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

   1 /*
   2  * WINED3D draw functions
   3  *
   4  * Copyright 2002-2004 Jason Edmeades
   5  * Copyright 2002-2004 Raphael Junqueira
   6  * Copyright 2004 Christian Costa
   7  * Copyright 2005 Oliver Stieber
   8  * Copyright 2006 Henri Verbeet
   9  * Copyright 2007 Stefan Dösinger for CodeWeavers
  10  *
  11  * This library is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public
  13  * License as published by the Free Software Foundation; either
  14  * version 2.1 of the License, or (at your option) any later version.
  15  *
  16  * This library is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with this library; if not, write to the Free Software
  23  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  24  */
  25
  26 #include "config.h"
  27 #include "wined3d_private.h"
  28
  29 WINE_DEFAULT_DEBUG_CHANNEL(d3d_draw);
  30 #define GLINFO_LOCATION This->adapter->gl_info
  31
  32 #include <stdio.h>
  33 #include <math.h>
  34
  35 #if 0 /* TODO */
  36 extern IWineD3DVertexShaderImpl*            VertexShaders[64];
  37 extern IWineD3DVertexDeclarationImpl*       VertexShaderDeclarations[64];
  38 extern IWineD3DPixelShaderImpl*             PixelShaders[64];
  39
  40 #undef GL_VERSION_1_4 /* To be fixed, caused by mesa headers */
  41 #endif
  42
  43 /* Issues the glBegin call for gl given the primitive type and count */
  44 static DWORD primitiveToGl(WINED3DPRIMITIVETYPE PrimitiveType,
  45                     DWORD            NumPrimitives,
  46                     GLenum          *primType)
  47 {
  48     DWORD   NumVertexes = NumPrimitives;
  49
  50     switch (PrimitiveType) {
  51     case WINED3DPT_POINTLIST:
  52         TRACE("POINTS\n");
  53         *primType   = GL_POINTS;
  54         NumVertexes = NumPrimitives;
  55         break;
  56
  57     case WINED3DPT_LINELIST:
  58         TRACE("LINES\n");
  59         *primType   = GL_LINES;
  60         NumVertexes = NumPrimitives * 2;
  61         break;
  62
  63     case WINED3DPT_LINESTRIP:
  64         TRACE("LINE_STRIP\n");
  65         *primType   = GL_LINE_STRIP;
  66         NumVertexes = NumPrimitives + 1;
  67         break;
  68
  69     case WINED3DPT_TRIANGLELIST:
  70         TRACE("TRIANGLES\n");
  71         *primType   = GL_TRIANGLES;
  72         NumVertexes = NumPrimitives * 3;
  73         break;
  74
  75     case WINED3DPT_TRIANGLESTRIP:
  76         TRACE("TRIANGLE_STRIP\n");
  77         *primType   = GL_TRIANGLE_STRIP;
  78         NumVertexes = NumPrimitives + 2;
  79         break;
  80
  81     case WINED3DPT_TRIANGLEFAN:
  82         TRACE("TRIANGLE_FAN\n");
  83         *primType   = GL_TRIANGLE_FAN;
  84         NumVertexes = NumPrimitives + 2;
  85         break;
  86
  87     default:
  88         FIXME("Unhandled primitive\n");
  89         *primType    = GL_POINTS;
  90         break;
  91     }
  92     return NumVertexes;
  93 }
  94
  95 static BOOL fixed_get_input(
  96     BYTE usage, BYTE usage_idx,
  97     unsigned int* regnum) {
  98
  99     *regnum = -1;
 100
 101     /* Those positions must have the order in the
 102      * named part of the strided data */
 103
 104     if ((usage == WINED3DDECLUSAGE_POSITION || usage == WINED3DDECLUSAGE_POSITIONT) && usage_idx == 0)
 105         *regnum = 0;
 106     else if (usage == WINED3DDECLUSAGE_BLENDWEIGHT && usage_idx == 0)
 107         *regnum = 1;
 108     else if (usage == WINED3DDECLUSAGE_BLENDINDICES && usage_idx == 0)
 109         *regnum = 2;
 110     else if (usage == WINED3DDECLUSAGE_NORMAL && usage_idx == 0)
 111         *regnum = 3;
 112     else if (usage == WINED3DDECLUSAGE_PSIZE && usage_idx == 0)
 113         *regnum = 4;
 114     else if (usage == WINED3DDECLUSAGE_COLOR && usage_idx == 0)
 115         *regnum = 5;
 116     else if (usage == WINED3DDECLUSAGE_COLOR && usage_idx == 1)
 117         *regnum = 6;
 118     else if (usage == WINED3DDECLUSAGE_TEXCOORD && usage_idx < WINED3DDP_MAXTEXCOORD)
 119         *regnum = 7 + usage_idx;
 120     else if ((usage == WINED3DDECLUSAGE_POSITION || usage == WINED3DDECLUSAGE_POSITIONT) && usage_idx == 1)
 121         *regnum = 7 + WINED3DDP_MAXTEXCOORD;
 122     else if (usage == WINED3DDECLUSAGE_NORMAL && usage_idx == 1)
 123         *regnum = 8 + WINED3DDP_MAXTEXCOORD;
 124     else if (usage == WINED3DDECLUSAGE_TANGENT && usage_idx == 0)
 125         *regnum = 9 + WINED3DDP_MAXTEXCOORD;
 126     else if (usage == WINED3DDECLUSAGE_BINORMAL && usage_idx == 0)
 127         *regnum = 10 + WINED3DDP_MAXTEXCOORD;
 128     else if (usage == WINED3DDECLUSAGE_TESSFACTOR && usage_idx == 0)
 129         *regnum = 11 + WINED3DDP_MAXTEXCOORD;
 130     else if (usage == WINED3DDECLUSAGE_FOG && usage_idx == 0)
 131         *regnum = 12 + WINED3DDP_MAXTEXCOORD;
 132     else if (usage == WINED3DDECLUSAGE_DEPTH && usage_idx == 0)
 133         *regnum = 13 + WINED3DDP_MAXTEXCOORD;
 134     else if (usage == WINED3DDECLUSAGE_SAMPLE && usage_idx == 0)
 135         *regnum = 14 + WINED3DDP_MAXTEXCOORD;
 136
 137     if (*regnum == -1) {
 138         FIXME("Unsupported input stream [usage=%s, usage_idx=%u]\n",
 139             debug_d3ddeclusage(usage), usage_idx);
 140         return FALSE;
 141     }
 142     return TRUE;
 143 }
 144
 145 void primitiveDeclarationConvertToStridedData(
 146      IWineD3DDevice *iface,
 147      BOOL useVertexShaderFunction,
 148      WineDirect3DVertexStridedData *strided,
 149      BOOL *fixup) {
 150
 151      /* We need to deal with frequency data!*/
 152
 153     BYTE  *data    = NULL;
 154     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 155     IWineD3DVertexDeclarationImpl* vertexDeclaration = (IWineD3DVertexDeclarationImpl *)This->stateBlock->vertexDecl;
 156     int i;
 157     WINED3DVERTEXELEMENT *element;
 158     DWORD stride;
 159     int reg;
 160     DWORD numPreloadStreams = This->stateBlock->streamIsUP ? 0 : vertexDeclaration->num_streams;
 161     DWORD *streams = vertexDeclaration->streams;
 162
 163     /* Check for transformed vertices, disable vertex shader if present */
 164     strided->u.s.position_transformed = vertexDeclaration->position_transformed;
 165     if(vertexDeclaration->position_transformed) {
 166         useVertexShaderFunction = FALSE;
 167     }
 168
 169     /* Translate the declaration into strided data */
 170     for (i = 0 ; i < vertexDeclaration->declarationWNumElements - 1; ++i) {
 171         GLint streamVBO = 0;
 172         BOOL stride_used;
 173         unsigned int idx;
 174
 175         element = vertexDeclaration->pDeclarationWine + i;
 176         TRACE("%p Element %p (%d of %d)\n", vertexDeclaration->pDeclarationWine,
 177             element,  i + 1, vertexDeclaration->declarationWNumElements - 1);
 178
 179         if (This->stateBlock->streamSource[element->Stream] == NULL)
 180             continue;
 181
 182         stride  = This->stateBlock->streamStride[element->Stream];
 183         if (This->stateBlock->streamIsUP) {
 184             TRACE("Stream is up %d, %p\n", element->Stream, This->stateBlock->streamSource[element->Stream]);
 185             streamVBO = 0;
 186             data    = (BYTE *)This->stateBlock->streamSource[element->Stream];
 187         } else {
 188             TRACE("Stream isn't up %d, %p\n", element->Stream, This->stateBlock->streamSource[element->Stream]);
 189             data    = IWineD3DVertexBufferImpl_GetMemory(This->stateBlock->streamSource[element->Stream], 0, &streamVBO);
 190
 191             /* Can't use vbo's if the base vertex index is negative. OpenGL doesn't accept negative offsets
 192              * (or rather offsets bigger than the vbo, because the pointer is unsigned), so use system memory
 193              * sources. In most sane cases the pointer - offset will still be > 0, otherwise it will wrap
 194              * around to some big value. Hope that with the indices, the driver wraps it back internally. If
 195              * not, drawStridedSlow is needed, including a vertex buffer path.
