(6, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3272     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3273     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3274     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3275     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3276     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3277     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3278     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3279     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3280     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3281     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3282     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3283     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3284     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3285     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3286     {STATE_TEXTURESTAGE(7, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3287     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3288     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3289     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3290     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3291     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3292     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3293     {STATE_SAMPLER(0),                                    { STATE_SAMPLER(0),                                   sampler_texdim          }, 0                               },
3294     {STATE_SAMPLER(1),                                    { STATE_SAMPLER(1),                                   sampler_texdim          }, 0                               },
3295     {STATE_SAMPLER(2),                                    { STATE_SAMPLER(2),                                   sampler_texdim          }, 0                               },
3296     {STATE_SAMPLER(3),                                    { STATE_SAMPLER(3),                                   sampler_texdim          }, 0                               },
3297     {STATE_SAMPLER(4),                                    { STATE_SAMPLER(4),                                   sampler_texdim          }, 0                               },
3298     {STATE_SAMPLER(5),                                    { STATE_SAMPLER(5),                                   sampler_texdim          }, 0                               },
3299     {STATE_SAMPLER(6),                                    { STATE_SAMPLER(6),                                   sampler_texdim          }, 0                               },
3300     {STATE_SAMPLER(7),                                    { STATE_SAMPLER(7),                                   sampler_texdim          }, 0                               },
3301     {STATE_PIXELSHADER,                                   { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3302     {STATE_RENDER(WINED3DRS_FOGENABLE),                   { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3303     {STATE_RENDER(WINED3DRS_FOGTABLEMODE),                { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3304     {STATE_RENDER(WINED3DRS_FOGVERTEXMODE),               { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3305     {STATE_RENDER(WINED3DRS_FOGSTART),                    { STATE_RENDER(WINED3DRS_FOGSTART),                   state_fogstartend       }, 0                               },
3306     {STATE_RENDER(WINED3DRS_FOGEND),                      { STATE_RENDER(WINED3DRS_FOGSTART),                   state_fogstartend       }, 0                               },
3307     {STATE_RENDER(WINED3DRS_SRGBWRITEENABLE),             { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3308     {STATE_RENDER(WINED3DRS_FOGCOLOR),                    { STATE_RENDER(WINED3DRS_FOGCOLOR),                   state_fogcolor          }, 0                               },
3309     {STATE_RENDER(WINED3DRS_FOGDENSITY),                  { STATE_RENDER(WINED3DRS_FOGDENSITY),                 state_fogdensity        }, 0                               },
3310     {STATE_TEXTURESTAGE(0,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(0, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3311     {STATE_TEXTURESTAGE(1,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(1, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3312     {STATE_TEXTURESTAGE(2,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(2, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3313     {STATE_TEXTURESTAGE(3,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(3, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3314     {STATE_TEXTURESTAGE(4,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(4, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3315     {STATE_TEXTURESTAGE(5,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(5, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3316     {STATE_TEXTURESTAGE(6,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(6, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3317     {STATE_TEXTURESTAGE(7,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(7, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3318     {STATE_RENDER(WINED3DRS_SPECULARENABLE),              { STATE_RENDER(WINED3DRS_SPECULARENABLE),             state_arb_specularenable}, 0                               },
