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

   1 /*
   2  * Pixel and vertex shaders implementation using ARB_vertex_program
   3  * and ARB_fragment_program GL extensions.
   4  *
   5  * Copyright 2002-2003 Jason Edmeades
   6  * Copyright 2002-2003 Raphael Junqueira
   7  * Copyright 2004 Christian Costa
   8  * Copyright 2005 Oliver Stieber
   9  * Copyright 2006 Ivan Gyurdiev
  10  * Copyright 2006 Jason Green
  11  * Copyright 2006 Henri Verbeet
  12  * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
  13  * Copyright 2009 Henri Verbeet for CodeWeavers
  14  *
  15  * This library is free software; you can redistribute it and/or
  16  * modify it under the terms of the GNU Lesser General Public
  17  * License as published by the Free Software Foundation; either
  18  * version 2.1 of the License, or (at your option) any later version.
  19  *
  20  * This library is distributed in the hope that it will be useful,
  21  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  22  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  23  * Lesser General Public License for more details.
  24  *
  25  * You should have received a copy of the GNU Lesser General Public
  26  * License along with this library; if not, write to the Free Software
  27  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  28  */
  29
  30 #include "config.h"
  31
  32 #include <math.h>
  33 #include <stdio.h>
  34
  35 #include "wined3d_private.h"
  36
  37 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  38 WINE_DECLARE_DEBUG_CHANNEL(d3d_constants);
  39 WINE_DECLARE_DEBUG_CHANNEL(d3d_caps);
  40 WINE_DECLARE_DEBUG_CHANNEL(d3d);
  41
  42 #define GLINFO_LOCATION      (*gl_info)
  43
  44 /* We have to subtract any other PARAMs that we might use in our shader programs.
  45  * ATI seems to count 2 implicit PARAMs when we use fog and NVIDIA counts 1,
  46  * and we reference one row of the PROJECTION matrix which counts as 1 PARAM. */
  47 #define ARB_SHADER_RESERVED_VS_CONSTS 3
  48
  49 /* The arb shader only loads the bump mapping environment matrix into the shader if it finds
  50  * a free constant to do that, so only reduce the number of available constants by 2 for the fog states.
  51  */
  52 #define ARB_SHADER_RESERVED_PS_CONSTS 2
  53
  54 /* Internally used shader constants. Applications can use constants 0 to GL_LIMITS(vshader_constantsF) - 1,
  55  * so upload them above that
  56  */
  57 #define ARB_SHADER_PRIVCONST_BASE (GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS)
  58 #define ARB_SHADER_PRIVCONST_POS ARB_SHADER_PRIVCONST_BASE + 0
  59
  60 /* ARB_program_shader private data */
  61 struct shader_arb_priv {
  62     GLuint                  current_vprogram_id;
  63     GLuint                  current_fprogram_id;
  64     GLuint                  depth_blt_vprogram_id;
  65     GLuint                  depth_blt_fprogram_id[tex_type_count];
  66     BOOL                    use_arbfp_fixed_func;
  67     struct hash_table_t     *fragment_shaders;
  68 };
  69
  70 /********************************************************
  71  * ARB_[vertex/fragment]_program helper functions follow
  72  ********************************************************/
  73
  74 /**
  75  * Loads floating point constants into the currently set ARB_vertex/fragment_program.
  76  * When constant_list == NULL, it will load all the constants.
  77  *
  78  * @target_type should be either GL_VERTEX_PROGRAM_ARB (for vertex shaders)
  79  *  or GL_FRAGMENT_PROGRAM_ARB (for pixel shaders)
  80  */
  81 static unsigned int shader_arb_load_constantsF(IWineD3DBaseShaderImpl* This, const WineD3D_GL_Info *gl_info,
  82         GLuint target_type, unsigned int max_constants, const float *constants, char *dirty_consts)
  83 {
  84     local_constant* lconst;
  85     DWORD i, j;
  86     unsigned int ret;
  87
  88     if (TRACE_ON(d3d_shader)) {
  89         for(i = 0; i < max_constants; i++) {
  90             if(!dirty_consts[i]) continue;
  91             TRACE_(d3d_constants)("Loading constants %i: %f, %f, %f, %f\n", i,
  92                         constants[i * 4 + 0], constants[i * 4 + 1],
  93                         constants[i * 4 + 2], constants[i * 4 + 3]);
  94         }
  95     }
  96     /* In 1.X pixel shaders constants are implicitly clamped in the range [-1;1] */
  97     if (target_type == GL_FRAGMENT_PROGRAM_ARB && This->baseShader.reg_maps.shader_version.major == 1)
  98     {
  99         float lcl_const[4];
 100         for(i = 0; i < max_constants; i++) {
 101             if(!dirty_consts[i]) continue;
 102             dirty_consts[i] = 0;
 103
 104             j = 4 * i;
 105             if(constants[j + 0] > 1.0) lcl_const[0] = 1.0;
 106             else if(constants[j + 0] < -1.0) lcl_const[0] = -1.0;
 107             else lcl_const[0] = constants[j + 0];
 108
 109             if(constants[j + 1] > 1.0) lcl_const[1] = 1.0;
 110             else if(constants[j + 1] < -1.0) lcl_const[1] = -1.0;
 111             else lcl_const[1] = constants[j + 1];
 112
 113             if(constants[j + 2] > 1.0) lcl_const[2] = 1.0;
 114             else if(constants[j + 2] < -1.0) lcl_const[2] = -1.0;
 115             else lcl_const[2] = constants[j + 2];
 116
 117             if(constants[j + 3] > 1.0) lcl_const[3] = 1.0;
 118             else if(constants[j + 3] < -1.0) lcl_const[3] = -1.0;
 119             else lcl_const[3] = constants[j + 3];
 120
 121             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, lcl_const));
 122         }
 123     } else {
 124         if(GL_SUPPORT(EXT_GPU_PROGRAM_PARAMETERS)) {
 125             /* TODO: Benchmark if we're better of with finding the dirty constants ourselves,
 126              * or just reloading *all* constants at once
 127              *
 128             GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, 0, max_constants, constants));
 129              */
 130             for(i = 0; i < max_constants; i++) {
 131                 if(!dirty_consts[i]) continue;
 132
 133                 /* Find the next block of dirty constants */
 134                 dirty_consts[i] = 0;
 135                 j = i;
 136                 for(i++; (i < max_constants) && dirty_consts[i]; i++) {
 137                     dirty_consts[i] = 0;
 138                 }
 139
 140                 GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, j, i - j, constants + (j * 4)));
 141             }
 142         } else {
 143             for(i = 0; i < max_constants; i++) {
 144                 if(dirty_consts[i]) {
 145                     dirty_consts[i] = 0;
 146                     GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, constants + (i * 4)));
 147                 }
 148             }
 149         }
 150     }
 151     checkGLcall("glProgramEnvParameter4fvARB()");
 152
 153     /* Load immediate constants */
 154     if(This->baseShader.load_local_constsF) {
 155         if (TRACE_ON(d3d_shader)) {
 156             LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 157                 GLfloat* values = (GLfloat*)lconst->value;
 158                 TRACE_(d3d_constants)("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
 159                         values[0], values[1], values[2], values[3]);
 160             }
 161         }
 162         /* Immediate constants are clamped for 1.X shaders at loading times */
 163         ret = 0;
 164         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 165             dirty_consts[lconst->idx] = 1; /* Dirtify so the non-immediate constant overwrites it next time */
 166             ret = max(ret, lconst->idx + 1);
 167             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, lconst->idx, (GLfloat*)lconst->value));
 168         }
 169         checkGLcall("glProgramEnvParameter4fvARB()");
 170         return ret; /* The loaded immediate constants need reloading for the next shader */
 171     } else {
 172         return 0; /* No constants are dirty now */
 173     }
 174 }
 175
 176 /**
 177  * Loads the texture dimensions for NP2 fixup into the currently set ARB_[vertex/fragment]_programs.
 178  */
 179 static void shader_arb_load_np2fixup_constants(
 180     IWineD3DDevice* device,
 181     char usePixelShader,
 182     char useVertexShader) {
 183     /* not implemented */
 184 }
 185
 186 /**
 187  * Loads the app-supplied constants into the currently set ARB_[vertex/fragment]_programs.
 188  *
 189  * We only support float constants in ARB at the moment, so don't
 190  * worry about the Integers or Booleans
 191  */
 192 static void shader_arb_load_constants(
 193     IWineD3DDevice* device,
 194     char usePixelShader,
 195     char useVertexShader) {
 196
 197     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
 198     IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
 199     const WineD3D_GL_Info *gl_info = &deviceImpl->adapter->gl_info;
 200     unsigned char i;
 201
 202     if (useVertexShader) {
 203         IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
 204
 205         /* Load DirectX 9 float constants for vertex shader */
 206         deviceImpl->highest_dirty_vs_const = shader_arb_load_constantsF(
 207                 vshader, gl_info, GL_VERTEX_PROGRAM_ARB,
 208                 deviceImpl->highest_dirty_vs_const,
 209                 stateBlock->vertexShaderConstantF,
 210                 deviceImpl->activeContext->vshader_const_dirty);
 211
 212         /* Upload the position fixup */
 213         GL_EXTCALL(glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, ARB_SHADER_PRIVCONST_POS, deviceImpl->posFixup));
 214     }
 215
 216     if (usePixelShader) {
 217
 218         IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
 219         IWineD3DPixelShaderImpl *psi = (IWineD3DPixelShaderImpl *) pshader;
 220
 221         /* Load DirectX 9 float constants for pixel shader */
 222         deviceImpl->highest_dirty_ps_const = shader_arb_load_constantsF(
 223                 pshader, gl_info, GL_FRAGMENT_PROGRAM_ARB,
 224                 deviceImpl->highest_dirty_ps_const,
 225                 stateBlock->pixelShaderConstantF,
 226                 deviceImpl->activeContext->pshader_const_dirty);
 227
 228         for(i = 0; i < psi->numbumpenvmatconsts; i++) {
 229             /* The state manager takes care that this function is always called if the bump env matrix changes
 230              */
 231             const float *data = (const float *)&stateBlock->textureState[(int) psi->bumpenvmatconst[i].texunit][WINED3DTSS_BUMPENVMAT00];
 232             GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->bumpenvmatconst[i].const_num, data));
 233             deviceImpl->activeContext->pshader_const_dirty[psi->bumpenvmatconst[i].const_num] = 1;
 234
 235             if (psi->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED)
 236             {
 237                 /* WINED3DTSS_BUMPENVLSCALE and WINED3DTSS_BUMPENVLOFFSET are next to each other.
 238                  * point gl to the scale, and load 4 floats. x = scale, y = offset, z and w are junk, we
 239                  * don't care about them. The pointers are valid for sure because the stateblock is bigger.
 240                  * (they're WINED3DTSS_TEXTURETRANSFORMFLAGS and WINED3DTSS_ADDRESSW, so most likely 0 or NaN
 241                  */
 242                 const float *scale = (const float *)&stateBlock->textureState[(int) psi->luminanceconst[i].texunit][WINED3DTSS_BUMPENVLSCALE];
 243                 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->luminanceconst[i].const_num, scale));
 244                 deviceImpl->activeContext->pshader_const_dirty[psi->luminanceconst[i].const_num] = 1;
 245             }
 246         }
 247     }
 248 }
 249
 250 static void shader_arb_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count)
 251 {
 252     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 253
 254     /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
 255      * context. On a context switch the old context will be fully dirtified */
 256     memset(This->activeContext->vshader_const_dirty + start, 1,
 257             sizeof(*This->activeContext->vshader_const_dirty) * count);
 258     This->highest_dirty_vs_const = max(This->highest_dirty_vs_const, start + count + 1);
 259 }
 260
 261 static void shader_arb_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count)
 262 {
 263     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 264
 265     /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
 266      * context. On a context switch the old context will be fully dirtified */
 267     memset(This->activeContext->pshader_const_dirty + start, 1,
 268             sizeof(*This->activeContext->pshader_const_dirty) * count);
 269     This->highest_dirty_ps_const = max(This->highest_dirty_ps_const, start + count + 1);
 270 }
 271
 272 static DWORD *local_const_mapping(IWineD3DBaseShaderImpl *This)
 273 {
 274     DWORD *ret;
 275     DWORD idx = 0;
 276     const local_constant *lconst;
 277
 278     if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) return NULL;
 279
 280     ret = HeapAlloc(GetProcessHeap(), 0, sizeof(DWORD) * This->baseShader.limits.temporary);
 281     if(!ret) {
 282         ERR("Out of memory\n");
 283         return NULL;
 284     }
 285
 286     LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 287         ret[lconst->idx] = idx++;
 288     }
 289     return ret;
 290 }
 291
 292 /* Generate the variable & register declarations for the ARB_vertex_program output target */
 293 static void shader_generate_arb_declarations(IWineD3DBaseShader *iface, const shader_reg_maps *reg_maps,
 294         SHADER_BUFFER *buffer, const WineD3D_GL_Info *gl_info, DWORD *lconst_map)
 295 {
 296     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 297     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
 298     DWORD i, cur;
 299     char pshader = shader_is_pshader_version(reg_maps->shader_version.type);
 300     unsigned max_constantsF = min(This->baseShader.limits.constant_float,
 301             (pshader ? GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS :
 302                        GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS));
 303     UINT extra_constants_needed = 0;
 304     const local_constant *lconst;
 305
 306     /* Temporary Output register */
 307     shader_addline(buffer, "TEMP TMP_OUT;\n");
 308
 309     for(i = 0; i < This->baseShader.limits.temporary; i++) {
 310         if (reg_maps->temporary[i])
 311             shader_addline(buffer, "TEMP R%u;\n", i);
 312     }
 313
 314     for (i = 0; i < This->baseShader.limits.address; i++) {
 315         if (reg_maps->address[i])
 316             shader_addline(buffer, "ADDRESS A%d;\n", i);
 317     }
 318
 319     if(pshader && reg_maps->shader_version.major == 1 && reg_maps->shader_version.minor <= 3) {
 320         for(i = 0; i < This->baseShader.limits.texcoord; i++) {
 321             if (reg_maps->texcoord[i] && pshader)
 322                 shader_addline(buffer,"TEMP T%u;\n", i);
 323         }
 324     }
 325
 326     for(i = 0; i < (sizeof(reg_maps->bumpmat) / sizeof(reg_maps->bumpmat[0])); i++) {
 327         IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *) This;
 328         if(!reg_maps->bumpmat[i]) continue;
 329
 330         cur = ps->numbumpenvmatconsts;
 331         ps->bumpenvmatconst[cur].const_num = -1;
 332         ps->bumpenvmatconst[cur].texunit = i;
 333         ps->luminanceconst[cur].const_num = -1;
 334         ps->luminanceconst[cur].texunit = i;
 335
 336         /* If the shader does not use all available constants, use the next free constant to load the bump mapping environment matrix from
 337          * the stateblock into the shader. If no constant is available don't load, texbem will then just sample the texture without applying
 338          * bump mapping.