 196              */
 197             if(This->stateBlock->loadBaseVertexIndex < 0) {
 198                 WARN("loadBaseVertexIndex is < 0 (%d), not using vbos\n", This->stateBlock->loadBaseVertexIndex);
 199                 streamVBO = 0;
 200                 data = ((IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[element->Stream])->resource.allocatedMemory;
 201                 if((UINT_PTR)data < -This->stateBlock->loadBaseVertexIndex * stride) {
 202                     FIXME("System memory vertex data load offset is negative!\n");
 203                 }
 204             }
 205
 206             if(fixup) {
 207                 if( streamVBO != 0) *fixup = TRUE;
 208                 else if(*fixup && !useVertexShaderFunction &&
 209                        (element->Usage == WINED3DDECLUSAGE_COLOR ||
 210                         element->Usage == WINED3DDECLUSAGE_POSITIONT)) {
 211                     /* This may be bad with the fixed function pipeline */
 212                     FIXME("Missing vbo streams with unfixed colors or transformed position, expect problems\n");
 213                 }
 214             }
 215         }
 216         data += element->Offset;
 217         reg = element->Reg;
 218
 219         TRACE("Offset %d Stream %d UsageIndex %d\n", element->Offset, element->Stream, element->UsageIndex);
 220
 221         if (useVertexShaderFunction)
 222             stride_used = vshader_get_input(This->stateBlock->vertexShader,
 223                 element->Usage, element->UsageIndex, &idx);
 224         else
 225             stride_used = fixed_get_input(element->Usage, element->UsageIndex, &idx);
 226
 227         if (stride_used) {
 228             TRACE("Loaded %s array %u [usage=%s, usage_idx=%u, "
 229                     "stream=%u, offset=%u, stride=%u, type=%s, VBO=%u]\n",
 230                     useVertexShaderFunction? "shader": "fixed function", idx,
 231                     debug_d3ddeclusage(element->Usage), element->UsageIndex,
 232                     element->Stream, element->Offset, stride, debug_d3ddecltype(element->Type), streamVBO);
 233
 234             strided->u.input[idx].lpData = data;
 235             strided->u.input[idx].dwType = element->Type;
 236             strided->u.input[idx].dwStride = stride;
 237             strided->u.input[idx].VBO = streamVBO;
 238             strided->u.input[idx].streamNo = element->Stream;
 239         }
 240     }
 241     /* Now call PreLoad on all the vertex buffers. In the very rare case
 242      * that the buffers stopps converting PreLoad will dirtify the VDECL again.
 243      * The vertex buffer can now use the strided structure in the device instead of finding its
 244      * own again.
 245      *
 246      * NULL streams won't be recorded in the array, UP streams won't be either. A stream is only
 247      * once in there.
 248      */
 249     for(i=0; i < numPreloadStreams; i++) {
 250         IWineD3DVertexBuffer *vb = This->stateBlock->streamSource[streams[i]];
 251         if(vb) {
 252             IWineD3DVertexBuffer_PreLoad(vb);
 253         }
 254     }
 255 }
 256
 257 static void drawStridedFast(IWineD3DDevice *iface,UINT numberOfVertices, GLenum glPrimitiveType,
 258                      const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx, ULONG startVertex) {
 259     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 260
 261     if (idxSize != 0 /* This crashes sometimes!*/) {
 262         TRACE("(%p) : glElements(%x, %d, %d, ...)\n", This, glPrimitiveType, numberOfVertices, minIndex);
 263         idxData = idxData == (void *)-1 ? NULL : idxData;
 264 #if 1
 265         glDrawElements(glPrimitiveType, numberOfVertices, idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 266                      (const char *)idxData+(idxSize * startIdx));
 267         checkGLcall("glDrawElements");
 268 #else /* using drawRangeElements may be faster */
 269
 270         glDrawRangeElements(glPrimitiveType, minIndex, minIndex + numberOfVertices - 1, numberOfVertices,
 271                       idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 272                       (const char *)idxData+(idxSize * startIdx));
 273         checkGLcall("glDrawRangeElements");
 274 #endif
 275
 276     } else {
 277
 278         /* Note first is now zero as we shuffled along earlier */
 279         TRACE("(%p) : glDrawArrays(%x, 0, %d)\n", This, glPrimitiveType, numberOfVertices);
 280         glDrawArrays(glPrimitiveType, startVertex, numberOfVertices);
 281         checkGLcall("glDrawArrays");
 282
 283     }
 284
 285     return;
 286 }
 287
 288 /*
 289  * Actually draw using the supplied information.
 290  * Slower GL version which extracts info about each vertex in turn
 291  */
 292
 293 static void drawStridedSlow(IWineD3DDevice *iface, WineDirect3DVertexStridedData *sd,
 294                      UINT NumVertexes, GLenum glPrimType,
 295                      const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx, ULONG startVertex) {
 296
 297     unsigned int               textureNo    = 0;
 298     const WORD                *pIdxBufS     = NULL;
 299     const DWORD               *pIdxBufL     = NULL;
 300     LONG                       vx_index;
 301     DWORD diffuseColor = 0xFFFFFFFF;       /* Diffuse Color              */
 302     DWORD specularColor = 0;               /* Specular Color             */
 303     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 304     UINT *streamOffset = This->stateBlock->streamOffset;
 305     long                      SkipnStrides = startVertex + This->stateBlock->loadBaseVertexIndex;
 306     BOOL                      pixelShader = use_ps(This);
 307
 308     BYTE *texCoords[WINED3DDP_MAXTEXCOORD];
 309     BYTE *diffuse = NULL, *specular = NULL, *normal = NULL, *position = NULL;
 310
 311     TRACE("Using slow vertex array code\n");
 312
 313     /* Variable Initialization */
 314     if (idxSize != 0) {
 315         /* Immediate mode drawing can't make use of indices in a vbo - get the data from the index buffer.
 316          * If the index buffer has no vbo(not supported or other reason), or with user pointer drawing
 317          * idxData will be != NULL
 318          */
 319         if(idxData == NULL) {
 320             idxData = ((IWineD3DIndexBufferImpl *) This->stateBlock->pIndexData)->resource.allocatedMemory;
 321         }
 322
 323         if (idxSize == 2) pIdxBufS = (const WORD *) idxData;
 324         else pIdxBufL = (const DWORD *) idxData;
 325     }
 326
 327     /* Adding the stream offset once is cheaper than doing it every iteration. Do not modify the strided data, it is a pointer
 328      * to the strided Data in the device and might be needed intact on the next draw
 329      */
 330     for (textureNo = 0; textureNo < GL_LIMITS(texture_stages); ++textureNo) {
 331         if(sd->u.s.texCoords[textureNo].lpData) {
 332             texCoords[textureNo] = sd->u.s.texCoords[textureNo].lpData + streamOffset[sd->u.s.texCoords[textureNo].streamNo];
 333         } else {
 334             texCoords[textureNo] = NULL;
 335         }
 336     }
 337     if(sd->u.s.diffuse.lpData) {
 338         diffuse = sd->u.s.diffuse.lpData + streamOffset[sd->u.s.diffuse.streamNo];
 339     }
 340     if(sd->u.s.specular.lpData) {
 341         specular = sd->u.s.specular.lpData + streamOffset[sd->u.s.specular.streamNo];
 342     }
 343     if(sd->u.s.normal.lpData) {
 344         normal = sd->u.s.normal.lpData + streamOffset[sd->u.s.normal.streamNo];
 345     }
 346     if(sd->u.s.position.lpData) {
 347         position = sd->u.s.position.lpData + streamOffset[sd->u.s.position.streamNo];
 348     }
 349
 350     /* The texture coordinate types are not so easy to map into a common function signature - we're
 351      * not using the vector functions here
 352      */
 353     if(FIXME_ON(d3d_draw)) {
 354         for (textureNo = 0; textureNo < GL_LIMITS(textures); ++textureNo) {
 355             DWORD type = sd->u.s.texCoords[textureNo].dwType;
 356             if (sd->u.s.texCoords[textureNo].lpData &&
 357                 type != WINED3DDECLTYPE_FLOAT1 &&
 358                 type != WINED3DDECLTYPE_FLOAT2 &&
 359                 type != WINED3DDECLTYPE_FLOAT3 &&
 360                 type != WINED3DDECLTYPE_FLOAT4) {
 361                 FIXME("Implement fixed function texture coordinates from %s\n", debug_d3ddecltype(type));
 362             }
 363         }
 364         if(specular && This->stateBlock->renderState[WINED3DRS_FOGENABLE] &&
 365            (This->stateBlock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || sd->u.s.position_transformed )&&
 366            This->stateBlock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
 367             if(GL_SUPPORT(EXT_FOG_COORD) && sd->u.s.specular.dwType != WINED3DDECLTYPE_D3DCOLOR) {
 368                 FIXME("Implement fog coordinates from %s\n", debug_d3ddecltype(sd->u.s.specular.dwType));
 369             }
 370         }
 371         if(This->activeContext->num_untracked_materials && sd->u.s.diffuse.dwType != WINED3DDECLTYPE_D3DCOLOR) {
 372             FIXME("Implement diffuse color tracking from %s\n", debug_d3ddecltype(sd->u.s.diffuse.dwType));
 373         }
 374     }
 375
 376     /* Start drawing in GL */
 377     VTRACE(("glBegin(%x)\n", glPrimType));
 378     glBegin(glPrimType);
 379
 380     /* Default settings for data that is not passed */
 381     if (sd->u.s.normal.lpData == NULL) {
 382         glNormal3f(0, 0, 0);
 383     }
 384     if(sd->u.s.diffuse.lpData == NULL) {
 385         glColor4f(1.0f, 1.0f, 1.0f, 1.0f);
 386     }
 387     if(sd->u.s.specular.lpData == NULL) {
 388         if (GL_SUPPORT(EXT_SECONDARY_COLOR)) {
 389             GL_EXTCALL(glSecondaryColor3fEXT)(0, 0, 0);
 390         }
 391     }
 392
 393     /* We shouldn't start this function if any VBO is involved. Should I put a safety check here?