3319     {0 /* Terminate */,                                   { 0,                                                  0                       }, 0                               },
3320 };
3321
3322 const struct fragment_pipeline arbfp_fragment_pipeline = {
3323     arbfp_enable,
3324     arbfp_get_caps,
3325     arbfp_alloc,
3326     arbfp_free,
3327     shader_arb_color_fixup_supported,
3328     arbfp_fragmentstate_template,
3329     TRUE /* We can disable projected textures */
3330 };
3331
3332 #define GLINFO_LOCATION device->adapter->gl_info
3333
3334 struct arbfp_blit_priv {
3335     GLenum yuy2_rect_shader, yuy2_2d_shader;
3336     GLenum uyvy_rect_shader, uyvy_2d_shader;
3337     GLenum yv12_rect_shader, yv12_2d_shader;
3338 };
3339
3340 static HRESULT arbfp_blit_alloc(IWineD3DDevice *iface) {
3341     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3342     device->blit_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct arbfp_blit_priv));
3343     if(!device->blit_priv) {
3344         ERR("Out of memory\n");
3345         return E_OUTOFMEMORY;
3346     }
3347     return WINED3D_OK;
3348 }
3349 static void arbfp_blit_free(IWineD3DDevice *iface) {
3350     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3351     struct arbfp_blit_priv *priv = device->blit_priv;
3352
3353     ENTER_GL();
3354     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_rect_shader));
3355     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_2d_shader));
3356     GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_rect_shader));
3357     GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_2d_shader));
3358     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_rect_shader));
3359     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_2d_shader));
3360     checkGLcall("Delete yuv programs\n");
3361     LEAVE_GL();
3362 }
3363
3364 static BOOL gen_planar_yuv_read(SHADER_BUFFER *buffer, enum yuv_fixup yuv_fixup, GLenum textype, char *luminance)
3365 {
3366     char chroma;
3367     const char *tex, *texinstr;
3368
3369     if (yuv_fixup == YUV_FIXUP_UYVY) {
3370         chroma = 'x';
3371         *luminance = 'w';
3372     } else {
3373         chroma = 'w';
3374         *luminance = 'x';
3375     }
3376     switch(textype) {
3377         case GL_TEXTURE_2D:             tex = "2D";     texinstr = "TXP"; break;
3378         case GL_TEXTURE_RECTANGLE_ARB:  tex = "RECT";   texinstr = "TEX"; break;
3379         default:
3380             /* This is more tricky than just replacing the texture type - we have to navigate
3381              * properly in the texture to find the correct chroma values
3382              */
3383             FIXME("Implement yuv correction for non-2d, non-rect textures\n");
3384             return FALSE;
3385     }
3386
3387     /* First we have to read the chroma values. This means we need at least two pixels(no filtering),
3388      * or 4 pixels(with filtering). To get the unmodified chromas, we have to rid ourselves of the
3389      * filtering when we sample the texture.
3390      *
3391      * These are the rules for reading the chroma:
3392      *
3393      * Even pixel: Cr
3394      * Even pixel: U
3395      * Odd pixel: V
3396      *
3397      * So we have to get the sampling x position in non-normalized coordinates in integers
3398      */
3399     if(textype != GL_TEXTURE_RECTANGLE_ARB) {
3400         shader_addline(buffer, "MUL texcrd.xy, fragment.texcoord[0], size.x;\n");
3401         shader_addline(buffer, "MOV texcrd.w, size.x;\n");
3402     } else {
3403         shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3404     }
3405     /* We must not allow filtering between pixel x and x+1, this would mix U and V
3406      * Vertical filtering is ok. However, bear in mind that the pixel center is at
3407      * 0.5, so add 0.5.
3408      */
3409     shader_addline(buffer, "FLR texcrd.x, texcrd.x;\n");
3410     shader_addline(buffer, "ADD texcrd.x, texcrd.x, coef.y;\n");
3411
3412     /* Divide the x coordinate by 0.5 and get the fraction. This gives 0.25 and 0.75 for the
3413      * even and odd pixels respectively
3414      */
3415     shader_addline(buffer, "MUL texcrd2, texcrd, coef.y;\n");
3416     shader_addline(buffer, "FRC texcrd2, texcrd2;\n");
3417
3418     /* Sample Pixel 1 */
3419     shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3420
3421     /* Put the value into either of the chroma values */
3422     shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3423     shader_addline(buffer, "MUL chroma.x, luminance.%c, temp.x;\n", chroma);
3424     shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3425     shader_addline(buffer, "MUL chroma.y, luminance.%c, temp.x;\n", chroma);
3426
3427     /* Sample pixel 2. If we read an even pixel(SLT above returned 1), sample
3428      * the pixel right to the current one. Otherwise, sample the left pixel.
3429      * Bias and scale the SLT result to -1;1 and add it to the texcrd.x.