 339          */
 340         if(max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS) {
 341             ps->bumpenvmatconst[cur].const_num = max_constantsF + extra_constants_needed;
 342             shader_addline(buffer, "PARAM bumpenvmat%d = program.env[%d];\n",
 343                            i, ps->bumpenvmatconst[cur].const_num);
 344             extra_constants_needed++;
 345
 346             if(reg_maps->luminanceparams && max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS) {
 347                 ((IWineD3DPixelShaderImpl *)This)->luminanceconst[cur].const_num = max_constantsF + extra_constants_needed;
 348                 shader_addline(buffer, "PARAM luminance%d = program.env[%d];\n",
 349                                i, ps->luminanceconst[cur].const_num);
 350                 extra_constants_needed++;
 351             } else if(reg_maps->luminanceparams) {
 352                 FIXME("No free constant to load the luminance parameters\n");
 353             }
 354         } else {
 355             FIXME("No free constant found to load environment bump mapping matrix into the shader. texbem instruction will not apply bump mapping\n");
 356         }
 357
 358         ps->numbumpenvmatconsts = cur + 1;
 359     }
 360
 361     if(device->stateBlock->renderState[WINED3DRS_SRGBWRITEENABLE] && pshader) {
 362         shader_addline(buffer, "PARAM srgb_consts1 = {%f, %f, %f, %f};\n",
 363                        srgb_mul_low, srgb_cmp, srgb_pow, srgb_mul_high);
 364         shader_addline(buffer, "PARAM srgb_consts2 = {%f, %f, %f, %f};\n",
 365                        srgb_sub_high, 0.0, 0.0, 0.0);
 366     }
 367
 368     /* Load local constants using the program-local space,
 369      * this avoids reloading them each time the shader is used
 370      */
 371     if(lconst_map) {
 372         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 373             shader_addline(buffer, "PARAM C%u = program.local[%u];\n", lconst->idx,
 374                            lconst_map[lconst->idx]);
 375         }
 376     }
 377
 378     /* we use the array-based constants array if the local constants are marked for loading,
 379      * because then we use indirect addressing, or when the local constant list is empty,
 380      * because then we don't know if we're using indirect addressing or not. If we're hardcoding
 381      * local constants do not declare the loaded constants as an array because ARB compilers usually
 382      * do not optimize unused constants away
 383      */
 384     if(This->baseShader.reg_maps.usesrelconstF) {
 385         /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
 386         shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
 387                     max_constantsF, max_constantsF - 1);
 388     } else {
 389         for(i = 0; i < max_constantsF; i++) {
 390             DWORD idx, mask;
 391             idx = i >> 5;
 392             mask = 1 << (i & 0x1f);
 393             if(!shader_constant_is_local(This, i) && (This->baseShader.reg_maps.constf[idx] & mask)) {
 394                 shader_addline(buffer, "PARAM C%d = program.env[%d];\n",i, i);
 395             }
 396         }
 397     }
 398 }
 399
 400 static const char * const shift_tab[] = {
 401     "dummy",     /*  0 (none) */
 402     "coefmul.x", /*  1 (x2)   */
 403     "coefmul.y", /*  2 (x4)   */
 404     "coefmul.z", /*  3 (x8)   */
 405     "coefmul.w", /*  4 (x16)  */
 406     "dummy",     /*  5 (x32)  */
 407     "dummy",     /*  6 (x64)  */
 408     "dummy",     /*  7 (x128) */
 409     "dummy",     /*  8 (d256) */
 410     "dummy",     /*  9 (d128) */
 411     "dummy",     /* 10 (d64)  */
 412     "dummy",     /* 11 (d32)  */
 413     "coefdiv.w", /* 12 (d16)  */
 414     "coefdiv.z", /* 13 (d8)   */
 415     "coefdiv.y", /* 14 (d4)   */
 416     "coefdiv.x"  /* 15 (d2)   */
 417 };
 418
 419 static void shader_arb_get_write_mask(const struct wined3d_shader_instruction *ins,
 420         const struct wined3d_shader_dst_param *dst, char *write_mask)
 421 {
 422     char *ptr = write_mask;
 423     char vshader = shader_is_vshader_version(ins->ctx->reg_maps->shader_version.type);
 424
 425     if (vshader && dst->reg.type == WINED3DSPR_ADDR)
 426     {
 427         *ptr++ = '.';
 428         *ptr++ = 'x';
 429     }
 430     else if (dst->write_mask != WINED3DSP_WRITEMASK_ALL)
 431     {
 432         *ptr++ = '.';
 433         if (dst->write_mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
 434         if (dst->write_mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
 435         if (dst->write_mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
 436         if (dst->write_mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
 437     }
 438
 439     *ptr = '\0';
 440 }
 441
 442 static void shader_arb_get_swizzle(const struct wined3d_shader_src_param *param, BOOL fixup, char *swizzle_str)
 443 {
 444     /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
 445      * but addressed as "rgba". To fix this we need to swap the register's x
 446      * and z components. */
 447     const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
 448     char *ptr = swizzle_str;
 449
 450     /* swizzle bits fields: wwzzyyxx */
 451     DWORD swizzle = param->swizzle;
 452     DWORD swizzle_x = swizzle & 0x03;
 453     DWORD swizzle_y = (swizzle >> 2) & 0x03;
 454     DWORD swizzle_z = (swizzle >> 4) & 0x03;
 455     DWORD swizzle_w = (swizzle >> 6) & 0x03;
 456
 457     /* If the swizzle is the default swizzle (ie, "xyzw"), we don't need to
 458      * generate a swizzle string. Unless we need to our own swizzling. */
 459     if (swizzle != WINED3DSP_NOSWIZZLE || fixup)
 460     {
 461         *ptr++ = '.';
 462         if (swizzle_x == swizzle_y && swizzle_x == swizzle_z && swizzle_x == swizzle_w) {
 463             *ptr++ = swizzle_chars[swizzle_x];
 464         } else {
 465             *ptr++ = swizzle_chars[swizzle_x];
 466             *ptr++ = swizzle_chars[swizzle_y];
 467             *ptr++ = swizzle_chars[swizzle_z];
 468             *ptr++ = swizzle_chars[swizzle_w];
 469         }
 470     }
 471
 472     *ptr = '\0';
 473 }
 474
 475 static void shader_arb_get_register_name(IWineD3DBaseShader *iface,
 476         const struct wined3d_shader_register *reg, char *register_name, BOOL *is_color)
 477 {
 478     /* oPos, oFog and oPts in D3D */
 479     static const char * const rastout_reg_names[] = {"TMP_OUT", "result.fogcoord", "result.pointsize"};
 480     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
 481     BOOL pshader = shader_is_pshader_version(This->baseShader.reg_maps.shader_version.type);
 482
 483     *is_color = FALSE;
 484
 485     switch (reg->type)
 486     {
 487         case WINED3DSPR_TEMP:
 488             sprintf(register_name, "R%u", reg->idx);
 489             break;
 490
 491         case WINED3DSPR_INPUT:
 492             if (pshader)
 493             {
 494                 if (reg->idx == 0) strcpy(register_name, "fragment.color.primary");
 495                 else strcpy(register_name, "fragment.color.secondary");
 496             }
 497             else
 498             {
 499                 if (((IWineD3DVertexShaderImpl *)This)->cur_args->swizzle_map & (1 << reg->idx)) *is_color = TRUE;
 500                 sprintf(register_name, "vertex.attrib[%u]", reg->idx);
 501             }
 502             break;
 503
 504         case WINED3DSPR_CONST:
 505             if (!pshader && reg->rel_addr)
 506             {
 507                 UINT rel_offset = ((IWineD3DVertexShaderImpl *)This)->rel_offset;
 508                 if (reg->idx >= rel_offset)
 509                     sprintf(register_name, "C[A0.x + %u]", reg->idx - rel_offset);
 510                 else
 511                     sprintf(register_name, "C[A0.x - %u]", -reg->idx + rel_offset);
 512             }
 513             else
 514             {
 515                 if (This->baseShader.reg_maps.usesrelconstF)
 516                     sprintf(register_name, "C[%u]", reg->idx);
 517                 else
 518                     sprintf(register_name, "C%u", reg->idx);
 519             }
 520             break;
 521
 522         case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
 523             if (pshader) {
 524                 if(This->baseShader.reg_maps.shader_version.major == 1 &&
 525                    This->baseShader.reg_maps.shader_version.minor <= 3) {
 526                     /* In ps <= 1.3, Tx is a temporary register as destination to all instructions,
 527                      * and as source to most instructions. For some instructions it is the texcoord
 528                      * input. Those instructions know about the special use
 529                      */
 530                     sprintf(register_name, "T%u", reg->idx);
 531                 } else {
 532                     /* in ps 1.4 and 2.x Tx is always a (read-only) varying */
 533                     sprintf(register_name, "fragment.texcoord[%u]", reg->idx);
 534                 }
 535             }
 536             else  sprintf(register_name, "A%u", reg->idx);
 537             break;
 538
 539         case WINED3DSPR_COLOROUT:
 540             if (reg->idx == 0)
 541             {
 542                 strcpy(register_name, "TMP_COLOR");
 543             }
 544             else
 545             {
 546                 /* TODO: See GL_ARB_draw_buffers */
 547                 FIXME("Unsupported write to render target %u\n", reg->idx);
 548                 sprintf(register_name, "unsupported_register");
 549             }
 550             break;
 551
 552         case WINED3DSPR_RASTOUT:
 553             sprintf(register_name, "%s", rastout_reg_names[reg->idx]);
 554             break;
 555
 556         case WINED3DSPR_DEPTHOUT:
 557             strcpy(register_name, "result.depth");
 558             break;
 559
 560         case WINED3DSPR_ATTROUT:
 561             if (pshader) sprintf(register_name, "oD[%u]", reg->idx);
 562             else if (reg->idx == 0) strcpy(register_name, "result.color.primary");
 563             else strcpy(register_name, "result.color.secondary");
 564             break;
 565
 566         case WINED3DSPR_TEXCRDOUT:
 567             if (pshader) sprintf(register_name, "oT[%u]", reg->idx);
 568             else sprintf(register_name, "result.texcoord[%u]", reg->idx);
 569             break;
 570
 571         default:
 572             FIXME("Unhandled register type %#x[%u]\n", reg->type, reg->idx);
 573             sprintf(register_name, "unrecognized_register[%u]", reg->idx);
 574             break;
 575     }
 576 }
 577
 578 static void shader_arb_get_dst_param(const struct wined3d_shader_instruction *ins,
 579         const struct wined3d_shader_dst_param *wined3d_dst, char *str)
 580 {
 581     char register_name[255];
 582     char write_mask[6];
 583     BOOL is_color;
 584
 585     shader_arb_get_register_name(ins->ctx->shader, &wined3d_dst->reg, register_name, &is_color);
 586     strcpy(str, register_name);
 587
 588     shader_arb_get_write_mask(ins, wined3d_dst, write_mask);
 589     strcat(str, write_mask);
 590 }
 591
 592 static const char *shader_arb_get_fixup_swizzle(enum fixup_channel_source channel_source)
 593 {
 594     switch(channel_source)
 595     {
 596         case CHANNEL_SOURCE_ZERO: return "0";
 597         case CHANNEL_SOURCE_ONE: return "1";
 598         case CHANNEL_SOURCE_X: return "x";
 599         case CHANNEL_SOURCE_Y: return "y";
 600         case CHANNEL_SOURCE_Z: return "z";
 601         case CHANNEL_SOURCE_W: return "w";
 602         default:
 603             FIXME("Unhandled channel source %#x\n", channel_source);
 604             return "undefined";
 605     }
 606 }
 607
 608 static void gen_color_correction(SHADER_BUFFER *buffer, const char *reg, DWORD dst_mask,
 609                                  const char *one, const char *two, struct color_fixup_desc fixup)
 610 {
 611     DWORD mask;
 612
 613     if (is_yuv_fixup(fixup))
 614     {
 615         enum yuv_fixup yuv_fixup = get_yuv_fixup(fixup);
 616         FIXME("YUV fixup (%#x) not supported\n", yuv_fixup);
 617         return;
 618     }
 619
 620     mask = 0;
 621     if (fixup.x_source != CHANNEL_SOURCE_X) mask |= WINED3DSP_WRITEMASK_0;
 622     if (fixup.y_source != CHANNEL_SOURCE_Y) mask |= WINED3DSP_WRITEMASK_1;
 623     if (fixup.z_source != CHANNEL_SOURCE_Z) mask |= WINED3DSP_WRITEMASK_2;
 624     if (fixup.w_source != CHANNEL_SOURCE_W) mask |= WINED3DSP_WRITEMASK_3;
 625     mask &= dst_mask;
 626
 627     if (mask)
 628     {
 629         shader_addline(buffer, "SWZ %s, %s, %s, %s, %s, %s;\n", reg, reg,
 630                 shader_arb_get_fixup_swizzle(fixup.x_source), shader_arb_get_fixup_swizzle(fixup.y_source),
 631                 shader_arb_get_fixup_swizzle(fixup.z_source), shader_arb_get_fixup_swizzle(fixup.w_source));
 632     }
 633
 634     mask = 0;
 635     if (fixup.x_sign_fixup) mask |= WINED3DSP_WRITEMASK_0;
 636     if (fixup.y_sign_fixup) mask |= WINED3DSP_WRITEMASK_1;
 637     if (fixup.z_sign_fixup) mask |= WINED3DSP_WRITEMASK_2;
 638     if (fixup.w_sign_fixup) mask |= WINED3DSP_WRITEMASK_3;
 639     mask &= dst_mask;
 640
 641     if (mask)
 642     {
 643         char reg_mask[6];
 644         char *ptr = reg_mask;
 645
 646         if (mask != WINED3DSP_WRITEMASK_ALL)
 647         {
 648             *ptr++ = '.';
 649             if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
 650             if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
 651             if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
 652             if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
 653         }
 654         *ptr = '\0';
 655
 656         shader_addline(buffer, "MAD %s%s, %s, %s, -%s;\n", reg, reg_mask, reg, two, one);
 657     }
 658 }
 659
 660 static void shader_hw_sample(const struct wined3d_shader_instruction *ins, DWORD sampler_idx,
 661         const char *dst_str, const char *coord_reg, BOOL projected, BOOL bias)
 662 {
 663     SHADER_BUFFER *buffer = ins->ctx->buffer;
 664     DWORD sampler_type = ins->ctx->reg_maps->sampler_type[sampler_idx];
 665     const char *tex_type;
 666     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)ins->ctx->shader;
 667     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
 668
 669     switch(sampler_type) {
 670         case WINED3DSTT_1D:
 671             tex_type = "1D";
 672             break;
 673
 674         case WINED3DSTT_2D:
 675             if(device->stateBlock->textures[sampler_idx] &&
 676                IWineD3DBaseTexture_GetTextureDimensions(device->stateBlock->textures[sampler_idx]) == GL_TEXTURE_RECTANGLE_ARB) {
 677                 tex_type = "RECT";
 678             } else {
 679                 tex_type = "2D";
 680             }
 681             if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version.type))
 682             {
 683                const IWineD3DPixelShaderImpl* const ps = (const IWineD3DPixelShaderImpl*)This;
 684                if(ps->cur_args->np2_fixup & (1 << sampler_idx)) {
 685                    FIXME("NP2 texcoord fixup is currently not implemented in ARB mode (use GLSL instead).\n");
 686                }
 687             }
 688             break;
 689
 690         case WINED3DSTT_VOLUME:
 691             tex_type = "3D";
 692             break;
 693
 694         case WINED3DSTT_CUBE:
 695             tex_type = "CUBE";
 696             break;
 697
 698         default:
 699             ERR("Unexpected texture type %d\n", sampler_type);
 700             tex_type = "";
 701     }
 702
 703     if (bias) {
 704         /* Shouldn't be possible, but let's check for it */
 705         if(projected) FIXME("Biased and Projected texture sampling\n");
 706         /* TXB takes the 4th component of the source vector automatically, as d3d. Nothing more to do */
 707         shader_addline(buffer, "TXB %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 708     } else if (projected) {
 709         shader_addline(buffer, "TXP %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 710     } else {
 711         shader_addline(buffer, "TEX %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 712     }
 713
 714     if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version.type))
 715     {
 716         IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
 717         gen_color_correction(buffer, dst_str, ins->dst[0].write_mask,
 718                 "one", "coefmul.x", ps->cur_args->color_fixup[sampler_idx]);
 719     }
 720 }
 721
 722 static void shader_arb_get_src_param(const struct wined3d_shader_instruction *ins,
 723         const struct wined3d_shader_src_param *src, unsigned int tmpreg, char *outregstr)
 724 {
 725     /* Generate a line that does the input modifier computation and return the input register to use */
 726     BOOL is_color = FALSE;
 727     char regstr[256];
 728     char swzstr[20];
 729     int insert_line;
 730     SHADER_BUFFER *buffer = ins->ctx->buffer;
 731
 732     /* Assume a new line will be added */
 733     insert_line = 1;
 734
 735     /* Get register name */
 736     shader_arb_get_register_name(ins->ctx->shader, &src->reg, regstr, &is_color);
 737     shader_arb_get_swizzle(src, is_color, swzstr);
 738
 739     switch (src->modifiers)
 740     {
 741     case WINED3DSPSM_NONE:
 742         sprintf(outregstr, "%s%s", regstr, swzstr);
 743         insert_line = 0;
 744         break;
 745     case WINED3DSPSM_NEG:
 746         sprintf(outregstr, "-%s%s", regstr, swzstr);
 747         insert_line = 0;
 748         break;
 749     case WINED3DSPSM_BIAS:
 750         shader_addline(buffer, "ADD T%c, %s, -coefdiv.x;\n", 'A' + tmpreg, regstr);
 751         break;
 752     case WINED3DSPSM_BIASNEG:
 753         shader_addline(buffer, "ADD T%c, -%s, coefdiv.x;\n", 'A' + tmpreg, regstr);
 754         break;
 755     case WINED3DSPSM_SIGN:
 756         shader_addline(buffer, "MAD T%c, %s, coefmul.x, -one.x;\n", 'A' + tmpreg, regstr);
 757         break;
 758     case WINED3DSPSM_SIGNNEG:
 759         shader_addline(buffer, "MAD T%c, %s, -coefmul.x, one.x;\n", 'A' + tmpreg, regstr);
 760         break;
 761     case WINED3DSPSM_COMP:
 762         shader_addline(buffer, "SUB T%c, one.x, %s;\n", 'A' + tmpreg, regstr);
 763         break;
 764     case WINED3DSPSM_X2:
 765         shader_addline(buffer, "ADD T%c, %s, %s;\n", 'A' + tmpreg, regstr, regstr);
 766         break;
 767     case WINED3DSPSM_X2NEG:
 768         shader_addline(buffer, "ADD T%c, -%s, -%s;\n", 'A' + tmpreg, regstr, regstr);
 769         break;
 770     case WINED3DSPSM_DZ:
 771         shader_addline(buffer, "RCP T%c, %s.z;\n", 'A' + tmpreg, regstr);
 772         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 773         break;
 774     case WINED3DSPSM_DW:
 775         shader_addline(buffer, "RCP T%c, %s.w;\n", 'A' + tmpreg, regstr);
 776         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 777         break;
 778     default:
 779         sprintf(outregstr, "%s%s", regstr, swzstr);
 780         insert_line = 0;
 781     }
 782
 783     /* Return modified or original register, with swizzle */
 784     if (insert_line)
 785         sprintf(outregstr, "T%c%s", 'A' + tmpreg, swzstr);
 786 }
 787
 788 static const char *shader_arb_get_modifier(const struct wined3d_shader_instruction *ins)
 789 {
 790     DWORD mod;
 791     const char *ret = "";
 792     if (!ins->dst_count) return "";
 793
 794     mod = ins->dst[0].modifiers;
 795     if(mod & WINED3DSPDM_SATURATE) {
 796         ret = "_SAT";
 797         mod &= ~WINED3DSPDM_SATURATE;
 798     }
 799     if(mod & WINED3DSPDM_PARTIALPRECISION) {
 800         FIXME("Unhandled modifier WINED3DSPDM_PARTIALPRECISION\n");
 801         mod &= ~WINED3DSPDM_PARTIALPRECISION;
 802     }
 803     if(mod & WINED3DSPDM_MSAMPCENTROID) {
 804         FIXME("Unhandled modifier WINED3DSPDM_MSAMPCENTROID\n");
 805         mod &= ~WINED3DSPDM_MSAMPCENTROID;
 806     }
 807     if(mod) {
 808         FIXME("Unknown modifiers 0x%08x\n", mod);
 809     }
 810     return ret;
 811 }
 812
 813 static void pshader_hw_bem(const struct wined3d_shader_instruction *ins)
 814 {
 815     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
 816     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 817     SHADER_BUFFER *buffer = ins->ctx->buffer;
 818     char dst_name[50];
 819     char src_name[2][50];
 820     DWORD sampler_code = dst->reg.idx;
 821     BOOL has_bumpmat = FALSE;
 822     int i;
 823
 824     for(i = 0; i < This->numbumpenvmatconsts; i++) {
 825         if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
 826                 && This->bumpenvmatconst[i].texunit == sampler_code)
 827         {
 828             has_bumpmat = TRUE;
 829             break;
 830         }
 831     }
 832
 833     shader_arb_get_src_param(ins, &ins->src[0], 0, src_name[0]);
 834     shader_arb_get_src_param(ins, &ins->src[1], 1, src_name[1]);
 835
 836     if(has_bumpmat) {
 837         /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
 838         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", sampler_code);
 839         shader_addline(buffer, "DP3 TMP.r, TMP2, %s;\n", src_name[1]);
 840         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", sampler_code);
 841         shader_addline(buffer, "DP3 TMP.g, TMP2, %s;\n", src_name[1]);
 842
 843         shader_addline(buffer, "ADD %s, %s, TMP;\n", dst_name, src_name[0]);
 844     } else {
 845         shader_addline(buffer, "MOV %s, %s;\n", dst_name, src_name[0]);
 846     }
 847 }
 848
 849 static void pshader_hw_cnd(const struct wined3d_shader_instruction *ins)
 850 {
 851     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 852     SHADER_BUFFER *buffer = ins->ctx->buffer;
 853     char dst_name[50];
 854     char src_name[3][50];
 855     DWORD shader_version = WINED3D_SHADER_VERSION(ins->ctx->reg_maps->shader_version.major,
 856             ins->ctx->reg_maps->shader_version.minor);
 857
 858     shader_arb_get_dst_param(ins, dst, dst_name);
 859     shader_arb_get_src_param(ins, &ins->src[1], 1, src_name[1]);
 860
 861     /* The coissue flag changes the semantic of the cnd instruction in <= 1.3 shaders */
 862     if (shader_version <= WINED3D_SHADER_VERSION(1, 3) && ins->coissue)
 863     {
 864         shader_addline(buffer, "MOV%s %s, %s;\n", shader_arb_get_modifier(ins), dst_name, src_name[1]);
 865     } else {
 866         shader_arb_get_src_param(ins, &ins->src[0], 0, src_name[0]);
 867         shader_arb_get_src_param(ins, &ins->src[2], 2, src_name[2]);
 868         shader_addline(buffer, "ADD TMP, -%s, coefdiv.x;\n", src_name[0]);
 869         shader_addline(buffer, "CMP%s %s, TMP, %s, %s;\n",
 870                        shader_arb_get_modifier(ins), dst_name, src_name[1], src_name[2]);
 871     }
 872 }
 873
 874 static void pshader_hw_cmp(const struct wined3d_shader_instruction *ins)
 875 {
 876     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 877     SHADER_BUFFER *buffer = ins->ctx->buffer;
 878     char dst_name[50];
 879     char src_name[3][50];
 880
 881     shader_arb_get_dst_param(ins, dst, dst_name);
 882
 883     /* Generate input register names (with modifiers) */
 884     shader_arb_get_src_param(ins, &ins->src[0], 0, src_name[0]);
 885     shader_arb_get_src_param(ins, &ins->src[1], 1, src_name[1]);
 886     shader_arb_get_src_param(ins, &ins->src[2], 2, src_name[2]);
 887
 888     shader_addline(buffer, "CMP%s %s, %s, %s, %s;\n", shader_arb_get_modifier(ins), dst_name,
 889                    src_name[0], src_name[2], src_name[1]);
 890 }
 891
 892 /** Process the WINED3DSIO_DP2ADD instruction in ARB.