 394      * Guess it's not necessary(we crash then anyway) and would only eat CPU time
 395      */
 396
 397     /* For each primitive */
 398     for (vx_index = 0; vx_index < NumVertexes; ++vx_index) {
 399
 400         /* Initialize diffuse color */
 401         diffuseColor = 0xFFFFFFFF;
 402
 403         /* Blending data and Point sizes are not supported by this function. They are not supported by the fixed
 404          * function pipeline at all. A Fixme for them is printed after decoding the vertex declaration
 405          */
 406
 407         /* For indexed data, we need to go a few more strides in */
 408         if (idxData != NULL) {
 409
 410             /* Indexed so work out the number of strides to skip */
 411             if (idxSize == 2) {
 412                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufS[startIdx+vx_index]));
 413                 SkipnStrides = pIdxBufS[startIdx + vx_index] + This->stateBlock->loadBaseVertexIndex;
 414             } else {
 415                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufL[startIdx+vx_index]));
 416                 SkipnStrides = pIdxBufL[startIdx + vx_index] + This->stateBlock->loadBaseVertexIndex;
 417             }
 418         }
 419
 420         /* Texture coords --------------------------- */
 421         for (textureNo = 0; textureNo < GL_LIMITS(texture_stages); ++textureNo) {
 422
 423             if (!GL_SUPPORT(ARB_MULTITEXTURE) && textureNo > 0) {
 424                 FIXME("Program using multiple concurrent textures which this opengl implementation doesn't support\n");
 425                 continue ;
 426             }
 427
 428             /* Query tex coords */
 429             if (This->stateBlock->textures[textureNo] != NULL || pixelShader) {
 430
 431                 int    coordIdx = This->stateBlock->textureState[textureNo][WINED3DTSS_TEXCOORDINDEX];
 432                 int texture_idx = This->texUnitMap[textureNo];
 433                 float *ptrToCoords = NULL;
 434                 float  s = 0.0, t = 0.0, r = 0.0, q = 0.0;
 435
 436                 if (coordIdx > 7) {
 437                     VTRACE(("tex: %d - Skip tex coords, as being system generated\n", textureNo));
 438                     continue;
 439                 } else if (coordIdx < 0) {
 440                     FIXME("tex: %d - Coord index %d is less than zero, expect a crash.\n", textureNo, coordIdx);
 441                     continue;
 442                 }
 443
 444                 ptrToCoords = (float *)(texCoords[coordIdx] + (SkipnStrides * sd->u.s.texCoords[coordIdx].dwStride));
 445                 if (texCoords[coordIdx] == NULL) {
 446                     TRACE("tex: %d - Skipping tex coords, as no data supplied\n", textureNo);
 447                     if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 448                         GL_EXTCALL(glMultiTexCoord4fARB(GL_TEXTURE0_ARB + texture_idx, 0, 0, 0, 1));
 449                     } else {
 450                         glTexCoord4f(0, 0, 0, 1);
 451                     }
 452                     continue;
 453                 } else {
 454                     int coordsToUse = sd->u.s.texCoords[coordIdx].dwType + 1; /* 0 == WINED3DDECLTYPE_FLOAT1 etc */
 455
 456                     if (texture_idx == -1) continue;
 457
 458                     /* The coords to supply depend completely on the fvf / vertex shader */
 459                     switch (coordsToUse) {
 460                     case 4: q = ptrToCoords[3]; /* drop through */
 461                     case 3: r = ptrToCoords[2]; /* drop through */
 462                     case 2: t = ptrToCoords[1]; /* drop through */
 463                     case 1: s = ptrToCoords[0];
 464                     }
 465
 466                     switch (coordsToUse) {   /* Supply the provided texture coords */
 467                     case WINED3DTTFF_COUNT1:
 468                         VTRACE(("tex:%d, s=%f\n", textureNo, s));
 469                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 470                             GL_EXTCALL(glMultiTexCoord1fARB(GL_TEXTURE0_ARB + texture_idx, s));
 471                         } else {
 472                             glTexCoord1f(s);
 473                         }
 474                         break;
 475                     case WINED3DTTFF_COUNT2:
 476                         VTRACE(("tex:%d, s=%f, t=%f\n", textureNo, s, t));
 477                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 478                             GL_EXTCALL(glMultiTexCoord2fARB(GL_TEXTURE0_ARB + texture_idx, s, t));
 479                         } else {
 480                             glTexCoord2f(s, t);
 481                         }
 482                         break;
 483                     case WINED3DTTFF_COUNT3:
 484                         VTRACE(("tex:%d, s=%f, t=%f, r=%f\n", textureNo, s, t, r));
 485                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 486                             GL_EXTCALL(glMultiTexCoord3fARB(GL_TEXTURE0_ARB + texture_idx, s, t, r));
 487                         } else {
 488                             glTexCoord3f(s, t, r);
 489                         }
 490                         break;
 491                     case WINED3DTTFF_COUNT4:
 492                         VTRACE(("tex:%d, s=%f, t=%f, r=%f, q=%f\n", textureNo, s, t, r, q));
 493                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 494                             GL_EXTCALL(glMultiTexCoord4fARB(GL_TEXTURE0_ARB + texture_idx, s, t, r, q));
 495                         } else {
 496                             glTexCoord4f(s, t, r, q);
 497                         }
 498                         break;
 499                     default:
 500                         FIXME("Should not get here as coordsToUse is two bits only (%x)!\n", coordsToUse);
 501                     }
 502                 }
 503             }
 504         } /* End of textures */
 505
 506         /* Diffuse -------------------------------- */
 507         if (diffuse) {
 508             DWORD *ptrToCoords = (DWORD *)(diffuse + (SkipnStrides * sd->u.s.diffuse.dwStride));
 509
 510             diffuse_funcs[sd->u.s.diffuse.dwType]((void *) ptrToCoords);
 511             if(This->activeContext->num_untracked_materials) {
 512                 unsigned char i;
 513                 float color[4];
 514
 515                 diffuseColor = ptrToCoords[0];
 516                 color[0] = D3DCOLOR_B_R(diffuseColor) / 255.0;
 517                 color[1] = D3DCOLOR_B_G(diffuseColor) / 255.0;
 518                 color[2] = D3DCOLOR_B_B(diffuseColor) / 255.0;
 519                 color[3] = D3DCOLOR_B_A(diffuseColor) / 255.0;
 520
 521                 for(i = 0; i < This->activeContext->num_untracked_materials; i++) {
 522                     glMaterialfv(GL_FRONT_AND_BACK, This->activeContext->untracked_materials[i], color);
 523                 }
 524             }
 525         }
 526
 527         /* Specular ------------------------------- */
 528         if (specular) {
 529             DWORD *ptrToCoords = (DWORD *)(specular + (SkipnStrides * sd->u.s.specular.dwStride));
 530
 531             /* special case where the fog density is stored in the specular alpha channel */
 532             if(This->stateBlock->renderState[WINED3DRS_FOGENABLE] &&
 533               (This->stateBlock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || sd->u.s.position.dwType == WINED3DDECLTYPE_FLOAT4 )&&
 534               This->stateBlock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
 535                 if(GL_SUPPORT(EXT_FOG_COORD)) {
 536                     specularColor = ptrToCoords[0];
 537                     GL_EXTCALL(glFogCoordfEXT(specularColor >> 24));
 538                 } else {
 539                     static BOOL warned = FALSE;
 540                     if(!warned) {
 541                         /* TODO: Use the fog table code from old ddraw */
 542                         FIXME("Implement fog for transformed vertices in software\n");
 543                         warned = TRUE;
 544                     }
 545                 }
 546             }
 547
 548             specular_funcs[sd->u.s.specular.dwType]((void *) ptrToCoords);
 549         }
 550
 551         /* Normal -------------------------------- */
 552         if (normal != NULL) {
 553             float *ptrToCoords = (float *)(normal + (SkipnStrides * sd->u.s.normal.dwStride));
 554             normal_funcs[sd->u.s.normal.dwType](ptrToCoords);
 555         }
 556
 557         /* Position -------------------------------- */
 558         if (position) {
 559             float *ptrToCoords = (float *)(position + (SkipnStrides * sd->u.s.position.dwStride));
 560             position_funcs[sd->u.s.position.dwType](ptrToCoords);
 561         }
 562
 563         /* For non indexed mode, step onto next parts */
 564         if (idxData == NULL) {
 565             ++SkipnStrides;
 566         }
 567     }
 568
 569     glEnd();
 570     checkGLcall("glEnd and previous calls");
 571 }
 572
 573 static inline void send_attribute(IWineD3DDeviceImpl *This, const DWORD type, const UINT index, const void *ptr) {
 574     switch(type) {
 575         case WINED3DDECLTYPE_FLOAT1:
 576             GL_EXTCALL(glVertexAttrib1fvARB(index, (float *) ptr));
 577             break;
 578         case WINED3DDECLTYPE_FLOAT2:
 579             GL_EXTCALL(glVertexAttrib2fvARB(index, (float *) ptr));
 580             break;
 581         case WINED3DDECLTYPE_FLOAT3:
 582             GL_EXTCALL(glVertexAttrib3fvARB(index, (float *) ptr));
 583             break;
 584         case WINED3DDECLTYPE_FLOAT4:
 585             GL_EXTCALL(glVertexAttrib4fvARB(index, (float *) ptr));
 586             break;
 587
 588         case WINED3DDECLTYPE_UBYTE4:
 589             GL_EXTCALL(glVertexAttrib4ubvARB(index, ptr));
 590             break;
 591         case WINED3DDECLTYPE_UBYTE4N:
 592         case WINED3DDECLTYPE_D3DCOLOR:
 593             GL_EXTCALL(glVertexAttrib4NubvARB(index, ptr));
 594             break;
 595
 596         case WINED3DDECLTYPE_SHORT2:
 597             GL_EXTCALL(glVertexAttrib4svARB(index, (GLshort *) ptr));
 598             break;