3430      */
3431     shader_addline(buffer, "MAD temp.x, temp.x, coef.z, -coef.x;\n");
3432     shader_addline(buffer, "ADD texcrd.x, texcrd, temp.x;\n");
3433     shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3434
3435     /* Put the value into the other chroma */
3436     shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3437     shader_addline(buffer, "MAD chroma.y, luminance.%c, temp.x, chroma.y;\n", chroma);
3438     shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3439     shader_addline(buffer, "MAD chroma.x, luminance.%c, temp.x, chroma.x;\n", chroma);
3440
3441     /* TODO: If filtering is enabled, sample a 2nd pair of pixels left or right of
3442      * the current one and lerp the two U and V values
3443      */
3444
3445     /* This gives the correctly filtered luminance value */
3446     shader_addline(buffer, "TEX luminance, fragment.texcoord[0], texture[0], %s;\n", tex);
3447
3448     return TRUE;
3449 }
3450
3451 static BOOL gen_yv12_read(SHADER_BUFFER *buffer, GLenum textype, char *luminance)
3452 {
3453     const char *tex;
3454
3455     switch(textype) {
3456         case GL_TEXTURE_2D:             tex = "2D";     break;
3457         case GL_TEXTURE_RECTANGLE_ARB:  tex = "RECT";   break;
3458         default:
3459             FIXME("Implement yv12 correction for non-2d, non-rect textures\n");
3460             return FALSE;
3461     }
3462
3463     /* YV12 surfaces contain a WxH sized luminance plane, followed by a (W/2)x(H/2)
3464      * V and a (W/2)x(H/2) U plane, each with 8 bit per pixel. So the effective
3465      * bitdepth is 12 bits per pixel. Since the U and V planes have only half the
3466      * pitch of the luminance plane, the packing into the gl texture is a bit
3467      * unfortunate. If the whole texture is interpreted as luminance data it looks
3468      * approximately like this:
3469      *
3470      *        +----------------------------------+----
3471      *        |                                  |
3472      *        |                                  |
3473      *        |                                  |
3474      *        |                                  |
3475      *        |                                  |   2
3476      *        |            LUMINANCE             |   -
3477      *        |                                  |   3
3478      *        |                                  |
3479      *        |                                  |
3480      *        |                                  |
3481      *        |                                  |
3482      *        +----------------+-----------------+----
3483      *        |                |                 |
3484      *        |  U even rows   |  U odd rows     |
3485      *        |                |                 |   1
3486      *        +----------------+------------------   -
3487      *        |                |                 |   3
3488      *        |  V even rows   |  V odd rows     |
3489      *        |                |                 |
3490      *        +----------------+-----------------+----
3491      *        |                |                 |
3492      *        |     0.5        |       0.5       |
3493      *
3494      * So it appears as if there are 4 chroma images, but in fact the odd rows
3495      * in the chroma images are in the same row as the even ones. So its is
3496      * kinda tricky to read
3497      *
3498      * When reading from rectangle textures, keep in mind that the input y coordinates
3499      * go from 0 to d3d_height, whereas the opengl texture height is 1.5 * d3d_height
3500      */
3501     shader_addline(buffer, "PARAM yv12_coef = {%f, %f, %f, %f};\n",
3502                    2.0 / 3.0, 1.0 / 6.0, (2.0 / 3.0) + (1.0 / 6.0), 1.0 / 3.0);
3503
3504     shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3505     /* the chroma planes have only half the width */
3506     shader_addline(buffer, "MUL texcrd.x, texcrd.x, coef.y;\n");
3507
3508     /* The first value is between 2/3 and 5/6th of the texture's height, so scale+bias
3509      * the coordinate. Also read the right side of the image when reading odd lines
3510      *
3511      * Don't forget to clamp the y values in into the range, otherwise we'll get filtering
3512      * bleeding
3513      */
3514     if(textype == GL_TEXTURE_2D) {
3515
3516         shader_addline(buffer, "RCP chroma.w, size.y;\n");
3517
3518         shader_addline(buffer, "MUL texcrd2.y, texcrd.y, size.y;\n");
3519
3520         shader_addline(buffer, "FLR texcrd2.y, texcrd2.y;\n");
3521         shader_addline(buffer, "MAD texcrd.y, texcrd.y, yv12_coef.y, yv12_coef.x;\n");
3522
3523         /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3524         shader_addline(buffer, "ADD texcrd2.x, texcrd2.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3525         shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3526         shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3527         shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3528
3529         /* clamp, keep the half pixel origin in mind */
3530         shader_addline(buffer, "MAD temp.y, coef.y, chroma.w, yv12_coef.x;\n");
3531         shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3532         shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.z;\n");