 893  * dst = dot2(src0, src1) + src2 */
 894 static void pshader_hw_dp2add(const struct wined3d_shader_instruction *ins)
 895 {
 896     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 897     SHADER_BUFFER *buffer = ins->ctx->buffer;
 898     char dst_name[50];
 899     char src_name[3][50];
 900
 901     shader_arb_get_dst_param(ins, dst, dst_name);
 902     shader_arb_get_src_param(ins, &ins->src[0], 0, src_name[0]);
 903     shader_arb_get_src_param(ins, &ins->src[1], 1, src_name[1]);
 904     shader_arb_get_src_param(ins, &ins->src[2], 2, src_name[2]);
 905
 906     /* Emulate a DP2 with a DP3 and 0.0 */
 907     shader_addline(buffer, "MOV TMP, %s;\n", src_name[0]);
 908     shader_addline(buffer, "MOV TMP.z, 0.0;\n");
 909     shader_addline(buffer, "DP3 TMP2, TMP, %s;\n", src_name[1]);
 910     shader_addline(buffer, "ADD%s %s, TMP2, %s;\n", shader_arb_get_modifier(ins), dst_name, src_name[2]);
 911 }
 912
 913 /* Map the opcode 1-to-1 to the GL code */
 914 static void shader_hw_map2gl(const struct wined3d_shader_instruction *ins)
 915 {
 916     SHADER_BUFFER *buffer = ins->ctx->buffer;
 917     const char *instruction;
 918     char arguments[256], dst_str[50];
 919     unsigned int i;
 920     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 921
 922     switch (ins->handler_idx)
 923     {
 924         case WINED3DSIH_ABS: instruction = "ABS"; break;
 925         case WINED3DSIH_ADD: instruction = "ADD"; break;
 926         case WINED3DSIH_CRS: instruction = "XPD"; break;
 927         case WINED3DSIH_DP3: instruction = "DP3"; break;
 928         case WINED3DSIH_DP4: instruction = "DP4"; break;
 929         case WINED3DSIH_DST: instruction = "DST"; break;
 930         case WINED3DSIH_EXP: instruction = "EX2"; break;
 931         case WINED3DSIH_EXPP: instruction = "EXP"; break;
 932         case WINED3DSIH_FRC: instruction = "FRC"; break;
 933         case WINED3DSIH_LIT: instruction = "LIT"; break;
 934         case WINED3DSIH_LOG: instruction = "LG2"; break;
 935         case WINED3DSIH_LOGP: instruction = "LOG"; break;
 936         case WINED3DSIH_LRP: instruction = "LRP"; break;
 937         case WINED3DSIH_MAD: instruction = "MAD"; break;
 938         case WINED3DSIH_MAX: instruction = "MAX"; break;
 939         case WINED3DSIH_MIN: instruction = "MIN"; break;
 940         case WINED3DSIH_MOV: instruction = "MOV"; break;
 941         case WINED3DSIH_MUL: instruction = "MUL"; break;
 942         case WINED3DSIH_POW: instruction = "POW"; break;
 943         case WINED3DSIH_SGE: instruction = "SGE"; break;
 944         case WINED3DSIH_SLT: instruction = "SLT"; break;
 945         case WINED3DSIH_SUB: instruction = "SUB"; break;
 946         default: instruction = "";
 947             FIXME("Unhandled opcode %#x\n", ins->handler_idx);
 948             break;
 949     }
 950
 951     /* Note that shader_arb_add_dst_param() adds spaces. */
 952     arguments[0] = '\0';
 953     shader_arb_get_dst_param(ins, dst, dst_str);
 954     for (i = 0; i < ins->src_count; ++i)
 955     {
 956         char operand[100];
 957         strcat(arguments, ", ");
 958         shader_arb_get_src_param(ins, &ins->src[i], i, operand);
 959         strcat(arguments, operand);
 960     }
 961     shader_addline(buffer, "%s%s %s%s;\n", instruction, shader_arb_get_modifier(ins), dst_str, arguments);
 962 }
 963
 964 static void shader_hw_nop(const struct wined3d_shader_instruction *ins)
 965 {
 966     SHADER_BUFFER *buffer = ins->ctx->buffer;
 967     shader_addline(buffer, "NOP;\n");
 968 }
 969
 970 static void shader_hw_mov(const struct wined3d_shader_instruction *ins)
 971 {
 972     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)ins->ctx->shader;
 973
 974     if ((ins->ctx->reg_maps->shader_version.major == 1
 975             && !shader_is_pshader_version(ins->ctx->reg_maps->shader_version.type)
 976             && ins->dst[0].reg.type == WINED3DSPR_ADDR)
 977             || ins->handler_idx == WINED3DSIH_MOVA)
 978     {
 979         SHADER_BUFFER *buffer = ins->ctx->buffer;
 980         char src0_param[256];
 981
 982         if (ins->handler_idx == WINED3DSIH_MOVA)
 983             FIXME("mova should round\n");
 984
 985         src0_param[0] = '\0';
 986         if (((IWineD3DVertexShaderImpl *)shader)->rel_offset)
 987         {
 988             shader_arb_get_src_param(ins, &ins->src[0], 0, src0_param);
 989             shader_addline(buffer, "ADD TMP.x, %s, helper_const.z;\n", src0_param);
 990             shader_addline(buffer, "ARL A0.x, TMP.x;\n");
 991         }
 992         else
 993         {
 994             /* Apple's ARB_vertex_program implementation does not accept an ARL source argument
 995              * with more than one component. Thus replicate the first source argument over all
 996              * 4 components. For example, .xyzw -> .x (or better: .xxxx), .zwxy -> .z, etc) */
 997             struct wined3d_shader_src_param tmp_src = ins->src[0];
 998             tmp_src.swizzle = (tmp_src.swizzle & 0x3) * 0x55;
 999             shader_arb_get_src_param(ins, &tmp_src, 0, src0_param);
1000             shader_addline(buffer, "ARL A0.x, %s;\n", src0_param);
1001         }
1002     }
1003     else
1004     {
1005         shader_hw_map2gl(ins);
1006     }
1007 }
1008
1009 static void pshader_hw_texkill(const struct wined3d_shader_instruction *ins)
1010 {
1011     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1012     SHADER_BUFFER *buffer = ins->ctx->buffer;
1013     char reg_dest[40];
1014
1015     /* No swizzles are allowed in d3d's texkill. PS 1.x ignores the 4th component as documented,
1016      * but >= 2.0 honors it(undocumented, but tested by the d3d9 testsuit)
1017      */
1018     shader_arb_get_dst_param(ins, dst, reg_dest);
1019
1020     if (ins->ctx->reg_maps->shader_version.major >= 2)
1021     {
1022         /* The arb backend doesn't claim ps 2.0 support, but try to eat what the app feeds to us */
1023         shader_arb_get_dst_param(ins, dst, reg_dest);
1024         shader_addline(buffer, "KIL %s;\n", reg_dest);
1025     } else {
1026         /* ARB fp doesn't like swizzles on the parameter of the KIL instruction. To mask the 4th component,
1027          * copy the register into our general purpose TMP variable, overwrite .w and pass TMP to KIL
1028          *
1029          * ps_1_3 shaders use the texcoord incarnation of the Tx register. ps_1_4 shaders can use the same,
1030          * or pass in any temporary register(in shader phase 2)
1031          */
1032         if(ins->ctx->reg_maps->shader_version.minor <= 3) {
1033             sprintf(reg_dest, "fragment.texcoord[%u]", dst->reg.idx);
1034         } else {
1035             shader_arb_get_dst_param(ins, dst, reg_dest);
1036         }
1037         shader_addline(buffer, "SWZ TMP, %s, x, y, z, 1;\n", reg_dest);
1038         shader_addline(buffer, "KIL TMP;\n");
1039     }
1040 }
1041
1042 static void pshader_hw_tex(const struct wined3d_shader_instruction *ins)
1043 {
1044     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1045     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1046     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1047     DWORD shader_version = WINED3D_SHADER_VERSION(ins->ctx->reg_maps->shader_version.major,
1048             ins->ctx->reg_maps->shader_version.minor);
1049     BOOL projected = FALSE, bias = FALSE;
1050     struct wined3d_shader_src_param src;
1051
1052     char reg_dest[40];
1053     char reg_coord[40];
1054     DWORD reg_sampler_code;
1055
1056     /* All versions have a destination register */
1057     shader_arb_get_dst_param(ins, dst, reg_dest);
1058
1059     /* 1.0-1.4: Use destination register number as texture code.
1060        2.0+: Use provided sampler number as texure code. */
1061     if (shader_version < WINED3D_SHADER_VERSION(2,0))
1062         reg_sampler_code = dst->reg.idx;
1063     else
1064         reg_sampler_code = ins->src[1].reg.idx;
1065
1066     /* 1.0-1.3: Use the texcoord varying.
1067        1.4+: Use provided coordinate source register. */
1068     if (shader_version < WINED3D_SHADER_VERSION(1,4))
1069         sprintf(reg_coord, "fragment.texcoord[%u]", reg_sampler_code);
1070     else {
1071         /* TEX is the only instruction that can handle DW and DZ natively */
1072         src = ins->src[0];
1073         if(src.modifiers == WINED3DSPSM_DW) src.modifiers = WINED3DSPSM_NONE;
1074         if(src.modifiers == WINED3DSPSM_DZ) src.modifiers = WINED3DSPSM_NONE;
1075         shader_arb_get_src_param(ins, &src, 0, reg_coord);
1076     }
1077
1078     /* projection flag:
1079      * 1.1, 1.2, 1.3: Use WINED3DTSS_TEXTURETRANSFORMFLAGS
1080      * 1.4: Use WINED3DSPSM_DZ or WINED3DSPSM_DW on src[0]
1081      * 2.0+: Use WINED3DSI_TEXLD_PROJECT on the opcode
1082      */
1083     if (shader_version < WINED3D_SHADER_VERSION(1,4))
1084     {
1085         DWORD flags = 0;
1086         if(reg_sampler_code < MAX_TEXTURES) {
1087             flags = deviceImpl->stateBlock->textureState[reg_sampler_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1088         }
1089         if (flags & WINED3DTTFF_PROJECTED) {
1090             projected = TRUE;
1091         }
1092     }
1093     else if (shader_version < WINED3D_SHADER_VERSION(2,0))
1094     {
1095         DWORD src_mod = ins->src[0].modifiers;
1096         if (src_mod == WINED3DSPSM_DZ) {
1097             /* TXP cannot handle DZ natively, so move the z coordinate to .w. reg_coord is a read-only
1098              * varying register, so we need a temp reg
1099              */
1100             shader_addline(ins->ctx->buffer, "SWZ TA, %s, x, y, z, z;\n", reg_coord);
1101             strcpy(reg_coord, "TA");
1102             projected = TRUE;
1103         } else if(src_mod == WINED3DSPSM_DW) {
1104             projected = TRUE;
1105         }
1106     } else {
1107         if (ins->flags & WINED3DSI_TEXLD_PROJECT) projected = TRUE;
1108         if (ins->flags & WINED3DSI_TEXLD_BIAS) bias = TRUE;
1109     }
1110     shader_hw_sample(ins, reg_sampler_code, reg_dest, reg_coord, projected, bias);
1111 }
1112
1113 static void pshader_hw_texcoord(const struct wined3d_shader_instruction *ins)
1114 {
1115     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1116     SHADER_BUFFER *buffer = ins->ctx->buffer;
1117     DWORD shader_version = WINED3D_SHADER_VERSION(ins->ctx->reg_maps->shader_version.major,
1118             ins->ctx->reg_maps->shader_version.minor);
1119
1120     char tmp[20];
1121     shader_arb_get_write_mask(ins, dst, tmp);
1122     if (shader_version != WINED3D_SHADER_VERSION(1,4))
1123     {
1124         DWORD reg = dst->reg.idx;
1125         shader_addline(buffer, "MOV_SAT T%u%s, fragment.texcoord[%u];\n", reg, tmp, reg);
1126     } else {
1127         char reg_src[40];
1128
1129         shader_arb_get_src_param(ins, &ins->src[0], 0, reg_src);
1130         shader_addline(buffer, "MOV R%u%s, %s;\n", dst->reg.idx, tmp, reg_src);
1131    }
1132 }
1133
1134 static void pshader_hw_texreg2ar(const struct wined3d_shader_instruction *ins)
1135 {
1136      SHADER_BUFFER *buffer = ins->ctx->buffer;
1137      IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1138      IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1139      DWORD flags;
1140
1141      DWORD reg1 = ins->dst[0].reg.idx;
1142      char dst_str[50];
1143      char src_str[50];
1144
1145      /* Note that texreg2ar treats Tx as a temporary register, not as a varying */
1146      shader_arb_get_dst_param(ins, &ins->dst[0], dst_str);
1147      shader_arb_get_src_param(ins, &ins->src[0], 0, src_str);
1148      shader_addline(buffer, "MOV TMP.x, %s.w;\n", src_str);
1149      shader_addline(buffer, "MOV TMP.y, %s.x;\n", src_str);
1150      flags = reg1 < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg1][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1151      shader_hw_sample(ins, reg1, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1152 }
1153
1154 static void pshader_hw_texreg2gb(const struct wined3d_shader_instruction *ins)
1155 {
1156      SHADER_BUFFER *buffer = ins->ctx->buffer;
1157
1158      DWORD reg1 = ins->dst[0].reg.idx;
1159      char dst_str[50];
1160      char src_str[50];
1161
1162      /* Note that texreg2gb treats Tx as a temporary register, not as a varying */
1163      shader_arb_get_dst_param(ins, &ins->dst[0], dst_str);
1164      shader_arb_get_src_param(ins, &ins->src[0], 0, src_str);
1165      shader_addline(buffer, "MOV TMP.x, %s.y;\n", src_str);
1166      shader_addline(buffer, "MOV TMP.y, %s.z;\n", src_str);
1167      shader_hw_sample(ins, reg1, dst_str, "TMP", FALSE, FALSE);
1168 }
1169
1170 static void pshader_hw_texreg2rgb(const struct wined3d_shader_instruction *ins)
1171 {
1172     DWORD reg1 = ins->dst[0].reg.idx;
1173     char dst_str[50];
1174     char src_str[50];
1175
1176     /* Note that texreg2rg treats Tx as a temporary register, not as a varying */
1177     shader_arb_get_dst_param(ins, &ins->dst[0], dst_str);
1178     shader_arb_get_src_param(ins, &ins->src[0], 0, src_str);
1179     shader_hw_sample(ins, reg1, dst_str, src_str, FALSE, FALSE);
1180 }
1181
1182 static void pshader_hw_texbem(const struct wined3d_shader_instruction *ins)
1183 {
1184     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1185     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1186     BOOL has_bumpmat = FALSE;
1187     BOOL has_luminance = FALSE;
1188     int i;
1189
1190     SHADER_BUFFER *buffer = ins->ctx->buffer;
1191
1192     char reg_coord[40], dst_reg[50];
1193     DWORD reg_dest_code;
1194
1195     /* All versions have a destination register. The Tx where the texture coordinates come
1196      * from is the varying incarnation of the texture register
1197      */
1198     reg_dest_code = dst->reg.idx;
1199     shader_arb_get_dst_param(ins, &ins->dst[0], dst_reg);
1200     sprintf(reg_coord, "fragment.texcoord[%u]", reg_dest_code);
1201
1202     for(i = 0; i < This->numbumpenvmatconsts; i++) {
1203         if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1204                 && reg_dest_code == This->bumpenvmatconst[i].texunit)
1205         {
1206             has_bumpmat = TRUE;
1207             break;
1208         }
1209     }
1210     for(i = 0; i < This->numbumpenvmatconsts; i++) {
1211         if (This->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1212                 && reg_dest_code == This->luminanceconst[i].texunit)
1213         {
1214             has_luminance = TRUE;
1215             break;
1216         }
1217     }
1218
1219     if(has_bumpmat) {
1220         DWORD src = ins->src[0].reg.idx;
1221
1222         /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed
1223          * The Tx in which the perturbation map is stored is the tempreg incarnation of the texture register
1224          */
1225         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", reg_dest_code);
1226         shader_addline(buffer, "DP3 TMP.x, TMP2, T%u;\n", src);
1227         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", reg_dest_code);
1228         shader_addline(buffer, "DP3 TMP.y, TMP2, T%u;\n", src);
1229
1230         /* with projective textures, texbem only divides the static texture coord, not the displacement,
1231          * so we can't let the GL handle this.
1232          */
1233         if (((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS]
1234               & WINED3DTTFF_PROJECTED) {
1235             shader_addline(buffer, "RCP TMP2.w, %s.w;\n", reg_coord);
1236             shader_addline(buffer, "MUL TMP2.xy, %s, TMP2.w;\n", reg_coord);
1237             shader_addline(buffer, "ADD TMP.xy, TMP, TMP2;\n");
1238         } else {
1239             shader_addline(buffer, "ADD TMP.xy, TMP, %s;\n", reg_coord);
1240         }
1241
1242         shader_hw_sample(ins, reg_dest_code, dst_reg, "TMP", FALSE, FALSE);
1243
1244         if (ins->handler_idx == WINED3DSIH_TEXBEML && has_luminance)
1245         {
1246             shader_addline(buffer, "MAD TMP, T%u.z, luminance%d.x, luminance%d.y;\n",
1247                            src, reg_dest_code, reg_dest_code);
1248             shader_addline(buffer, "MUL %s, %s, TMP;\n", dst_reg, dst_reg);
1249         }
1250
1251     } else {
1252         DWORD tf;
1253         if(reg_dest_code < MAX_TEXTURES) {
1254             tf = ((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1255         } else {
1256             tf = 0;
1257         }
1258         /* Without a bump matrix loaded, just sample with the unmodified coordinates */
1259         shader_hw_sample(ins, reg_dest_code, dst_reg, reg_coord, tf & WINED3DTTFF_PROJECTED, FALSE);
1260     }
1261 }
1262
1263 static void pshader_hw_texm3x2pad(const struct wined3d_shader_instruction *ins)
1264 {
1265     DWORD reg = ins->dst[0].reg.idx;
1266     SHADER_BUFFER *buffer = ins->ctx->buffer;
1267     char src0_name[50];
1268
1269     shader_arb_get_src_param(ins, &ins->src[0], 0, src0_name);
1270     shader_addline(buffer, "DP3 TMP.x, fragment.texcoord[%u], %s;\n", reg, src0_name);
1271 }
1272
1273 static void pshader_hw_texm3x2tex(const struct wined3d_shader_instruction *ins)
1274 {
1275     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1276     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1277     DWORD flags;
1278     DWORD reg = ins->dst[0].reg.idx;
1279     SHADER_BUFFER *buffer = ins->ctx->buffer;
1280     char dst_str[50];
1281     char src0_name[50];
1282
1283     shader_arb_get_dst_param(ins, &ins->dst[0], dst_str);
1284     shader_arb_get_src_param(ins, &ins->src[0], 0, src0_name);
1285     shader_addline(buffer, "DP3 TMP.y, fragment.texcoord[%u], %s;\n", reg, src0_name);
1286     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1287     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1288 }
1289
1290 static void pshader_hw_texm3x3pad(const struct wined3d_shader_instruction *ins)
1291 {
1292     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1293     DWORD reg = ins->dst[0].reg.idx;
1294     SHADER_BUFFER *buffer = ins->ctx->buffer;
1295     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1296     char src0_name[50];
1297
1298     shader_arb_get_src_param(ins, &ins->src[0], 0, src0_name);
1299     shader_addline(buffer, "DP3 TMP.%c, fragment.texcoord[%u], %s;\n", 'x' + current_state->current_row, reg, src0_name);
1300     current_state->texcoord_w[current_state->current_row++] = reg;
1301 }
1302
1303 static void pshader_hw_texm3x3tex(const struct wined3d_shader_instruction *ins)
1304 {
1305     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1306     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1307     DWORD flags;
1308     DWORD reg = ins->dst[0].reg.idx;
1309     SHADER_BUFFER *buffer = ins->ctx->buffer;
1310     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1311     char dst_str[50];
1312     char src0_name[50];
1313
1314     shader_arb_get_src_param(ins, &ins->src[0], 0, src0_name);
1315     shader_addline(buffer, "DP3 TMP.z, fragment.texcoord[%u], %s;\n", reg, src0_name);
1316
1317     /* Sample the texture using the calculated coordinates */
1318     shader_arb_get_dst_param(ins, &ins->dst[0], dst_str);
1319     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1320     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1321     current_state->current_row = 0;
1322 }
1323
1324 static void pshader_hw_texm3x3vspec(const struct wined3d_shader_instruction *ins)
1325 {
1326     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1327     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1328     DWORD flags;
1329     DWORD reg = ins->dst[0].reg.idx;
1330     SHADER_BUFFER *buffer = ins->ctx->buffer;
1331     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1332     char dst_str[50];
1333     char src0_name[50];
1334
1335     shader_arb_get_src_param(ins, &ins->src[0], 0, src0_name);
1336     shader_addline(buffer, "DP3 TMP.z, fragment.texcoord[%u], %s;\n", reg, src0_name);
1337
1338     /* Construct the eye-ray vector from w coordinates */
1339     shader_addline(buffer, "MOV TMP2.x, fragment.texcoord[%u].w;\n", current_state->texcoord_w[0]);
1340     shader_addline(buffer, "MOV TMP2.y, fragment.texcoord[%u].w;\n", current_state->texcoord_w[1]);
1341     shader_addline(buffer, "MOV TMP2.z, fragment.texcoord[%u].w;\n", reg);
1342
1343     /* Calculate reflection vector
1344      */
1345     shader_addline(buffer, "DP3 TMP.w, TMP, TMP2;\n");
1346     /* The .w is ignored when sampling, so I can use TMP2.w to calculate dot(N, N) */
1347     shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1348     shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1349     shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1350     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1351     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -TMP2;\n");
1352
1353     /* Sample the texture using the calculated coordinates */
1354     shader_arb_get_dst_param(ins, &ins->dst[0], dst_str);
1355     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1356     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1357     current_state->current_row = 0;
1358 }
1359
1360 static void pshader_hw_texm3x3spec(const struct wined3d_shader_instruction *ins)
1361 {
1362     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1363     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1364     DWORD flags;
1365     DWORD reg = ins->dst[0].reg.idx;
1366     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1367     SHADER_BUFFER *buffer = ins->ctx->buffer;
1368     char dst_str[50];
1369     char src0_name[50];
1370     char src1_name[50];
1371
1372     shader_arb_get_src_param(ins, &ins->src[0], 0, src0_name);
1373     shader_arb_get_src_param(ins, &ins->src[0], 1, src1_name);
1374     shader_addline(buffer, "DP3 TMP.z, fragment.texcoord[%u], %s;\n", reg, src0_name);
1375
1376     /* Calculate reflection vector.