 599         case WINED3DDECLTYPE_SHORT4:
 600             GL_EXTCALL(glVertexAttrib4svARB(index, (GLshort *) ptr));
 601             break;
 602
 603         case WINED3DDECLTYPE_SHORT2N:
 604         {
 605             GLshort s[4] = {((short *) ptr)[0], ((short *) ptr)[1], 0, 1};
 606             GL_EXTCALL(glVertexAttrib4NsvARB(index, s));
 607             break;
 608         }
 609         case WINED3DDECLTYPE_USHORT2N:
 610         {
 611             GLushort s[4] = {((unsigned short *) ptr)[0], ((unsigned short *) ptr)[1], 0, 1};
 612             GL_EXTCALL(glVertexAttrib4NusvARB(index, s));
 613             break;
 614         }
 615         case WINED3DDECLTYPE_SHORT4N:
 616             GL_EXTCALL(glVertexAttrib4NsvARB(index, (GLshort *) ptr));
 617             break;
 618         case WINED3DDECLTYPE_USHORT4N:
 619             GL_EXTCALL(glVertexAttrib4NusvARB(index, (GLushort *) ptr));
 620             break;
 621
 622         case WINED3DDECLTYPE_UDEC3:
 623             FIXME("Unsure about WINED3DDECLTYPE_UDEC3\n");
 624             /*glVertexAttrib3usvARB(instancedData[j], (GLushort *) ptr); Does not exist */
 625             break;
 626         case WINED3DDECLTYPE_DEC3N:
 627             FIXME("Unsure about WINED3DDECLTYPE_DEC3N\n");
 628             /*glVertexAttrib3NusvARB(instancedData[j], (GLushort *) ptr); Does not exist */
 629             break;
 630
 631         case WINED3DDECLTYPE_FLOAT16_2:
 632             /* Are those 16 bit floats. C doesn't have a 16 bit float type. I could read the single bits and calculate a 4
 633              * byte float according to the IEEE standard
 634              */
 635             if (GL_SUPPORT(NV_HALF_FLOAT)) {
 636                 GL_EXTCALL(glVertexAttrib2hvNV(index, (GLhalfNV *)ptr));
 637             } else {
 638                 float x = float_16_to_32(((unsigned short *) ptr) + 0);
 639                 float y = float_16_to_32(((unsigned short *) ptr) + 1);
 640                 GL_EXTCALL(glVertexAttrib2fARB(index, x, y));
 641             }
 642             break;
 643         case WINED3DDECLTYPE_FLOAT16_4:
 644             if (GL_SUPPORT(NV_HALF_FLOAT)) {
 645                 GL_EXTCALL(glVertexAttrib4hvNV(index, (GLhalfNV *)ptr));
 646             } else {
 647                 float x = float_16_to_32(((unsigned short *) ptr) + 0);
 648                 float y = float_16_to_32(((unsigned short *) ptr) + 1);
 649                 float z = float_16_to_32(((unsigned short *) ptr) + 2);
 650                 float w = float_16_to_32(((unsigned short *) ptr) + 3);
 651                 GL_EXTCALL(glVertexAttrib4fARB(index, x, y, z, w));
 652             }
 653             break;
 654
 655         case WINED3DDECLTYPE_UNUSED:
 656         default:
 657             ERR("Unexpected attribute declaration: %d\n", type);
 658             break;
 659     }
 660 }
 661
 662 static void drawStridedSlowVs(IWineD3DDevice *iface, WineDirect3DVertexStridedData *sd, UINT numberOfVertices,
 663                               GLenum glPrimitiveType, const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx,
 664                               ULONG startVertex) {
 665
 666     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 667     long                      SkipnStrides = startVertex + This->stateBlock->loadBaseVertexIndex;
 668     const WORD                *pIdxBufS     = NULL;
 669     const DWORD               *pIdxBufL     = NULL;
 670     LONG                       vx_index;
 671     int i;
 672     IWineD3DStateBlockImpl *stateblock = (IWineD3DStateBlockImpl *) This->stateBlock;
 673     BYTE *ptr;
 674
 675     if (idxSize != 0) {
 676         /* Immediate mode drawing can't make use of indices in a vbo - get the data from the index buffer.
 677          * If the index buffer has no vbo(not supported or other reason), or with user pointer drawing
 678          * idxData will be != NULL
 679          */
 680         if(idxData == NULL) {
 681             idxData = ((IWineD3DIndexBufferImpl *) stateblock->pIndexData)->resource.allocatedMemory;
 682         }
 683
 684         if (idxSize == 2) pIdxBufS = (const WORD *) idxData;
 685         else pIdxBufL = (const DWORD *) idxData;
 686     }
 687
 688     /* Start drawing in GL */
 689     VTRACE(("glBegin(%x)\n", glPrimType));
 690     glBegin(glPrimitiveType);
 691
 692     for (vx_index = 0; vx_index < numberOfVertices; ++vx_index) {
 693         if (idxData != NULL) {
 694
 695             /* Indexed so work out the number of strides to skip */
 696             if (idxSize == 2) {
 697                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufS[startIdx+vx_index]));
 698                 SkipnStrides = pIdxBufS[startIdx + vx_index] + stateblock->loadBaseVertexIndex;
 699             } else {
 700                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufL[startIdx+vx_index]));
 701                 SkipnStrides = pIdxBufL[startIdx + vx_index] + stateblock->loadBaseVertexIndex;
 702             }
 703         }
 704
 705         for(i = MAX_ATTRIBS - 1; i >= 0; i--) {
 706             if(!sd->u.input[i].lpData) continue;
 707
 708             ptr = sd->u.input[i].lpData +
 709                   sd->u.input[i].dwStride * SkipnStrides +
 710                   stateblock->streamOffset[sd->u.input[i].streamNo];
 711
 712             send_attribute(This, sd->u.input[i].dwType, i, ptr);
 713         }
 714         SkipnStrides++;
 715     }
 716
 717     glEnd();
 718 }
 719
 720 static void depth_blt(IWineD3DDevice *iface, GLuint texture) {
 721     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 722     GLint old_binding = 0;
 723
 724     glPushAttrib(GL_ENABLE_BIT | GL_DEPTH_BUFFER_BIT | GL_COLOR_BUFFER_BIT);
 725
 726     glDisable(GL_CULL_FACE);
 727     glEnable(GL_BLEND);
 728     glDisable(GL_ALPHA_TEST);
 729     glDisable(GL_SCISSOR_TEST);
 730     glDisable(GL_STENCIL_TEST);
 731     glEnable(GL_DEPTH_TEST);
 732     glDepthFunc(GL_ALWAYS);
 733     glBlendFunc(GL_ZERO, GL_ONE);
 734
 735     GL_EXTCALL(glActiveTextureARB(GL_TEXTURE0_ARB));
 736     glGetIntegerv(GL_TEXTURE_BINDING_2D, &old_binding);
 737     glBindTexture(GL_TEXTURE_2D, texture);
 738     glEnable(GL_TEXTURE_2D);
 739
 740     This->shader_backend->shader_select_depth_blt(iface);
 741
 742     glBegin(GL_TRIANGLE_STRIP);
 743     glVertex2f(-1.0f, -1.0f);
 744     glVertex2f(1.0f, -1.0f);
 745     glVertex2f(-1.0f, 1.0f);
 746     glVertex2f(1.0f, 1.0f);
 747     glEnd();
 748
 749     glBindTexture(GL_TEXTURE_2D, old_binding);
 750
 751     glPopAttrib();
 752
 753     /* Reselect the old shaders. There doesn't seem to be any glPushAttrib bit for arb shaders,
 754      * and this seems easier and more efficient than providing the shader backend with a private
 755      * storage to read and restore the old shader settings
 756      */
 757     This->shader_backend->shader_select(iface, use_ps(This), use_vs(This));
 758 }
 759
 760 static void depth_copy(IWineD3DDevice *iface) {
 761     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 762     IWineD3DSurfaceImpl *depth_stencil = (IWineD3DSurfaceImpl *)This->auto_depth_stencil_buffer;
 763
 764     /* Only copy the depth buffer if there is one. */
 765     if (!depth_stencil) return;
 766
 767     /* TODO: Make this work for modes other than FBO */
 768     if (wined3d_settings.offscreen_rendering_mode != ORM_FBO) return;
 769
 770     if (depth_stencil->current_renderbuffer) {
 771         FIXME("Not supported with fixed up depth stencil\n");
 772         return;
 773     }
 774
 775     if (This->render_offscreen) {
 776         static GLuint tmp_texture = 0;
 777         GLint old_binding = 0;
 778
 779         TRACE("Copying onscreen depth buffer to offscreen surface\n");
 780
 781         if (!tmp_texture) {
 782             glGenTextures(1, &tmp_texture);
 783         }
 784
 785         /* Note that we use depth_blt here as well, rather than glCopyTexImage2D
 786          * directly on the FBO texture. That's because we need to flip. */
 787         GL_EXTCALL(glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, 0));
 788         glGetIntegerv(GL_TEXTURE_BINDING_2D, &old_binding);
 789         glBindTexture(GL_TEXTURE_2D, tmp_texture);
 790         glCopyTexImage2D(depth_stencil->glDescription.target,
 791                 depth_stencil->glDescription.level,
 792                 depth_stencil->glDescription.glFormatInternal,
 793                 0,
 794                 0,
 795                 depth_stencil->currentDesc.Width,
 796                 depth_stencil->currentDesc.Height,
 797                 0);
 798         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 799         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 800         glTexParameteri(GL_TEXTURE_2D, GL_DEPTH_TEXTURE_MODE_ARB, GL_LUMINANCE);
 801         glBindTexture(GL_TEXTURE_2D, old_binding);
 802
 803         GL_EXTCALL(glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, This->fbo));
 804         checkGLcall("glBindFramebuffer()");
 805         depth_blt(iface, tmp_texture);
 806         checkGLcall("depth_blt");
 807     } else {
 808         TRACE("Copying offscreen surface to onscreen depth buffer\n");
 809
 810         GL_EXTCALL(glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, 0));
 811         checkGLcall("glBindFramebuffer()");
 812         depth_blt(iface, depth_stencil->glDescription.textureName);
 813         checkGLcall("depth_blt");
 814     }
 815 }
 816
 817 static inline void drawStridedInstanced(IWineD3DDevice *iface, WineDirect3DVertexStridedData *sd, UINT numberOfVertices,
 818                                  GLenum glPrimitiveType, const void *idxData, short idxSize, ULONG minIndex,
 819                                  ULONG startIdx, ULONG startVertex) {
 820     UINT numInstances = 0;
 821     int numInstancedAttribs = 0, i, j;
 822     UINT instancedData[sizeof(sd->u.input) / sizeof(sd->u.input[0]) /* 16 */];
 823     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 824     IWineD3DStateBlockImpl *stateblock = This->stateBlock;
 825
 826     if (idxSize == 0) {
 827         /* This is a nasty thing. MSDN says no hardware supports that and apps have to use software vertex processing.