3533         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3534     } else {
3535         /* Read from [size - size+size/4] */
3536         shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3537         shader_addline(buffer, "MAD texcrd.y, texcrd.y, coef.w, size.y;\n");
3538
3539         /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3540         shader_addline(buffer, "ADD texcrd2.x, texcrd.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3541         shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3542         shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3543         shader_addline(buffer, "MUL texcrd2.x, texcrd2.x, size.x;\n");
3544         shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3545
3546         /* Make sure to read exactly from the pixel center */
3547         shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3548         shader_addline(buffer, "ADD texcrd.y, texcrd.y, coef.y;\n");
3549
3550         /* Clamp */
3551         shader_addline(buffer, "MAD temp.y, size.y, coef.w, size.y;\n");
3552         shader_addline(buffer, "ADD temp.y, temp.y, -coef.y;\n");
3553         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3554         shader_addline(buffer, "ADD temp.y, size.y, -coef.y;\n");
3555         shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3556     }
3557     /* Read the texture, put the result into the output register */
3558     shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3559     shader_addline(buffer, "MOV chroma.x, temp.w;\n");
3560
3561     /* The other chroma value is 1/6th of the texture lower, from 5/6th to 6/6th
3562      * No need to clamp because we're just reusing the already clamped value from above
3563      */
3564     if(textype == GL_TEXTURE_2D) {
3565         shader_addline(buffer, "ADD texcrd.y, texcrd.y, yv12_coef.y;\n");
3566     } else {
3567         shader_addline(buffer, "MAD texcrd.y, size.y, coef.w, texcrd.y;\n");
3568     }
3569     shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3570     shader_addline(buffer, "MOV chroma.y, temp.w;\n");
3571
3572     /* Sample the luminance value. It is in the top 2/3rd of the texture, so scale the y coordinate.
3573      * Clamp the y coordinate to prevent the chroma values from bleeding into the sampled luminance
3574      * values due to filtering
3575      */
3576     shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3577     if(textype == GL_TEXTURE_2D) {
3578         /* Multiply the y coordinate by 2/3 and clamp it */
3579         shader_addline(buffer, "MUL texcrd.y, texcrd.y, yv12_coef.x;\n");
3580         shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.x;\n");
3581         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3582         shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3583     } else {
3584         /* Reading from texture_rectangles is pretty straightforward, just use the unmodified
3585          * texture coordinate. It is still a good idea to clamp it though, since the opengl texture
3586          * is bigger
3587          */
3588         shader_addline(buffer, "ADD temp.x, size.y, -coef.y;\n");
3589         shader_addline(buffer, "MIN texcrd.y, texcrd.y, size.x;\n");
3590         shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3591     }
3592     *luminance = 'a';
3593
3594     return TRUE;
3595 }
3596
3597 static GLuint gen_yuv_shader(IWineD3DDeviceImpl *device, enum yuv_fixup yuv_fixup, GLenum textype)
3598 {
3599     GLenum shader;
3600     SHADER_BUFFER buffer;
3601     char luminance_component;
3602     struct arbfp_blit_priv *priv = device->blit_priv;
3603
3604     /* Shader header */
3605     shader_buffer_init(&buffer);
3606
3607     ENTER_GL();
3608     GL_EXTCALL(glGenProgramsARB(1, &shader));
3609     checkGLcall("GL_EXTCALL(glGenProgramsARB(1, &shader))");
3610     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3611     checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3612     LEAVE_GL();
3613     if(!shader) {
3614         shader_buffer_free(&buffer);
3615         return 0;
3616     }
3617
3618     /* The YUY2 and UYVY formats contain two pixels packed into a 32 bit macropixel,
3619      * giving effectively 16 bit per pixel. The color consists of a luminance(Y) and
3620      * two chroma(U and V) values. Each macropixel has two luminance values, one for
3621      * each single pixel it contains, and one U and one V value shared between both
3622      * pixels.
3623      *
3624      * The data is loaded into an A8L8 texture. With YUY2, the luminance component
3625      * contains the luminance and alpha the chroma. With UYVY it is vice versa. Thus
3626      * take the format into account when generating the read swizzles
3627      *
3628      * Reading the Y value is straightforward - just sample the texture. The hardware
3629      * takes care of filtering in the horizontal and vertical direction.
3630      *
3631      * Reading the U and V values is harder. We have to avoid filtering horizontally,
3632      * because that would mix the U and V values of one pixel or two adjacent pixels.
3633      * Thus floor the texture coordinate and add 0.5 to get an unfiltered read,
3634      * regardless of the filtering setting. Vertical filtering works automatically
3635      * though - the U and V values of two rows are mixed nicely.