1377      *
1378      *               dot(N, E)
1379      * TMP.xyz = 2 * --------- * N - E
1380      *               dot(N, N)
1381      *
1382      * Which normalizes the normal vector
1383      */
1384     shader_addline(buffer, "DP3 TMP.w, TMP, %s;\n", src1_name);
1385     shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1386     shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1387     shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1388     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1389     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -%s;\n", src1_name);
1390
1391     /* Sample the texture using the calculated coordinates */
1392     shader_arb_get_dst_param(ins, &ins->dst[0], dst_str);
1393     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1394     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1395     current_state->current_row = 0;
1396 }
1397
1398 static void pshader_hw_texdepth(const struct wined3d_shader_instruction *ins)
1399 {
1400     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1401     SHADER_BUFFER *buffer = ins->ctx->buffer;
1402     char dst_name[50];
1403
1404     /* texdepth has an implicit destination, the fragment depth value. It's only parameter,
1405      * which is essentially an input, is the destination register because it is the first
1406      * parameter. According to the msdn, this must be register r5, but let's keep it more flexible
1407      * here(writemasks/swizzles are not valid on texdepth)
1408      */
1409     shader_arb_get_dst_param(ins, dst, dst_name);
1410
1411     /* According to the msdn, the source register(must be r5) is unusable after
1412      * the texdepth instruction, so we're free to modify it
1413      */
1414     shader_addline(buffer, "MIN %s.y, %s.y, one.y;\n", dst_name, dst_name);
1415
1416     /* How to deal with the special case dst_name.g == 0? if r != 0, then
1417      * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1418      * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1419      */
1420     shader_addline(buffer, "RCP %s.y, %s.y;\n", dst_name, dst_name);
1421     shader_addline(buffer, "MUL TMP.x, %s.x, %s.y;\n", dst_name, dst_name);
1422     shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1423     shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1424 }
1425
1426 /** Process the WINED3DSIO_TEXDP3TEX instruction in ARB:
1427  * Take a 3-component dot product of the TexCoord[dstreg] and src,
1428  * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1429 static void pshader_hw_texdp3tex(const struct wined3d_shader_instruction *ins)
1430 {
1431     SHADER_BUFFER *buffer = ins->ctx->buffer;
1432     DWORD sampler_idx = ins->dst[0].reg.idx;
1433     char src0[50];
1434     char dst_str[50];
1435
1436     shader_arb_get_src_param(ins, &ins->src[0], 0, src0);
1437     shader_addline(buffer, "MOV TMP, 0.0;\n");
1438     shader_addline(buffer, "DP3 TMP.x, fragment.texcoord[%u], %s;\n", sampler_idx, src0);
1439
1440     shader_arb_get_dst_param(ins, &ins->dst[0], dst_str);
1441     shader_hw_sample(ins, sampler_idx, dst_str, "TMP", FALSE /* Only one coord, can't be projected */, FALSE);
1442 }
1443
1444 /** Process the WINED3DSIO_TEXDP3 instruction in ARB:
1445  * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1446 static void pshader_hw_texdp3(const struct wined3d_shader_instruction *ins)
1447 {
1448     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1449     char src0[50];
1450     char dst_str[50];
1451     SHADER_BUFFER *buffer = ins->ctx->buffer;
1452
1453     /* Handle output register */
1454     shader_arb_get_dst_param(ins, dst, dst_str);
1455     shader_arb_get_src_param(ins, &ins->src[0], 0, src0);
1456     shader_addline(buffer, "DP3 %s, fragment.texcoord[%u], %s;\n", dst_str, dst->reg.idx, src0);
1457 }
1458
1459 /** Process the WINED3DSIO_TEXM3X3 instruction in ARB
1460  * Perform the 3rd row of a 3x3 matrix multiply */
1461 static void pshader_hw_texm3x3(const struct wined3d_shader_instruction *ins)
1462 {
1463     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1464     SHADER_BUFFER *buffer = ins->ctx->buffer;
1465     char dst_str[50];
1466     char src0[50];
1467
1468     shader_arb_get_dst_param(ins, dst, dst_str);
1469     shader_arb_get_src_param(ins, &ins->src[0], 0, src0);
1470     shader_addline(buffer, "DP3 TMP.z, fragment.texcoord[%u], %s;\n", dst->reg.idx, src0);
1471     shader_addline(buffer, "MOV %s, TMP;\n", dst_str);
1472 }
1473
1474 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in ARB:
1475  * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1476  * Calculate tmp0.y = TexCoord[dstreg] . src.xyz;  (tmp0.x has already been calculated)
1477  * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1478  */
1479 static void pshader_hw_texm3x2depth(const struct wined3d_shader_instruction *ins)
1480 {
1481     SHADER_BUFFER *buffer = ins->ctx->buffer;
1482     DWORD dst_reg = ins->dst[0].reg.idx;
1483     char src0[50];
1484
1485     shader_arb_get_src_param(ins, &ins->src[0], 0, src0);
1486     shader_addline(buffer, "DP3 TMP.y, fragment.texcoord[%u], %s;\n", dst_reg, src0);
1487
1488     /* How to deal with the special case dst_name.g == 0? if r != 0, then
1489      * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1490      * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1491      */
1492     shader_addline(buffer, "RCP TMP.y, TMP.y;\n");
1493     shader_addline(buffer, "MUL TMP.x, TMP.x, TMP.y;\n");
1494     shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1495     shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1496 }
1497
1498 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1499     Vertex/Pixel shaders to ARB_vertex_program codes */
1500 static void shader_hw_mnxn(const struct wined3d_shader_instruction *ins)
1501 {
1502     int i;
1503     int nComponents = 0;
1504     struct wined3d_shader_dst_param tmp_dst = {{0}};
1505     struct wined3d_shader_src_param tmp_src[2] = {{{0}}};
1506     struct wined3d_shader_instruction tmp_ins;
1507
1508     memset(&tmp_ins, 0, sizeof(tmp_ins));
1509
1510     /* Set constants for the temporary argument */
1511     tmp_ins.ctx = ins->ctx;
1512     tmp_ins.dst_count = 1;
1513     tmp_ins.dst = &tmp_dst;
1514     tmp_ins.src_count = 2;
1515     tmp_ins.src = tmp_src;
1516
1517     switch(ins->handler_idx)
1518     {
1519         case WINED3DSIH_M4x4:
1520             nComponents = 4;
1521             tmp_ins.handler_idx = WINED3DSIH_DP4;
1522             break;
1523         case WINED3DSIH_M4x3:
1524             nComponents = 3;
1525             tmp_ins.handler_idx = WINED3DSIH_DP4;
1526             break;
1527         case WINED3DSIH_M3x4:
1528             nComponents = 4;
1529             tmp_ins.handler_idx = WINED3DSIH_DP3;
1530             break;
1531         case WINED3DSIH_M3x3:
1532             nComponents = 3;
1533             tmp_ins.handler_idx = WINED3DSIH_DP3;
1534             break;
1535         case WINED3DSIH_M3x2:
1536             nComponents = 2;
1537             tmp_ins.handler_idx = WINED3DSIH_DP3;
1538             break;
1539         default:
1540             FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1541             break;
1542     }
1543
1544     tmp_dst = ins->dst[0];
1545     tmp_src[0] = ins->src[0];
1546     tmp_src[1] = ins->src[1];
1547     for (i = 0; i < nComponents; i++) {
1548         tmp_dst.write_mask = WINED3DSP_WRITEMASK_0 << i;
1549         shader_hw_map2gl(&tmp_ins);
1550         ++tmp_src[1].reg.idx;
1551     }
1552 }
1553
1554 static void shader_hw_rsq_rcp(const struct wined3d_shader_instruction *ins)
1555 {
1556     SHADER_BUFFER *buffer = ins->ctx->buffer;
1557     const char *instruction;
1558
1559     char dst[50];
1560     char src[50];
1561
1562     switch(ins->handler_idx)
1563     {
1564         case WINED3DSIH_RSQ: instruction = "RSQ"; break;
1565         case WINED3DSIH_RCP: instruction = "RCP"; break;
1566         default: instruction = "";
1567             FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1568             break;
1569     }
1570
1571     shader_arb_get_dst_param(ins, &ins->dst[0], dst); /* Destination */
1572     shader_arb_get_src_param(ins, &ins->src[0], 0, src);
1573     if (ins->src[0].swizzle == WINED3DSP_NOSWIZZLE)
1574     {
1575         /* Dx sdk says .x is used if no swizzle is given, but our test shows that
1576          * .w is used
1577          */
1578         strcat(src, ".w");
1579     }
1580
1581     shader_addline(buffer, "%s%s %s, %s;\n", instruction, shader_arb_get_modifier(ins), dst, src);
1582 }
1583
1584 static void shader_hw_nrm(const struct wined3d_shader_instruction *ins)
1585 {
1586     SHADER_BUFFER *buffer = ins->ctx->buffer;
1587     char dst_name[50];
1588     char src_name[50];
1589
1590     shader_arb_get_dst_param(ins, &ins->dst[0], dst_name);
1591     shader_arb_get_src_param(ins, &ins->src[0], 0, src_name);
1592     shader_addline(buffer, "DP3 TMP, %s, %s;\n", src_name, src_name);
1593     shader_addline(buffer, "RSQ TMP, TMP.x;\n");
1594     /* dst.w = src[0].w * 1 / (src.x^2 + src.y^2 + src.z^2)^(1/2) according to msdn*/
1595     shader_addline(buffer, "MUL%s %s, %s, TMP;\n", shader_arb_get_modifier(ins), dst_name,
1596                    src_name);
1597 }
1598
1599 static void shader_hw_sincos(const struct wined3d_shader_instruction *ins)
1600 {
1601     /* This instruction exists in ARB, but the d3d instruction takes two extra parameters which
1602      * must contain fixed constants. So we need a separate function to filter those constants and
1603      * can't use map2gl
1604      */
1605     SHADER_BUFFER *buffer = ins->ctx->buffer;
1606     char dst_name[50];
1607     char src_name[50];
1608
1609     shader_arb_get_dst_param(ins, &ins->dst[0], dst_name);
1610     shader_arb_get_src_param(ins, &ins->src[0], 0, src_name);
1611     shader_addline(buffer, "SCS%s %s, %s;\n", shader_arb_get_modifier(ins), dst_name,
1612                    src_name);
1613 }
1614
1615 static GLuint create_arb_blt_vertex_program(const WineD3D_GL_Info *gl_info)
1616 {
1617     GLuint program_id = 0;
1618     const char *blt_vprogram =
1619         "!!ARBvp1.0\n"
1620         "PARAM c[1] = { { 1, 0.5 } };\n"
1621         "MOV result.position, vertex.position;\n"
1622         "MOV result.color, c[0].x;\n"
1623         "MOV result.texcoord[0], vertex.texcoord[0];\n"
1624         "END\n";
1625
1626     GL_EXTCALL(glGenProgramsARB(1, &program_id));
1627     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, program_id));
1628     GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_vprogram), blt_vprogram));
1629
1630     if (glGetError() == GL_INVALID_OPERATION) {
1631         GLint pos;
1632         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1633         FIXME("Vertex program error at position %d: %s\n", pos,
1634             debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1635     }
1636
1637     return program_id;
1638 }
1639
1640 static GLuint create_arb_blt_fragment_program(const WineD3D_GL_Info *gl_info, enum tex_types tex_type)
1641 {
1642     GLuint program_id = 0;
1643     static const char * const blt_fprograms[tex_type_count] =
1644     {
1645         /* tex_1d */
1646         NULL,
1647         /* tex_2d */
1648         "!!ARBfp1.0\n"
1649         "TEMP R0;\n"
1650         "TEX R0.x, fragment.texcoord[0], texture[0], 2D;\n"
1651         "MOV result.depth.z, R0.x;\n"
1652         "END\n",
1653         /* tex_3d */
1654         NULL,
1655         /* tex_cube */
1656         "!!ARBfp1.0\n"
1657         "TEMP R0;\n"
1658         "TEX R0.x, fragment.texcoord[0], texture[0], CUBE;\n"
1659         "MOV result.depth.z, R0.x;\n"
1660         "END\n",
1661         /* tex_rect */
1662         "!!ARBfp1.0\n"
1663         "TEMP R0;\n"
1664         "TEX R0.x, fragment.texcoord[0], texture[0], RECT;\n"
1665         "MOV result.depth.z, R0.x;\n"
1666         "END\n",
1667     };
1668
1669     if (!blt_fprograms[tex_type])
1670     {
1671         FIXME("tex_type %#x not supported\n", tex_type);
1672         tex_type = tex_2d;
1673     }
1674
1675     GL_EXTCALL(glGenProgramsARB(1, &program_id));
1676     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, program_id));
1677     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_fprograms[tex_type]), blt_fprograms[tex_type]));
1678
1679     if (glGetError() == GL_INVALID_OPERATION) {
1680         GLint pos;
1681         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1682         FIXME("Fragment program error at position %d: %s\n", pos,
1683             debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1684     }
1685
1686     return program_id;
1687 }
1688
1689 static void shader_arb_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
1690     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1691     struct shader_arb_priv *priv = This->shader_priv;
1692     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1693
1694     if (useVS) {
1695         struct vs_compile_args compile_args;
1696
1697         TRACE("Using vertex shader\n");
1698         find_vs_compile_args((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, This->stateBlock, &compile_args);
1699         priv->current_vprogram_id = find_gl_vshader((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, &compile_args);
1700
1701         /* Bind the vertex program */
1702         GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1703         checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id);");
1704
1705         /* Enable OpenGL vertex programs */
1706         glEnable(GL_VERTEX_PROGRAM_ARB);
1707         checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1708         TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1709     } else if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
1710         priv->current_vprogram_id = 0;
1711         glDisable(GL_VERTEX_PROGRAM_ARB);
1712         checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1713     }
1714
1715     if (usePS) {
1716         struct ps_compile_args compile_args;
1717         TRACE("Using pixel shader\n");
1718         find_ps_compile_args((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader, This->stateBlock, &compile_args);
1719         priv->current_fprogram_id = find_gl_pshader((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader,
1720                                                     &compile_args);
1721
1722         /* Bind the fragment program */
1723         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1724         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id);");
1725
1726         if(!priv->use_arbfp_fixed_func) {
1727             /* Enable OpenGL fragment programs */
1728             glEnable(GL_FRAGMENT_PROGRAM_ARB);
1729             checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1730         }
1731         TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1732     } else if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM) && !priv->use_arbfp_fixed_func) {
1733         /* Disable only if we're not using arbfp fixed function fragment processing. If this is used,
1734          * keep GL_FRAGMENT_PROGRAM_ARB enabled, and the fixed function pipeline will bind the fixed function
1735          * replacement shader
1736          */
1737         glDisable(GL_FRAGMENT_PROGRAM_ARB);
1738         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1739         priv->current_fprogram_id = 0;
1740     }
1741 }
1742
1743 static void shader_arb_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {
1744     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1745     struct shader_arb_priv *priv = This->shader_priv;
1746     GLuint *blt_fprogram = &priv->depth_blt_fprogram_id[tex_type];
1747     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1748
1749     if (!priv->depth_blt_vprogram_id) priv->depth_blt_vprogram_id = create_arb_blt_vertex_program(gl_info);
1750     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->depth_blt_vprogram_id));
1751     glEnable(GL_VERTEX_PROGRAM_ARB);
1752
1753     if (!*blt_fprogram) *blt_fprogram = create_arb_blt_fragment_program(gl_info, tex_type);
1754     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, *blt_fprogram));
1755     glEnable(GL_FRAGMENT_PROGRAM_ARB);
1756 }
1757
1758 static void shader_arb_deselect_depth_blt(IWineD3DDevice *iface) {
1759     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1760     struct shader_arb_priv *priv = This->shader_priv;
1761     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1762
1763     if (priv->current_vprogram_id) {
1764         GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1765         checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vertexShader->prgId);");
1766
1767         glEnable(GL_VERTEX_PROGRAM_ARB);
1768         checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1769
1770         TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1771     } else {
1772         glDisable(GL_VERTEX_PROGRAM_ARB);
1773         checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1774     }
1775
1776     if (priv->current_fprogram_id) {
1777         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1778         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, pixelShader->prgId);");
1779
1780         glEnable(GL_FRAGMENT_PROGRAM_ARB);
1781         checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1782
1783         TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1784     } else {
1785         glDisable(GL_FRAGMENT_PROGRAM_ARB);
1786         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1787     }
1788 }
1789
1790 static void shader_arb_destroy(IWineD3DBaseShader *iface) {
1791     IWineD3DBaseShaderImpl *baseShader = (IWineD3DBaseShaderImpl *) iface;
1792     const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)baseShader->baseShader.device)->adapter->gl_info;
1793
1794     if (shader_is_pshader_version(baseShader->baseShader.reg_maps.shader_version.type))
1795     {
1796         IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *) iface;
1797         UINT i;
1798
1799         ENTER_GL();
1800         for(i = 0; i < This->num_gl_shaders; i++) {
1801             GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1802             checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1803         }
1804         LEAVE_GL();
1805         HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1806         This->gl_shaders = NULL;
1807         This->num_gl_shaders = 0;
1808         This->shader_array_size = 0;
1809     } else {
1810         IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *) iface;
1811         UINT i;
1812
1813         ENTER_GL();
1814         for(i = 0; i < This->num_gl_shaders; i++) {
1815             GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1816             checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1817         }
1818         LEAVE_GL();
1819         HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1820         This->gl_shaders = NULL;
1821         This->num_gl_shaders = 0;
1822         This->shader_array_size = 0;
1823     }
1824 }
1825
1826 static HRESULT shader_arb_alloc(IWineD3DDevice *iface) {
1827     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1828     This->shader_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
1829     return WINED3D_OK;
1830 }
1831
1832 static void shader_arb_free(IWineD3DDevice *iface) {
1833     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1834     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1835     struct shader_arb_priv *priv = This->shader_priv;
1836     int i;
1837
1838     ENTER_GL();
1839     if(priv->depth_blt_vprogram_id) {
1840         GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_vprogram_id));
1841     }
1842     for (i = 0; i < tex_type_count; ++i) {
1843         if (priv->depth_blt_fprogram_id[i]) {
1844             GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_fprogram_id[i]));
1845         }
1846     }
1847     LEAVE_GL();
1848
1849     HeapFree(GetProcessHeap(), 0, This->shader_priv);
1850 }
1851
1852 static BOOL shader_arb_dirty_const(IWineD3DDevice *iface) {
1853     return TRUE;
1854 }
1855
1856 static void arbfp_add_sRGB_correction(SHADER_BUFFER *buffer, const char *fragcolor, const char *tmp1,
1857                                       const char *tmp2, const char *tmp3, const char *tmp4) {
1858     /* Perform sRGB write correction. See GLX_EXT_framebuffer_sRGB */
1859
1860     /* Calculate the > 0.0031308 case */
1861     shader_addline(buffer, "POW %s.x, %s.x, srgb_consts1.z;\n", tmp1, fragcolor);
1862     shader_addline(buffer, "POW %s.y, %s.y, srgb_consts1.z;\n", tmp1, fragcolor);
1863     shader_addline(buffer, "POW %s.z, %s.z, srgb_consts1.z;\n", tmp1, fragcolor);
1864     shader_addline(buffer, "MUL %s, %s, srgb_consts1.w;\n", tmp1, tmp1);
1865     shader_addline(buffer, "SUB %s, %s, srgb_consts2.x;\n", tmp1, tmp1);
1866     /* Calculate the < case */
1867     shader_addline(buffer, "MUL %s, srgb_consts1.x, %s;\n", tmp2, fragcolor);
1868     /* Get 1.0 / 0.0 masks for > 0.0031308 and < 0.0031308 */
1869     shader_addline(buffer, "SLT %s, srgb_consts1.y, %s;\n", tmp3, fragcolor);
1870     shader_addline(buffer, "SGE %s, srgb_consts1.y, %s;\n", tmp4, fragcolor);
1871     /* Store the components > 0.0031308 in the destination */
1872     shader_addline(buffer, "MUL %s, %s, %s;\n", fragcolor, tmp1, tmp3);
1873     /* Add the components that are < 0.0031308 */
1874     shader_addline(buffer, "MAD result.color.xyz, %s, %s, %s;\n", tmp2, tmp4, fragcolor);
1875     /* [0.0;1.0] clamping. Not needed, this is done implicitly */
1876 }
1877
1878 static GLuint shader_arb_generate_pshader(IWineD3DPixelShader *iface,
1879         SHADER_BUFFER *buffer, const struct ps_compile_args *args)
1880 {
1881     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1882     const shader_reg_maps* reg_maps = &This->baseShader.reg_maps;
1883     CONST DWORD *function = This->baseShader.function;
1884     const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)This->baseShader.device)->adapter->gl_info;
1885     const local_constant *lconst;
1886     GLuint retval;
1887     const char *fragcolor;
1888     DWORD *lconst_map = local_const_mapping((IWineD3DBaseShaderImpl *) This);
1889
1890     /*  Create the hw ARB shader */
1891     shader_addline(buffer, "!!ARBfp1.0\n");
1892
1893     if (reg_maps->shader_version.major < 3)
1894     {
1895         switch(args->fog) {
1896             case FOG_OFF:
1897                 break;
1898             case FOG_LINEAR:
1899                 shader_addline(buffer, "OPTION ARB_fog_linear;\n");
1900                 break;
1901             case FOG_EXP:
1902                 shader_addline(buffer, "OPTION ARB_fog_exp;\n");
1903                 break;
1904             case FOG_EXP2:
1905                 shader_addline(buffer, "OPTION ARB_fog_exp2;\n");
1906                 break;
1907         }
1908     }
1909
1910     shader_addline(buffer, "TEMP TMP;\n");     /* Used in matrix ops */
1911     shader_addline(buffer, "TEMP TMP2;\n");    /* Used in matrix ops */
1912     shader_addline(buffer, "TEMP TA;\n");      /* Used for modifiers */
1913     shader_addline(buffer, "TEMP TB;\n");      /* Used for modifiers */
1914     shader_addline(buffer, "TEMP TC;\n");      /* Used for modifiers */
1915     shader_addline(buffer, "PARAM coefdiv = { 0.5, 0.25, 0.125, 0.0625 };\n");
1916     shader_addline(buffer, "PARAM coefmul = { 2, 4, 8, 16 };\n");