 828          * We don't support this for now
 829          *
 830          * Shouldn't be too hard to support with opengl, in theory just call glDrawArrays instead of drawElements.
 831          * But the StreamSourceFreq value has a different meaning in that situation.
 832          */
 833         FIXME("Non-indexed instanced drawing is not supported\n");
 834         return;
 835     }
 836
 837     TRACE("(%p) : glElements(%x, %d, %d, ...)\n", This, glPrimitiveType, numberOfVertices, minIndex);
 838     idxData = idxData == (void *)-1 ? NULL : idxData;
 839
 840     /* First, figure out how many instances we have to draw */
 841     for(i = 0; i < MAX_STREAMS; i++) {
 842         /* Look at all non-instanced streams */
 843         if(!(stateblock->streamFlags[i] & WINED3DSTREAMSOURCE_INSTANCEDATA) &&
 844            stateblock->streamSource[i]) {
 845             int inst = stateblock->streamFreq[i];
 846
 847             if(numInstances && inst != numInstances) {
 848                 ERR("Two streams specify a different number of instances. Got %d, new is %d\n", numInstances, inst);
 849             }
 850             numInstances = inst;
 851         }
 852     }
 853
 854     for(i = 0; i < sizeof(sd->u.input) / sizeof(sd->u.input[0]); i++) {
 855         if(stateblock->streamFlags[sd->u.input[i].streamNo] & WINED3DSTREAMSOURCE_INSTANCEDATA) {
 856             instancedData[numInstancedAttribs] = i;
 857             numInstancedAttribs++;
 858         }
 859     }
 860
 861     /* now draw numInstances instances :-) */
 862     for(i = 0; i < numInstances; i++) {
 863         /* Specify the instanced attributes using immediate mode calls */
 864         for(j = 0; j < numInstancedAttribs; j++) {
 865             BYTE *ptr = sd->u.input[instancedData[j]].lpData +
 866                         sd->u.input[instancedData[j]].dwStride * i +
 867                         stateblock->streamOffset[sd->u.input[instancedData[j]].streamNo];
 868             if(sd->u.input[instancedData[j]].VBO) {
 869                 IWineD3DVertexBufferImpl *vb = (IWineD3DVertexBufferImpl *) stateblock->streamSource[sd->u.input[instancedData[j]].streamNo];
 870                 ptr += (long) vb->resource.allocatedMemory;
 871             }
 872
 873             send_attribute(This, sd->u.input[instancedData[j]].dwType, instancedData[j], ptr);
 874         }
 875
 876         glDrawElements(glPrimitiveType, numberOfVertices, idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 877                     (const char *)idxData+(idxSize * startIdx));
 878         checkGLcall("glDrawElements");
 879     }
 880 }
 881
 882 static inline void remove_vbos(IWineD3DDeviceImpl *This, WineDirect3DVertexStridedData *s) {
 883     unsigned char i;
 884     IWineD3DVertexBufferImpl *vb;
 885
 886     if(s->u.s.position.VBO) {
 887         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.position.streamNo];
 888         s->u.s.position.VBO = 0;
 889         s->u.s.position.lpData = (BYTE *) ((unsigned long) s->u.s.position.lpData + (unsigned long) vb->resource.allocatedMemory);
 890     }
 891     if(s->u.s.blendWeights.VBO) {
 892         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.blendWeights.streamNo];
 893         s->u.s.blendWeights.VBO = 0;
 894         s->u.s.blendWeights.lpData = (BYTE *) ((unsigned long) s->u.s.blendWeights.lpData + (unsigned long) vb->resource.allocatedMemory);
 895     }
 896     if(s->u.s.blendMatrixIndices.VBO) {
 897         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.blendMatrixIndices.streamNo];
 898         s->u.s.blendMatrixIndices.VBO = 0;
 899         s->u.s.blendMatrixIndices.lpData = (BYTE *) ((unsigned long) s->u.s.blendMatrixIndices.lpData + (unsigned long) vb->resource.allocatedMemory);
 900     }
 901     if(s->u.s.normal.VBO) {
 902         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.normal.streamNo];
 903         s->u.s.normal.VBO = 0;
 904         s->u.s.normal.lpData = (BYTE *) ((unsigned long) s->u.s.normal.lpData + (unsigned long) vb->resource.allocatedMemory);
 905     }
 906     if(s->u.s.pSize.VBO) {
 907         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.pSize.streamNo];
 908         s->u.s.pSize.VBO = 0;
 909         s->u.s.pSize.lpData = (BYTE *) ((unsigned long) s->u.s.pSize.lpData + (unsigned long) vb->resource.allocatedMemory);
 910     }
 911     if(s->u.s.diffuse.VBO) {
 912         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.diffuse.streamNo];
 913         s->u.s.diffuse.VBO = 0;
 914         s->u.s.diffuse.lpData = (BYTE *) ((unsigned long) s->u.s.diffuse.lpData + (unsigned long) vb->resource.allocatedMemory);
 915     }
 916     if(s->u.s.specular.VBO) {
 917         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.specular.streamNo];
 918         s->u.s.specular.VBO = 0;
 919         s->u.s.specular.lpData = (BYTE *) ((unsigned long) s->u.s.specular.lpData + (unsigned long) vb->resource.allocatedMemory);
 920     }
 921     for(i = 0; i < WINED3DDP_MAXTEXCOORD; i++) {
 922         if(s->u.s.texCoords[i].VBO) {
 923             vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.texCoords[i].streamNo];
 924             s->u.s.texCoords[i].VBO = 0;
 925             s->u.s.texCoords[i].lpData = (BYTE *) ((unsigned long) s->u.s.texCoords[i].lpData + (unsigned long) vb->resource.allocatedMemory);
 926         }
 927     }
 928     if(s->u.s.position2.VBO) {
 929         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.position2.streamNo];
 930         s->u.s.position2.VBO = 0;
 931         s->u.s.position2.lpData = (BYTE *) ((unsigned long) s->u.s.position2.lpData + (unsigned long) vb->resource.allocatedMemory);
 932     }
 933     if(s->u.s.normal2.VBO) {
 934         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.normal2.streamNo];
 935         s->u.s.normal2.VBO = 0;
 936         s->u.s.normal2.lpData = (BYTE *) ((unsigned long) s->u.s.normal2.lpData + (unsigned long) vb->resource.allocatedMemory);
 937     }
 938     if(s->u.s.tangent.VBO) {
 939         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.tangent.streamNo];
 940         s->u.s.tangent.VBO = 0;
 941         s->u.s.tangent.lpData = (BYTE *) ((unsigned long) s->u.s.tangent.lpData + (unsigned long) vb->resource.allocatedMemory);
 942     }
 943     if(s->u.s.binormal.VBO) {
 944         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.binormal.streamNo];
 945         s->u.s.binormal.VBO = 0;
 946         s->u.s.binormal.lpData = (BYTE *) ((unsigned long) s->u.s.binormal.lpData + (unsigned long) vb->resource.allocatedMemory);
 947     }
 948     if(s->u.s.tessFactor.VBO) {
 949         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.tessFactor.streamNo];
 950         s->u.s.tessFactor.VBO = 0;
 951         s->u.s.tessFactor.lpData = (BYTE *) ((unsigned long) s->u.s.tessFactor.lpData + (unsigned long) vb->resource.allocatedMemory);
 952     }
 953     if(s->u.s.fog.VBO) {
 954         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.fog.streamNo];
 955         s->u.s.fog.VBO = 0;
 956         s->u.s.fog.lpData = (BYTE *) ((unsigned long) s->u.s.fog.lpData + (unsigned long) vb->resource.allocatedMemory);
 957     }
 958     if(s->u.s.depth.VBO) {
 959         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.depth.streamNo];
 960         s->u.s.depth.VBO = 0;
 961         s->u.s.depth.lpData = (BYTE *) ((unsigned long) s->u.s.depth.lpData + (unsigned long) vb->resource.allocatedMemory);
 962     }
 963     if(s->u.s.sample.VBO) {
 964         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.sample.streamNo];
 965         s->u.s.sample.VBO = 0;
 966         s->u.s.sample.lpData = (BYTE *) ((unsigned long) s->u.s.sample.lpData + (unsigned long) vb->resource.allocatedMemory);
 967     }
 968 }
 969
 970 /* Routine common to the draw primitive and draw indexed primitive routines */
 971 void drawPrimitive(IWineD3DDevice *iface,
 972                    int PrimitiveType,
 973                    long NumPrimitives,
 974                    /* for Indexed: */
 975                    long  StartVertexIndex,
 976                    UINT  numberOfVertices,
 977                    long  StartIdx,
 978                    short idxSize,
 979                    const void *idxData,
 980                    int   minIndex) {
 981
 982     IWineD3DDeviceImpl           *This = (IWineD3DDeviceImpl *)iface;
 983     IWineD3DSwapChain            *swapchain;
 984     IWineD3DBaseTexture          *texture = NULL;
 985     IWineD3DSurfaceImpl          *target;
 986     int i;
 987
 988     /* Signals other modules that a drawing is in progress and the stateblock finalized */
 989     This->isInDraw = TRUE;
 990
 991     /* Invalidate the back buffer memory so LockRect will read it the next time */