3636      *
3637      * Appart of avoiding filtering issues, the code has to know which value it just
3638      * read, and where it can find the other one. To determine this, it checks if
3639      * it sampled an even or odd pixel, and shifts the 2nd read accordingly.
3640      *
3641      * Handling horizontal filtering of U and V values requires reading a 2nd pair
3642      * of pixels, extracting U and V and mixing them. This is not implemented yet.
3643      *
3644      * An alternative implementation idea is to load the texture as A8R8G8B8 texture,
3645      * with width / 2. This way one read gives all 3 values, finding U and V is easy
3646      * in an unfiltered situation. Finding the luminance on the other hand requires
3647      * finding out if it is an odd or even pixel. The real drawback of this approach
3648      * is filtering. This would have to be emulated completely in the shader, reading
3649      * up two 2 packed pixels in up to 2 rows and interpolating both horizontally and
3650      * vertically. Beyond that it would require adjustments to the texture handling
3651      * code to deal with the width scaling
3652      */
3653     shader_addline(&buffer, "!!ARBfp1.0\n");
3654     shader_addline(&buffer, "TEMP luminance;\n");
3655     shader_addline(&buffer, "TEMP temp;\n");
3656     shader_addline(&buffer, "TEMP chroma;\n");
3657     shader_addline(&buffer, "TEMP texcrd;\n");
3658     shader_addline(&buffer, "TEMP texcrd2;\n");
3659     shader_addline(&buffer, "PARAM coef = {1.0, 0.5, 2.0, 0.25};\n");
3660     shader_addline(&buffer, "PARAM yuv_coef = {1.403, 0.344, 0.714, 1.770};\n");
3661     shader_addline(&buffer, "PARAM size = program.local[0];\n");
3662
3663     switch (yuv_fixup)
3664     {
3665         case YUV_FIXUP_UYVY:
3666         case YUV_FIXUP_YUY2:
3667             if (!gen_planar_yuv_read(&buffer, yuv_fixup, textype, &luminance_component))
3668             {
3669                 shader_buffer_free(&buffer);
3670                 return 0;
3671             }
3672             break;
3673
3674         case YUV_FIXUP_YV12:
3675             if (!gen_yv12_read(&buffer, textype, &luminance_component))
3676             {
3677                 shader_buffer_free(&buffer);
3678                 return 0;
3679             }
3680             break;
3681
3682         default:
3683             FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3684             shader_buffer_free(&buffer);
3685             return 0;
3686     }
3687
3688     /* Calculate the final result. Formula is taken from
3689      * http://www.fourcc.org/fccyvrgb.php. Note that the chroma
3690      * ranges from -0.5 to 0.5
3691      */
3692     shader_addline(&buffer, "SUB chroma.xy, chroma, coef.y;\n");
3693
3694     shader_addline(&buffer, "MAD result.color.x, chroma.x, yuv_coef.x, luminance.%c;\n", luminance_component);
3695     shader_addline(&buffer, "MAD temp.x, -chroma.y, yuv_coef.y, luminance.%c;\n", luminance_component);
3696     shader_addline(&buffer, "MAD result.color.y, -chroma.x, yuv_coef.z, temp.x;\n");
3697     shader_addline(&buffer, "MAD result.color.z, chroma.y, yuv_coef.w, luminance.%c;\n", luminance_component);
3698     shader_addline(&buffer, "END\n");
3699
3700     ENTER_GL();
3701     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3702
3703     if (glGetError() == GL_INVALID_OPERATION) {
3704         GLint pos;
3705         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3706         FIXME("Fragment program error at position %d: %s\n", pos,
3707               debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3708     }
3709     shader_buffer_free(&buffer);
3710     LEAVE_GL();
3711
3712     switch (yuv_fixup)
3713     {
3714         case YUV_FIXUP_YUY2:
3715             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yuy2_rect_shader = shader;
3716             else priv->yuy2_2d_shader = shader;
3717             break;
3718
3719         case YUV_FIXUP_UYVY:
3720             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->uyvy_rect_shader = shader;
3721             else priv->uyvy_2d_shader = shader;
3722             break;
3723
3724         case YUV_FIXUP_YV12:
3725             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yv12_rect_shader = shader;
3726             else priv->yv12_2d_shader = shader;
3727             break;