1917     shader_addline(buffer, "PARAM one = { 1.0, 1.0, 1.0, 1.0 };\n");
1918
1919     if (reg_maps->shader_version.major < 2)
1920     {
1921         fragcolor = "R0";
1922     } else {
1923         shader_addline(buffer, "TEMP TMP_COLOR;\n");
1924         fragcolor = "TMP_COLOR";
1925     }
1926
1927     /* Base Declarations */
1928     shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION, lconst_map);
1929
1930     /* Base Shader Body */
1931     shader_generate_main((IWineD3DBaseShader *)This, buffer, reg_maps, function);
1932
1933     if(args->srgb_correction) {
1934         arbfp_add_sRGB_correction(buffer, fragcolor, "TMP", "TMP2", "TA", "TB");
1935     }
1936     shader_addline(buffer, "MOV result.color, %s;\n", fragcolor);
1937     shader_addline(buffer, "END\n");
1938
1939     /* TODO: change to resource.glObjectHandle or something like that */
1940     GL_EXTCALL(glGenProgramsARB(1, &retval));
1941
1942     TRACE("Creating a hw pixel shader, prg=%d\n", retval);
1943     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, retval));
1944
1945     TRACE("Created hw pixel shader, prg=%d\n", retval);
1946     /* Create the program and check for errors */
1947     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
1948                buffer->bsize, buffer->buffer));
1949
1950     if (glGetError() == GL_INVALID_OPERATION) {
1951         GLint errPos;
1952         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
1953         FIXME("HW PixelShader Error at position %d: %s\n",
1954               errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1955         retval = 0;
1956     }
1957
1958     /* Load immediate constants */
1959     if(lconst_map) {
1960         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
1961             const float *value = (const float *)lconst->value;
1962             GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, lconst_map[lconst->idx], value));
1963             checkGLcall("glProgramLocalParameter4fvARB");
1964         }
1965         HeapFree(GetProcessHeap(), 0, lconst_map);
1966     }
1967
1968     return retval;
1969 }
1970
1971 static GLuint shader_arb_generate_vshader(IWineD3DVertexShader *iface,
1972         SHADER_BUFFER *buffer, const struct vs_compile_args *args)
1973 {
1974     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
1975     const shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
1976     CONST DWORD *function = This->baseShader.function;
1977     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)This->baseShader.device;
1978     const WineD3D_GL_Info *gl_info = &device->adapter->gl_info;
1979     const local_constant *lconst;
1980     GLuint ret;
1981     DWORD *lconst_map = local_const_mapping((IWineD3DBaseShaderImpl *) This);
1982
1983     /*  Create the hw ARB shader */
1984     shader_addline(buffer, "!!ARBvp1.0\n");
1985     shader_addline(buffer, "PARAM helper_const = { 2.0, -1.0, %d.0, 0.0 };\n", This->rel_offset);
1986
1987     /* Mesa supports only 95 constants */
1988     if (GL_VEND(MESA) || GL_VEND(WINE))
1989         This->baseShader.limits.constant_float =
1990                 min(95, This->baseShader.limits.constant_float);
1991
1992     shader_addline(buffer, "TEMP TMP;\n");
1993
1994     /* Base Declarations */
1995     shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION, lconst_map);
1996
1997     /* We need a constant to fixup the final position */
1998     shader_addline(buffer, "PARAM posFixup = program.env[%d];\n", ARB_SHADER_PRIVCONST_POS);
1999
2000     /* Initialize output parameters. GL_ARB_vertex_program does not require special initialization values
2001      * for output parameters. D3D in theory does not do that either, but some applications depend on a
2002      * proper initialization of the secondary color, and programs using the fixed function pipeline without
2003      * a replacement shader depend on the texcoord.w being set properly.
2004      *
2005      * GL_NV_vertex_program defines that all output values are initialized to {0.0, 0.0, 0.0, 1.0}. This
2006      * assertion is in effect even when using GL_ARB_vertex_program without any NV specific additions. So
2007      * skip this if NV_vertex_program is supported. Otherwise, initialize the secondary color. For the tex-
2008      * coords, we have a flag in the opengl caps. Many cards do not require the texcoord being set, and
2009      * this can eat a number of instructions, so skip it unless this cap is set as well
2010      */
2011     if(!GL_SUPPORT(NV_VERTEX_PROGRAM)) {
2012         shader_addline(buffer, "MOV result.color.secondary, -helper_const.wwwy;\n");
2013
2014         if((GLINFO_LOCATION).set_texcoord_w && !device->frag_pipe->ffp_proj_control) {
2015             int i;
2016             for(i = 0; i < min(8, MAX_REG_TEXCRD); i++) {
2017                 if(This->baseShader.reg_maps.texcoord_mask[i] != 0 &&
2018                 This->baseShader.reg_maps.texcoord_mask[i] != WINED3DSP_WRITEMASK_ALL) {
2019                     shader_addline(buffer, "MOV result.texcoord[%u].w, -helper_const.y;\n", i);
2020                 }
2021             }
2022         }
2023     }
2024
2025     /* Base Shader Body */
2026     shader_generate_main((IWineD3DBaseShader *)This, buffer, reg_maps, function);
2027
2028     /* The D3DRS_FOGTABLEMODE render state defines if the shader-generated fog coord is used
2029      * or if the fragment depth is used. If the fragment depth is used(FOGTABLEMODE != NONE),
2030      * the fog frag coord is thrown away. If the fog frag coord is used, but not written by
2031      * the shader, it is set to 0.0(fully fogged, since start = 1.0, end = 0.0)
2032      */
2033     if(args->fog_src == VS_FOG_Z) {
2034         shader_addline(buffer, "MOV result.fogcoord, TMP_OUT.z;\n");
2035     } else if (!reg_maps->fog) {
2036         shader_addline(buffer, "MOV result.fogcoord, helper_const.w;\n");
2037     }
2038
2039     /* Write the final position.
2040      *
2041      * OpenGL coordinates specify the center of the pixel while d3d coords specify
2042      * the corner. The offsets are stored in z and w in posFixup. posFixup.y contains
2043      * 1.0 or -1.0 to turn the rendering upside down for offscreen rendering. PosFixup.x
2044      * contains 1.0 to allow a mad, but arb vs swizzles are too restricted for that.
2045      */
2046     shader_addline(buffer, "MUL TMP, posFixup, TMP_OUT.w;\n");
2047     shader_addline(buffer, "ADD TMP_OUT.x, TMP_OUT.x, TMP.z;\n");
2048     shader_addline(buffer, "MAD TMP_OUT.y, TMP_OUT.y, posFixup.y, TMP.w;\n");
2049
2050     /* Z coord [0;1]->[-1;1] mapping, see comment in transform_projection in state.c
2051      * and the glsl equivalent
2052      */
2053     shader_addline(buffer, "MAD TMP_OUT.z, TMP_OUT.z, helper_const.x, -TMP_OUT.w;\n");
2054
2055     shader_addline(buffer, "MOV result.position, TMP_OUT;\n");
2056
2057     shader_addline(buffer, "END\n");
2058
2059     /* TODO: change to resource.glObjectHandle or something like that */
2060     GL_EXTCALL(glGenProgramsARB(1, &ret));
2061
2062     TRACE("Creating a hw vertex shader, prg=%d\n", ret);
2063     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, ret));
2064
2065     TRACE("Created hw vertex shader, prg=%d\n", ret);
2066     /* Create the program and check for errors */
2067     GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2068                buffer->bsize, buffer->buffer));
2069
2070     if (glGetError() == GL_INVALID_OPERATION) {
2071         GLint errPos;
2072         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2073         FIXME("HW VertexShader Error at position %d: %s\n",
2074               errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2075         ret = -1;
2076     } else {
2077         /* Load immediate constants */
2078         if(lconst_map) {
2079             LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2080                 const float *value = (const float *)lconst->value;
2081                 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_VERTEX_PROGRAM_ARB, lconst_map[lconst->idx], value));
2082             }
2083         }
2084     }
2085     HeapFree(GetProcessHeap(), 0, lconst_map);
2086
2087     return ret;
2088 }
2089
2090 static void shader_arb_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct shader_caps *pCaps)
2091 {
2092     /* We don't have an ARB fixed function pipeline yet, so let the none backend set its caps,
2093      * then overwrite the shader specific ones
2094      */
2095     none_shader_backend.shader_get_caps(devtype, gl_info, pCaps);
2096
2097     if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
2098         pCaps->VertexShaderVersion = WINED3DVS_VERSION(1,1);
2099         TRACE_(d3d_caps)("Hardware vertex shader version 1.1 enabled (ARB_PROGRAM)\n");
2100         pCaps->MaxVertexShaderConst = GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS;
2101     }
2102
2103     if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM)) {
2104         pCaps->PixelShaderVersion    = WINED3DPS_VERSION(1,4);
2105         pCaps->PixelShader1xMaxValue = 8.0;
2106         TRACE_(d3d_caps)("Hardware pixel shader version 1.4 enabled (ARB_PROGRAM)\n");
2107         pCaps->MaxPixelShaderConst = GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS;
2108     }
2109 }
2110
2111 static BOOL shader_arb_color_fixup_supported(struct color_fixup_desc fixup)
2112 {
2113     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
2114     {
2115         TRACE("Checking support for color_fixup:\n");
2116         dump_color_fixup_desc(fixup);
2117     }
2118
2119     /* We support everything except YUV conversions. */
2120     if (!is_yuv_fixup(fixup))
2121     {
2122         TRACE("[OK]\n");
2123         return TRUE;
2124     }
2125
2126     TRACE("[FAILED]\n");
2127     return FALSE;
2128 }
2129
2130 static void shader_arb_add_instruction_modifiers(const struct wined3d_shader_instruction *ins) {
2131     DWORD shift;
2132     char write_mask[20], regstr[50];
2133     SHADER_BUFFER *buffer = ins->ctx->buffer;
2134     BOOL is_color = FALSE;
2135     const struct wined3d_shader_dst_param *dst;
2136
2137     if (!ins->dst_count) return;
2138
2139     dst = &ins->dst[0];
2140     shift = dst->shift;
2141     if(shift == 0) return; /* Saturate alone is handled by the instructions */
2142
2143     shader_arb_get_write_mask(ins, dst, write_mask);
2144     shader_arb_get_register_name(ins->ctx->shader, &dst->reg, regstr, &is_color);
2145
2146     /* Generate a line that does the output modifier computation
2147      * FIXME: _SAT vs shift? _SAT alone is already handled in the instructions, if this
2148      * maps problems in e.g. _d4_sat modify shader_arb_get_modifier
2149      */
2150     shader_addline(buffer, "MUL%s %s%s, %s, %s;\n", shader_arb_get_modifier(ins),
2151                    regstr, write_mask, regstr, shift_tab[shift]);
2152 }
2153
2154 static const SHADER_HANDLER shader_arb_instruction_handler_table[WINED3DSIH_TABLE_SIZE] =
2155 {
2156     /* WINED3DSIH_ABS           */ shader_hw_map2gl,
2157     /* WINED3DSIH_ADD           */ shader_hw_map2gl,
2158     /* WINED3DSIH_BEM           */ pshader_hw_bem,
2159     /* WINED3DSIH_BREAK         */ NULL,
2160     /* WINED3DSIH_BREAKC        */ NULL,
2161     /* WINED3DSIH_BREAKP        */ NULL,
2162     /* WINED3DSIH_CALL          */ NULL,
2163     /* WINED3DSIH_CALLNZ        */ NULL,
2164     /* WINED3DSIH_CMP           */ pshader_hw_cmp,
2165     /* WINED3DSIH_CND           */ pshader_hw_cnd,
2166     /* WINED3DSIH_CRS           */ shader_hw_map2gl,
2167     /* WINED3DSIH_DCL           */ NULL,
2168     /* WINED3DSIH_DEF           */ NULL,
2169     /* WINED3DSIH_DEFB          */ NULL,
2170     /* WINED3DSIH_DEFI          */ NULL,
2171     /* WINED3DSIH_DP2ADD        */ pshader_hw_dp2add,
2172     /* WINED3DSIH_DP3           */ shader_hw_map2gl,
2173     /* WINED3DSIH_DP4           */ shader_hw_map2gl,
2174     /* WINED3DSIH_DST           */ shader_hw_map2gl,
2175     /* WINED3DSIH_DSX           */ NULL,
2176     /* WINED3DSIH_DSY           */ NULL,
2177     /* WINED3DSIH_ELSE          */ NULL,
2178     /* WINED3DSIH_ENDIF         */ NULL,
2179     /* WINED3DSIH_ENDLOOP       */ NULL,
2180     /* WINED3DSIH_ENDREP        */ NULL,
2181     /* WINED3DSIH_EXP           */ shader_hw_map2gl,
2182     /* WINED3DSIH_EXPP          */ shader_hw_map2gl,
2183     /* WINED3DSIH_FRC           */ shader_hw_map2gl,
2184     /* WINED3DSIH_IF            */ NULL,
2185     /* WINED3DSIH_IFC           */ NULL,
2186     /* WINED3DSIH_LABEL         */ NULL,
2187     /* WINED3DSIH_LIT           */ shader_hw_map2gl,
2188     /* WINED3DSIH_LOG           */ shader_hw_map2gl,
2189     /* WINED3DSIH_LOGP          */ shader_hw_map2gl,
2190     /* WINED3DSIH_LOOP          */ NULL,
2191     /* WINED3DSIH_LRP           */ shader_hw_map2gl,
2192     /* WINED3DSIH_M3x2          */ shader_hw_mnxn,
2193     /* WINED3DSIH_M3x3          */ shader_hw_mnxn,
2194     /* WINED3DSIH_M3x4          */ shader_hw_mnxn,
2195     /* WINED3DSIH_M4x3          */ shader_hw_mnxn,
2196     /* WINED3DSIH_M4x4          */ shader_hw_mnxn,
2197     /* WINED3DSIH_MAD           */ shader_hw_map2gl,
2198     /* WINED3DSIH_MAX           */ shader_hw_map2gl,
2199     /* WINED3DSIH_MIN           */ shader_hw_map2gl,
2200     /* WINED3DSIH_MOV           */ shader_hw_mov,
2201     /* WINED3DSIH_MOVA          */ shader_hw_mov,
2202     /* WINED3DSIH_MUL           */ shader_hw_map2gl,
2203     /* WINED3DSIH_NOP           */ shader_hw_nop,
2204     /* WINED3DSIH_NRM           */ shader_hw_nrm,
2205     /* WINED3DSIH_PHASE         */ NULL,
2206     /* WINED3DSIH_POW           */ shader_hw_map2gl,
2207     /* WINED3DSIH_RCP           */ shader_hw_rsq_rcp,
2208     /* WINED3DSIH_REP           */ NULL,
2209     /* WINED3DSIH_RET           */ NULL,
2210     /* WINED3DSIH_RSQ           */ shader_hw_rsq_rcp,
2211     /* WINED3DSIH_SETP          */ NULL,
2212     /* WINED3DSIH_SGE           */ shader_hw_map2gl,
2213     /* WINED3DSIH_SGN           */ NULL,
2214     /* WINED3DSIH_SINCOS        */ shader_hw_sincos,
2215     /* WINED3DSIH_SLT           */ shader_hw_map2gl,
2216     /* WINED3DSIH_SUB           */ shader_hw_map2gl,
2217     /* WINED3DSIH_TEX           */ pshader_hw_tex,
2218     /* WINED3DSIH_TEXBEM        */ pshader_hw_texbem,
2219     /* WINED3DSIH_TEXBEML       */ pshader_hw_texbem,
2220     /* WINED3DSIH_TEXCOORD      */ pshader_hw_texcoord,
2221     /* WINED3DSIH_TEXDEPTH      */ pshader_hw_texdepth,
2222     /* WINED3DSIH_TEXDP3        */ pshader_hw_texdp3,
2223     /* WINED3DSIH_TEXDP3TEX     */ pshader_hw_texdp3tex,
2224     /* WINED3DSIH_TEXKILL       */ pshader_hw_texkill,
2225     /* WINED3DSIH_TEXLDD        */ NULL,
2226     /* WINED3DSIH_TEXLDL        */ NULL,
2227     /* WINED3DSIH_TEXM3x2DEPTH  */ pshader_hw_texm3x2depth,
2228     /* WINED3DSIH_TEXM3x2PAD    */ pshader_hw_texm3x2pad,
2229     /* WINED3DSIH_TEXM3x2TEX    */ pshader_hw_texm3x2tex,
2230     /* WINED3DSIH_TEXM3x3       */ pshader_hw_texm3x3,
2231     /* WINED3DSIH_TEXM3x3DIFF   */ NULL,
2232     /* WINED3DSIH_TEXM3x3PAD    */ pshader_hw_texm3x3pad,
2233     /* WINED3DSIH_TEXM3x3SPEC   */ pshader_hw_texm3x3spec,
2234     /* WINED3DSIH_TEXM3x3TEX    */ pshader_hw_texm3x3tex,
2235     /* WINED3DSIH_TEXM3x3VSPEC  */ pshader_hw_texm3x3vspec,
2236     /* WINED3DSIH_TEXREG2AR     */ pshader_hw_texreg2ar,
2237     /* WINED3DSIH_TEXREG2GB     */ pshader_hw_texreg2gb,
2238     /* WINED3DSIH_TEXREG2RGB    */ pshader_hw_texreg2rgb,
2239 };
2240
2241 const shader_backend_t arb_program_shader_backend = {
2242     shader_arb_instruction_handler_table,
2243     shader_arb_select,
2244     shader_arb_select_depth_blt,
2245     shader_arb_deselect_depth_blt,
2246     shader_arb_update_float_vertex_constants,
2247     shader_arb_update_float_pixel_constants,
2248     shader_arb_load_constants,
2249     shader_arb_load_np2fixup_constants,
2250     shader_arb_destroy,
2251     shader_arb_alloc,
2252     shader_arb_free,
2253     shader_arb_dirty_const,
2254     shader_arb_generate_pshader,
2255     shader_arb_generate_vshader,
2256     shader_arb_get_caps,
2257     shader_arb_color_fixup_supported,
2258     shader_arb_add_instruction_modifiers,
2259 };
2260
2261 /* ARB_fragment_program fixed function pipeline replacement definitions */
2262 #define ARB_FFP_CONST_TFACTOR           0
2263 #define ARB_FFP_CONST_SPECULAR_ENABLE   ((ARB_FFP_CONST_TFACTOR) + 1)
2264 #define ARB_FFP_CONST_CONSTANT(i)       ((ARB_FFP_CONST_SPECULAR_ENABLE) + 1 + i)
2265 #define ARB_FFP_CONST_BUMPMAT(i)        ((ARB_FFP_CONST_CONSTANT(7)) + 1 + i)
2266 #define ARB_FFP_CONST_LUMINANCE(i)      ((ARB_FFP_CONST_BUMPMAT(7)) + 1 + i)
2267
2268 struct arbfp_ffp_desc
2269 {
2270     struct ffp_frag_desc parent;
2271     GLuint shader;
2272     unsigned int num_textures_used;
2273 };
2274
2275 static void arbfp_enable(IWineD3DDevice *iface, BOOL enable) {
2276     ENTER_GL();
2277     if(enable) {
2278         glEnable(GL_FRAGMENT_PROGRAM_ARB);
2279         checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
2280     } else {
2281         glDisable(GL_FRAGMENT_PROGRAM_ARB);
2282         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
2283     }
2284     LEAVE_GL();
2285 }
2286
2287 static HRESULT arbfp_alloc(IWineD3DDevice *iface) {
2288     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2289     struct shader_arb_priv *priv;
2290     /* Share private data between the shader backend and the pipeline replacement, if both
2291      * are the arb implementation. This is needed to figure out whether ARBfp should be disabled
2292      * if no pixel shader is bound or not
2293      */
2294     if(This->shader_backend == &arb_program_shader_backend) {
2295         This->fragment_priv = This->shader_priv;
2296     } else {
2297         This->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
2298         if(!This->fragment_priv) return E_OUTOFMEMORY;
2299     }
2300     priv = This->fragment_priv;
2301     priv->fragment_shaders = hash_table_create(ffp_frag_program_key_hash, ffp_frag_program_key_compare);
2302     priv->use_arbfp_fixed_func = TRUE;
2303     return WINED3D_OK;
2304 }
2305
2306 static void arbfp_free_ffpshader(void *value, void *gli) {
2307     const WineD3D_GL_Info *gl_info = gli;
2308     struct arbfp_ffp_desc *entry_arb = value;
2309
2310     ENTER_GL();
2311     GL_EXTCALL(glDeleteProgramsARB(1, &entry_arb->shader));
2312     checkGLcall("glDeleteProgramsARB(1, &entry_arb->shader)");
2313     HeapFree(GetProcessHeap(), 0, entry_arb);
2314     LEAVE_GL();
2315 }
2316
2317 static void arbfp_free(IWineD3DDevice *iface) {
2318     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2319     struct shader_arb_priv *priv = This->fragment_priv;
2320
2321     hash_table_destroy(priv->fragment_shaders, arbfp_free_ffpshader, &This->adapter->gl_info);
2322     priv->use_arbfp_fixed_func = FALSE;
2323
2324     if(This->shader_backend != &arb_program_shader_backend) {
2325         HeapFree(GetProcessHeap(), 0, This->fragment_priv);
2326     }
2327 }
2328
2329 static void arbfp_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct fragment_caps *caps)
2330 {
2331     caps->TextureOpCaps =  WINED3DTEXOPCAPS_DISABLE                     |
2332                            WINED3DTEXOPCAPS_SELECTARG1                  |
2333                            WINED3DTEXOPCAPS_SELECTARG2                  |
2334                            WINED3DTEXOPCAPS_MODULATE4X                  |
2335                            WINED3DTEXOPCAPS_MODULATE2X                  |
2336                            WINED3DTEXOPCAPS_MODULATE                    |
2337                            WINED3DTEXOPCAPS_ADDSIGNED2X                 |
2338                            WINED3DTEXOPCAPS_ADDSIGNED                   |
2339                            WINED3DTEXOPCAPS_ADD                         |
2340                            WINED3DTEXOPCAPS_SUBTRACT                    |
2341                            WINED3DTEXOPCAPS_ADDSMOOTH                   |
2342                            WINED3DTEXOPCAPS_BLENDCURRENTALPHA           |
2343                            WINED3DTEXOPCAPS_BLENDFACTORALPHA            |
2344                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHA           |
2345                            WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA           |
2346                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM         |
2347                            WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR      |
2348                            WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA      |
2349                            WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA   |
2350                            WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR   |
2351                            WINED3DTEXOPCAPS_DOTPRODUCT3                 |
2352                            WINED3DTEXOPCAPS_MULTIPLYADD                 |
2353                            WINED3DTEXOPCAPS_LERP                        |
2354                            WINED3DTEXOPCAPS_BUMPENVMAP                  |
2355                            WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE;
2356
2357     /* TODO: Implement WINED3DTEXOPCAPS_PREMODULATE */
2358
2359     caps->MaxTextureBlendStages   = 8;
2360     caps->MaxSimultaneousTextures = min(GL_LIMITS(fragment_samplers), 8);
2361
2362     caps->PrimitiveMiscCaps |= WINED3DPMISCCAPS_TSSARGTEMP;
2363 }
2364 #undef GLINFO_LOCATION
2365
2366 #define GLINFO_LOCATION stateblock->wineD3DDevice->adapter->gl_info
2367 static void state_texfactor_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2368     float col[4];
2369     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2370
2371     /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2372      * application provided constants
2373      */
2374     if(device->shader_backend == &arb_program_shader_backend) {
2375         if (use_ps(stateblock)) return;
2376
2377         device = stateblock->wineD3DDevice;
2378         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_TFACTOR] = 1;
2379         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_TFACTOR + 1);
2380     }
2381
2382     D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
2383     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col));
2384     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col)");
2385
2386 }
2387
2388 static void state_arb_specularenable(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2389     float col[4];
2390     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2391
2392     /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2393      * application provided constants
2394      */
2395     if(device->shader_backend == &arb_program_shader_backend) {
2396         if (use_ps(stateblock)) return;
2397
2398         device = stateblock->wineD3DDevice;
2399         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_SPECULAR_ENABLE] = 1;
2400         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_SPECULAR_ENABLE + 1);
2401     }
2402
2403     if(stateblock->renderState[WINED3DRS_SPECULARENABLE]) {
2404         /* The specular color has no alpha */
2405         col[0] = 1.0; col[1] = 1.0;
2406         col[2] = 1.0; col[3] = 0.0;
2407     } else {
2408         col[0] = 0.0; col[1] = 0.0;
2409         col[2] = 0.0; col[3] = 0.0;
2410     }
2411     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col));
2412     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col)");
2413 }