 992     for(i = 0; i < GL_LIMITS(buffers); i++) {
 993         target = (IWineD3DSurfaceImpl *) This->render_targets[i];
 994
 995         /* TODO: Only do all that if we're going to change anything
 996          * Texture container dirtification does not work quite right yet
 997          */
 998         if(target /*&& target->Flags & (SFLAG_INTEXTURE | SFLAG_INSYSMEM)*/) {
 999             swapchain = NULL;
1000             texture = NULL;
1001
1002             if(i == 0) {
1003                 IWineD3DSurface_GetContainer((IWineD3DSurface *) target, &IID_IWineD3DSwapChain, (void **)&swapchain);
1004
1005                 /* Need the surface in the drawable! */
1006                 IWineD3DSurface_LoadLocation((IWineD3DSurface *) target, SFLAG_INDRAWABLE, NULL);
1007
1008                 /* TODO: Move fbo logic to ModifyLocation */
1009                 IWineD3DSurface_ModifyLocation((IWineD3DSurface *) target, SFLAG_INDRAWABLE, TRUE);
1010                 if(swapchain) {
1011                     /* Onscreen target. Invalidate system memory copy and texture copy */
1012                     IWineD3DSwapChain_Release(swapchain);
1013                 } else if(wined3d_settings.offscreen_rendering_mode != ORM_FBO) {
1014                     /* Non-FBO target: Invalidate system copy, texture copy and dirtify the container */
1015                     /* TODO: Move container dirtification to ModifyLocation */
1016                     IWineD3DSurface_GetContainer((IWineD3DSurface *) target, &IID_IWineD3DBaseTexture, (void **)&texture);
1017
1018                     if(texture) {
1019                         IWineD3DBaseTexture_SetDirty(texture, TRUE);
1020                         IWineD3DTexture_Release(texture);
1021                     }
1022                 } else {
1023                     /* FBO offscreen target. Texture == Drawable */
1024                     target->Flags |= SFLAG_INTEXTURE;
1025                 }
1026             } else {
1027                 /* Must be an fbo render target */
1028                 IWineD3DSurface_ModifyLocation((IWineD3DSurface *) target, SFLAG_INDRAWABLE, TRUE);
1029                 target->Flags |=  SFLAG_INTEXTURE;
1030             }
1031         }
1032     }
1033
1034     /* Ok, we will be updating the screen from here onwards so grab the lock */
1035
1036     if (wined3d_settings.offscreen_rendering_mode == ORM_FBO) {
1037         ENTER_GL();
1038         apply_fbo_state(iface);
1039         LEAVE_GL();
1040     }
1041
1042     ActivateContext(This, This->render_targets[0], CTXUSAGE_DRAWPRIM);
1043     ENTER_GL();
1044
1045     if (This->depth_copy_state == WINED3D_DCS_COPY) {
1046         depth_copy(iface);
1047     }
1048     This->depth_copy_state = WINED3D_DCS_INITIAL;
1049
1050     {
1051         GLenum glPrimType;
1052         BOOL emulation = FALSE;
1053         WineDirect3DVertexStridedData *strided = &This->strided_streams;
1054         WineDirect3DVertexStridedData stridedlcl;
1055         /* Ok, Work out which primitive is requested and how many vertexes that
1056            will be                                                              */
1057         UINT calculatedNumberOfindices = primitiveToGl(PrimitiveType, NumPrimitives, &glPrimType);
1058         if (numberOfVertices == 0 )
1059             numberOfVertices = calculatedNumberOfindices;
1060
1061         if(!use_vs(This)) {
1062             if(!This->strided_streams.u.s.position_transformed && This->activeContext->num_untracked_materials &&
1063                 This->stateBlock->renderState[WINED3DRS_LIGHTING]) {
1064                 FIXME("Using software emulation because not all material properties could be tracked\n");
1065                 emulation = TRUE;
1066             }
1067             else if(This->activeContext->fog_coord && This->stateBlock->renderState[WINED3DRS_FOGENABLE]) {
1068                 /* Either write a pipeline replacement shader or convert the specular alpha from unsigned byte
1069                  * to a float in the vertex buffer
1070                  */
1071                 FIXME("Using software emulation because manual fog coordinates are provided\n");
1072                 emulation = TRUE;
1073             }
1074
1075             if(emulation) {
1076                 strided = &stridedlcl;
1077                 memcpy(&stridedlcl, &This->strided_streams, sizeof(stridedlcl));
1078                 remove_vbos(This, &stridedlcl);
1079             }
1080         }
1081
1082         if (This->useDrawStridedSlow || emulation) {
1083             /* Immediate mode drawing */
1084             if(use_vs(This)) {
1085                 FIXME("Using immediate mode with vertex shaders for half float emulation\n");
1086                 drawStridedSlowVs(iface, strided, calculatedNumberOfindices, glPrimType,
1087                                   idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1088             } else {
1089                 drawStridedSlow(iface, strided, calculatedNumberOfindices,
1090                                 glPrimType, idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1091             }
1092         } else if(This->instancedDraw) {
1093             /* Instancing emulation with mixing immediate mode and arrays */
1094             drawStridedInstanced(iface, &This->strided_streams, calculatedNumberOfindices, glPrimType,
1095                             idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1096         } else {
1097             drawStridedFast(iface, calculatedNumberOfindices, glPrimType,
1098                             idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1099         }
1100     }
1101
1102     /* Finshed updating the screen, restore lock */
1103     LEAVE_GL();
1104     TRACE("Done all gl drawing\n");
1105
1106     /* Diagnostics */
1107 #ifdef SHOW_FRAME_MAKEUP
1108     {
1109         static long int primCounter = 0;
1110         /* NOTE: set primCounter to the value reported by drawprim
1111            before you want to to write frame makeup to /tmp */
1112         if (primCounter >= 0) {
1113             WINED3DLOCKED_RECT r;
1114             char buffer[80];
1115             IWineD3DSurface_LockRect(This->renderTarget, &r, NULL, WINED3DLOCK_READONLY);
1116             sprintf(buffer, "/tmp/backbuffer_%d.tga", primCounter);
1117             TRACE("Saving screenshot %s\n", buffer);
1118             IWineD3DSurface_SaveSnapshot(This->renderTarget, buffer);
1119             IWineD3DSurface_UnlockRect(This->renderTarget);
1120
1121 #ifdef SHOW_TEXTURE_MAKEUP
1122            {
1123             IWineD3DSurface *pSur;
1124             int textureNo;
1125             for (textureNo = 0; textureNo < MAX_COMBINED_SAMPLERS; ++textureNo) {
1126                 if (This->stateBlock->textures[textureNo] != NULL) {
1127                     sprintf(buffer, "/tmp/texture_%p_%d_%d.tga", This->stateBlock->textures[textureNo], primCounter, textureNo);
1128                     TRACE("Saving texture %s\n", buffer);
1129                     if (IWineD3DBaseTexture_GetType(This->stateBlock->textures[textureNo]) == WINED3DRTYPE_TEXTURE) {
1130                             IWineD3DTexture_GetSurfaceLevel((IWineD3DTexture *)This->stateBlock->textures[textureNo], 0, &pSur);
1131                             IWineD3DSurface_SaveSnapshot(pSur, buffer);
1132                             IWineD3DSurface_Release(pSur);
1133                     } else  {
1134                         FIXME("base Texture isn't of type texture %d\n", IWineD3DBaseTexture_GetType(This->stateBlock->textures[textureNo]));
1135                     }
1136                 }
1137             }
1138            }
1139 #endif
1140         }
1141         TRACE("drawprim #%d\n", primCounter);
1142         ++primCounter;
1143     }
1144 #endif
1145
1146     /* Control goes back to the device, stateblock values may change again */
1147     This->isInDraw = FALSE;
1148 }
1149
1150 static void normalize_normal(float *n) {
1151     float length = n[0] * n[0] + n[1] * n[1] + n[2] * n[2];
1152     if(length == 0.0) return;
1153     length = sqrt(length);
1154     n[0] = n[0] / length;
1155     n[1] = n[1] / length;
1156     n[2] = n[2] / length;
1157 }
1158
1159 /* Tesselates a high order rectangular patch into single triangles using gl evaluators
1160  *
1161  * The problem is that OpenGL does not offer a direct way to return the tesselated primitives,
1162  * and they can't be sent off for rendering directly either. Tesselating is slow, so we want
1163  * to chache the patches in a vertex buffer. But more importantly, gl can't bind generated
1164  * attributes to numbered shader attributes, so we have to store them and rebind them as needed
1165  * in drawprim.