3728     }
3729
3730     return shader;
3731 }
3732
3733 static HRESULT arbfp_blit_set(IWineD3DDevice *iface, const struct GlPixelFormatDesc *format_desc,
3734         GLenum textype, UINT width, UINT height)
3735 {
3736     GLenum shader;
3737     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3738     float size[4] = {width, height, 1, 1};
3739     struct arbfp_blit_priv *priv = device->blit_priv;
3740     enum yuv_fixup yuv_fixup;
3741
3742     if (!is_yuv_fixup(format_desc->color_fixup))
3743     {
3744         TRACE("Fixup:\n");
3745         dump_color_fixup_desc(format_desc->color_fixup);
3746         /* Don't bother setting up a shader for unconverted formats */
3747         ENTER_GL();
3748         glEnable(textype);
3749         checkGLcall("glEnable(textype)");
3750         LEAVE_GL();
3751         return WINED3D_OK;
3752     }
3753
3754     yuv_fixup = get_yuv_fixup(format_desc->color_fixup);
3755
3756     switch(yuv_fixup)
3757     {
3758         case YUV_FIXUP_YUY2:
3759             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yuy2_rect_shader : priv->yuy2_2d_shader;
3760             break;
3761
3762         case YUV_FIXUP_UYVY:
3763             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->uyvy_rect_shader : priv->uyvy_2d_shader;
3764             break;
3765
3766         case YUV_FIXUP_YV12:
3767             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yv12_rect_shader : priv->yv12_2d_shader;
3768             break;
3769
3770         default:
3771             FIXME("Unsupported YUV fixup %#x, not setting a shader\n", yuv_fixup);
3772             ENTER_GL();
3773             glEnable(textype);
3774             checkGLcall("glEnable(textype)");
3775             LEAVE_GL();
3776             return E_NOTIMPL;
3777     }
3778
3779     if (!shader) shader = gen_yuv_shader(device, yuv_fixup, textype);
3780
3781     ENTER_GL();
3782     glEnable(GL_FRAGMENT_PROGRAM_ARB);
3783     checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
3784     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3785     checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3786     GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, 0, size));
3787     checkGLcall("glProgramLocalParameter4fvARB");
3788     LEAVE_GL();
3789
3790     return WINED3D_OK;
3791 }
3792
3793 static void arbfp_blit_unset(IWineD3DDevice *iface) {
3794     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3795
3796     ENTER_GL();
3797     glDisable(GL_FRAGMENT_PROGRAM_ARB);
3798     checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
3799     glDisable(GL_TEXTURE_2D);
3800     checkGLcall("glDisable(GL_TEXTURE_2D)");
3801     if(GL_SUPPORT(ARB_TEXTURE_CUBE_MAP)) {
3802         glDisable(GL_TEXTURE_CUBE_MAP_ARB);
3803         checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
3804     }
3805     if(GL_SUPPORT(ARB_TEXTURE_RECTANGLE)) {
3806         glDisable(GL_TEXTURE_RECTANGLE_ARB);
3807         checkGLcall("glDisable(GL_TEXTURE_RECTANGLE_ARB)");
3808     }
3809     LEAVE_GL();
3810 }
3811
3812 static BOOL arbfp_blit_color_fixup_supported(struct color_fixup_desc fixup)
3813 {
3814     enum yuv_fixup yuv_fixup;
3815
3816     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
3817     {
3818         TRACE("Checking support for fixup:\n");
3819         dump_color_fixup_desc(fixup);
3820     }
3821
3822     if (is_identity_fixup(fixup))
3823     {
3824         TRACE("[OK]\n");
3825         return TRUE;
3826     }
3827
3828     /* We only support YUV conversions. */
3829     if (!is_yuv_fixup(fixup))
3830     {
3831         TRACE("[FAILED]\n");
3832         return FALSE;
3833     }
3834
3835     yuv_fixup = get_yuv_fixup(fixup);
3836     switch(yuv_fixup)
3837     {
3838         case YUV_FIXUP_YUY2:
3839         case YUV_FIXUP_UYVY:
3840         case YUV_FIXUP_YV12:
3841             TRACE("[OK]\n");
3842             return TRUE;
3843
3844         default:
3845             FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3846             TRACE("[FAILED]\n");
3847             return FALSE;
3848     }
3849 }
3850
3851 const struct blit_shader arbfp_blit = {
3852     arbfp_blit_alloc,
3853     arbfp_blit_free,
3854     arbfp_blit_set,
3855     arbfp_blit_unset,
3856     arbfp_blit_color_fixup_supported,
3857 };
3858
3859 #undef GLINFO_LOCATION