2414
2415 static void set_bumpmat_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2416     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2417     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2418     float mat[2][2];
2419
2420     if (use_ps(stateblock))
2421     {
2422         if(stage != 0 &&
2423            ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.bumpmat[stage]) {
2424             /* The pixel shader has to know the bump env matrix. Do a constants update if it isn't scheduled
2425              * anyway
2426              */
2427             if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2428                 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2429             }
2430         }
2431
2432         if(device->shader_backend == &arb_program_shader_backend) {
2433             /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2434             return;
2435         }
2436     } else if(device->shader_backend == &arb_program_shader_backend) {
2437         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_BUMPMAT(stage)] = 1;
2438         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_BUMPMAT(stage) + 1);
2439     }
2440
2441     mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
2442     mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
2443     mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
2444     mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
2445
2446     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0]));
2447     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0])");
2448 }
2449
2450 static void tex_bumpenvlum_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2451     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2452     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2453     float param[4];
2454
2455     if (use_ps(stateblock))
2456     {
2457         if(stage != 0 &&
2458            ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.luminanceparams[stage]) {
2459             /* The pixel shader has to know the luminance offset. Do a constants update if it
2460              * isn't scheduled anyway
2461              */
2462             if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2463                 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2464             }
2465         }
2466
2467         if(device->shader_backend == &arb_program_shader_backend) {
2468             /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2469             return;
2470         }
2471     } else if(device->shader_backend == &arb_program_shader_backend) {
2472         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_LUMINANCE(stage)] = 1;
2473         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_LUMINANCE(stage) + 1);
2474     }
2475
2476     param[0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLSCALE]);
2477     param[1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLOFFSET]);
2478     param[2] = 0.0;
2479     param[3] = 0.0;
2480
2481     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param));
2482     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param)");
2483 }
2484
2485 static const char *get_argreg(SHADER_BUFFER *buffer, DWORD argnum, unsigned int stage, DWORD arg) {
2486     const char *ret;
2487
2488     if(arg == ARG_UNUSED) return "unused"; /* This is the marker for unused registers */
2489
2490     switch(arg & WINED3DTA_SELECTMASK) {
2491         case WINED3DTA_DIFFUSE:
2492             ret = "fragment.color.primary"; break;
2493
2494         case WINED3DTA_CURRENT:
2495             if(stage == 0) ret = "fragment.color.primary";
2496             else ret = "ret";
2497             break;
2498
2499         case WINED3DTA_TEXTURE:
2500             switch(stage) {
2501                 case 0: ret = "tex0"; break;
2502                 case 1: ret = "tex1"; break;
2503                 case 2: ret = "tex2"; break;
2504                 case 3: ret = "tex3"; break;
2505                 case 4: ret = "tex4"; break;
2506                 case 5: ret = "tex5"; break;
2507                 case 6: ret = "tex6"; break;
2508                 case 7: ret = "tex7"; break;
2509                 default: ret = "unknown texture";
2510             }
2511             break;
2512
2513         case WINED3DTA_TFACTOR:
2514             ret = "tfactor"; break;
2515
2516         case WINED3DTA_SPECULAR:
2517             ret = "fragment.color.secondary"; break;
2518
2519         case WINED3DTA_TEMP:
2520             ret = "tempreg"; break;
2521
2522         case WINED3DTA_CONSTANT:
2523             FIXME("Implement perstage constants\n");
2524             switch(stage) {
2525                 case 0: ret = "const0"; break;
2526                 case 1: ret = "const1"; break;
2527                 case 2: ret = "const2"; break;
2528                 case 3: ret = "const3"; break;
2529                 case 4: ret = "const4"; break;
2530                 case 5: ret = "const5"; break;
2531                 case 6: ret = "const6"; break;
2532                 case 7: ret = "const7"; break;
2533                 default: ret = "unknown constant";
2534             }
2535             break;
2536
2537         default:
2538             return "unknown";
2539     }
2540
2541     if(arg & WINED3DTA_COMPLEMENT) {
2542         shader_addline(buffer, "SUB arg%u, const.x, %s;\n", argnum, ret);
2543         if(argnum == 0) ret = "arg0";
2544         if(argnum == 1) ret = "arg1";
2545         if(argnum == 2) ret = "arg2";
2546     }
2547     if(arg & WINED3DTA_ALPHAREPLICATE) {
2548         shader_addline(buffer, "MOV arg%u, %s.w;\n", argnum, ret);
2549         if(argnum == 0) ret = "arg0";
2550         if(argnum == 1) ret = "arg1";
2551         if(argnum == 2) ret = "arg2";
2552     }
2553     return ret;
2554 }
2555
2556 static void gen_ffp_instr(SHADER_BUFFER *buffer, unsigned int stage, BOOL color, BOOL alpha,
2557                           DWORD dst, DWORD op, DWORD dw_arg0, DWORD dw_arg1, DWORD dw_arg2) {
2558     const char *dstmask, *dstreg, *arg0, *arg1, *arg2;
2559     unsigned int mul = 1;
2560     BOOL mul_final_dest = FALSE;
2561
2562     if(color && alpha) dstmask = "";
2563     else if(color) dstmask = ".xyz";
2564     else dstmask = ".w";
2565
2566     if(dst == tempreg) dstreg = "tempreg";
2567     else dstreg = "ret";
2568
2569     arg0 = get_argreg(buffer, 0, stage, dw_arg0);
2570     arg1 = get_argreg(buffer, 1, stage, dw_arg1);
2571     arg2 = get_argreg(buffer, 2, stage, dw_arg2);
2572
2573     switch(op) {
2574         case WINED3DTOP_DISABLE:
2575             if(stage == 0) shader_addline(buffer, "MOV %s%s, fragment.color.primary;\n", dstreg, dstmask);
2576             break;
2577
2578         case WINED3DTOP_SELECTARG2:
2579             arg1 = arg2;
2580         case WINED3DTOP_SELECTARG1:
2581             shader_addline(buffer, "MOV %s%s, %s;\n", dstreg, dstmask, arg1);
2582             break;
2583
2584         case WINED3DTOP_MODULATE4X:
2585             mul = 2;
2586         case WINED3DTOP_MODULATE2X:
2587             mul *= 2;
2588             if(strcmp(dstreg, "result.color") == 0) {
2589                 dstreg = "ret";
2590                 mul_final_dest = TRUE;
2591             }
2592         case WINED3DTOP_MODULATE:
2593             shader_addline(buffer, "MUL %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2594             break;
2595
2596         case WINED3DTOP_ADDSIGNED2X:
2597             mul = 2;
2598             if(strcmp(dstreg, "result.color") == 0) {
2599                 dstreg = "ret";
2600                 mul_final_dest = TRUE;
2601             }
2602         case WINED3DTOP_ADDSIGNED:
2603             shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2604             arg2 = "arg2";
2605         case WINED3DTOP_ADD:
2606             shader_addline(buffer, "ADD_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2607             break;
2608
2609         case WINED3DTOP_SUBTRACT:
2610             shader_addline(buffer, "SUB_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2611             break;
2612
2613         case WINED3DTOP_ADDSMOOTH:
2614             shader_addline(buffer, "SUB arg1, const.x, %s;\n", arg1);
2615             shader_addline(buffer, "MAD_SAT %s%s, arg1, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2616             break;
2617
2618         case WINED3DTOP_BLENDCURRENTALPHA:
2619             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_CURRENT);
2620             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2621             break;
2622         case WINED3DTOP_BLENDFACTORALPHA:
2623             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TFACTOR);
2624             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2625             break;
2626         case WINED3DTOP_BLENDTEXTUREALPHA:
2627             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2628             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2629             break;
2630         case WINED3DTOP_BLENDDIFFUSEALPHA:
2631             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_DIFFUSE);
2632             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2633             break;
2634
2635         case WINED3DTOP_BLENDTEXTUREALPHAPM:
2636             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2637             shader_addline(buffer, "SUB arg0.w, const.x, %s.w;\n", arg0);
2638             shader_addline(buffer, "MAD_SAT %s%s, %s, arg0.w, %s;\n", dstreg, dstmask, arg2, arg1);
2639             break;
2640
2641         /* D3DTOP_PREMODULATE ???? */
2642
2643         case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
2644             shader_addline(buffer, "SUB arg0.w, const.x, %s;\n", arg1);
2645             shader_addline(buffer, "MAD_SAT %s%s, arg0.w, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2646             break;
2647         case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
2648             shader_addline(buffer, "MAD_SAT %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg1);
2649             break;
2650         case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
2651             shader_addline(buffer, "SUB arg0, const.x, %s;\n", arg1);
2652             shader_addline(buffer, "MAD_SAT %s%s, arg0, %s, %s.w;\n", dstreg, dstmask, arg2, arg1);
2653             break;
2654         case WINED3DTOP_MODULATECOLOR_ADDALPHA:
2655             shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s.w;\n", dstreg, dstmask, arg1, arg2, arg1);
2656             break;
2657
2658         case WINED3DTOP_DOTPRODUCT3:
2659             mul = 4;
2660             if(strcmp(dstreg, "result.color") == 0) {
2661                 dstreg = "ret";
2662                 mul_final_dest = TRUE;
2663             }
2664             shader_addline(buffer, "SUB arg1, %s, const.w;\n", arg1);
2665             shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2666             shader_addline(buffer, "DP3_SAT %s%s, arg1, arg2;\n", dstreg, dstmask);
2667             break;
2668
2669         case WINED3DTOP_MULTIPLYADD:
2670             shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg0);
2671             break;
2672
2673         case WINED3DTOP_LERP:
2674             /* The msdn is not quite right here */
2675             shader_addline(buffer, "LRP %s%s, %s, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2676             break;
2677
2678         case WINED3DTOP_BUMPENVMAP:
2679         case WINED3DTOP_BUMPENVMAPLUMINANCE:
2680             /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
2681             break;
2682
2683         default:
2684             FIXME("Unhandled texture op %08x\n", op);
2685     }
2686
2687     if(mul == 2) {
2688         shader_addline(buffer, "MUL_SAT %s%s, %s, const.y;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2689     } else if(mul == 4) {
2690         shader_addline(buffer, "MUL_SAT %s%s, %s, const.z;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2691     }
2692 }
2693
2694 /* The stateblock is passed for GLINFO_LOCATION */
2695 static GLuint gen_arbfp_ffp_shader(const struct ffp_frag_settings *settings, IWineD3DStateBlockImpl *stateblock)
2696 {
2697     unsigned int stage;
2698     SHADER_BUFFER buffer;
2699     BOOL tex_read[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2700     BOOL bump_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2701     BOOL luminance_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2702     const char *textype;
2703     const char *instr, *sat;
2704     char colorcor_dst[8];
2705     GLuint ret;
2706     DWORD arg0, arg1, arg2;
2707     BOOL tempreg_used = FALSE, tfactor_used = FALSE;
2708     BOOL op_equal;
2709     const char *final_combiner_src = "ret";
2710
2711     /* Find out which textures are read */
2712     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2713         if(settings->op[stage].cop == WINED3DTOP_DISABLE) break;
2714         arg0 = settings->op[stage].carg0 & WINED3DTA_SELECTMASK;
2715         arg1 = settings->op[stage].carg1 & WINED3DTA_SELECTMASK;
2716         arg2 = settings->op[stage].carg2 & WINED3DTA_SELECTMASK;
2717         if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2718         if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2719         if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2720
2721         if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) tex_read[stage] = TRUE;
2722         if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHAPM) tex_read[stage] = TRUE;
2723         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP) {
2724             bump_used[stage] = TRUE;
2725             tex_read[stage] = TRUE;
2726         }
2727         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2728             bump_used[stage] = TRUE;
2729             tex_read[stage] = TRUE;
2730             luminance_used[stage] = TRUE;
2731         } else if(settings->op[stage].cop == WINED3DTOP_BLENDFACTORALPHA) {
2732             tfactor_used = TRUE;
2733         }
2734
2735         if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2736             tfactor_used = TRUE;
2737         }
2738
2739         if(settings->op[stage].dst == tempreg) tempreg_used = TRUE;
2740         if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2741             tempreg_used = TRUE;
2742         }
2743
2744         if(settings->op[stage].aop == WINED3DTOP_DISABLE) continue;
2745         arg0 = settings->op[stage].aarg0 & WINED3DTA_SELECTMASK;
2746         arg1 = settings->op[stage].aarg1 & WINED3DTA_SELECTMASK;
2747         arg2 = settings->op[stage].aarg2 & WINED3DTA_SELECTMASK;
2748         if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2749         if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2750         if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2751
2752         if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2753             tempreg_used = TRUE;
2754         }
2755         if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2756             tfactor_used = TRUE;
2757         }
2758     }
2759
2760     /* Shader header */
2761     shader_buffer_init(&buffer);
2762
2763     shader_addline(&buffer, "!!ARBfp1.0\n");
2764
2765     switch(settings->fog) {
2766         case FOG_OFF:                                                         break;
2767         case FOG_LINEAR: shader_addline(&buffer, "OPTION ARB_fog_linear;\n"); break;
2768         case FOG_EXP:    shader_addline(&buffer, "OPTION ARB_fog_exp;\n");    break;
2769         case FOG_EXP2:   shader_addline(&buffer, "OPTION ARB_fog_exp2;\n");   break;
2770         default: FIXME("Unexpected fog setting %d\n", settings->fog);
2771     }
2772
2773     shader_addline(&buffer, "PARAM const = {1, 2, 4, 0.5};\n");
2774     shader_addline(&buffer, "TEMP TMP;\n");
2775     shader_addline(&buffer, "TEMP ret;\n");
2776     if(tempreg_used || settings->sRGB_write) shader_addline(&buffer, "TEMP tempreg;\n");
2777     shader_addline(&buffer, "TEMP arg0;\n");
2778     shader_addline(&buffer, "TEMP arg1;\n");
2779     shader_addline(&buffer, "TEMP arg2;\n");
2780     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2781         if(!tex_read[stage]) continue;
2782         shader_addline(&buffer, "TEMP tex%u;\n", stage);
2783         if(!bump_used[stage]) continue;
2784         shader_addline(&buffer, "PARAM bumpmat%u = program.env[%u];\n", stage, ARB_FFP_CONST_BUMPMAT(stage));
2785         if(!luminance_used[stage]) continue;
2786         shader_addline(&buffer, "PARAM luminance%u = program.env[%u];\n", stage, ARB_FFP_CONST_LUMINANCE(stage));
2787     }
2788     if(tfactor_used) {
2789         shader_addline(&buffer, "PARAM tfactor = program.env[%u];\n", ARB_FFP_CONST_TFACTOR);
2790     }
2791         shader_addline(&buffer, "PARAM specular_enable = program.env[%u];\n", ARB_FFP_CONST_SPECULAR_ENABLE);
2792
2793     if(settings->sRGB_write) {
2794         shader_addline(&buffer, "PARAM srgb_consts1 = {%f, %f, %f, %f};\n",
2795                        srgb_mul_low, srgb_cmp, srgb_pow, srgb_mul_high);
2796         shader_addline(&buffer, "PARAM srgb_consts2 = {%f, %f, %f, %f};\n",
2797                        srgb_sub_high, 0.0, 0.0, 0.0);
2798     }
2799
2800     /* Generate texture sampling instructions) */
2801     for(stage = 0; stage < MAX_TEXTURES && settings->op[stage].cop != WINED3DTOP_DISABLE; stage++) {
2802         if(!tex_read[stage]) continue;
2803
2804         switch(settings->op[stage].tex_type) {
2805             case tex_1d:                    textype = "1D";     break;
2806             case tex_2d:                    textype = "2D";     break;
2807             case tex_3d:                    textype = "3D";     break;
2808             case tex_cube:                  textype = "CUBE";   break;
2809             case tex_rect:                  textype = "RECT";   break;
2810             default: textype = "unexpected_textype";   break;
2811         }
2812
2813         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP ||
2814            settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2815             sat = "";
2816         } else {
2817             sat = "_SAT";
2818         }
2819
2820         if(settings->op[stage].projected == proj_none) {
2821             instr = "TEX";
2822         } else if(settings->op[stage].projected == proj_count4 ||
2823                   settings->op[stage].projected == proj_count3) {
2824             instr = "TXP";
2825         } else {
2826             FIXME("Unexpected projection mode %d\n", settings->op[stage].projected);
2827             instr = "TXP";
2828         }
2829
2830         if(stage > 0 &&
2831            (settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
2832             settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
2833             shader_addline(&buffer, "SWZ arg1, bumpmat%u, x, z, 0, 0;\n", stage - 1);
2834             shader_addline(&buffer, "DP3 ret.x, arg1, tex%u;\n", stage - 1);
2835             shader_addline(&buffer, "SWZ arg1, bumpmat%u, y, w, 0, 0;\n", stage - 1);
2836             shader_addline(&buffer, "DP3 ret.y, arg1, tex%u;\n", stage - 1);
2837
2838             /* with projective textures, texbem only divides the static texture coord, not the displacement,
2839              * so multiply the displacement with the dividing parameter before passing it to TXP
2840              */
2841             if (settings->op[stage].projected != proj_none) {
2842                 if(settings->op[stage].projected == proj_count4) {
2843                     shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].w;\n", stage);
2844                     shader_addline(&buffer, "MUL ret.xyz, ret, fragment.texcoord[%u].w, fragment.texcoord[%u];\n", stage, stage);
2845                 } else {
2846                     shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].z;\n", stage);
2847                     shader_addline(&buffer, "MAD ret.xyz, ret, fragment.texcoord[%u].z, fragment.texcoord[%u];\n", stage, stage);
2848                 }
2849             } else {
2850                 shader_addline(&buffer, "ADD ret, ret, fragment.texcoord[%u];\n", stage);
2851             }
2852
2853             shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2854                            instr, sat, stage, stage, textype);
2855             if(settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2856                 shader_addline(&buffer, "MAD_SAT ret.x, tex%u.z, luminance%u.x, luminance%u.y;\n",
2857                                stage - 1, stage - 1, stage - 1);
2858                 shader_addline(&buffer, "MUL tex%u, tex%u, ret.x;\n", stage, stage);
2859             }
2860         } else if(settings->op[stage].projected == proj_count3) {
2861             shader_addline(&buffer, "MOV ret, fragment.texcoord[%u];\n", stage);
2862             shader_addline(&buffer, "MOV ret.w, ret.z;\n");
2863             shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2864                             instr, sat, stage, stage, textype);
2865         } else {
2866             shader_addline(&buffer, "%s%s tex%u, fragment.texcoord[%u], texture[%u], %s;\n",
2867                             instr, sat, stage, stage, stage, textype);
2868         }
2869
2870         sprintf(colorcor_dst, "tex%u", stage);
2871         gen_color_correction(&buffer, colorcor_dst, WINED3DSP_WRITEMASK_ALL, "const.x", "const.y",
2872                 settings->op[stage].color_fixup);
2873     }
2874
2875     /* Generate the main shader */
2876     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2877         if(settings->op[stage].cop == WINED3DTOP_DISABLE) {
2878             if(stage == 0) {
2879                 final_combiner_src = "fragment.color.primary";
2880             }
2881             break;
2882         }
2883
2884         if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2885            settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2886             op_equal = settings->op[stage].carg1 == settings->op[stage].aarg1;
2887         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2888                   settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2889             op_equal = settings->op[stage].carg1 == settings->op[stage].aarg2;
2890         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2891                   settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2892             op_equal = settings->op[stage].carg2 == settings->op[stage].aarg1;
2893         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2894                   settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2895             op_equal = settings->op[stage].carg2 == settings->op[stage].aarg2;
2896         } else {
2897             op_equal = settings->op[stage].aop   == settings->op[stage].cop &&
2898                        settings->op[stage].carg0 == settings->op[stage].aarg0 &&
2899                        settings->op[stage].carg1 == settings->op[stage].aarg1 &&
2900                        settings->op[stage].carg2 == settings->op[stage].aarg2;
2901         }
2902
2903         if(settings->op[stage].aop == WINED3DTOP_DISABLE) {
2904             gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2905                           settings->op[stage].cop, settings->op[stage].carg0,
2906                           settings->op[stage].carg1, settings->op[stage].carg2);
2907             if(stage == 0) {
2908                 shader_addline(&buffer, "MOV ret.w, fragment.color.primary.w;\n");