1166  *
1167  * To read back, the opengl feedback mode is used. This creates a proplem because we want
1168  * untransformed, unlit vertices, but feedback runs everything through transform and lighting.
1169  * Thus disable lighting and set identity matrices to get unmodified colors and positions.
1170  * To overcome clipping find the biggest x, y and z values of the vertices in the patch and scale
1171  * them to [-1.0;+1.0] and set the viewport up to scale them back.
1172  *
1173  * Normals are more tricky: Draw white vertices with 3 directional lights, and calculate the
1174  * resulting colors back to the normals.
1175  *
1176  * NOTE: This function activates a context for blitting, modifies matrices & viewport, but
1177  * does not restore it because normally a draw follows immediately afterwards. The caller is
1178  * responsible of taking care that either the gl states are restored, or the context activated
1179  * for drawing to reset the lastWasBlit flag.
1180  */
1181 HRESULT tesselate_rectpatch(IWineD3DDeviceImpl *This,
1182                             struct WineD3DRectPatch *patch) {
1183     unsigned int i, j, num_quads, out_vertex_size, buffer_size, d3d_out_vertex_size;
1184     float max_x = 0.0, max_y = 0.0, max_z = 0.0, neg_z = 0.0;
1185     WineDirect3DVertexStridedData strided;
1186     BYTE *data;
1187     WINED3DRECTPATCH_INFO *info = &patch->RectPatchInfo;
1188     DWORD vtxStride;
1189     GLenum feedback_type;
1190     GLfloat *feedbuffer;
1191
1192     /* First, locate the position data. This is provided in a vertex buffer in the stateblock.
1193      * Beware of vbos
1194      */
1195     memset(&strided, 0, sizeof(strided));
1196     primitiveDeclarationConvertToStridedData((IWineD3DDevice *) This, FALSE, &strided, NULL);
1197     if(strided.u.s.position.VBO) {
1198         IWineD3DVertexBufferImpl *vb;
1199         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[strided.u.s.position.streamNo];
1200         strided.u.s.position.lpData = (BYTE *) ((unsigned long) strided.u.s.position.lpData +
1201                                                 (unsigned long) vb->resource.allocatedMemory);
1202     }
1203     vtxStride = strided.u.s.position.dwStride;
1204     data = strided.u.s.position.lpData +
1205            vtxStride * info->Stride * info->StartVertexOffsetHeight +
1206            vtxStride * info->StartVertexOffsetWidth;
1207
1208     /* Not entirely sure about what happens with transformed vertices */
1209     if(strided.u.s.position_transformed) {
1210         FIXME("Transformed position in rectpatch generation\n");
1211     }
1212     if(vtxStride % sizeof(GLfloat)) {
1213         /* glMap2f reads vertex sizes in GLfloats, the d3d stride is in bytes.
1214          * I don't see how the stride could not be a multiple of 4, but make sure
1215          * to check it
1216          */
1217         ERR("Vertex stride is not a multiple of sizeof(GLfloat)\n");
1218     }
1219     if(info->Basis != WINED3DBASIS_BEZIER) {
1220         FIXME("Basis is %s, how to handle this?\n", debug_d3dbasis(info->Basis));
1221     }
1222     if(info->Degree != WINED3DDEGREE_CUBIC) {
1223         FIXME("Degree is %s, how to handle this?\n", debug_d3ddegree(info->Degree));
1224     }
1225
1226     /* First, get the boundary cube of the input data */
1227     for(j = 0; j < info->Height; j++) {
1228         for(i = 0; i < info->Width; i++) {
1229             float *v = (float *) (data + vtxStride * i + vtxStride * info->Stride * j);
1230             if(fabs(v[0]) > max_x) max_x = fabs(v[0]);
1231             if(fabs(v[1]) > max_y) max_y = fabs(v[1]);
1232             if(fabs(v[2]) > max_z) max_z = fabs(v[2]);
1233             if(v[2] < neg_z) neg_z = v[2];
1234         }
1235     }
1236
1237     /* This needs some improvements in the vertex decl code */
1238     FIXME("Cannot find data to generate. Only generating position and normals\n");
1239     patch->has_normals = TRUE;
1240     patch->has_texcoords = FALSE;
1241
1242     /* Simply activate the context for blitting. This disables all the things we don't want and
1243      * takes care of dirtifying. Dirtifying is preferred over pushing / popping, since drawing the
1244      * patch (as opposed to normal draws) will most likely need different changes anyway
1245      */
1246     ActivateContext(This, This->lastActiveRenderTarget, CTXUSAGE_BLIT);
1247     ENTER_GL();
1248
1249     glMatrixMode(GL_PROJECTION);
1250     checkGLcall("glMatrixMode(GL_PROJECTION)");
1251     glLoadIdentity();
1252     checkGLcall("glLoadIndentity()");
1253     glScalef(1 / (max_x) , 1 / (max_y), max_z == 0 ? 1 : 1 / ( 2 * max_z));
1254     glTranslatef(0, 0, 0.5);
1255     checkGLcall("glScalef");
1256     glViewport(-max_x, -max_y, 2 * (max_x), 2 * (max_y));
1257     checkGLcall("glViewport");
1258
1259     /* Some states to take care of. If we're in wireframe opengl will produce lines, and confuse
1260      * our feedback buffer parser
1261      */
1262     glPolygonMode(GL_FRONT_AND_BACK, GL_FILL);
1263     checkGLcall("glPolygonMode(GL_FRONT_AND_BACK, GL_FILL)");
1264     IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_FILLMODE));
1265     if(patch->has_normals) {
1266         float black[4] = {0, 0, 0, 0};
1267         float red[4]   = {1, 0, 0, 0};
1268         float green[4] = {0, 1, 0, 0};
1269         float blue[4]  = {0, 0, 1, 0};
1270         float white[4] = {1, 1, 1, 1};
1271         glEnable(GL_LIGHTING);
1272         checkGLcall("glEnable(GL_LIGHTING)");
1273         glLightModelfv(GL_LIGHT_MODEL_AMBIENT, black);
1274         checkGLcall("glLightModel for MODEL_AMBIENT");
1275         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_AMBIENT));
1276
1277         for(i = 3; i < GL_LIMITS(lights); i++) {
1278             glDisable(GL_LIGHT0 + i);
1279             checkGLcall("glDisable(GL_LIGHT0 + i)");
1280             IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(i));
1281         }
1282
1283         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(0));
1284         glLightfv(GL_LIGHT0, GL_DIFFUSE, red);
1285         glLightfv(GL_LIGHT0, GL_SPECULAR, black);
1286         glLightfv(GL_LIGHT0, GL_AMBIENT, black);
1287         glLightfv(GL_LIGHT0, GL_POSITION, red);
1288         glEnable(GL_LIGHT0);
1289         checkGLcall("Setting up light 1\n");
1290         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(1));
1291         glLightfv(GL_LIGHT1, GL_DIFFUSE, green);
1292         glLightfv(GL_LIGHT1, GL_SPECULAR, black);
1293         glLightfv(GL_LIGHT1, GL_AMBIENT, black);
1294         glLightfv(GL_LIGHT1, GL_POSITION, green);
1295         glEnable(GL_LIGHT1);
1296         checkGLcall("Setting up light 2\n");
1297         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(2));
1298         glLightfv(GL_LIGHT2, GL_DIFFUSE, blue);
1299         glLightfv(GL_LIGHT2, GL_SPECULAR, black);
1300         glLightfv(GL_LIGHT2, GL_AMBIENT, black);
1301         glLightfv(GL_LIGHT2, GL_POSITION, blue);
1302         glEnable(GL_LIGHT2);
1303         checkGLcall("Setting up light 3\n");
1304
1305         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_MATERIAL);
1306         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_COLORVERTEX));
1307         glDisable(GL_COLOR_MATERIAL);
1308         glMaterialfv(GL_FRONT_AND_BACK, GL_EMISSION, black);
1309         glMaterialfv(GL_FRONT_AND_BACK, GL_SPECULAR, black);
1310         glMaterialfv(GL_FRONT_AND_BACK, GL_DIFFUSE, white);
1311         checkGLcall("Setting up materials\n");
1312     }
1313
1314     /* Enable the needed maps.
1315      * GL_MAP2_VERTEX_3 is needed for positional data.
1316      * GL_AUTO_NORMAL to generate normals from the position. Do not use GL_MAP2_NORMAL.
1317      * GL_MAP2_TEXTURE_COORD_4 for texture coords
1318      */
1319     num_quads = ceilf(patch->numSegs[0]) * ceilf(patch->numSegs[1]);
1320     out_vertex_size = 3 /* position */;
1321     d3d_out_vertex_size = 3;
1322     glEnable(GL_MAP2_VERTEX_3);
1323     if(patch->has_normals && patch->has_texcoords) {
1324         FIXME("Texcoords not handled yet\n");
1325         feedback_type = GL_3D_COLOR_TEXTURE;
1326         out_vertex_size += 8;
1327         d3d_out_vertex_size += 7;
1328         glEnable(GL_AUTO_NORMAL);
1329         glEnable(GL_MAP2_TEXTURE_COORD_4);
1330     } else if(patch->has_texcoords) {
1331         FIXME("Texcoords not handled yet\n");
1332         feedback_type = GL_3D_COLOR_TEXTURE;
1333         out_vertex_size += 7;
1334         d3d_out_vertex_size += 4;
1335         glEnable(GL_MAP2_TEXTURE_COORD_4);
1336     } else if(patch->has_normals) {
1337         feedback_type = GL_3D_COLOR;
1338         out_vertex_size += 4;
1339         d3d_out_vertex_size += 3;
1340         glEnable(GL_AUTO_NORMAL);
1341     } else {
1342         feedback_type = GL_3D;
1343     }
1344     checkGLcall("glEnable vertex attrib generation");
1345
1346     buffer_size = num_quads * out_vertex_size * 2 /* triangle list */ * 3 /* verts per tri */
1347                    + 4 * num_quads /* 2 triangle markers per quad + num verts in tri */;
1348     feedbuffer = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, buffer_size * sizeof(float) * 8);
1349
1350     glMap2f(GL_MAP2_VERTEX_3,
1351             0, 1, vtxStride / sizeof(float), info->Width,
1352             0, 1, info->Stride * vtxStride / sizeof(float), info->Height,
1353             (float *) data);
1354     checkGLcall("glMap2f");
1355     if(patch->has_texcoords) {
1356         glMap2f(GL_MAP2_TEXTURE_COORD_4,
1357                 0, 1, vtxStride / sizeof(float), info->Width,
1358                 0, 1, info->Stride * vtxStride / sizeof(float), info->Height,
1359                 (float *) data);
1360         checkGLcall("glMap2f");
1361     }
1362     glMapGrid2f(ceilf(patch->numSegs[0]), 0.0, 1.0, ceilf(patch->numSegs[1]), 0.0, 1.0);
1363     checkGLcall("glMapGrid2f");
1364
1365     glFeedbackBuffer(buffer_size * 2, feedback_type, feedbuffer);
1366     checkGLcall("glFeedbackBuffer");
1367     glRenderMode(GL_FEEDBACK);
1368
1369     glEvalMesh2(GL_FILL, 0, ceilf(patch->numSegs[0]), 0, ceilf(patch->numSegs[1]));
1370     checkGLcall("glEvalMesh2\n");
1371
1372     i = glRenderMode(GL_RENDER);
1373     if(i == -1) {
1374         ERR("Feedback failed. Expected %d elements back\n", buffer_size);
1375         Sleep(10000);
1376         HeapFree(GetProcessHeap(), 0, feedbuffer);
1377         return WINED3DERR_DRIVERINTERNALERROR;
1378     } else if(i != buffer_size) {
1379         ERR("Unexpected amount of elements returned. Expected %d, got %d\n", buffer_size, i);
1380         Sleep(10000);
1381         HeapFree(GetProcessHeap(), 0, feedbuffer);
1382         return WINED3DERR_DRIVERINTERNALERROR;
1383     } else {
1384         TRACE("Got %d elements as expected\n", i);
1385     }
1386
1387     HeapFree(GetProcessHeap(), 0, patch->mem);
1388     patch->mem = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, num_quads * 6 * d3d_out_vertex_size * sizeof(float) * 8);
1389     i = 0;
1390     for(j = 0; j < buffer_size; j += (3 /* num verts */ * out_vertex_size + 2 /* tri marker */)) {
1391         if(feedbuffer[j] != GL_POLYGON_TOKEN) {
1392             ERR("Unexpected token: %f\n", feedbuffer[j]);
1393             continue;
1394         }
1395         if(feedbuffer[j + 1] != 3) {
1396             ERR("Unexpected polygon: %f corners\n", feedbuffer[j + 1]);
1397             continue;
1398         }
1399         /* Somehow there are different ideas about back / front facing, so fix up the
1400          * vertex order
1401          */
1402         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 2 + 2]; /* x, triangle 2 */
1403         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 2 + 3]; /* y, triangle 2 */
1404         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 2 + 4] - 0.5) * 4 * max_z; /* z, triangle 3 */
1405         if(patch->has_normals) {
1406             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 2 + 5];
1407             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 2 + 6];
1408             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 2 + 7];
1409         }
1410         i += d3d_out_vertex_size;
1411
1412         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 1 + 2]; /* x, triangle 2 */
1413         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 1 + 3]; /* y, triangle 2 */
1414         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 1 + 4] - 0.5) * 4 * max_z; /* z, triangle 2 */
1415         if(patch->has_normals) {
1416             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 1 + 5];
1417             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 1 + 6];
1418             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 1 + 7];
1419         }
1420         i += d3d_out_vertex_size;
1421
1422         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 0 + 2]; /* x, triangle 1 */
1423         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 0 + 3]; /* y, triangle 1 */
1424         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 0 + 4] - 0.5) * 4 * max_z; /* z, triangle 1 */
1425         if(patch->has_normals) {
1426             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 0 + 5];
1427             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 0 + 6];
1428             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 0 + 7];
1429         }
1430         i += d3d_out_vertex_size;
1431     }
1432
1433     if(patch->has_normals) {
1434         /* Now do the same with reverse light directions */
1435         float x[4] = {-1,  0,  0, 0};
1436         float y[4] = { 0, -1,  0, 0};
1437         float z[4] = { 0,  0, -1, 0};
1438         glLightfv(GL_LIGHT0, GL_POSITION, x);
1439         glLightfv(GL_LIGHT1, GL_POSITION, y);
1440         glLightfv(GL_LIGHT2, GL_POSITION, z);
1441         checkGLcall("Setting up reverse light directions\n");
1442
1443         glRenderMode(GL_FEEDBACK);
1444         checkGLcall("glRenderMode(GL_FEEDBACK)");
1445         glEvalMesh2(GL_FILL, 0, ceilf(patch->numSegs[0]), 0, ceilf(patch->numSegs[1]));
1446         checkGLcall("glEvalMesh2\n");
1447         i = glRenderMode(GL_RENDER);
1448         checkGLcall("glRenderMode(GL_RENDER)");
1449
1450         i = 0;
1451         for(j = 0; j < buffer_size; j += (3 /* num verts */ * out_vertex_size + 2 /* tri marker */)) {
1452             if(feedbuffer[j] != GL_POLYGON_TOKEN) {
1453                 ERR("Unexpected token: %f\n", feedbuffer[j]);
1454                 continue;
1455             }
1456             if(feedbuffer[j + 1] != 3) {
1457                 ERR("Unexpected polygon: %f corners\n", feedbuffer[j + 1]);
1458                 continue;
1459             }
1460             if(patch->mem[i + 3] == 0.0)
1461                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 2 + 5];
1462             if(patch->mem[i + 4] == 0.0)
1463                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 2 + 6];
1464             if(patch->mem[i + 5] == 0.0)
1465                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 2 + 7];
1466             normalize_normal(patch->mem + i + 3);
1467             i += d3d_out_vertex_size;
1468
1469             if(patch->mem[i + 3] == 0.0)
1470                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 1 + 5];
1471             if(patch->mem[i + 4] == 0.0)
1472                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 1 + 6];
1473             if(patch->mem[i + 5] == 0.0)
1474                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 1 + 7];
1475             normalize_normal(patch->mem + i + 3);
1476             i += d3d_out_vertex_size;
1477
1478             if(patch->mem[i + 3] == 0.0)
1479                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 0 + 5];
1480             if(patch->mem[i + 4] == 0.0)
1481                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 0 + 6];
1482             if(patch->mem[i + 5] == 0.0)
1483                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 0 + 7];
1484             normalize_normal(patch->mem + i + 3);
1485             i += d3d_out_vertex_size;
1486         }
1487     }
1488
1489     glDisable(GL_MAP2_VERTEX_3);
1490     glDisable(GL_AUTO_NORMAL);
1491     glDisable(GL_MAP2_NORMAL);
1492     glDisable(GL_MAP2_TEXTURE_COORD_4);
1493     checkGLcall("glDisable vertex attrib generation");
1494     LEAVE_GL();
1495
1496     HeapFree(GetProcessHeap(), 0, feedbuffer);
1497
1498     vtxStride = 3 * sizeof(float);
1499     if(patch->has_normals) {
1500         vtxStride += 3 * sizeof(float);
1501     }
1502     if(patch->has_texcoords) {
1503         vtxStride += 4 * sizeof(float);
1504     }
1505     memset(&patch->strided, 0, sizeof(&patch->strided));
1506     patch->strided.u.s.position.lpData = (BYTE *) patch->mem;
1507     patch->strided.u.s.position.dwStride = vtxStride;
1508     patch->strided.u.s.position.dwType = WINED3DDECLTYPE_FLOAT3;
1509     patch->strided.u.s.position.streamNo = 255;
1510
1511     if(patch->has_normals) {
1512         patch->strided.u.s.normal.lpData = (BYTE *) patch->mem + 3 * sizeof(float) /* pos */;
1513         patch->strided.u.s.normal.dwStride = vtxStride;
1514         patch->strided.u.s.normal.dwType = WINED3DDECLTYPE_FLOAT3;
1515         patch->strided.u.s.normal.streamNo = 255;
1516     }
1517     if(patch->has_texcoords) {
1518         patch->strided.u.s.texCoords[0].lpData = (BYTE *) patch->mem + 3 * sizeof(float) /* pos */;
1519         if(patch->has_normals) {
1520             patch->strided.u.s.texCoords[0].lpData += 3 * sizeof(float);
1521         }
1522         patch->strided.u.s.texCoords[0].dwStride = vtxStride;
1523         patch->strided.u.s.texCoords[0].dwType = WINED3DDECLTYPE_FLOAT4;
1524         /* MAX_STREAMS index points to an unused element in stateblock->streamOffsets which
1525          * always remains set to 0. Windows uses stream 255 here, but this is not visible to the
1526          * application.
1527          */
1528         patch->strided.u.s.texCoords[0].streamNo = MAX_STREAMS;
1529     }
1530
1531     return WINED3D_OK;
1532 }