2909             }
2910         } else if(op_equal) {
2911             gen_ffp_instr(&buffer, stage, TRUE, TRUE, settings->op[stage].dst,
2912                           settings->op[stage].cop, settings->op[stage].carg0,
2913                           settings->op[stage].carg1, settings->op[stage].carg2);
2914         } else {
2915             gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2916                           settings->op[stage].cop, settings->op[stage].carg0,
2917                           settings->op[stage].carg1, settings->op[stage].carg2);
2918             gen_ffp_instr(&buffer, stage, FALSE, TRUE, settings->op[stage].dst,
2919                           settings->op[stage].aop, settings->op[stage].aarg0,
2920                           settings->op[stage].aarg1, settings->op[stage].aarg2);
2921         }
2922     }
2923
2924     if(settings->sRGB_write) {
2925         shader_addline(&buffer, "MAD ret, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
2926         arbfp_add_sRGB_correction(&buffer, "ret", "arg0", "arg1", "arg2", "tempreg");
2927         shader_addline(&buffer, "MOV result.color.w, ret.w;\n");
2928     } else {
2929         shader_addline(&buffer, "MAD result.color, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
2930     }
2931
2932     /* Footer */
2933     shader_addline(&buffer, "END\n");
2934
2935     /* Generate the shader */
2936     GL_EXTCALL(glGenProgramsARB(1, &ret));
2937     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, ret));
2938     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
2939
2940     if (glGetError() == GL_INVALID_OPERATION) {
2941         GLint pos;
2942         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
2943         FIXME("Fragment program error at position %d: %s\n", pos,
2944               debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2945     }
2946     shader_buffer_free(&buffer);
2947     return ret;
2948 }
2949
2950 static void fragment_prog_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2951     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2952     struct shader_arb_priv *priv = device->fragment_priv;
2953     BOOL use_pshader = use_ps(stateblock);
2954     BOOL use_vshader = use_vs(stateblock);
2955     struct ffp_frag_settings settings;
2956     const struct arbfp_ffp_desc *desc;
2957     unsigned int i;
2958
2959     TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
2960
2961     if(isStateDirty(context, STATE_RENDER(WINED3DRS_FOGENABLE))) {
2962         if(!use_pshader && device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
2963             /* Reload fixed function constants since they collide with the pixel shader constants */
2964             for(i = 0; i < MAX_TEXTURES; i++) {
2965                 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
2966             }
2967             state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
2968             state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
2969         } else if(use_pshader && !isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
2970             device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
2971         }
2972         return;
2973     }
2974
2975     if(!use_pshader) {
2976         /* Find or create a shader implementing the fixed function pipeline settings, then activate it */
2977         gen_ffp_frag_op(stateblock, &settings, FALSE);
2978         desc = (const struct arbfp_ffp_desc *)find_ffp_frag_shader(priv->fragment_shaders, &settings);
2979         if(!desc) {
2980             struct arbfp_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
2981             if (!new_desc)
2982             {
2983                 ERR("Out of memory\n");
2984                 return;
2985             }
2986             new_desc->num_textures_used = 0;
2987             for(i = 0; i < GL_LIMITS(texture_stages); i++) {
2988                 if(settings.op[i].cop == WINED3DTOP_DISABLE) break;
2989                 new_desc->num_textures_used = i;
2990             }
2991
2992             memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
2993             new_desc->shader = gen_arbfp_ffp_shader(&settings, stateblock);
2994             add_ffp_frag_shader(priv->fragment_shaders, &new_desc->parent);
2995             TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
2996             desc = new_desc;
2997         }
2998
2999         /* Now activate the replacement program. GL_FRAGMENT_PROGRAM_ARB is already active(however, note the
3000          * comment above the shader_select call below). If e.g. GLSL is active, the shader_select call will
3001          * deactivate it.
3002          */
3003         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader));
3004         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader)");
3005         priv->current_fprogram_id = desc->shader;
3006
3007         if(device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3008             /* Reload fixed function constants since they collide with the pixel shader constants */
3009             for(i = 0; i < MAX_TEXTURES; i++) {
3010                 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3011             }
3012             state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3013             state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3014         }
3015         context->last_was_pshader = FALSE;
3016     } else {
3017         context->last_was_pshader = TRUE;
3018     }
3019
3020     /* Finally, select the shader. If a pixel shader is used, it will be set and enabled by the shader backend.
3021      * If this shader backend is arbfp(most likely), then it will simply overwrite the last fixed function replace-
3022      * ment shader. If the shader backend is not ARB, it currently is important that the opengl implementation
3023      * type overwrites GL_ARB_fragment_program. This is currently the case with GLSL. If we really want to use
3024      * atifs or nvrc pixel shaders with arb fragment programs we'd have to disable GL_FRAGMENT_PROGRAM_ARB here
3025      *
3026      * Don't call shader_select if the vertex shader is dirty, because it will be called later on by the vertex
3027      * shader handler
3028      */
3029     if(!isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3030         device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3031
3032         if (!isStateDirty(context, STATE_VERTEXSHADERCONSTANT) && (use_vshader || use_pshader)) {
3033             device->StateTable[STATE_VERTEXSHADERCONSTANT].apply(STATE_VERTEXSHADERCONSTANT, stateblock, context);
3034         }
3035     }
3036     if(use_pshader) {
3037         device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
3038     }
3039 }
3040
3041 /* We can't link the fog states to the fragment state directly since the vertex pipeline links them
3042  * to FOGENABLE. A different linking in different pipeline parts can't be expressed in the combined
3043  * state table, so we need to handle that with a forwarding function. The other invisible side effect
3044  * is that changing the fog start and fog end(which links to FOGENABLE in vertex) results in the
3045  * fragment_prog_arbfp function being called because FOGENABLE is dirty, which calls this function here
3046  */
3047 static void state_arbfp_fog(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3048     enum fogsource new_source;
3049
3050     TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
3051
3052     if(!isStateDirty(context, STATE_PIXELSHADER)) {
3053         fragment_prog_arbfp(state, stateblock, context);
3054     }
3055
3056     if(!stateblock->renderState[WINED3DRS_FOGENABLE]) return;
3057
3058     if(stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
3059         if(use_vs(stateblock)) {
3060             new_source = FOGSOURCE_VS;
3061         } else {
3062             if(stateblock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || context->last_was_rhw) {
3063                 new_source = FOGSOURCE_COORD;
3064             } else {
3065                 new_source = FOGSOURCE_FFP;
3066             }
3067         }
3068     } else {
3069         new_source = FOGSOURCE_FFP;
3070     }
3071     if(new_source != context->fog_source) {
3072         context->fog_source = new_source;
3073         state_fogstartend(STATE_RENDER(WINED3DRS_FOGSTART), stateblock, context);
3074     }
3075 }
3076
3077 static void textransform(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3078     if(!isStateDirty(context, STATE_PIXELSHADER)) {
3079         fragment_prog_arbfp(state, stateblock, context);
3080     }
3081 }
3082
3083 #undef GLINFO_LOCATION
3084
3085 static const struct StateEntryTemplate arbfp_fragmentstate_template[] = {
3086     {STATE_RENDER(WINED3DRS_TEXTUREFACTOR),               { STATE_RENDER(WINED3DRS_TEXTUREFACTOR),              state_texfactor_arbfp   }, 0                               },
3087     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3088     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3089     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3090     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3091     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3092     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3093     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3094     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3095     {STATE_TEXTURESTAGE(0, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3096     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3097     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3098     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3099     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3100     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3101     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3102     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3103     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3104     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3105     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3106     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3107     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3108     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3109     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3110     {STATE_TEXTURESTAGE(1, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3111     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3112     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3113     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3114     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3115     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3116     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3117     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3118     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3119     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3120     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3121     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3122     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3123     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3124     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3125     {STATE_TEXTURESTAGE(2, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3126     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3127     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3128     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3129     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3130     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3131     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3132     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3133     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3134     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3135     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3136     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3137     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3138     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3139     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3140     {STATE_TEXTURESTAGE(3, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3141     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3142     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3143     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3144     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3145     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3146     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3147     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3148     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3149     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3150     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3151     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3152     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3153     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3154     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3155     {STATE_TEXTURESTAGE(4, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3156     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3157     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3158     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3159     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3160     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3161     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3162     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3163     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3164     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3165     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3166     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3167     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3168     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3169     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3170     {STATE_TEXTURESTAGE(5, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3171     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3172     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3173     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3174     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3175     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3176     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3177     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3178     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3179     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3180     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3181     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3182     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3183     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3184     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3185     {STATE_TEXTURESTAGE(6, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3186     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3187     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3188     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3189     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3190     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3191     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3192     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3193     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3194     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3195     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3196     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3197     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3198     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3199     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3200     {STATE_TEXTURESTAGE(7, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3201     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3202     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3203     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3204     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3205     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3206     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3207     {STATE_SAMPLER(0),                                    { STATE_SAMPLER(0),                                   sampler_texdim          }, 0                               },
3208     {STATE_SAMPLER(1),                                    { STATE_SAMPLER(1),                                   sampler_texdim          }, 0                               },
3209     {STATE_SAMPLER(2),                                    { STATE_SAMPLER(2),                                   sampler_texdim          }, 0                               },
3210     {STATE_SAMPLER(3),                                    { STATE_SAMPLER(3),                                   sampler_texdim          }, 0                               },
3211     {STATE_SAMPLER(4),                                    { STATE_SAMPLER(4),                                   sampler_texdim          }, 0                               },
3212     {STATE_SAMPLER(5),                                    { STATE_SAMPLER(5),                                   sampler_texdim          }, 0                               },
3213     {STATE_SAMPLER(6),                                    { STATE_SAMPLER(6),                                   sampler_texdim          }, 0                               },
3214     {STATE_SAMPLER(7),                                    { STATE_SAMPLER(7),                                   sampler_texdim          }, 0                               },
3215     {STATE_PIXELSHADER,                                   { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3216     {STATE_RENDER(WINED3DRS_FOGENABLE),                   { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3217     {STATE_RENDER(WINED3DRS_FOGTABLEMODE),                { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3218     {STATE_RENDER(WINED3DRS_FOGVERTEXMODE),               { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3219     {STATE_RENDER(WINED3DRS_FOGSTART),                    { STATE_RENDER(WINED3DRS_FOGSTART),                   state_fogstartend       }, 0                               },
3220     {STATE_RENDER(WINED3DRS_FOGEND),                      { STATE_RENDER(WINED3DRS_FOGSTART),                   state_fogstartend       }, 0                               },
3221     {STATE_RENDER(WINED3DRS_SRGBWRITEENABLE),             { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3222     {STATE_RENDER(WINED3DRS_FOGCOLOR),                    { STATE_RENDER(WINED3DRS_FOGCOLOR),                   state_fogcolor          }, 0                               },
3223     {STATE_RENDER(WINED3DRS_FOGDENSITY),                  { STATE_RENDER(WINED3DRS_FOGDENSITY),                 state_fogdensity        }, 0                               },
3224     {STATE_TEXTURESTAGE(0,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(0, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3225     {STATE_TEXTURESTAGE(1,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(1, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3226     {STATE_TEXTURESTAGE(2,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(2, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3227     {STATE_TEXTURESTAGE(3,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(3, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3228     {STATE_TEXTURESTAGE(4,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(4, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3229     {STATE_TEXTURESTAGE(5,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(5, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3230     {STATE_TEXTURESTAGE(6,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(6, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3231     {STATE_TEXTURESTAGE(7,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(7, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3232     {STATE_RENDER(WINED3DRS_SPECULARENABLE),              { STATE_RENDER(WINED3DRS_SPECULARENABLE),             state_arb_specularenable}, 0                               },
3233     {0 /* Terminate */,                                   { 0,                                                  0                       }, 0                               },
3234 };
3235
3236 const struct fragment_pipeline arbfp_fragment_pipeline = {
3237     arbfp_enable,
3238     arbfp_get_caps,
3239     arbfp_alloc,
3240     arbfp_free,
3241     shader_arb_color_fixup_supported,
3242     arbfp_fragmentstate_template,
3243     TRUE /* We can disable projected textures */
3244 };
3245
3246 #define GLINFO_LOCATION device->adapter->gl_info
3247
3248 struct arbfp_blit_priv {
3249     GLenum yuy2_rect_shader, yuy2_2d_shader;
3250     GLenum uyvy_rect_shader, uyvy_2d_shader;
3251     GLenum yv12_rect_shader, yv12_2d_shader;
3252 };
3253
3254 static HRESULT arbfp_blit_alloc(IWineD3DDevice *iface) {
3255     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3256     device->blit_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct arbfp_blit_priv));
3257     if(!device->blit_priv) {
3258         ERR("Out of memory\n");
3259         return E_OUTOFMEMORY;
3260     }
3261     return WINED3D_OK;
3262 }
3263 static void arbfp_blit_free(IWineD3DDevice *iface) {
3264     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3265     struct arbfp_blit_priv *priv = device->blit_priv;
3266
3267     ENTER_GL();
3268     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_rect_shader));
3269     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_2d_shader));
3270     GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_rect_shader));
3271     GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_2d_shader));
3272     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_rect_shader));
3273     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_2d_shader));
3274     checkGLcall("Delete yuv programs\n");
3275     LEAVE_GL();
3276 }
3277
3278 static BOOL gen_planar_yuv_read(SHADER_BUFFER *buffer, enum yuv_fixup yuv_fixup, GLenum textype, char *luminance)
3279 {
3280     char chroma;
3281     const char *tex, *texinstr;
3282
3283     if (yuv_fixup == YUV_FIXUP_UYVY) {
3284         chroma = 'x';
3285         *luminance = 'w';
3286     } else {
3287         chroma = 'w';
3288         *luminance = 'x';
3289     }
3290     switch(textype) {
3291         case GL_TEXTURE_2D:             tex = "2D";     texinstr = "TXP"; break;
3292         case GL_TEXTURE_RECTANGLE_ARB:  tex = "RECT";   texinstr = "TEX"; break;
3293         default:
3294             /* This is more tricky than just replacing the texture type - we have to navigate
3295              * properly in the texture to find the correct chroma values
3296              */
3297             FIXME("Implement yuv correction for non-2d, non-rect textures\n");
3298             return FALSE;
3299     }
3300
3301     /* First we have to read the chroma values. This means we need at least two pixels(no filtering),
3302      * or 4 pixels(with filtering). To get the unmodified chromas, we have to rid ourselves of the
3303      * filtering when we sample the texture.
3304      *
3305      * These are the rules for reading the chroma:
3306      *
3307      * Even pixel: Cr
3308      * Even pixel: U
3309      * Odd pixel: V
3310      *
3311      * So we have to get the sampling x position in non-normalized coordinates in integers
3312      */
3313     if(textype != GL_TEXTURE_RECTANGLE_ARB) {
3314         shader_addline(buffer, "MUL texcrd.xy, fragment.texcoord[0], size.x;\n");
3315         shader_addline(buffer, "MOV texcrd.w, size.x;\n");
3316     } else {
3317         shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3318     }
3319     /* We must not allow filtering between pixel x and x+1, this would mix U and V
3320      * Vertical filtering is ok. However, bear in mind that the pixel center is at
3321      * 0.5, so add 0.5.
3322      */
3323     shader_addline(buffer, "FLR texcrd.x, texcrd.x;\n");
3324     shader_addline(buffer, "ADD texcrd.x, texcrd.x, coef.y;\n");
3325
3326     /* Divide the x coordinate by 0.5 and get the fraction. This gives 0.25 and 0.75 for the
3327      * even and odd pixels respectively
3328      */
3329     shader_addline(buffer, "MUL texcrd2, texcrd, coef.y;\n");
3330     shader_addline(buffer, "FRC texcrd2, texcrd2;\n");
3331
3332     /* Sample Pixel 1 */
3333     shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3334
3335     /* Put the value into either of the chroma values */
3336     shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3337     shader_addline(buffer, "MUL chroma.x, luminance.%c, temp.x;\n", chroma);
3338     shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3339     shader_addline(buffer, "MUL chroma.y, luminance.%c, temp.x;\n", chroma);
3340
3341     /* Sample pixel 2. If we read an even pixel(SLT above returned 1), sample
3342      * the pixel right to the current one. Otherwise, sample the left pixel.
3343      * Bias and scale the SLT result to -1;1 and add it to the texcrd.x.
3344      */
3345     shader_addline(buffer, "MAD temp.x, temp.x, coef.z, -coef.x;\n");
3346     shader_addline(buffer, "ADD texcrd.x, texcrd, temp.x;\n");
3347     shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3348
3349     /* Put the value into the other chroma */
3350     shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3351     shader_addline(buffer, "MAD chroma.y, luminance.%c, temp.x, chroma.y;\n", chroma);
3352     shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3353     shader_addline(buffer, "MAD chroma.x, luminance.%c, temp.x, chroma.x;\n", chroma);
3354
3355     /* TODO: If filtering is enabled, sample a 2nd pair of pixels left or right of
3356      * the current one and lerp the two U and V values
3357      */
3358
3359     /* This gives the correctly filtered luminance value */
3360     shader_addline(buffer, "TEX luminance, fragment.texcoord[0], texture[0], %s;\n", tex);
3361
3362     return TRUE;
3363 }
3364
3365 static BOOL gen_yv12_read(SHADER_BUFFER *buffer, GLenum textype, char *luminance)
3366 {
3367     const char *tex;
3368
3369     switch(textype) {
3370         case GL_TEXTURE_2D:             tex = "2D";     break;
3371         case GL_TEXTURE_RECTANGLE_ARB:  tex = "RECT";   break;
3372         default:
3373             FIXME("Implement yv12 correction for non-2d, non-rect textures\n");
3374             return FALSE;
3375     }
3376
3377     /* YV12 surfaces contain a WxH sized luminance plane, followed by a (W/2)x(H/2)
3378      * V and a (W/2)x(H/2) U plane, each with 8 bit per pixel. So the effective
3379      * bitdepth is 12 bits per pixel. Since the U and V planes have only half the
3380      * pitch of the luminance plane, the packing into the gl texture is a bit
3381      * unfortunate. If the whole texture is interpreted as luminance data it looks
3382      * approximately like this:
3383      *
3384      *        +----------------------------------+----
3385      *        |                                  |
3386      *        |                                  |
3387      *        |                                  |
3388      *        |                                  |
3389      *        |                                  |   2
3390      *        |            LUMINANCE             |   -
3391      *        |                                  |   3
3392      *        |                                  |
3393      *        |                                  |
3394      *        |                                  |
3395      *        |                                  |
3396      *        +----------------+-----------------+----
3397      *        |                |                 |
3398      *        |  U even rows   |  U odd rows     |
3399      *        |                |                 |   1
3400      *        +----------------+------------------   -
3401      *        |                |                 |   3
3402      *        |  V even rows   |  V odd rows     |
3403      *        |                |                 |
3404      *        +----------------+-----------------+----
3405      *        |                |                 |
3406      *        |     0.5        |       0.5       |
3407      *
3408      * So it appears as if there are 4 chroma images, but in fact the odd rows
3409      * in the chroma images are in the same row as the even ones. So its is
3410      * kinda tricky to read
3411      *
3412      * When reading from rectangle textures, keep in mind that the input y coordinates
3413      * go from 0 to d3d_height, whereas the opengl texture height is 1.5 * d3d_height
3414      */
3415     shader_addline(buffer, "PARAM yv12_coef = {%f, %f, %f, %f};\n",
3416                    2.0 / 3.0, 1.0 / 6.0, (2.0 / 3.0) + (1.0 / 6.0), 1.0 / 3.0);
3417
3418     shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3419     /* the chroma planes have only half the width */
3420     shader_addline(buffer, "MUL texcrd.x, texcrd.x, coef.y;\n");
3421
3422     /* The first value is between 2/3 and 5/6th of the texture's height, so scale+bias
3423      * the coordinate. Also read the right side of the image when reading odd lines
3424      *
3425      * Don't forget to clamp the y values in into the range, otherwise we'll get filtering
3426      * bleeding
3427      */
3428     if(textype == GL_TEXTURE_2D) {
3429
3430         shader_addline(buffer, "RCP chroma.w, size.y;\n");
3431
3432         shader_addline(buffer, "MUL texcrd2.y, texcrd.y, size.y;\n");
3433
3434         shader_addline(buffer, "FLR texcrd2.y, texcrd2.y;\n");
3435         shader_addline(buffer, "MAD texcrd.y, texcrd.y, yv12_coef.y, yv12_coef.x;\n");
3436
3437         /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3438         shader_addline(buffer, "ADD texcrd2.x, texcrd2.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3439         shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3440         shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3441         shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3442
3443         /* clamp, keep the half pixel origin in mind */
3444         shader_addline(buffer, "MAD temp.y, coef.y, chroma.w, yv12_coef.x;\n");
3445         shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3446         shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.z;\n");
3447         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3448     } else {
3449         /* Read from [size - size+size/4] */
3450         shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3451         shader_addline(buffer, "MAD texcrd.y, texcrd.y, coef.w, size.y;\n");
3452
3453         /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3454         shader_addline(buffer, "ADD texcrd2.x, texcrd.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3455         shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3456         shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3457         shader_addline(buffer, "MUL texcrd2.x, texcrd2.x, size.x;\n");
3458         shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3459
3460         /* Make sure to read exactly from the pixel center */
3461         shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3462         shader_addline(buffer, "ADD texcrd.y, texcrd.y, coef.y;\n");
3463
3464         /* Clamp */
3465         shader_addline(buffer, "MAD temp.y, size.y, coef.w, size.y;\n");
3466         shader_addline(buffer, "ADD temp.y, temp.y, -coef.y;\n");
3467         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3468         shader_addline(buffer, "ADD temp.y, size.y, -coef.y;\n");
3469         shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3470     }
3471     /* Read the texture, put the result into the output register */
3472     shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3473     shader_addline(buffer, "MOV chroma.x, temp.w;\n");
3474
3475     /* The other chroma value is 1/6th of the texture lower, from 5/6th to 6/6th
3476      * No need to clamp because we're just reusing the already clamped value from above
3477      */
3478     if(textype == GL_TEXTURE_2D) {
3479         shader_addline(buffer, "ADD texcrd.y, texcrd.y, yv12_coef.y;\n");
3480     } else {
3481         shader_addline(buffer, "MAD texcrd.y, size.y, coef.w, texcrd.y;\n");
3482     }
3483     shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3484     shader_addline(buffer, "MOV chroma.y, temp.w;\n");
3485
3486     /* Sample the luminance value. It is in the top 2/3rd of the texture, so scale the y coordinate.
3487      * Clamp the y coordinate to prevent the chroma values from bleeding into the sampled luminance
3488      * values due to filtering
3489      */
3490     shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3491     if(textype == GL_TEXTURE_2D) {
3492         /* Multiply the y coordinate by 2/3 and clamp it */
3493         shader_addline(buffer, "MUL texcrd.y, texcrd.y, yv12_coef.x;\n");
3494         shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.x;\n");
3495         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3496         shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3497     } else {
3498         /* Reading from texture_rectangles is pretty straightforward, just use the unmodified
3499          * texture coordinate. It is still a good idea to clamp it though, since the opengl texture
3500          * is bigger
3501          */
3502         shader_addline(buffer, "ADD temp.x, size.y, -coef.y;\n");
3503         shader_addline(buffer, "MIN texcrd.y, texcrd.y, size.x;\n");
3504         shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3505     }
3506     *luminance = 'a';
3507
3508     return TRUE;
3509 }
3510
3511 static GLuint gen_yuv_shader(IWineD3DDeviceImpl *device, enum yuv_fixup yuv_fixup, GLenum textype)
3512 {
3513     GLenum shader;
3514     SHADER_BUFFER buffer;
3515     char luminance_component;
3516     struct arbfp_blit_priv *priv = device->blit_priv;
3517
3518     /* Shader header */
3519     shader_buffer_init(&buffer);
3520
3521     ENTER_GL();
3522     GL_EXTCALL(glGenProgramsARB(1, &shader));
3523     checkGLcall("GL_EXTCALL(glGenProgramsARB(1, &shader))");
3524     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3525     checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3526     LEAVE_GL();
3527     if(!shader) {
3528         shader_buffer_free(&buffer);
3529         return 0;
3530     }
3531
3532     /* The YUY2 and UYVY formats contain two pixels packed into a 32 bit macropixel,
3533      * giving effectively 16 bit per pixel. The color consists of a luminance(Y) and
3534      * two chroma(U and V) values. Each macropixel has two luminance values, one for
3535      * each single pixel it contains, and one U and one V value shared between both
3536      * pixels.
3537      *
3538      * The data is loaded into an A8L8 texture. With YUY2, the luminance component
3539      * contains the luminance and alpha the chroma. With UYVY it is vice versa. Thus
3540      * take the format into account when generating the read swizzles
3541      *
3542      * Reading the Y value is straightforward - just sample the texture. The hardware
3543      * takes care of filtering in the horizontal and vertical direction.
3544      *
3545      * Reading the U and V values is harder. We have to avoid filtering horizontally,
3546      * because that would mix the U and V values of one pixel or two adjacent pixels.
3547      * Thus floor the texture coordinate and add 0.5 to get an unfiltered read,
3548      * regardless of the filtering setting. Vertical filtering works automatically
3549      * though - the U and V values of two rows are mixed nicely.
3550      *
3551      * Appart of avoiding filtering issues, the code has to know which value it just
3552      * read, and where it can find the other one. To determine this, it checks if
3553      * it sampled an even or odd pixel, and shifts the 2nd read accordingly.
3554      *
3555      * Handling horizontal filtering of U and V values requires reading a 2nd pair
3556      * of pixels, extracting U and V and mixing them. This is not implemented yet.
3557      *
3558      * An alternative implementation idea is to load the texture as A8R8G8B8 texture,
3559      * with width / 2. This way one read gives all 3 values, finding U and V is easy
3560      * in an unfiltered situation. Finding the luminance on the other hand requires
3561      * finding out if it is an odd or even pixel. The real drawback of this approach
3562      * is filtering. This would have to be emulated completely in the shader, reading
3563      * up two 2 packed pixels in up to 2 rows and interpolating both horizontally and
3564      * vertically. Beyond that it would require adjustments to the texture handling
3565      * code to deal with the width scaling
3566      */
3567     shader_addline(&buffer, "!!ARBfp1.0\n");
3568     shader_addline(&buffer, "TEMP luminance;\n");
3569     shader_addline(&buffer, "TEMP temp;\n");
3570     shader_addline(&buffer, "TEMP chroma;\n");
3571     shader_addline(&buffer, "TEMP texcrd;\n");
3572     shader_addline(&buffer, "TEMP texcrd2;\n");
3573     shader_addline(&buffer, "PARAM coef = {1.0, 0.5, 2.0, 0.25};\n");
3574     shader_addline(&buffer, "PARAM yuv_coef = {1.403, 0.344, 0.714, 1.770};\n");
3575     shader_addline(&buffer, "PARAM size = program.local[0];\n");
3576
3577     switch (yuv_fixup)
3578     {
3579         case YUV_FIXUP_UYVY:
3580         case YUV_FIXUP_YUY2:
3581             if (!gen_planar_yuv_read(&buffer, yuv_fixup, textype, &luminance_component))
3582             {
3583                 shader_buffer_free(&buffer);
3584                 return 0;
3585             }
3586             break;
3587
3588         case YUV_FIXUP_YV12:
3589             if (!gen_yv12_read(&buffer, textype, &luminance_component))
3590             {
3591                 shader_buffer_free(&buffer);
3592                 return 0;
3593             }
3594             break;
3595
3596         default:
3597             FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3598             shader_buffer_free(&buffer);
3599             return 0;
3600     }
3601
3602     /* Calculate the final result. Formula is taken from
3603      * http://www.fourcc.org/fccyvrgb.php. Note that the chroma
3604      * ranges from -0.5 to 0.5
3605      */
3606     shader_addline(&buffer, "SUB chroma.xy, chroma, coef.y;\n");
3607
3608     shader_addline(&buffer, "MAD result.color.x, chroma.x, yuv_coef.x, luminance.%c;\n", luminance_component);
3609     shader_addline(&buffer, "MAD temp.x, -chroma.y, yuv_coef.y, luminance.%c;\n", luminance_component);
3610     shader_addline(&buffer, "MAD result.color.y, -chroma.x, yuv_coef.z, temp.x;\n");
3611     shader_addline(&buffer, "MAD result.color.z, chroma.y, yuv_coef.w, luminance.%c;\n", luminance_component);
3612     shader_addline(&buffer, "END\n");
3613
3614     ENTER_GL();
3615     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3616
3617     if (glGetError() == GL_INVALID_OPERATION) {
3618         GLint pos;
3619         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3620         FIXME("Fragment program error at position %d: %s\n", pos,
3621               debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3622     }
3623     shader_buffer_free(&buffer);
3624     LEAVE_GL();
3625
3626     switch (yuv_fixup)
3627     {
3628         case YUV_FIXUP_YUY2:
3629             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yuy2_rect_shader = shader;
3630             else priv->yuy2_2d_shader = shader;
3631             break;
3632
3633         case YUV_FIXUP_UYVY:
3634             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->uyvy_rect_shader = shader;
3635             else priv->uyvy_2d_shader = shader;
3636             break;
3637
3638         case YUV_FIXUP_YV12:
3639             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yv12_rect_shader = shader;
3640             else priv->yv12_2d_shader = shader;
3641             break;
3642     }
3643
3644     return shader;
3645 }
3646
3647 static HRESULT arbfp_blit_set(IWineD3DDevice *iface, const struct GlPixelFormatDesc *format_desc,
3648         GLenum textype, UINT width, UINT height)
3649 {
3650     GLenum shader;
3651     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3652     float size[4] = {width, height, 1, 1};
3653     struct arbfp_blit_priv *priv = device->blit_priv;
3654     enum yuv_fixup yuv_fixup;
3655
3656     if (!is_yuv_fixup(format_desc->color_fixup))
3657     {
3658         TRACE("Fixup:\n");
3659         dump_color_fixup_desc(format_desc->color_fixup);
3660         /* Don't bother setting up a shader for unconverted formats */
3661         ENTER_GL();
3662         glEnable(textype);
3663         checkGLcall("glEnable(textype)");
3664         LEAVE_GL();
3665         return WINED3D_OK;
3666     }
3667
3668     yuv_fixup = get_yuv_fixup(format_desc->color_fixup);
3669
3670     switch(yuv_fixup)
3671     {
3672         case YUV_FIXUP_YUY2:
3673             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yuy2_rect_shader : priv->yuy2_2d_shader;
3674             break;
3675
3676         case YUV_FIXUP_UYVY:
3677             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->uyvy_rect_shader : priv->uyvy_2d_shader;
3678             break;
3679
3680         case YUV_FIXUP_YV12:
3681             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yv12_rect_shader : priv->yv12_2d_shader;
3682             break;
3683
3684         default:
3685             FIXME("Unsupported YUV fixup %#x, not setting a shader\n", yuv_fixup);
3686             ENTER_GL();
3687             glEnable(textype);
3688             checkGLcall("glEnable(textype)");
3689             LEAVE_GL();
3690             return E_NOTIMPL;
3691     }
3692
3693     if (!shader) shader = gen_yuv_shader(device, yuv_fixup, textype);
3694
3695     ENTER_GL();
3696     glEnable(GL_FRAGMENT_PROGRAM_ARB);
3697     checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
3698     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3699     checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3700     GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, 0, size));
3701     checkGLcall("glProgramLocalParameter4fvARB");
3702     LEAVE_GL();
3703
3704     return WINED3D_OK;
3705 }
3706
3707 static void arbfp_blit_unset(IWineD3DDevice *iface) {
3708     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3709
3710     ENTER_GL();
3711     glDisable(GL_FRAGMENT_PROGRAM_ARB);
3712     checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
3713     glDisable(GL_TEXTURE_2D);
3714     checkGLcall("glDisable(GL_TEXTURE_2D)");
3715     if(GL_SUPPORT(ARB_TEXTURE_CUBE_MAP)) {
3716         glDisable(GL_TEXTURE_CUBE_MAP_ARB);
3717         checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
3718     }
3719     if(GL_SUPPORT(ARB_TEXTURE_RECTANGLE)) {
3720         glDisable(GL_TEXTURE_RECTANGLE_ARB);
3721         checkGLcall("glDisable(GL_TEXTURE_RECTANGLE_ARB)");
3722     }
3723     LEAVE_GL();
3724 }
3725
3726 static BOOL arbfp_blit_color_fixup_supported(struct color_fixup_desc fixup)
3727 {
3728     enum yuv_fixup yuv_fixup;
3729
3730     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
3731     {
3732         TRACE("Checking support for fixup:\n");
3733         dump_color_fixup_desc(fixup);
3734     }
3735
3736     if (is_identity_fixup(fixup))
3737     {
3738         TRACE("[OK]\n");
3739         return TRUE;
3740     }
3741
3742     /* We only support YUV conversions. */
3743     if (!is_yuv_fixup(fixup))
3744     {
3745         TRACE("[FAILED]\n");
3746         return FALSE;
3747     }
3748
3749     yuv_fixup = get_yuv_fixup(fixup);
3750     switch(yuv_fixup)
3751     {
3752         case YUV_FIXUP_YUY2:
3753         case YUV_FIXUP_UYVY:
3754         case YUV_FIXUP_YV12:
3755             TRACE("[OK]\n");
3756             return TRUE;
3757
3758         default:
3759             FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3760             TRACE("[FAILED]\n");
3761             return FALSE;
3762     }
3763 }
3764
3765 const struct blit_shader arbfp_blit = {
3766     arbfp_blit_alloc,
3767     arbfp_blit_free,
3768     arbfp_blit_set,
3769     arbfp_blit_unset,
3770     arbfp_blit_color_fixup_supported,
3771 };
3772
3773 #undef GLINFO_LOCATION